Fekadu Moreda

Laboratory of Hydrology
Faculty of Applied Sciences, VUB

Pleinlaan 2, 1050 Brussels, Belgium

Conceptual rainfall - runoff models for different time steps
with special consideration for semi-arid and arid catchments

by

Fekadu Moreda

Promotor :
Prof. dr. ir. Willy Bauwens
Thesis submitted in fulfilment of the
requirements for the degree of Doctor of
Philosophy in Applied Sciences

Brussels, July 1999

ii

ABSTRACT

This study investigates the applicability of conceptual monthly water balance models (MWBM)
originally developed for humid climates to semi-arid and arid regions. These models are based
on a simplified representation of the hydrological processes and the balances of incoming and
outgoing water from the catchment on a monthly time base. Based on the experience of the
monthly water balance models, two new models, the 10-day Water Balance Models (DWBM)
and Parsimonious Daily Rainfall-Runoff Model (PDRRM) are developed. The time step of
modelling is reduced to 10 days and one day to account for the high temporal variability of
hydrological variables such as rainfall and soil moisture for semi-arid and arid catchments. A
stepwise parameter optimization procedure is implemented to obtain base flow parameters from
hydrograph analysis. Parameter optimization is achieved by using two optimization algorithms:
the VA05A (Harwell Subroutine Library, 1974) and the Shuffled Complex Algorithm (Duan, et
al., 1992).

A total of 20 catchments from West Africa, North East and South Africa and China are used for
testing the models. The catchments used are characterized by having an extended dry period
followed by short rainy period. For most catchments, the potential evaporation exceeds the
amount of rainfall received for a large period of the year. Though for some catchments the
annual rainfall is as high as 1000 mm, due to the short duration of the rainfall and high potential
evaporation, the runoff that reaches the outlet is only 10-20 % of the rainfall.

The study of the monthly water balance models showed that a slight modification of the
structure of the model concerning the temporal variability of the hydrological improves the
performance of the models. Moreover, it is shown that a particular representation of the
evaporation equation is justified and some of the discrete parameters were narrowed to a
definite range of values, indicating the possible regionalization of the forms of the general
equations governing the models.

The investigation on the 10 days and daily models shows that a relatively small number of
parameters are sufficient to represent the rainfall-runoff relations. Further, it demonstrates that a
routine devised to compute the recession coefficient for 10-days and daily rainfall-runoff model
is useful to incorporate a priori knowledge of a catchment in hydrological modelling. The
stepwise parameter optimization can be extended to daily rainfall runoff models by

iii
hierarchically determining parameters such as the recession coefficient from hydrograph
analysis and evaporation parameters from long-term water balance models. This will ease the
competition of numerous parameters in the minimization of the objective function in standard
optimization procedures.

Further, the developed models are implemented to extend flow data from meteorological data,
to study influences of variability of flow regimes, determination of low-flow-duration curves
and reservoir designs and operations.

The developed models are implemented in a software package (VUBMOD for Windows) for
data entering, model calibration and simulation. The package facilitates the visualization of both
numerical and graphical outputs for analyses. The package runs on personal computers in
Windows (95/98) and Microsoft Windows NT operating systems.

iv
ACKNOWLEDGMENTS

I would like to express my gratitude to my promoter Prof. W. Bauwens for his consistent
guidance and invaluable suggestions without which the realization of this work would have
been difficult.

This work has been a continuation of the research, which has been undertaken at the Laboratory
of Hydrology Vrije Universiteit Brussel to develop conceptual monthly water balance models.
The works of Prof. A Van der Beken, Prof. G. Vandewiele, Dr. Xu-Yu-Chong Dr. Ni-Lar-Win
and many others have been a springboard for this research. My acknowledgement extends to all
of them.

I am very grateful to Prof. A. Van der Beken Director of Laboratory of Hydrology and Prof. F.
De Smedt, Chairman of the IUPWARE for their keen interest academically and also for giving
me an opportunity to work as a part-time at the Laboratory. My thanks also go to Mr. O.
Batelaan for his friendship and encouragement. I have been very happy working with them,
supporting the logistic and program co-ordination for the IUPWARE (Inter-University
Programme of Water Resources Engineering) and also in the organization of four series of
summer courses (1994-1997) in Microcomputer Application in Water resources Engineering.

I am indebted to Prof. J.J. Peters, Prof. G. Vandewiele, Prof. J. Marien and Dr. G. Demaree for
their excellent lectures and personal discussions on various scientific topics. I am grateful to Dr.
Z. Wang and Mr. A. Debebe for their valuable suggestions and discussions on this research and
above all for their friendship.

I wish to acknowledge the support I received from all the technical staff of the Laboratory of
Hydrology: Mrs J.Vermeulen (former secretary), Mr. E. Verbeken, Mr. E. Van den Storm, Mr.
L. Dode, Mr, J. Debruyne, and Ms. Peggy Reniers. I am also grateful to Ms. Howson Louise for
her time reading and correcting this manuscript. I wish to express my gratitude to Mrs C.
Fronteau, Dr. F. Mwanuzi, Ms. Gemma Manache, Mr. A. Sbai, Mr. T. Belayneh Mr. T. Abebe
and all former and present IUPHY/IUPWARE alumni from the different corners of the world
for their valuable friendship.
I am thankful to the member of the jury: Prof. A. Van der Beken, Prof. F. De Smedt. Prof,
J.Marien, Prof. J. Feyen, (Catholic University Leuven) Prof. J.W. Delleur, Purdue University,
USA. for their willingness to read and evaluate this thesis.

v

I would like to thank the Ethiopian Ministry of Water Resources and the Zambian Government
Department of Water Affairs and the Meteorological Department for providing data of the
Awash and the Kafu catchments respectively. I would like to thank Prof. D. A Hughes (Rhodes
University South Africa) for the data of the Tati catchment in Botswana. I wish to express my
thanks to Dr. E. Servat, l'Institut Franais de Recherche Scientifique pour le Development et
Coopration (ORSTOM) for the data on West African catchments and discussions on the
validity of the models. Thanks are also due to Dr Q. Duan, Department of Hydrology and Water
Resources, University of Arizona, Tucson, for making available the subroutines of the Shuffled
Complex Evolution Algorithm.

The financial support to attend international conferences by the Research and Development
Office, Vrije Univesiteit Brussels, is gratefully acknowledged. The last year of my study was
also sponsored by an award from the Rector-scholarship.

My earnest appreciation goes to friends and families I met in Brussels for their understanding,
continuous support and advise. Having life time friends from home, who have extended their
moral support for such a long time regardless of the distance has been a great help for the
success of this work.

Finally, my heartfelt thanks goes to my beloved parents Mr. Moreda Game and Mrs. Walala
Joro for their prayers, love and care. I am also thankful to my brothers: Eshetu, Berhan, Tesfaye,
Addisu, and sisters: Aberash, Tadelech, Mulunesh, Asnakech, Tigist, Mekdes and my little
niece Lelise for their true love. I am grateful to my brother in-law Dr. Mesret Yazachew for his
moral support and above all, making himself available for the family during my long absence.

Fekadu Moreda
Brussels, July 1999.

TABLE OF CONTENTS
Abstract I
Acknowledgement iii
Table of contents v
List of figures x
List of tables xiv
Abbreviations and symbols xvi

CHAPTER 1 INTRODUCTION

1.1 General 1
1.2 Problem definition 2
1.3 Objectives of the research 4
1.4 Structure of the thesis 5

CHAPTER 2 LITERATURE REVIEW ON RAINFALL RUNOFF
MODELLING

2.1 Components of the hydrological cycle 6
2.1.1 Precipitation 8
2.1.2 Evaporation and transpiration 8
2.1.3 Interception 9
2.1.4 Infiltration 9
2.1.5 Stream Flow 9
2.1.6 Hydrological characteristics of semi-arid and arid areas 10
2.2 System approach in catchment modelling 14
2.2.1 The system 14
2.2.2 Modelling concept and classification 15
2.3 Historical development of rainfall - runoff modelling 18
2.4 Time resolution of modelling 21
2.4.1 Common monthly water balance models 22
2.5 Review of the VUB monthly water balance models 24
2.5.1 Model hypothesis 24
2.5.2 Models for different input levels 29
2.5.3 Regional study of monthly water balance models 30
2.6 Summary 30
Contents

vii
CHAPTER 3 MODEL STRUCTURES FOR DIFFERENT MODELLING
TIME STEPS
3.1 Introduction 34
3.2 The monthly water balance model for semi-arid catchments 35
3.2.1 Structure of the model 35
3.2.2 Models for rainfall and temperature inputs 49
3.2.3 Parameter description 50
3.2.4 Initial soil moisture storage 51
3.3 A 10-day water balance model (DWBM) 52
3.3.1 Background 52
3.3.2 Description of the 10-day water balance model 53
3.3.3 Infiltration component 54
3.3.4 Evapotranspiration component 56
3.3.5 Percolation component 57
3.3.6 Flow components 58
3.3.7 The water balance 58
3.3.8 Parameter sets 59
3.3.9 Base flow recession parameter estimation 60
3.4 A parsimonious daily rainfall - runoff model 64
3.4.1 Background 64
3.4.2 Model structure 64
3.4.3 Properties of linear reservoirs 66
3.4.4 The Nash-Unit Hydrograph 67
3.4.5 Parameter sets 69
3.5 Data requirement and scope of application of the developed models 71

CHAPTER 4 STATISTICAL METHODS

4.1 Introduction 74
4.2 Data checking 74
4.3 Parameter estimation 77
4.3.1 Objective function 78
4.3.2 Optimization Algorithms 80
4.3.3 The VAO5A algorithm 84
4.3.4 The shuffled complex evolution algorithm 86
4.4 Model validation 90

Contents

viii
4.4.1 Checks on significance of a parameter 92
4.4.2 Graphical comparison of simulated and observed flows 92
4.4.3 Comparison of monthly means of observed and computed flow 92
4.4.4 Model Efficiency 93
4.4.5 Model Quality factor 93
4.4.6 Residual Analysis 95
4.4.7 Sensitivity to calibration period 96

CHAPTER 5 CASE STUDIES OF MONTHLY WATER BALANCE
MODELS FOR SEMI-ARID AND ARID CATCHMENTS

5.1 Introduction 98
5.2 Data description 98
5.2.1 Ethiopian catchments 98
5.2.2 Chinese catchments 105
5.2.3 West African Catchments 107
5.2.4 Tanzanian catchments 109
5.2.5 Zambian catchments 109
5.2.6 Botswana catchment 111
5.2.7 Common characteristics of the test catchments 112
5.3 Monthly model application 115
5.3.1 Initial soil moisture index 115
5.3.2 Parameter optimization 116
5.3.3 Selection of the forms of the model 116
5.3.4 Summary of the results for the test catchments 129
5.3.5 Reproducing flow regimes 132
5.3.6 Water balance variables 134
5.3.7 Sensitivity of the parameters for the calibration period and extrapolation test 137
5.4 Conclusions 138

Contents

ix
CHAPTER 6 CASE STUDIES OF THE 10-DAY WATER BALANCE
MODEL

6.1 Introduction 140
6.2 Data description 140
6.3 Detailed model tests with an example catchment 143
6.3.1 Parameter optimization 143
6.3.2 Comparison of observed and computed flow 145
6.3.3 Flow components 147
6.3.4 Intermediate water balance variables 149
6.3.5 Residual analyses 150
6.4 Summary of the results for the test catchments 154
6.4.1 General performance of the models 154
6.4.1 Reproducing flow regimes 154
6.4.2 Water Balance components 156
6.5 Concluding remarks 158

CHAPTER 7 CASE STUDIES OF DAILY RAINFALL RUNOFF
MODELS

7.2 Models from literature 161
7.2.1 The XNJ model 161
7.2.2 The Soil Moisture Accounting and Routing (SMAR) model. 161
7.2.3 The NAM model 163
7.3 Application of the rainfall-runoff models 164
7.3.1 Calibration of the models 164
7.3.2 Comparison of performance of the models 166
7.4 Comparison of the PDRRM with the 10 days and monthly models 173
7.5 Concluding remarks 175

Contents

x
CHAPTER 8 PRACTICAL APPLICATION OF THE DEVELOPED
MODEL

8.2 Influences of the variability of rainfall on flow regimes 179
8.2.1 Introduction 179
8.2.2 Annual rainfall variability 180
8.2.3 Influence of rainfall timing on flow regimes 182
8.2.4 Influence of long -term rainfall variability on flow regimes 183
8.2.5 Conclusions 185
8.3 Low flow - duration - return period curves 186
8.4 Reservoir capacity design 188

CHAPTER 9 SUMMARY CONCLUSIONS AND RECOMMENDATIONS

9.1 Summary 190
9.2 Conclusions 191
9.2.1 Monthly water balance models 191
9.2.2 10 days water balance models 192
9.2.3 Daily rainfall runoff models 192
9.2.4 Practical application of the models 193
9.3 Recommendations 194

REFERENCES 196

APPENDIX A THE VUBMOD FOR WINDOWS 208

APPENDIX B ORDERS OF MAGNITUDE OF THE MODEL
PARAMETERS 223

Contents

xi
LIST OF FIGURES

Figure 2.1 Elements of the hydrologic cycle (Chow et al., 1988) 7
Figure 2.2 Schematic representation of the land phase of the water Cycle (after Starosolszky, 1987) 10
Figure 2.3 Diagrammatic representation of a system 14
Figure 2.4 Model classification according to a priori knowledge. (after Todini,1988). 16
Figure 2.5 Classification of models (Chow et al 1988) 17
Figure 2. 6 Diagrammatic representation of the VUB monthly water balance model. 25

Figure 3.1 Conceptual representation of the monthly water balance variables. 36
Figure 3.2 Conceptual representation of the monthly water balance model. 36
Figure 3.3 Influences of evaporation parameter for evaporation equation 3.1. 40
Figure 3.4 Influences of evaporation parameter for evaporation equation 3.2 40
Figure 3.5 Typical monthly rainfall series of humid climate
(e.g. from KleineNete catchment, Belgium) 43
Figure 3. 6 Typical evolution of soil moisture storage in response to humid climate precipitation.
(e.g. from Kleine Nete catchment, Belgium) 43
Figure 3.7 Typical monthly rainfall series. (E.g. of Faleme River, Senegal) 44
Figure 3. 8 Typical evolution of soil moisture storage in response to strong
contrast between rainy and dry season (E.g. of Faleme River, Senegal) 44
Figure 3.9 Physical description of the flow components (Xu, 1988) 46
Figure 3.10 Conceptual representation of the 10-day water balance model. 54
Figure 3.11 Evolution of the infiltration rate 55
Figure 3.12 Relationship between infiltration and soil moisture storage. 56
Figure 3.13 Example of parallel recession curves (Mwambashi catchment, Zambia). 62
Figure 3.14 Example of normalized recession hydrographs (Mwambashi catchment, Zambia) 62
Figure 3.15 Flow-chart for recession parameter estimation 63
Figure 3.16 Conceptual representation of the daily rainfall runoff model. 65

Figure 4.1 Response surface A: one parameter model; B: two-parameter mode l81

Figure 5.1 Location of Awash River Basin 100
Figure 5.2 10-days moving average of rainfall at Addis Ababa and Sebeta stations. 102
Figure 5.3 Typical comparison of hydrograph of Awash River
at Melka Kuntre and Hombole stations (1988) 106
Figure 5.4 Over estimation of rating curves in extrapolation (Modjo River at Modjo) 106
Figure 5.5 Location of West African basins under study 108

Contents

xii
Figure 5.6 Location of Ruaha Catchment (Tanzania) 110
Figure 5.7 Location of Kafu catchments. (Zambia) 110
Figure 5.8 Location of Tati catchment, (Botswana). 111
Figure 5.9 Location of the study areas in Africa. 112
Figure 5.10 Long term monthly pan evaporation, precipitation and flow
for some of the study catchments 114
Figure 5.11 Time series of Input data (Boa Catchment, Ivory Coast) 117
Figure 5.12 Plot of the parameter and the sum of squares of error in the neighborhood
of the optimum value. (Boa catchment -modeled with MWBM-B) 121
Figure 5.13 Comparison of observed and modeled flow with the two
variants of monthly water balance models (Boa catchment) 122
Figure 5.14 Residual versus rainfall MWBM-A. (Boa Catchment, Ivory Coast) 124
Figure 5.15 Residual versus rainfall MWBM-B (Boa Catchment, Ivory Coast) 124
Figure 5.16 Residual versus computed flow MWBM-A. (Boa Catchment, Ivory Coast) 125
Figure 5.17 Residual versus computed flow MWBM-B. (Boa Catchment, Ivory Coast) 125
Figure 5.18 Residual versus Potential evapotranspiration MWBM-A.
(Boa Catchment, Ivory Coast) 126
Figure 5.19 Residual versus Potential evapotranspiration MWBM-B.
Figure 5.20 Auto correlation of residuals from MWBM-A. (Boa Catchment, Ivory Coast) 127
Figure 5.21 Time series plot of residuals. (Boa Catchment, Ivory Coast) 127
Figure 5.22 Comparison soil moisture evolution flow with the two
variants of monthly water balance models (Boa Catchment, Ivory Coast) 128
Figure 5.23 Comparison of the actual evaporation with the two
variants of monthly water balance models. (Boa Catchment, Ivory Coast) 128
Figure 5.24 Flow components using MWBM-A. (Boa Catchment, Ivory Coast). 129
Figure 5.25 Flow components using MWBM-B. (Boa Catchment, Ivory Coast). 129
Figure 5.26 Ranges of estimates of parameter for MWBM-A (left) and MWBM-B (right) 132
Figure 5.27 Comparison of seasonal observed and simulated monthly flows
using MWBM-A and MWBM-B. 133
Figure 5.28 Proportion of water balances variables for East African catchments 135
Figure 5.29 Proportion of water balances for West African catchments 135
Figure 5.30 Proportion of water balances for South African catchments 136
Figure 5.31 Proportion of water balances for Chines catchments 136

Figure 5.32 Comparison of observed and computed flow for the calibration period
(Awash -Hombole) 138
Figure 5.33 Comparison of observed and computed flow for the verification period

Contents

xiii
(Awash -Hombole) 138

Figure 6.1 10-day long-term mean of rainfall and flow for the study catchments 142
Figure 6.2 Time series of rainfall, evaporation and flow data for calibration (Awash -Hombole) 143
Figure 6.3 Plot of the parameter and the sum of squares of error in the neighborhood
of optimum value. (left: MWBM-B, and right DWBM). 146
Figure 6.4 Comparison of observed and modeled flow with MWBM and DWBM (Awash-Hombole) 147
Figure 6.5 Flow components using MWBM-B. (Awash-Hombole) 148
Figure 6.6 Flow components using DWBM (Awash-Hombole) 148
Figure 6.7 Comparison of soil moisture storage evolution with MWBM-B
and DWBM. (Awash-Hombole) 149
Figure 6. 8 Comparison of 10-day actual evaporation with MWBM-B
and DWBM. (Awash-Hombole) 149
Figure 6.9 Residual versus rainfall using MWBM-B.(Awash-Hombole) 150
Figure 6.10 Residual versus rainfall using DWBM. (Awash-Hombole) 150
Figure 6.11 Residual versus potential evapotranspiration using MWBM-B. (Awash-Hombole) 151
Figure 6.12 Residual versus potential evapotranspiration DWBM. (Awash-Hombole) 151
Figure 6.13 Residual versus computed flow using MWBM-B. (Awash-Hombole) 152
Figure 6.14 Residual versus computed flow using DWBM. (Awash-Hombole). 152
Figure 6.15 Time series plot of residuals. (Awash-Hombole) 153
Figure 6.16 Autocorrelation of residuals for DWBM. (Awash-Hombole) 153
Figure 6.17 Comparison of long-term 10 day mean observed and computed flow
using MWBM-A and DWBAM. 156
Figure 6.18 Proportion of water balance s using DWBM. 157
Figure 6.19 Comparison of monthly observed and computed flow using the MWBAM
and the DWBM. (Awash -Hombole) 159

Figure 7.1 A comparison of scatter plots of the observed and computed flow for
the four models (Awash River at Hombole). 169
Figure 7.2 Comparison of observed and computed hydrographs for
the Awash-Hombole catchment. (Typical year 1986.) 171
Figure 7.3 Comparison of observed and estimated flows for Awash-Melka catchment 172
Figure 7.4 Comparison of observed and estimated flows Berga catchment 172
Figure 7.5 Comparison of the daily rainfall-runoff and the 10-day water balance models
(Awash-Hombole) 174
Figure 7.6 Comparison of the daily rainfall-runoff, the 10-day and monthly
water balance models (Awash-Hombole) 174

Contents

xiv

Figure 8.1 Schematic representation of extension of flow. 177
Figure 8.2 Relationship between areal and Addis Ababa monthly rainfall. 178
Figure 8.3 Time series comparison of long term rainfall series and modeled annual flows 179
Figure 8.4 Annual variability of rainfall at Addis Ababa station. 181
Figure 8.5 Annual variability of Awash River flow at Hombole 182
Figure 8.6 Departure of annual and summer rainfall from long term means on decade basis. 184
Figure 8.7 Departure of annual and summer flow from long-term means on decade basis. 184
Figure 8.8 Departure of annual base flow and low flow from long term means on decade basis. 185
Figure 8.9 Low flow duration -return period relationships for Awash River at
Hombole (1963-1992) 187
Figure 8.10 Low flow duration -return period relationships for Awash River at
Hombole. (1900-1995) 187
Figure 8.11 Return period versus reservoir capacity using the extended flow series
for given demands. (Awash River at Hombole). 189

Contents

xv
LIST OF TABLES

Table 2.1 Relationship between time step of modelling and area of catchment 22
Table 2.2 Common monthly water balance models 24
Table 2.3 model types and input requirements 30

Table 3.1 Description of parameters of monthly water balance model. 50
Table 3.2 Ranges of values of hydraulic conductivity 55
Table 3.3 Parameters of the 10-days water balance model 60
Table 3.4 Description of parameters of the daily rainfall-runoff model 71

Table 4.1 Inputs of the VA05A optimization method 86
Table 4.2 Inputs of the Shuffled complex optimization method 89
Table 4.3 Some 5 % points of the Chi-square distribution with n degrees of freedom. 97

Table 5.1 Description of daily rainfall data in upper Awash River Basin 101
Table 5.2 Cross-correlation between monthly rainfall data of stations in
the upper Awash River Basin 102
Table 5.3 Pan evaporation daily data 103
Table 5.4 Daily River flow data 103
Table 5.5 Stage-discharge relationships rating curves for different periods 103
Table 5.6 Characteristics of the test catchments and available data on monthly time step 112
Table 5.7 Representation of different warming up period 115
Table 5.8 Influences of length of warming period on initial soil moisture storage
using MWBM-A (IR=2, b
1
=2, b
2
=2) 116
Table 5.9 Example application of the MWBM-A 117
Table 5.10 Example application of the MWBM-B 118
Table 5.11 Summary of results for all 18 variants of the model WBAM-A.
Table 5.12 Summary of results for all 18 variants of the WBAM-B.
Table 5.13 Parameter estimates for MWBM-A (IR=2, b
1
=b
2
=2.0) 131
Table 5.14 Parameter estimates for MWBM-B (IR=2, b
1
=b
2
=2.0) 131
Table 5.15 Comparison of parameter estimates and water balance variables
for the three sets of data (Awash catchment the discrete parameters: b
1
=2, b
2
=2) 137

Contents

xvi
Table 6.1 Test catchments for 10-day time step modelling 141
Table 6.2 Parameter estimates of the MWBM-B and DWBM for Awash -Hombole) 145
Table 6.3 Table Parameter estimates for the MWBM-B with 10-day time step 155
Table 6.4 Parameter estimates for the DWBM 155

Table 7.1 Description of XNJ model parameters (Zhao et al., 1980) 162
Table 7.2 Description of SMAR model parameters (OConnell et al., 1970) 163
Table 7.3 Parameters of the NAM model (Ref. Danish Hydraulic Institute, 1982) 164
Table 7.4 Calibration periods 164
Table 7.5 Model parameters: The XNJ Model 165
Table 7.6 The SMAR Model parameters 165
Table 7.7 The NAM Model parameters 166
Table 7.8 The PDRRM Model parameters 166
Table 7.9 Comparison of statistics of the studied models 168

Table 8.1 Rainfall Variability during the last 9 decades based on Addis Ababa observations. 181
Table 8.2 Cross correlation coefficients of rainfall and flow regimes 183

Contents

xvii
Abbreviations
Bias Bias of the residual series
DWBM Decade (10-day) water balance model
CI Confidence interval
Ef. Model Efficiency as defined by Nash and Sutcliffe, (1970)
HMLE Heteroscedastic Maximum Likelihood Estimator
HWCI Half width of confidence interval
MQ Model quality factor
MWBM Monthly water balance model
MWBM-A Monthly water balance model variant A (without upper limit for soil
moisture storage
MWBM-B Monthly water balance model variant B (without upper limit for soil
moisture storage)
NAM Danish Hydraulic Institute rainfall-runoff model
PDRRM Parsimonious daily rainfall -runoff model
RMSE Root mean square error
SHU-CA Shuffled complex Algorithm
SMAR Soil Moisture Accounting and Routing
VAO5A Optimiazation algorithm from Harwell Subroutine Library (1974)
XNJ XINJIANG Daily rainfall-runoff model

Symbols
(a
1
, a
2
) continuous parameters
b
t
base flow
b
1
, b
2
discrete parameters
c
t
temperature
d
t
computed total flow
e
t
Potential evaporation
f
t
fast flow
g
t
groundwater storage
h
t
humidity
h
f
ordinates fast flow unit hydrograph
h
s
ordinates of slow flow unit hydrograph
I
t
infiltration
k recession coefficient
l
t
percolation
m
t
soil moisture storage
n
t
net rain
p
t
precipitation
q
t
observed flow
r
t
actual evaporation
s
t
slow flow
w
t
available water

Contents

xviii

Chapter 1
1. INTRODUCTION

1.1 General ............................................................................................................................................1
1.2 Problem definition...........................................................................................................................2
1.3 Objectives of the research ...............................................................................................................4
1.4 Structure of the thesis......................................................................................................................5

1.1 General

Water resources are essential renewable resources that are the basis for existence and
development of a society. Proper utilization of these resources requires assessment and
management of the quantity and quality of the water resources both spatially and temporally.

Water crises caused by shortages, floods and diminishing water quality, among others, are
increasing in all parts of the world. The growth of population demands for increased domestic
water supplies and, at the same time, results with a higher consumption of water due to
expansion in agriculture and industry. Mismanagement and lack of knowledge about existing
water resources and the changing climatic conditions have consequences of an imbalance of
supply and demand of water. The problem is pronounced in semi-arid and arid areas where the
resources are limited.

Surface water being easy, direct and therefore less expensive to exploit in comparison to other
sources like groundwater or desalinization makes it the major source of water supply for
irrigation, industry and domestic uses. The surface water, in the form of lakes and river
discharge (runoff) is predominately obtained from rainfall after being generated by the rainfall -
1. Introduction

2
runoff processes. In order to make decisions for planning, design and control of water resource
systems, long runoff series are required. The latter are not often available with reasonable
length. On the other hand, for flood control and reservoir regulation future, flows shall be
forecasted with rainfall runoff models. A number of rainfall runoff models exist for generation
of flow, forecasting and other purposes.

Establishing a rainfall-runoff relationship is the central focus of hydrological modelling from its
simple form of unit hydrograph to rather complex models based on fully dynamic flow
equations. As the computing capabilities are increasing, the use of these models to simulate a
catchment became a standard. Models are generally used as utility in various areas of water
resource development, in assessing the available resources, in studying the impact of human
interference in an area such as land use change, deforestation and other hydraulic structure such
as dams and reservoirs.

1.2 Problem definition

Traditional lumped, conceptual rainfall-runoff models attempt to represent hydrological
processes by mathematical equations and multiple layers of storage. These equations involve a
large number of parameters, e.g., the Stanford Watershed Model IV (Crawford and Linsley,
1966) has 16 parameters, the SACRAMENTO model (Burnash et al., 1973) has 21 parameters,
and the XINANJIANG model (Zhao et al., 1980) has 15 parameters. In applying such models, a
modeler always faces the difficulty of obtaining a unique set of parameters that satisfies a
certain objective, usually minimizing the sum of squares of errors between measured and model
outputs. Moreover, high correlations between parameters hinder identification and result in
individual parameters that lack (statistical) significance. The numerous parameters make it
impossible to relate catchment characteristics to parameter values for regionalization of
parameters and application of the models to ungauged catchments. This prompts researchers to
develop physically based models such as the SHE model (Abott, 1986a, 1986b). Data
availability for model calibration remains the bottleneck for development in this direction.

For the reasons previously described, simple conceptual rainfall-runoff models that use readily
available data are applied for water resources assessment and management in areas where
hydrological data are limited. Monthly water balance models (Vandewiele et al., 1992) are

1. Introduction

3
examples of parsimonious models that are pragmatic for application. The models developed
have conceptual structures and inputs are monthly precipitation and evaporation or (mean
temperature and/or mean relative humidity). The outputs are typically monthly flow and soil
moisture index. The models are described with few parameters (3 to 6) and the inputs are
relatively easily available at catchment scales. These models have been successfully applied in
humid climate and for a few semi-arid catchments (see Xu, 1992, 1995, 1996, Vandewiele and
Ni-Lar-Win, 1998).

This study focuses in extending the application of these models to semi-arid and arid
catchments and also to reduce the time resolution of modelling from monthly to 10 days and
one day.

Compared to humid climate there are a lot of problems in modelling hydrological process of
arid and semi-arid regions, which cover a large part of the world. In general the two interrelated
underpinning problems in hydrological modelling in such region are: (1) the feebleness of the
model assumptions and simplifications which are inherent to any modelling in any region but
lead to over simplification of the variability in arid and semi-arid areas, (2) the limitation of data
availability as opposed to the temporal and spatial variability of the input to any physical or
conceptual (distributed or lumped) models. Therefore there is a need of research which
compromises the two problems by developing models which are not too simple to ignore the
processes and yet not requiring very much detail of data that are not available in real life
application.

Time resolution of modelling determines the complexity of modelling as much as the spatial
resolution of modelling. One can argue that if it is possible to model with a smaller time step,
the larger time steps can always be reconstructed by aggregations. Perhaps the very objectives
of modelling with larger time spans are to simplify the processes prevailing in a catchment and
availability of data to calibrate and eventually apply the model. Moreover, the time step of
modelling depends on the potential uses of a model. A flood warning scheme of a small
catchment requires modelling over shorter time spans such as minutes and hours where as a
monthly water balance model would be much more handy to extend a runoff series for the
purpose of reservoir capacity design. A decade model would also be good enough to study the
water availability of a river flow for irrigation diversion and to study the soil water evolution in
a soil, to manage crop water requirements.

1. Introduction

4

1.3 Objectives of the research

The objectives of this thesis are to investigate the applicability of the models originally
developed for humid climates to semi-arid and arid regions. Further, to study the peculiar
hydrological phenomena in such regions, either by modification of the existing model or
establishing new ones. The research also attempts to address the problem of data deficiency in
such regions and hence the models developed require readily available data on a catchment
scale.

The specific objectives are:

(1) Modifying the existing models for application to different climatic and basin
characteristics.
(2) Developing a conceptual 10-day water balance model and a daily rainfall-runoff model
by keeping the modelling philosophy as the monthly water balance models i.e. the
model being simple and parsimonious.
(3) Introducing step-wise parameter optimization by inferring some parameters from
hydrograph analysis.
(4) Coupling the models with modern parameter optimization techniques
(5) Testing the models
(6) Application of the model to real problems such as runoff extension for studies of long
term changes in flow regimes, low flow analysis and reservoir capacity design.
(7) Packaging the three models in a software system. This allows not only model
calibration and simulation but also graphical visualization with formats compatible for
exporting to other systems.

The present models are intended to satisfy the following guidelines for building conceptual
rainfall runoff models (Van der Beken, 1977)

The models should be representative of the regimes of a wide variety of streams
The models should be easily applicable to different basins with existing hydrological data

1. Introduction

5
The models should be physically sound so that, in addition to stream flows, estimates of
surface runoff, actual evaporation, soil moisture can be made
The calibration procedure should be automatic or, at least, fully reproducible without
ambiguity.

1.4 Structure of the thesis

This dissertation is subdivided into 9 chapters including introduction and the last chapter, which
present summary, conclusions and recommendations. The first four chapters will discuss the
general development of the models. Following this introduction a brief review of rainfall
modelling is given in Chapter 2. Also, a historical development of rainfall runoff modelling and
concepts of the VUB monthly water balance models are discussed in this chapter to lay a
background of the study. The core of the thesis is presented in Chapter 3, which treats the
Monthly Water Balance Model, (MWBM) for the semi-arid and arid case and gives the full
development of the Decade (10-day) Water Balance Model (DWBM). Finally, it introduces the
Parsimonious Daily Rainfall- Runoff Model, (PDRRM). The later two models have evolved
from the concepts of the monthly water balance models. The justifications and pitfalls of the
models developed are discussed. The fourth chapter discusses statistical methods applied in
parameter optimization and evaluation of results.

The next three chapters present the case studies to validate the models developed in Chapter 3.
Chapter 5 gives applications of the monthly water balance model to 20 catchments from 8
countries in Africa and Asia. Chapter 6 gives applications of the 10 days Water Balance model
to 8 catchments from Africa. In Chapter 7, the daily models are applied to a few catchments.
The results are compared to three classical conceptual rainfall runoff models namely XNJ,
(Zhao et al., 1980), SMAR (OConnell et al., 1970) and NAM (Danish Hydraulic Institute,
1982)

Chapter 8 presents the practical application of the developed model in three aspects: extension
of flow time series for studying long term variability of flow regimes, low flow analysis and
reservoir capacity design. Lastly, the summary and conclusions drawn from the study are given
in chapter 9.

Chapter 2

2 LITERATURE REVIEW
ON RAINFALL RUNOFF MODELLING

2.1 Components of the hydrological cycle ............................................................................................6
2.2 System approach in catchment modelling.....................................................................................14
2.3 Historical development of rainfall - runoff modelling ..................................................................18
2.4 Time resolution of modelling.......................................................................................................21
2.5 Review of the VUB monthly water balance models .....................................................................24
2.6 Summary .......................................................................................................................................30

2.1 Components of the hydrological cycle

The central focus of any hydro-meteorological study is the hydrological cycle shown in Figure
2.1. The hydrological cycle has no beginning or end and its many processes occur continuously
(Chow et al., 1988). In describing the cycle, the water evaporates from ocean and land surface
to become part of atmosphere; water vapour is transported and lifted in the atmosphere until it
condenses and precipitates on the land or the oceans. Precipitated water may be intercepted by
vegetation, becomes overland flow over the ground surface, infiltrate into the ground, flow
through the soil as subsurface flow and discharges into streams as surface runoff. The infiltrated
water may percolate deeper to recharge groundwater, later emerging as spring and seeping into
streams to form surface runoff and finally flowing into the sea or evaporating into the
atmosphere as the hydrological cycle continues.
2. Literature review

7

Figure 2.1 Elements of the hydrologic cycle (Chow et al., 1988)

It is noted that though the concept of the cycle seems simple, the phenomena are enormously
complex and intricate. It is not just one large cycle but it is rather composed of many
interrelated cycles of continental, regional and local extent. The major achievement and
objectives of the rainfall runoff modelling is thus to study a part (section) of the hydrological
cycle, namely the land phase of the hydrological cycle on a catchment scale. Then the problem
becomes to express the runoff from the catchment as a function of the rainfall and other
catchment characteristics.

Considering the land phase of the hydrological cycle, any conceptual model predicates its effort
on an expansion of the basic water balance or continuity equation that is:

dt
ds
O I = (2.1)

where I is the input to the system (Precipitation), and O is output from the system (evaporation,
stream and groundwater flow) and the ds/dt is the change in soil moisture. The main processes
encompassed are precipitation, evapotranspiration, interception, infiltration, subsurface flow and
streamflow. It is evident that before any modelling effort can be performed, one has to


8
understand the above physical processes, their extent of effect on the abstraction from or
addition of water to a catchment.

2.1.1 Precipitation

Precipitation is the input to the system of catchment, which may have different forms, rainfall,
storms, dew or any form of water landing from atmosphere. The amount of precipitation can be
defined as an accumulated total volume for any selected period. Precipitation as a function of
time and space is highly variable. Systematic averaging methods such as Thiessen polygon,
isohyte and reciprocal distance methods have been developed to account for variations in space
to obtain a representation of areal precipitation values from point observation. Singh and
Chowdhury, (1986) after comparing the various methods for calculating areal averages,
concluded that all methods give comparable results, especially when the time period is long. For
short time step records, the conversion of a point observation to an areal rainfall has a large
influence.

2.1.2 Evaporation and transpiration

Catchment evaporation demand is generally defined as that evaporation which would occur if
there were no deficiencies in the availability of moisture for evapotranspiration by that area's
particular plant regime. The two main factors influencing evaporation from an open water
surface are the supply of energy to provide latent heat of vaporization and the ability to transport
the vapour away from the evaporative surface: solar radiation and wind. Evapotranspiration
from land surface comprises evaporation directly from the soil and vegetation surface and
transpiration through plant leaves, in which water is abstracted from the sub soil. The third
factor is the supply of moisture at evaporative surface, which brought about the definition of
potential and actual evaporation. Evaporation involves a highly complex set of processes, which
themselves are influenced by factors dependent on the local conditions (land use, vegetation
cover, and meteorological variables). Mostly the potential evaporation is the quantity obtained
either by using some simple empirical formula such as Thornthwaite, (1948), Penman formula
(Penman, 1948) and a process-based model of Penman-Monteith (Monteith, 1965).

Since potential evaporation and evaporation from pans are governed by the same meteorological
factors they have strong correlation. The relation between them is often give as a simple ratio.


9
Burnash (1995) suggests using seasonal coefficients for converting pan data to potential
evaporation rather than a single coefficient. In conceptual rainfall runoff modelling one of the
two terms, pan evaporation and potential evapotranspiration are equally used as input, which
exerts energy to extract water from open surface or soil moisture storage.

2.1.3 Interception

The portion of rainfall intercepted by the vegetation and roofs before reaching the ground is
referred to as interception. The water, which is intercepted by the leaves of vegetation and roofs
eventually evaporates into atmosphere. The amount of interception could be significant in
densely vegetated areas such as tropical rainforests. Such forests maintain a relatively consistent
canopy and do not generally exhibit the seasonal range of interception encountered in areas
where deciduous trees are dominant. It is commonly understood that if the density of the
vegetation cover is sparse then this loss is insignificant.

2.1.4 Infiltration

The precipitation, which is not intercepted or evaporated from the land, will eventually infiltrate
into the soil or flow as overland flow. Infiltration is one of the most difficult hydrological
processes to quantify. The difficulty arises due to many physical factors affecting the rate of
infiltration such as rainfall intensity, initial moisture content, soil property, etc. Some
experimental and empirical formulas such as Horton (1939), Philip (1957), and others are
available to compute infiltration rates during a rainfall event. Depending on the soil strata, the
infiltrated water gradually percolates to the groundwater or either flows as subsurface flow
supplying river or springs within the catchment.

2.1.5 Stream Flow

The rainfall that exceeds the interception requirement and infiltration starts to accumulate on the
surface. Initially the excess water collects to fill depressions, until the surface detention
requirement is satisfied. There after when water begins to move down slope as a thin film and
tiny streams which eventually join to form bigger and bigger channels. This part of the stream
flow is termed as surface runoff. The infiltrated part of the rain may sometimes come as
subsurface runoff, which combined with the surface runoff, constitutes the direct runoff. Hence


10
the direct runoff is the result of the immediate response of a catchment to the input rainfall. The
stream flow consists of the direct runoff (which lasts for hours or days depending upon the
catchment size) and the base flow (that emerges from groundwater resources and also delayed
subsurface runoff). The above description of the processes at catchment scale is schematically
represented in Figure 2.2.

Soil Moisture
Capillary
Rise
Interception
Depression
Storage
Groundwater
Storage
Underground flow into
or out of the area
Channel
Precipitation
Surface
Runoff
Base
Flow
Stream Flow
Very Fast
Fast
Medium
Slow to
Very Slow
Interflow
Evaporation
Infiltration
Rainfall
Evaporation

Figure 2.2 Schematic representation of the land phase of the water Cycle (after Starosolszky,
1987)

2.1.6 Hydrological characteristics of semi-arid and arid areas

In the past, several hydrologists attempted to classify zones of the world in to humid, semi arid
and arid according to the climatological characteristics. One of the earliest indexes (Chow,
1964) used for classification is based upon the adequacy of precipitation in relation to the needs
of plants whereby the precipitation analyzed month by month is just adequate to supply all the
water for maximum evaporation and transpiration in the course of a year. UNESCO (1979)
attributes the arid and semiarid zones as the dry areas associated with annual potential
evaporation over 1000mm and further classifies according to the amount of annual rainfall they


11
receive to hyper-arid, arid semiarid and humid. Chow, (1964) suggests that in addition to
climatic characteristics other features of the land surface may also be used to delimit arid zones,
since the geomorphology, soils and vegetation have their own distinctive characteristics. The
objective of these classifications is mainly to study the peculiar characteristics common to a
region, which would help generalization and inference for climatological and hydrological
processes prevailing in these regions.

The hydrological processes operating in rainfall runoff transformation for the semi-arid and arid
areas differ from those in humid temperate. Some of the distinct properties manifested in semi-
arid and arid catchments are pointed out below.

2.1.6.1 Variability in time

The rainfall in semi-arid and arid catchments is characterized by a high variability of the small
amount received in space and time (Moore, 1989). A high percentage (about 80 percent of the
annual rainfall) is received within the rainy seasons during 3 to 6 months. Individual rainfall
events generally occur with high intensity and short duration storms. Verma (1979) points out
some of the particular features of the semiarid and arid hydrological processes as:

The marked seasonal variation in semiarid climates may require segregation of data by
season. A combination of hydrological factors common in one season of the year may be
virtually non-existent during another season.
A particular combination of factors may exist for only a few days in several years and may
render hydrological computation based on average values grossly erroneous.

Actual evaporation from semi-arid zones is a highly transient phenomenon with extreme
variation within a day because of the available water but not over a season. The transient nature
of evaporation is also controlled by the rapid growth of vegetation to climax followed by rapid
die-off (Moore, 1989).

2.1.6.2 Variability in space


12
In contrast to humid climate, hydrological processes in semiarid and arid regions often vary
greatly over different parts of a catchment. Especially in large catchments, the contributing area
could be localized at the upper part of the catchment. In such cases, computation of areal rainfall
in a lumped conceptual model leads to unrealistic average distribution over the whole area.
More over, the sparse vegetation cover and its sharp response to the first rain have an impact on
the evaporation process prevailing in such regions. The rivers in such regions are generally
characterized by having long periods of low flow regime.

Another distinct characteristic of such regions is that in some cases infiltration could be very
small due to outcropped rocks on the slopes of valleys whereas it could be high in areas with
fractured bedrock channels. There could also be the possibility of channel infiltration from the
bed of the rivers supplying the groundwater in lower valleys of the river (Sami, 1992). This fact
implies that especially during low flow regime the stream flow that originates from upstream
will be depleted by the channel bed before it reaches the outlet. Hence this phenomenon should
be accounted for in formulating models based on the water balance of a catchment

2.1.6.3 Hydrological modelling inadequacy

In arid regions, an important feature of the water balance is the high proportion of incoming
water which is returned to the atmosphere by evaporation from soil surface. In contrast to the
humid regions, where evaporation is limited by available energy (e.g. net radiation), in the arid
zone water availability is the dominant control over evaporation rates. Because of the sparse
density of vegetation, direct evaporation of water from the soil is of enhanced importance, and
frequently as much as half of the annual rainfall can be lost in this manner (Chow, 1964).

It is noted that the prevailing rainfall and evaporation mechanism in semi-arid and arid
catchments, associated with the thin and sparse vegetation cover, alter the runoff generation of
these regions in contrast to the humid regions. The runoff generated is mainly controlled by
infiltration excess and is frequently localized. The runoff generated on some of the slopes and
first order catchments may not always survive to contribute to the flow at the outlet of
catchments of sufficient size. Hughes (1995) numerates the possible reasons why deterministic
models can fail as tools for water resources estimation purposes, where failure implies the
model imperfection. Apart from erroneous data inputs and poor interpretation of model results,


13
the problems associated with the application of rainfall- runoff models to arid and semiarid
areas are:

Inadequate or inappropriate model representation of the prevailing catchment
processes
Inadequate representation of the spatial variability of runoff generation response to runoff.
While this problem can be masked by spatial lumping, it may be important if the effects are
non-linear and non-stationary
Inadequate representation of the spatial variability in rainfall input, either through lumping or
lack of spatial resolution in the available data.
Inadequate representation of the temporal variability in rainfall input through the use of a
coarse time interval model. This is not always a serious problem as long as the rainfall
mechanisms are reasonably consistent and the durations and intensities of the major rainfall
events are similar.
Inadequate estimation of parameter values. This problem may relate to the length of the
records available for calibration (Grgens, 1983) and the extent to which the rainfall-runoff
relationships reflected in the observed data are sufficiently representative to allow a suitable
parameter set to be quantified

In general the two interrelated underpinning problems in hydrological modelling in such regions
are:

1. the model assumptions and simplifications which are not always justified in modelling in
any region but over simplification of the variability in such areas,
2. The limitation of data availability as opposed to the temporal and special variability of the
input to any physical or conceptual (distributed or lumped) models.

This implies that any effort in modelling such region should consider and compromise the two
underlying problems that on one hand the model has to address the peculiar phenomena and at
the same time it should require limited input as only limited data are available.


14
2.2 System approach in catchment modelling

2.2.1 The system

The detailed processes that link the rainfall over the catchment to the stream flow, may be
studied by applying physical laws that are reasonably well known. However, the complexity of
the boundary conditions (i.e. the physical description of the catchment and the initial conditions
and distribution of the variables) makes a solution based on the direct application of the laws of
physics impracticable. Moreover, direct application of these laws requires subdividing the
catchment into homogenous and isotropic regions. The sub division depends on catchment
characteristics (soil type, land use, slope, vegetation cover, etc.) and these factors may also vary
in space and time. For these reasons, instead of exact representation of the processes effort is
directed to the construction of a model by using system concepts relating input and outputs.

A hydrologic system is defined as a structure or volume in a space, surrounded by a boundary
that accepts water and other inputs, operates on them internally, and produces them as outputs.
(Chow et al., 1988). Schematic representation of the system operation is shown in Figure 2.3
where the symbol represents a transformation between the input and the output.

Input Output

Figure 2.3 Diagrammatic representation of a system

The objective of the hydrologic system analysis is to study the system operation and to predict
its output. A hydrological system model is an approximation of the actual system. Its input and
output are measurable hydrological variables and its structure is the concept of the system
transformation.

The system analysis approach depends on the prior supposition of a general, flexible,
relationship (e.g. linear time invariant) whose expression may be obtained by the application of
systems analysis methods to the records. Examples in the hydrological context include the unit
hydrograph method, the linear difference equation models (Box and Jankins, 1976), the


15
constrained linear system model (Natale, and Todini, 1976), the linear perturbation model
(Nash, and Barsi, 1983) and many similar models are found in literature.

2.2.2 Modelling concept and classification

Based on the assumptions and concepts formulating the structure of the transformation
(Operator) the resulting models may have different forms. According to Clarke, 1973
mathematical models may be classified in to four main groups.
1) Stochastic - Conceptual
2) Stochastic - Empirical
3) Deterministic - Conceptual
4) Deterministic - Empirical

Any of which may be classified as linear or non-linear in the system theory sense. Group 1 and
2 may be linear or non-linear in the statistical regression sense. The models in any group may
further be classified as lumped, probability distributed or geometrically distributed.

If any of the variables is a random variable having a probability distribution, then the model is
stochastic rather than statistical. Conversely, if all the variables are free from random variations,
then the model is deterministic. The hydrological models are described as conceptual or
empirical according to whether the form of the model equation is, or is not suggested by
consideration of the physical processes acting up on the input variable to produce the output
variables. This distinction is, however artificial since many physical laws contain empirical
constants. Despite the artificiality of the distinction, it is one widely drawn in the literature, the
difference between 'conceptual' and 'empirical' corresponding approximately to the O' Connell's
(1966) terms 'synthetic' and 'analytic' whilst the term 'black box' used by some writers,
corresponds to 'empirical'.

Linearity and non-linearity in the system analysis sense is based on the principle of
superposition (Clarke 1973). A model is linear in the system analyses sense if the principle of
superposition holds. The model is linear in statistical regression sense if it is linear function of
the parameters to be estimated.


16
A lumped model ignores the spatial distribution of the input variables and parameters, which
characterises the physical process. A probability-distributed model describes the spatial
variation of the input variables without geometrical references to the point at which the input is
measured or estimated. On the other hand a geometrically distributed model expresses the
spatial variability in terms of the orientation of the network points relative to each other and
their spacing.

Todini (1988) classified models based on the level of the priori knowledge on the system under
study in terms of both model structure and parameters. Accordingly the recent classification
groups the model in the increasing level of a priori knowledge as: 1) purely stochastic, 2)
lumped integral, 3) distributed integral and 4) distributed differential. (See Figure 2.4) Chow et
al. (1988) give classification of model based on three objectives, model account for
randomness, time and space. Figure 2.5 shows possible combination of the model types.

A priori Knowledge
Models
S. L.I. D.I. D.D
S.

P.

P
a
r
a
m
e
t
e
r
s

S. = stochastic; L.I.=lumped integral;
D.I. = distributed integral; D.D.= distributed differential ;
P = physical
Figure 2.4 Model classification according to a priori knowledge. (after Todini,1988).


17

Figure 2.5 Classification of models (Chow et al., 1988)


18
2.3 Historical development of rainfall - runoff modelling

In this section a brief historical development of rainfall runoff modelling is discussed. It is
obvious that a large number of models have been developed and used both as research and as
operational tools. An exhaustive list and account is not the purpose of this study. We try to
evince the trend of development and the causes for such evolution. Therefore only some of the
models pertinent to the development of the conceptual modelling are discussed.

Though the speculation and awareness of hydrological cycle is dated BC, modern
methodologies relating rainfall to runoff originated during the last half of the 19th century. The
main objectives at that time was to determine the design discharges for sewerage design, land
reclamation and reservoir spillway design.

According to Dooge (1957, 1973), during the last part of the 19th century and earlier part of the
20th century, most engineers used empirical formulas derived for particular cases and applied
them to the other cases under the assumption that condition were similar enough. The engineers
at that time also used the 'rational method' which may be seen as the first attempt to approach
practically the problem of predicting runoff from rainfall. The methods were derived only for
small mountainous catchments. Later during 1920s many modifications were introduced to
extend the applicability of the rational formula to larger catchments.

The modified rational method based on the concept of isochrones, or lines of equal travel time,
can be seen as the first rainfall-runoff model based on a transfer function. The model parameters
were derived by means of topographic maps and the use of Mannings formula to evaluate the
different travel times.

Sherman (1932) introduced the concept of the unit hydrograph on the basis of the principle of
superposition. Although not yet known at that time, the superposition principle implied many
assumptions, i.e. the catchment behaves like a causative, linear time invariant system with
respect to the rainfall/surface runoff transformation.

At the end of 1930s and during 1940s a number of techniques were proposed in order to
improve the objectivity of methods and results, the techniques of statistical analysis were
invoked.


19

The real break through came in the fifties when hydrologists became aware of system
engineering approaches used for the analysis of complex dynamic systems. They realized that
the unit hydrographs was the solution of a causitive, linear time invariant system and the
response functions could be derived from input and output data by use of mathematical
techniques. Prior to the development of computer models, the Antecedent Index approach
(Kohler and Linsley, 1951) was a popular technique to determine runoff from rainfall. The
Antecedent Index system correlates prior precipitation and a second variable, generally a week
number to determine a "runoff curve". The "runoff curve" provides a technique for determining
the quantity of runoff, which would be produced from storm rainfall totals. By comparing
rainfall and runoff totals at fixed times throughout a storm, period volumes can be computed. A
unit graph is then applied to the period runoff volumes, an estimate of base flow is added, and
estimates of catchment streamflow are obtained.

The derivation of the unit hydrograph in the discrete form from input and output still remained a
big problem, due to the assumed linear behavior of the system and errors in the data. Shape and
volume constrained unit hydrograph with a smaller number of parameters was introduced to
overcome these problems (Nash, 1958,1959, 1960).

Box and Jenkins (1976) provided hydrologists with alternative methods of expressing the unit
hydrograph in terms of autoregressive moving average (ARMA). This school of thought
brought about the use of Artificial Neural Network (ANN) approach a method applied widely in
other fields of science and engineering to represent non-linear and dynamic system. An
example in the field of hydrology includes the recent work of Hsu et al., (1995). They made a
comparison of the ARMAX, the ANN and a conceptual model (the Sacramento soil moisture
accounting model, SAC-SMA). The comparison of the three models shows the superiority of
the performance of the ANN approach over the three others for a medium sized Leaf River
Basin near Collins Mississippi. However, the transportability of the model and internal structure
of the ARMA and ANN model and their relation to the physical process in a catchment remains
ambiguous.

In the 1960s, in search for more physical understanding of the rainfall runoff processes other
approaches were considered. With the advent of the computer capability the search for an
improved rainfall-runoff methodology accelerated rapidly. Several models, which represent the


20
single hydrologic cycle at catchment level by using interconnected conceptual elements, were
developed. Among the more widely used and reported lumped watershed models are Stanford
model IV (Crawford and Linsley, 1966), Sacramento River (Burnash et al., 1973), HEC-1 U.S.
Army Corps of Engineers (Rockwood and Nelson, 1966), Tank and so on (WMO, 1975). These
models are designed to approximate within their structures (in some physical realistic manner)
the general internal sub-processes and physical mechanisms which govern the hydrological
cycle. The main problem of these models was the fact that they have a large number of
parameters, which have to be optimized.

During the late seventies and beginning of the eighties, the real time forecasting model which
are based on recent updating and calibrating techniques. The Theoretical concepts used were
due to Kalman, (1960), Kalman and Bucy, (1961); Todini, (1978); O' Connell, (1985) and etc.

The lack of a one to-one relationship between the model and the reality gave rise to a
development of a so-called physical based models such as SHE (Systm Hydrologique
Europen) (Abbot, 1986). This model is believed to be the state-art of the physically based
model. A fourth generation model MIKE SHE (Refsgaard and Storm, 1995) is now being
developed to apply for integrated water resource management. Typical areas of applications of
this model are river basin planing, water supply, irrigation and drainage, effects of change of
land use, and other ecological factors influencing surface and groundwater system.

To minimize the uncertainty of the parameters, lumped conceptual models are extended to
handle catchments as composites of sub catchments, which are divided according to
homogeneity in topography and land use. A good example of such development are
TOPMODEL (Beven et al., 1979, Beven et al., 1995) and ARINO (Todini, 1996). The main
concept behind these models is to give detailed account of very important topographic and
hydraulic characteristics of a watershed. The TOPMODEL is developed on the basis that the
catchment storage is related to the local water table in which the main factor is the topographic
index known as Kirkby index (Beven and Kirkby, 1979). Therefore a basin is subdivided in to
segments according to topographic index. Segments with a similar value of index are assumed
to have the same response irrespective of their location in a catchment. One can argue that
though the topography or slope plays great role in flow generation there is still a gap of
understanding whether the effect of adjacent planes in catchment subdivision dominates how
flows are transferred from one subdivision to the other. At this stage it is interesting to consider


21
the subdivision of a catchment used in the ARNO model (Todini, 1995) which looks more
realistic. In the latter model, the catchment is divided into series of subbasins to each of which
the rainfall-runoff model is applied; this division takes place according to the natural subbasin
boundaries so that the subbasin closing section coincides with the cross-sections of interest
along the river and its tributary.

There is still a point of argument that once the model is described with a number of parameters,
they are mostly calibrated with the judgement of the model ability to reproduce observed
quantities (such as flow at the outlet). The problem is that when one has several subbasins, the
number of parameters to be optimized is multiplied by the number of subbasins. Optimizing
these parameters always solely depends on one time series (observed flow). Therefore the
parameters determined are very much correlated and competition among them is also very high.
One possible attempt to solve these problems is to obtain all the parameters from physical
properties of a catchment as for example in the SHE model. Which of course requires detailed
spatially distributed information over the catchment.

The availability and use of digital elevation model (DEM), geographical information system
(GIS) and remotely sensed data for inferring (soil moisture and vegetation index) somewhat
moderate the problems huge data requirement of distributed models.

In this study we focus on areas where even major water balance variables such as rainfall and
evaporation time series are sparsely and scarcely available. The scarcity of data in most of
developing countries are so alarming that one gets a single station for a catchment as large as
1000 km
2
. It is also possible a station from neighboring catchment should be considered for
those catchments, which have no observation at all. For such areas the pragmatic solution is to
apply conceptually sound lumped rainfall-runoff models with fewer parameters.
2.4 Time resolution of modelling

Hydrological processes occur at a wide range of scales, from unsaturated flow 1m soil profile to
floods in river system of a million square kilometers; from flashfloods of several minutes
duration to flow in aquifers over hundreds of years (Blschl and Sivapalan, 1995). There exist
rainfall runoff models which represents these processes using time spans a few minutes, hours,
days or even up to one year. Selection of a time step of modelling depends on:


22

Catchment input characteristics such as dominant storms in the area. An area with commonly
conventional type of storms with separate rainfall storms can only be modeled with time
steps less than a day.
The model structure in representing the time scale of hydrological processes. A model,
which accounts for infiltration rate of loose soil, should have computation in terms of
minutes. The time step is interrelated with the area of the catchment under study.
Starosolszky (1987) gives an approximate representation of this relationship (Table 2.1)
The scope (purpose) of the model

Table 2.1 Relationship between time step of modelling and area of catchment (Adapted from
Starosolszky (1987)

2.4.1 Common monthly water balance models

The main structure of a monthly water balance model attempts to model the hydrological
processes by conceptualizing the catchment as interconnected storages, through which the water
passes from input as rainfall to output as stream flow. The models at this scale do not reflect
implicitly the hydraulic part of the catchment process but rather depend on the empirical
relationship between storage and extraction rates and water balances


23

Monthly water balance models have been in hydrologic research as early as 1940s. The first
monthly water balance models were developed by Thornthwaite (1948) and later revised by
Thornthwaite and Mather (1955, 1957). A review of water balance model in common use is
given by Alley (1984) and recently by Xu and Singh, (1998). The models discussed are
summarized in Table 2. 2.

The available models use rainfall, temperature or evaporation either on monthly or daily time
step. Most of the models have few parameters compared to the daily rainfall runoff models. One
of the common features of these models is that most of the parameters (if not all) are estimated
by fitting the observed hydrologic data such as rainfall and stream flow (Xu and Singh, 1998).

The models that use rainfall and evaporation as input are usually found to be more realistic,
especially in reproducing seasonal flows and intermediate water balance variables. Attempts at
developing models with input of rainfall and temperature resulted either with over
parameterization or non-realistic storages as reported by Alley (1984). The VUB monthly water
balance models have been tested by more than 100 catchments in different climatic regions and
proved to satisfy most of the requirement at this scale (See Xu, 1992, 1995, 1996 Vandewiele
and Ni-Larwin, 1998). Compared to the other monthly water balance model with almost same
complexity the VUB monthly water balance models are reported performing better (Xu, 1994).
The overview of the VUB monthly water balance model is given in the following section.


24
Table 2. 2 Common monthly water balance models

Input Name/Reference
Number of
Storages
Number of
Parameters
T-model Thornthwaite
and Mather (1955)
2 2
T -Model Alley (1984) 2 3
Monthly
rainfall, Temperature
abcd-Model Thomas
(1981)
2 4
Pitman (1973,1978) 2 12
Roberts (1979) 2 8
Hughs (1982) 1 8
Salas et al. (1986) 11
VUBMOD Vandewiele et
al. (1992)
1 3
Monthly
Rainfall and
Evaporation
Makhlouf and Michel
(1994)
1 2
Haan (1972) 2 4
Kuczera (1983) 2 9
Daily
Rainfall and
Evaporation
McMahon and Mein,
Bouhton (1973)model
3 10

2.5 Review of the VUB monthly water balance models

This section reviews the development of a parsimonious water balance model and describes the
basic hypothesis and principles. Discussions on the applicability of the model to different
climatic regions and highlights of further room for improvement of the structure of the model
for the given time step are highlighted.

2.5.1 Model hypothesis

The VUB monthly water balance model was designed based on the conceptualization of a river
basin as a system and book keeping the water balance of various processes on a monthly time
base. The model is intended to be rather simple and requires readily available
hydrometeorological data. A water balance model on a monthly basis with a single reservoir
(Figure 2. 6) uses precipitation and evaporation rate as input and has flow of the river as output.
All the water balance terms are on monthly basis and expressed in mm. The model is generally
described with three basic equations.


25
1) Computation of the actual evaporation r
t
as a function of rainfall and soil moisture m
t
of
month t,
2) Computation of stream flow as a function of the amount of rainfall, p
t
and soil moisture
storage m
t
.
3) Computation of the water balance, at the end of the month.

Figure 2. 6 Diagrammatic representation of the VUB monthly water balance model.

Originally, Van der Beken (1977) proposed the above three relationships with the following set
of equations.

(2.2) ) 1 (
1 t
m a
t t
e e r
= 0
1
> a

where e
t
is potential evapotranspiration and a
1
is a parameter.

Here the actual evapotranspiration r
t
is represented as a function of potential evaporation or
pan-evaporation which ever is available. The actual evaporation increases as the soil moisture,


26
m
t
increases. The actual evaporation approaches potential evaporation, as the soil moisture is
very large. The values of parameter a
1
are high for sandy texture soils.

The surface outflow, d
t
is represented as follows

t t t
N a m a d
3 1 2
+ =
0 1 ,
3 2
a a (2.3)

The first term of the right side of Equation 2.3 represents the slow flow component and the
second term represents the fast flow component which is proportional to the net rainfall N
t
= P
t
-
r
t
. The parameter a
2
is expected to increase when the soil texture is more sandy and parameter
a
3
is expected to increase with degree of urbanization and average basin slope.

The water balance equation is written as:

t t t t t t t
T R d r P m m + =
+1
(2.4)

R
t
is a variable that takes care of the unaccounted inflow and outflow from the catchment if
little is known about this component it may be represented as

5 4
a m a R
t t
= 1 0
4
a (2.5) a
5
0 >

The variable T
t
represents the known amount of water supply or abstraction from the catchment.
Optimization of the above parameters was performed by minimizing the sum of error squares
i.e. the deviation of the observed discharge q
t
from the computed discharge d
t
. This model was
reported successfully applied to the catchments Grote Nete and Zwalm, located in North
Belgium (see Van der Beken, 1977 and Van der Beken and Byloos, 1977)

Based on the above general structure of the model, several researchers either applied the model
directly to new catchments or modified some of the equations. Among others Quasim, (1984)
introduced another evaporation equation whereby the actual evaporation is computed as a
fraction of the potential evaporation, depending on a ratio of the amount of rainfall to the
potential evaporation.


27

( e e r =
(2.6)
) 1
/
1 t
e
t
p a
t t
a
1
0 >

According to the above equation, for months without rainfall the actual evaporation is zero,
which is not necessarily true because evaporation from the soil moisture and transpiration from
plants may take place. This type of model may be limited to regions where rainfall in a month is
always greater than zero. Hassan (1984) suggested different forms of evaporation equations
which are based on the principle that the actual evaporation is a fraction of the potential
evaporation and represent the relationship which accounts for soil moisture, and the available
rainfall.

Substantial improvements of the monthly water balance model were carried out by, Vandewiele
et al. (1993), where a methodology for model formulation, parameter estimation and model
testing are developed.

After comparing different equations they suggested the following two equations for the
computation of the actual evapotranspiration.

] ), 1 ( min[
/
1 t
t
e
t
w
t t
w a e r = 0 1
1
< a (2.7)

] ), 1 ( min[
1
t
a
t
e
t t
e e w r
=
1
0 a (2.8)

In here the available water during the month, w
t,
is defined as the sum of the moisture storage
and precipitation during that month i.e.

t t t
p m w + =
1
(2.9)

It is pointed out that the difference between the above two equation is that in second equation
the actual evaporation approaches to total available water when evaporation is very high which
is not the case in first equation. These two equations seem to combine both the influences of the
available water and also the available energy for evaporation.


28
According to the evapotranspiration equations 2.7 and 2.8 the computation of the actual
evapotranspiration holds a time invariant parameter a
1.
In reality, because of the seasonal
variation of the level of vegetation growth and cover, this parameter should vary. Ni-Lar-Win
(1994) proposed to account for such variation and the parameter a
1
to have a variable value
represented as:

)] (
12
2
[
5 4 1
a t Sin a a a
t
+ =
(2.10)

Note that this increases the number of parameters from 3 to 5. A comparison between the results
of evaporation equation with constant and time varying parameter for 88 catchment from
(Belgium, China, Burma, and West Africa) showed a slight improvement of the model
performance for models with variable evaporation parameter. Dias (1993) applied Dissociated
models with the exponential function of evaporation equations in 7 West African and 3 Srilanka
catchments (Semi-arid catchments) and reported that the results are only as good as the previous
rather simple models.

The flow equation is also modified by introducing discrete parameters to relate flow with soil
moisture. Xu (1988), based on the concept of variable source area, showed that the fast flow
component is not only a function of effective rainfall but also the wetness of the soil. Hence the
flow equation is adjusted as follows:

t
b
t
b
t t
N m a m a d
2
1 3
1
1 2
+ = 1 , 0
3 2
a a (2.11)

and N
t
is also redefined as:

(2.12) ) 1 (
/
t
r
t
p
t t t
e r P N
=

Due to the high correlation between the pairs (a
1
, b
1
) and (a
2
, b
2
), specific discrete values were
assigned to (b
1
=0.5,1,2) and (b
2
=0.5,1,2). Each of these combinations would result with a
specific model for a catchment. The physical significance of these parameters and computation
of the initial values from catchment properties are studied by Xu (1988) and where a
2
is found


29
to depend upon the transmisivity of the aquifer of the catchment and a
3
depends mainly on the
slope and imperviousness of the catchment.

2.5.2 Models for different input levels

Owing to the lack of evapotranspiration data the monthly water balance models were modified
or augmented for different input levels. What is really modified in this cases is that the driving
force for water extraction from the catchment e
t
is represented as a function. The first attempt is
to represent the e
t
input series as a function of monthly mean temperature, c
t
and mean
humidity, h
t
which are the dominant factors for evaporation process. The function used is

(2.13)
t t t
h c a e
2
4
) (
+
=

where a
4
is a positive valued parameter.
c
t
=temperature in degree Celsius (the plus sign is to indicate that zero values are taken
for temperatures below freezing.
h
t
=relative humidity in per cent

If data on relative humidity is not available e
t
may be written as

2
4
) (
+
=
t t
c a e (2.14)

If only rainfall data is available,

)] (
12
2
[
5 4 6
a t Sin a a e
t
+ =
(2.15)

Equation 2.15 is justified since the e
t
series generally varies regularly in most region of the
world and it is nearly a periodic function with a more or less sinusoidal function.


30
The e
t
series generated by one of the equations 2.13, 2.14 and 2.15 is not intended to compute a
physically relevant potential evaporation, therefore these are only intermediate driving forces
computation for the water balance model.

Table 2. 3 model types and input requirements
MODEL
TYPE
INPUT DATA PARAMETERS
1 Precipitation and Potential evapotranspiration a
1
, a
2
, a
3

2 Precipitation, Temperature, and Relative humidity a
1
, a
2
, a
3,
a
4

3 Precipitation and Temperature a
1
,a
2
, a
3,
a
4

4 Precipitation

a
1
, a
2
, a
3 ,
a
4,
a
5,
a
6

2.5.3 Regional study of monthly water balance models

Along with the development of the monthly water balance models efforts were also directed to
relate parameter values to catchment properties. The model parameters could be linked to
physical basin characteristics such as for example lithology (Xu, 1988, Vandewiele et al.,
1991).

In the context of Belgium (Vandewiele and Atlabachew, 1995) the monthly water balance
model parameters could be regionalized, which enables deriving flows for ungauged catchment.
A method of Kriging and parameters from few neighbouring basins were used to obtain the
parameters for ungauged catchments. Further research along this line is necessary to analyze the
possibility for regionalization of the water balance models in other climatic regions.

2.6 Summary

In order to select among the plethora of different mathematical models available today, it is
possible to identify models according to a priori knowledge. The latter ranges from total
ignorance (pure stochastic models) to the full description of system dynamics based upon
differential equations describing the balance of mass and momentum. The final choice on which
model to use mainly depends on the purpose of application, the accuracy desired, the available
data and economics.


31

Conceptual models are widely used in water balance studies and chosen in this study
particularly due to:

their ability in representing the system in terms of parameters.
their potential to correlate the few parameters with hydrometeorological and physiographical
characteristics.
their ability to carry out computations on time scales varying from hours to years. The time
step chosen depends on the application of the model. Daily and monthly time steps are
common in water balance models for basins. Important to note is that the smaller the time
step, the more complicated the model becomes and the more stringent are requirements on
the data.
their use of readily available hydrometeorological data (as mean or totals of daily, weekly or
monthly values).

The physical based distributed models are appropriate to study land use change impacts on the
catchment system and also to investigate test catchments to prove hydrological hypotheses.
There are a number of sophisticated and detailed models appearing in the field of hydrology but,
for many catchments, the available data are far from the requirements of these models. In
addition arid and semi-arid catchments are very difficult to model as compared to humid areas
due to the large temporal variation of the variables and the limited data.

Therefore, the focal point of the ongoing research is to develop conceptual rainfall - runoff
models on a catchment scale and at different time steps. Hereby it will be attempted to
incorporate the specific features of semi-arid and arid catchments.

Since 1977, in an effort to develop simple conceptual models for limited available data, monthly
water balance models have been developed and tested for over 100 catchments from 8 countries
at the Laboratory of Hydrology of the Vrije Universiteit Brussel. Most of the test catchments lie
in humid climatic regions and a few in arid and semi-arid regions. The result of the study could
be summarized as follows:

1) the models perform satisfactorily in humid regions


32
2) model parameters of a few regions from Belgium could be related to lithology and also
geographically regionalized.
3) application to some arid catchments resulted with high correlation of residuals
4) the models show a low performance for intermittent flows.

Further areas of improvement addressed in this study are:

In semi-arid catchments, it is known that the potential evaporation is higher than the actual
evaporation. In all cases, the limiting factor for actual evaporation is the available water
rather than the potential evaporation. Model structures, which include this phenomenon,
should be investigated.

In the model formulation, flow components are expressed as a fraction of the soil moisture
storages of the previous month. This leads to under estimation of the runoff generation
during the transition from a very long dry period to a rainy period. This phenomenon often
prevails in semi-arid and arid climates. Therefore, a means of accounting this transitional
period should be investigated.

The model structure does not have an upper limit for the storage of the catchment. In reality,
the storages are limited to a saturated condition. Hence an upper limit should be included in
the structure.

In dealing with semi-arid and arid regions, due to the high variability of input variables,
there is a need for smaller time step of modelling. A 10- day time step can be considered as
a homogeneous time span for example for the soil moisture content. Hence a 10-day water
balance models can be developed based on the experience of the previous studies.

Further, the concept of the water balance can be extended to develop a parsimonious daily
rainfall runoff model at catchment scale. Smaller time step models are applicable in water
resources structure operation and also in flood forecasting.


33

Chapter 3
3 MODEL STRUCTURES FOR
DIFFERENT MODELLING TIME STEPS

3.1 Introduction...................................................................................................................................34
3.2 The monthly water balance model for semi-arid catchments ........................................................35
3.3 A 10-day water balance model (DWBM)......................................................................................52
3.4 A parsimonious daily rainfall - runoff model ................................................................................64
3.5 Data requirement and scope of application of the developed models ...........................................71

3.1 Introduction

Conceptual monthly water balance models have been developed and tested for a number of
catchments in humid climates and few arid and semi-arid regions at the Laboratory of
Hydrology, Free University of Brussels. These models are based on a simplified representation
of the hydrological processes and the balances of incoming and outgoing water from the
catchment on a monthly time step. This chapter presents further developments and
improvements of the model structure to extend their applicability to semi-arid and arid
catchments. Secondly, the chapter discusses the development of a new 10-day Water Balance
Model (DWBM) and of a Parsimonious Daily Rainfall-Runoff Model (PDRRM).
The last two models have evolved from the long experience obtained from the monthly time
step model as applied to different climatic regions. The methodology of parameter optimization
and the main philosophy of modelling are kept as for previous VUB monthly water balance
models.

3. Model structures for different time steps

35
3.2 The monthly water balance model for semi-arid catchments

3.2.1 Structure of the model

The models have generally three components.

(1) Computation of actual evapotranspiration, r
t
which is empirically related to the areal
potential evaporation, e
t
and the available water in the catchment expressed as the
storage index m
t
.
(2) Computation of flow, d
t
which is partitioned as fast flow, f
t
(a function of excess rainfall
and wetness of the catchment) and the slow flow, s
t
as a function of catchment storage.
(3) The water balance of the single storage certainly augmented by rainfall p
t
and depleted
by the flow of the river and evaporation.

All variables are in the same unit and are normally expressed as mm of water depth over the
surface area. The subscript t is used to represent the time horizon. For the variables: rainfall,
evaporation and flows the subscript t indicates the total amount of the variables during time t.
For the soil moisture storage variable m
t-1
indicates the state at the beginning of the particular
time (month in this case) and m
t
indicates the state at the end of the t. Figure 3.1 shows the
schematized water balance variables.


36

r
t
p
t

f
t

m
t-1
m
t
time
s
t

Figure 3.1 Conceptual representation of the monthly water balance variables.

Rainfall
Soil moisture &
Groundwater storage
Evapo-
transpiration
Fast flow
Slow flow
Total flow

Figure 3.2 Conceptual representation of the monthly water balance model.

Actual Evapotranspiration

Evaporation of water from a soil is dependent on many physical parameters of the soil and
hydrometeorological variables. On annual basis, evaporation extracts as much as 40 per cent to
90 percent of the water in humid and dry regions respectively. Practically, the actual
evaporation from a natural surface is the most difficult to measure. Two existing methods of
estimating the evaporation from land surfaces are based upon changes of the water content of
the soil. This can be done by periodically measuring the soil moisture profile and by comparing
successive profiles. The second method is by using a Lysimeter, a tank filled with soils in
which all conditions inside and around the tank are reproduced as exactly as possible. Note that


37
the first method has limitations because the computed difference can comprise drainage to the
bottom or lateral strata of the soil in addition to evaporation. The second method gives relatively
accurate data for a specific local condition. However installing, such an instrument to cover a
large heterogeneous catchment is not practical from economical point of view.

The obvious fact is that the actual evaporation from a heterogeneous land cover depends on a)
the climatological factors b) the land cover and c) the moisture availability. At present the
theoretical studies and observations, regarding explicit relationship among the factors are not
sufficient to adapt them to catchment studies because they are microscale and time span
oriented. This leads to incorporating the actual evaporation calculation as an intermediate step in
conceptual rainfall runoff models.

To compute actual evapotranspiration from known water balance variables, the following basic
assumptions are necessary:

a) the superimposed effect of climatological driving forces (factors) can be represented by the
potential evapotranspiration or pan evaporation;
b) for fairly sparse vegetation cover and rural catchments, the effect of interception may be
neglected.
c) at the larger scale of time step (decade and monthly) the actual evaporation is considered to
be that amount of water extracted from the soil moisture which has been filled with not only
rainfall of the current month but also the previous months. This overcomes the problem of
deficiency of water (rain-potential evaporation) due to a large value of the potential
evaporation, as compared to the rainfall amount during the time step. Further it can be
justified that rainfall events are rather short compared to continuous process of extraction of
water. Hence there should be a mechanism of delaying the water in storage.

Following the above assumptions, empirical relationships between the actual evaporation (r
t
)
and the other water balance variables (e
t
, p
t
, m
t
) have to be established. In most of moisture
accounting models such as the Soil Moisture Accounting and Routing Model known as Layer
model (O' Connell et al., 1970) evaporation is allowed to extract water at potential rate from the
most upper stack of soil layers. After the exhaustion of the upper layer, evaporation takes place
from the second layer at the potential rate multiplied by a parameter whose value is less than
unity. If the potential evaporation is not satisfied at the exhaustion of the second layer the


38
evaporation takes place at the potential rate multiplied by the square of the parameter and so on.
A similar approach is used by Xinanjiang model (Zhao et al., 1980).

At least for the monthly water balance model considered here, the layers are amalgamated to a
single storage. This assumption is fairly valid for a shallow unconfined aquifer beneath the root
zone. In the next section the necessity of sub-dividing this storage into two namely the soil
moisture and the deep groundwater storage is discussed with respect to smaller time step
resolutions.

Any empirical relation between actual evaporation and known variables should satisfy the
following requirements. Let the available water in the soil storage at time t be w
t
=m
t-1
+

p
t
, the
potential evaporation e
t
, and actual evaporation during the month t= r
t
,

a) r
t
increases with e
t
and w
t

b) r
t
=0 when w
t
=0 or e
t
=0.
c) r
t
e
t
and r
t
w
t

d) r
t
e
t
when w
t
.

Xu, (1992) tested seven equations (from Roberts, 1978) and two equations from the VUB
monthly water balance model. All the tested equations satisfy the above requirements. He
concluded that the latter two equations are overwhelmingly good. Though some variants of the
seven equations also perform well for the monthly water balance model, the two equations -
hereafter also refereed as evaporation IR=1 (Equation 3.1) and IR=2 (Equation 3.2) are used in
this study.

] ), 1 ( min[
/
1 t
t
e
t
w
t t
w a e r =
1
a 0 1 (3.1)
] ), 1 ( min[
1
t
a
t
e
t t
e e w r
=
1
0 a (3. 2)
where a
1
is a time invariant parameter, which represents the characteristics of a catchment under
study.

Basically in evaporation equation (3.1) the actual evaporation r
t
is computed as a portion of
potential evaporation which is bound by the available water in the catchment during month t.
Alternatively the actual evaporation may be computed as a portion of the available water which


39
in this case is limited to the potential evaporation equation (3.2). Their behavior in the range of
possible values and influences of the corresponding parameters are shown in Figure 3.3 and
Figure 3.4 for the first and second evaporation equations respectively.

It is evident that in semiarid and arid catchment, the potential evaporation is higher than the
actual evaporation for a large period of the year. Hence, the limiting factor for evaporation in
this region is the available water rather than the potential evaporation. Relations such as
equation (3.2) have therefore a better background since the actual evaporation is a fraction of
the available water rather than being a fraction of potential evaporation. Comparison of the
model outputs for most of the test catchments shows that equation (3.2) resulted with mostly
better and in a few cases equal model performance compared to equation (3.1).

Ni-Lar-Win (1994), and Vandewiele and Ni-Lar-Win (1998) proposed to vary the parameters of
evaporation seasonally with a sinusoidal function. Hughes (1997) and Hughes and Metzler
(1998) also introduced a potential vegetation growth factor as a function of the last two months
and current soil moisture states in the Pitman monthly rainfall runoff model. The modifications
are necessitated from the fact that a single wet month preceded by two dry months will therefore
have less effect than as sequence of moderately wet months. These different studies came
however to the conclusion that separation of the vegetation factor and soil moisture storage led
to difficulties in calibrating the models with more additional parameters.


40

Figure 3.3 Influences of evaporation parameter for evaporation equation 3.1. The left diagram
shows curves with w
t
=constant. The right diagram shows curves with e
t
=constant. (Vandewiele
et al., 1993)

Figure 3.4 Influences of evaporation parameter for evaporation equation 3.2. The left diagram
shows curves with w
t
=constant. The right diagram shows curves with e
t
=constant. (Vandewiele
et al., 1993)


41
Empirical equations relating the actual evapotranspiration with soil moisture storage and
potential evaporation are used for all the three types of models. The effect of growth of
vegetation has not been considered explicitly. It is assumed that this phenomenon is well
represented by the soil moisture available in a catchment.

Flow components

River flow is composed of several components such as, channel flow, surface runoff, interflow
and base flow. Modelling each component in a catchment scale is rather complicated and
requires a detailed representation. Especially in large time scale modelling, it is often impossible
to make a clear distinction between each component. In the present study, the flow components
are conceptually grouped in to two main domains namely the fast and slow flows. The fast flow
may be defined as the part of precipitation that reaches the outlet within the time step of
modeling, in this case a month. It is comprised of channel flow and surface flow and part of the
interflow. The slow flow, which is gradually released from the groundwater storage, may be
regarded as the "base flow" component.

The fast flow component

Fast flow is considered as the immediate response of a catchment to a given precipitation and
hence can be represented as the difference between rainfall and actual evaporation. The
equivalent term to the so-called net rainfall or effective rain n

is given as:

) 1 (
/
t
r t
p
t t t
e r p n
= (3.3)

The flow generated depends on the evaporation, excess rainfall and also the wetness of a
catchment, which is expressed as m
t-1
. The (+) sign indicates that m
t-1
is restricted to non -
negative value. One of the efficient relationships tested for different climatic regions is

(3.4)
t
b
t t
n m a f
1
1 3
) (
+
=


42
where a
3
and b
1
are parameters. It was observed that parameters a
3
and b
1
are highly correlated.
Hence it was suggested to limit the later parameter to fixed values (0.5, 1.0, 2.0) Experience
showed that one of the discrete values fits a given catchment.

The consequence of this equation is that a precipitation, which occurs after several dry months,
would produce a fairly lower magnitude of fast flow compared to that occuring in the middle of
the wet season. In essence this is a sort of variable source area concept, i.e. the wetter the
catchment the higher will be the generated surface runoff.

Referring to equation (3.4) the fast flow component is a function of the soil moisture storage at
the beginning of a month. In a temperate area where these formulations are proved to be
working efficiently, the moisture index rarely drops to a value of zero (see for e.g. Figure 3. 6).
Whereas in semiarid and arid regions since the dry seasons last for longer periods the index can
have a value about zero for more than one month (see for e.g. Figure 3.7). It is noted that in
these equation if the soil moisture storage at the beginning of a month is zero we would have no
fast flow for the next month for whatever amount of rainfall.

The rainfall of semi-arid and arid areas is associated with high intensity. Also, at the beginning
of the rainy season, the vegetation is normally dry and scarce which reduces the retention of the
rain. Though the initial infiltration rate is high, because of the larger intensity of the rain it is
more likely that it produces some fast flow. This process is not well represented by equation
(3.4) i.e. regardless of the amount of effective rain, the fast flow component is not generated if
the soil moisture index is zero. Hence for the fast flow equation (3.4) an average soil moisture
index of the month is used instead of the storage index at the beginning of the month. Therefore
Equation 3.4 is rewritten as:

t
b
t t
n m a f
1
3
= (3.5)

where
2
1 t t
t
m m +
=
m


43

0
50
100
150
200
250
0 12 24 36 48
Months starting January, 1983
P
r
e
c
i
p
i
t
a
t
i
o
n

(
m
m
)
60

Figure 3.5 Typical monthly rainfall series of humid climate (e.g. from Kleine Nete catchment,
Belgium)

0
50
100
150
200
250
0 12 24 36 48
months Starting January 1983
S
o
i
l

m
o
i
s
t
u
r
e

s
t
o
r
a
g
e

(
m
m
)
60

Figure 3. 6 Typical evolution of soil moisture storage in response to humid climate
precipitation. (e.g. from Kleine Nete catchment, Belgium)


44
0
50
100
150
200
250
300
0 12 24 36 48
Months starting January, 1981
P
r
e
c
i
p
i
t
a
t
i
o
n

(
m
m
)
60

Figure 3.7 Typical monthly rainfall series. (e.g. of Faleme River, Senegal)

0
50
100
150
200
250
300
350
400
0 12 24 36 48
S
o
i
l

m
o
i
s
t
u
r
e

s
t
o
r
a
g
e

(
m
m
)
60

Figure 3. 8 Typical evolution of soil moisture storage in response to strong contrast between
rainy and dry season (e.g. of Faleme River, Senegal)


45
To obtain the average soil moisture an iterative method is introduced which is performed for
every computation step. Firstly the initial
t
m

is assumed to be m
t-1
, then all computations are
executed and the resulting final storage, m
t
is computed. Next the average is computed and
compared with the previous value and the iteration will continue until the difference between
the two consecutive mean storages is smaller than the defined accuracy (0.01mm). This
modification does not add any other parameter but the computation time during calibration will
take more time since the iteration is done for the individual time steps of the computation. This
is not considered to be a limiting factor with the modern fast computing facilities.

The slow flow component

Let us consider the steady state of groundwater flow for schematic representation of a catchment
of Figure 3.9. From Darcy's law and Dupuit's assumptions, the groundwater flow component of
the river is proportional to the transmissivity and the hydraulic gradient. For the total flow
during a period of one month per unit area, this becomes:

t
o
b
L
L
h h T
Q
=
)
2
/( ) (
(3.6)

where: T is the transmissivity: the hydraulic conductivity times the average thickness of the
ground layer

t
= time step
h is the average water table elevation
h
o
is the average water level in the river channel
L is the average width of a cross section and L/2 is assumed to be the average distance
between the groundwater and the river.


46
G.W.T
Groundwater flow
evapotranspiration
Surface runoff
Precipitation
River
Water divide
h
0
L
h

Figure 3.9 Physical description of the flow components (Xu, 1988)

From Figure 3.9 it is clear that the groundwater per unit area that can drain to the river system is
proportional to the difference between the two heads (h-h
o
). The total amount of water available
in this layer is then:

) (
o
h h m = (3.7)

where m is the storage and ,is average effective porosity. Hence we have:

m
L
T
Q
t
b
2
2
= (3.8)

It follows that the groundwater contribution is largely a function of water table changes
reflected in T and other physical characteristics of the catchment. Two extremes can be
considered. First, if we assume water level fluctuations are very small compared to the
thickness, then T can be considered as constant and


47
m a Q
b 2
= with
2
2
2
L
T
a
t
= (3.9)

On the other hand if the thickness of the groundwater layer below the river h
o
is much smaller
than the water table drop h-h
o
, then

/ ) ( Km h h K T
o
= = (3.10)

where K is hydraulic conductivity of the layer.

2
2
m a Q
b
= with
2 2
2
2
L
K
a
t
= (3.11)

This analysis shows that for the intermediate conditions one has to look for optimum values by
changing all possible values. The factor L can be calculated as the catchment area divided by the
total length of the river. Therefore, for general condition the low flow component is represented
as :

2
2
) (
b
t t
m a s = (3. 12)

Like b
1
, b
2
is also restricted to have values of (0.5,1.0 and 2.0). Note also that the average
storage in a month is used here also for the same reasoning as for the fast flow case.

The total flow

Eventually the total flow during the month t, d
t
is given as the sum of the previous two
components equation (3.13). No delay of flows is considered, which limits the model to apply to
catchments with a concentration time not more than a month. As a rule of thumb, the upper area
limit is considered to be about 5000 km
2
. This is a very hypothetical limit, since other factors
such as the shape of the catchment and slope also affects the concentration time. A major
limiting factor is also the homogeneity of rainfall and evaporation inputs over the catchment
considered as a unit.


48
t t t
s f d + = (3.13)
The water balance

The water balances of the catchment i.e. the input to and the output from the catchment is
computed for month to obtain the soil moisture available at the end of each computation.

m m p r d
t t t t
= +
t

1
(3.14)

Maximum soil moisture storage

Though the described model is not a physical model whereby all the complex processes are
strictly represented, one would expect the storage index to have an upper limit by analogy: the
field capacity of the soil or, for shallow groundwater tables, the ground surface. An additional
parameter a
4
(maximum soil moisture storage) may be introduced whereby storage values
exceeding a
4
augment directly the fast flow and the soil moisture at the end of the month will
have the value of the upper limit, a
4
. Consequently considering the later modification, Equation
(3.5) may be written again as:

| | { }
f a m p r p r m a
t t
b
t t t t t
= +
+
3 4
1
1 ( ) exp( / ) ( ) (3.15)

This is to be considered as the second variant of the described model. The (+) sign indicates that
only non-negative values are considered. This modification also costs one additional parameter.
One should pay attention to the values of this parameter because of the fact that if very low
values are obtained by optimization and in situations where we have consecutive rainfall events,
the storage would likely be exceeded and the fast flow generated is going to be unrealistic. In
such cases, one has to gradually increase this value to obtain realistic upper limits. On the other
hand, if one starts with very high upper limits, there will not be any influence on the flow
generation. More explanation will be provided in the case study (Chapter 5).


49

3.2.2 Models for rainfall and temperature inputs

Previous studies on these types of modelling have shown that any attempt of computation of
evaporation form other meteorological data such as temperature and humidity within the model
structure adds one or two more parameters. It often results with unrealistic intermediate
variables.

In this study we recommend that in the case of an absence of longer potential evaporation:
a) If data for a number of years are available, calculate the mean monthly evaporation for the
12 months of the year, and use these values for all the calibration period.
b) If data of evaporation for few years and longer data of temperature and/ or relative humidity
are available establish a simple relationship on a monthly basis. Use this relationship to
derive potential evaporation, which eventually would be the input to the models.
c) If there is no data on evaporation but there is a time series of temperature, a simple formula
can be applied to derive potential evaporation which does not require many variables. One
of the earliest formulas are those proposed by Thornthwaite (1948) and Blaney and Criddle
(1950) several other formulas are available with different input requirements. Recently
Singh and Xu, (1997) have evaluated and generalized them into seven categories.

The Thornthwaite equation is applied here because of fewer data requirements. The method
computes the potential evaporation from series of mean temperature and average monthly
sunshine hours.

The equation calculates monthly potential evaporation as

a t d d
t
)
I
c 10
(
30
N
12
l
6 . 1 e = (3.16)

where e
t
= potential evaporation (mm),
l
d
= actual day length in hours,
N
d
= the number of days in the month,
c
t
= mean monthly air temperature (
0
c),


50
( )
514 . 1
12
1 t
t
5 / c I
=
= and (3.17)
6 2 3
10 ) 492390 1 . 77 657 . 0 (
+ = I I a

The factors in the formula can be easily obtained from tables. The advantage of using this
equation is to reduce the number of inter-related parameters in hydrological models.

3.2.3 Parameter description

The first variant of the monthly water balance model, which is denoted as MWBM-A, has three
parameters. The second variant, which is denoted as MWBM-B, has one more parameter, the
upper limit of soil moisture storage in addition to the three parameters.

Taking average values of water balance variables and substituting them in model equations one
can obtain reasonable orders of magnitudes of the parameters. The Order of magnitudes of the
four parameters are shown in Table 3.1. For detailed explanation on the scale of parameters see
Appendix B.

Table 3.1 Description of parameters of monthly water balance model.
Parameter Description Initial values recommended
range
IR=1 0.4 0.1-0.9 a
1
evaporation parameter
IR=2 0.004 0.0001-0.009
b
1
=0.5 0.4 0.01-0.9
b
1
=1.0 0.04 0.001-0.1
a
2
slow flow Parameter
b
1
=2.0 0.0004 0.00001-0.001
b
2
=0.5 0.04 0.01-0.1
b
2
=1.0 0.004 0.001-0.01
a
3
fast flow parameter
b
2
=2.0 0.0004 0.0001-0.001
a
1
for

(MWBM-B)
max. soil moisture storage 400 100-1000


51

3.2.4 Initial soil moisture storage

The soil moisture storage at every time step is calculated using the other known water balance
variables and hence is known as an intermediate variable. But while calibrating or simulating
the model one has to start from known soil moisture storage. In fact this value is not measurable
because it is a conceptual variable of the whole catchment. Two possible solutions are available
to handle this problem. The first solution is to consider the initial soil moisture as an additional
parameter and obtain the optimum value through optimization methods along the other model
parameters. This of course is against the major objectives towards developing parsimonious
models.

The other solution which is adopted as a standard in modelling practice in this field of
conceptual rainfall-runoff modelling is to start with a reasonable assumption of the initial soil
moisture storage and run the model for a few years referred to as a 'warming up period'. The
idea is that the model performance should not be accounted for during this period and at the end
of the warming period it is assumed that the effect of the initial value would disappear and the
catchment start to respond naturally. The length of the warming up period depends upon a) the
availability of data and b) the memory of a catchment. In case of short data, mean monthly
values of input variables (precipitation and potential evaporation) can be appended to the
beginning of the data, which would be used as a warming up period. It is recommended that at
least one year of actual inputs should be used as a part of the warming up period.


52

3.3 A 10-day water balance model (DWBM)

3.3.1 Background

This model is based on the experiences gained from the extensive application of the monthly
water balance models discussed in section 3.2. The objective of this study is generally to reduce
the time step of the model to account for the high variability of hydrological variables such as
rainfall and soil moisture in a month for semi-arid and arid catchments. The first logical attempt
was to use the same model structure as for the monthly time step, which has resulted with
unsatisfactory results.

Further attention was given to the use of this model in catchments where there is a contrast
between the wet and the dry season. Semi-arid and arid regions are characterized by having at
least two distinct seasons: the rainy season and the dry season where no rainfall (insignificant)
but still flow is available in smaller magnitude. It is then possible to treat the two situations
separately, or this feature may be used as an indication of some processes governing the two
seasons. This prompted us in modifying the existing model. Most of the methodologies would
be adopted from the parent model and new concepts specific to this model are detailed in the
following sections.

The other peculiarity of this study is that among the parameters describing the model, the base
flow parameter is obtained by hydrograph analysis. This model is especially suitable for
regions where two seasons, rainy and dry seasons, are easily identified to obtain the base flow
parameter through hydrograph analysis.


53

3.3.2 Description of the 10-day water balance model

A Parsimonious (5 parameters) lumped conceptual rainfall-runoff model for a time step of 10-
day is described. The new features of the actual model deal with the reduction of the time step
to 10-day and with the introduction of two distinct storages: a soil moisture storage and a
groundwater storage. The time step of the model is reduced to 10-day to account to some extent
for the temporal variability of rainfall. Obviously, the variability of separate storms can not be
handled neither by the 10-day time step and requires modelling with shorter time steps such as
hourly or daily. However, a 10-day time step could be sufficient to represent the uniform state
of a catchment.

The separation of the storages is necessitated by the fact that the upper soil moisture storage
controls the processes during the rainy season and the lower part predominantly controls the
gradual release of water during the dry season. The upper storage is augmented by rainfall
(infiltration) and depleted by evapotranspiration, interflow and percolation to the underlying
groundwater storage. The groundwater storage is augmented by percolation and depleted by
base flow. The actual evaporation and flow components are calculated on a 10-day time step
using empirical equations, which relate the rainfall and the average states of storage. The river
flow is represented as the sum of three-flow components: surface runoff (infiltration excess),
interflow and base flow. The diagrammatic representation of the structure of the model is shown
in
Figure 3.10. The mathematical formulation of the components of the model is described in the
following sections.

All the water balance variables considered are described in (mm) of depth of water over the
surface area of the catchment. The subscript t designates the time step where t=1, 2,3,
corresponds to the first, second and third decades (10-day) etc. All quantities are assumed to be
uniform over the catchment.


54
Rainfall
Infiltration
Soil moisture storage
Groundwater storage
Evapo-
transpiration
Fast flow
Percolation
Interflow
Base flow
Total flow

Figure 3.10 Conceptual representation of the 10-day water balance model.

3.3.3 Infiltration component

Generally the infiltration rate depends on soil characteristics, vegetation cover, initial moisture
content of the soil and intensity of the rain. Horton (1939) represents rate of infiltration as a
decaying function starting at a rate i
o
until it reaches a constant rate i
c
.

(3.18)
t
h
k
c o c t
e i i i i
+ = ) (

where k
h
is a decay constant. Similar observations are noted by Philip (1957) who proposes the
following evolution of the infiltration rate.

h t
k t i + =
2 / 1
2
1
(3.19)

where is the sorpitivity parameter and k
h
is hydraulic conductivity. The above and other
theoretical derivations of the infiltration process imply that under favorable conditions where
there is enough water supply from the surface, there will be at least a minimum value of
infiltration rate taking place. The minimum value corresponds to the saturated hydraulic
conductivity of the soil (Figure 3.11). The problems that arise in applying the above hypothesis


55
in water balance models for larger time steps (10-day or monthly) are: within the time step of
computation one can have several such curves during sequences of storms separated by one or
more dry days. Moreover, if the infiltration rate is integrated over a time step, a large amount of
potential infiltration is expected which would not be satisfied with rainfall within the time step.
One can see from Table 3.2 that the total potential infiltration in one day is very high for all soil
types (except clay). Daily rainfall could not satisfy even the lowest integrated infiltration
amount using the saturated hydraulic conductivity.

i
c
i
o
I
n
f
l
t
r
a
t
i
o
n

r
a
t
e
,

i
Time

Figure 3.11 Evolution of the infiltration rate

Table 3.2 Ranges of values of hydraulic conductivity: (Todd, 1980)

Soil Type m/day m/10day
Sand course 45 450
Sand 12 120
Sand fine 2.5 25
Silt 0.08 0.8
Clay 0.0002 0.002

Hence using one of the common infiltration equations at this time step is not justified. In this
study we preserve the decay of the infiltration rate through the control of the storage of the soil.
We assume that the amount of rain infiltrated to the subsoil at a given time step is equal to the
soil moisture deficit defined as the difference between the storage capacity and the current soil
moisture content.


56
)) ( , min(
1 4
=
t t t
m a p i > a (3.20) 0
4

where i
t
is actual infiltration in mm per time step, a
4
is the storage capacity of the upper storage
and m
t-1
represents the soil moisture content at the beginning of time step t. The actual amount is
always the minimum of the deficit and the rainfall p
t
. The effect of antecedent rainfall is taken
care of by the soil moisture storage i.e. the deficit calculated will be lower in cases the
catchment is already wet. The graphical explanation of the above concept is shown in Figure
3.12. The available rainfall p
t
controls the actual amount.

p
o
t
e
n
t
i
a
l

i
n
f
i
l
t
r
a
t
i
o
n
Time
s
o
i
l

m
o
i
s
t
u
r
e

m
t
storage capacity

Figure 3.12 Relationship between infiltration and soil moisture storage.

3.3.4 Evapotranspiration component

The actual evaporation r
t
is computed as a portion of the available water, while limited to the
potential evaporation. In this model the average soil moisture is introduced as described in
section 3.2. The same second evaporation equation of monthly water balance model is used
with a 10-day time step with a slight modification. Instead of the available water w
t
in the
previous case, the average soil moisture storage is used in the equation. This is intended to
consider the influence of variability of the soil moisture storage as in the flow computations.

] ), 1 ( min[
1
t
a
t
e
t t
e e m r
= a (3.21) 1
1

where
2
1 t t
t
m m
m
+
=


57
and a
1
is a time invariant parameter, which represents the characteristics of the catchment under
study. e
t
is the potential evaporation. It is evident that the predominant factor for evaporation in
semi-arid and arid regions is the available water rather than the potential evaporation (Moore,
1989). Hence the above equation is applicable to catchments where the potential evaporation is
very high during long periods and where mainly the available water controls the actual
evaporation

3.3.5 Percolation component

Percolation, l
t
, from the upper storage
,
to the groundwater is proportional to the upper soil
moisture storage. The maximum percolation will be obtained when the upper storage is
saturated. In some models such as the Stanford watershed IV Model (Crawford et al., 1966)
percolation between layers is assumed to be not only function of the upper storage but also to
the storage in the underlying layer. This will only make a difference when we consider the detail
of the unsaturated zone between the two layers as capillary fringes, which account for very
small part of the total storage.

The following empirical equation is used to represent the percolation at time t.

t t
m a l
2
= 1 0
2
a (3.22)

where a
2
is a percolation parameter when the upper storage is saturated.

An alternative to the above formulation was proposed by Moreda and Bauwens (1998). Hereby,
the percolation was expressed as a function the ratio of the available moisture storage and the
maximum capacity of the storage i.e. l
t
=a
2
m
t
/a
4
.

This results in a physically meaningful relation
that when the available water is approaches the upper limit of the moisture storage, the
percolation rate would be a
2
. It follows that parameter a
2
is analogous to the saturated hydraulic
conductivity. This model structure was tested for two catchments and reported to show good
performance. However this structure leads to problems for the parameter optimization, due to a
high correlation between a
2
and a
4
.


58
3.3.6 Flow components

Three flow components are identified. The fast flow component f
t
generated whenever the
rainfall exceeds the infiltration rate within a time step considered.

t t t
i p f = (3.23)

The interflow s
t
is calculated as a function of the infiltrated water and the product of the wetness
of the upper storage expressed as m
t
and parameter a
3
.

t t t
i m a s
3
= 1 0
3
a (3.24)

The base flow component is computed as the release of the lower storage g
t
which is controlled
by base flow parameter a
5
.
t t
g a b
5
= 0 1
5
a (3.25)
where
2
1 t t
t
g g
g
+
=

The total flow d
t
is then,

t t t t
b s f d + + = (3.26)

3.3.7 The water balance

At each time step, the water balance is computed for the two storages. The water balance of the
upper soil moisture storage is computed at each time step as:

t t t t t t
r l s i m m + =
1
(3.27)

and the groundwater balance is computed as,


59
t t t t
b l g g + =
1
(3. 28)

For medium sized catchments no delay of the flow components are considered, however for
large catchments delay of flow (routing) may be necessary.

In the equations the average storages are used. To obtain the average storages, a similar iterative
method as previously discussed for the monthly water balance models is adopted. The iteration
is extended for both the soil moisture storage and the groundwater storages.

3.3.8 Parameter sets

Table 3.3 lists the parameter set of the 10-day water balance model. The table also gives the
initial values for parameter optimization and approximate ranges of possible values for each
parameter. The ranges are obtained by considering the order of magnitude of different
components the water balance. Full description on the order of magnitude of parameters is
provided in Appendix B.


60

Table 3.3 Parameters of the 10-days water balance model.

Parameter Description Initial values
Range

a
1
evaporation 0.004 0.001-0.01
a
2
percolation 0.04 0.001-0.01
a
3
interflow parameter 0.004 0.001-0.01
a
4
max. soil moisture storage 400 100-1000
a
5
base flow recession 0.04 0.01- 0.9

The optimum value of the first four parameters is obtained by minimizing the sum of squares of
errors. For the last parameter, a priori knowledge on the catchment is used. The following
section explains the methodology developed and coupled to the model.

3.3.9 Base flow recession parameter estimation

During low flow periods, the discharge of a river may be represented by an exponential decay
function (Chow et al. 1988):

(3.29)
)
0
(
0
t t k
t t
e q q

=

It can be shown that applying the continuity equation for a linear reservoir, the flow of the river
is given as:

t
g a t q
5
) ( = (3.30)

where a
5
=-k is a decay coefficient and g
t
is the groundwater storage. To determine the recession
coefficient, the recession limbs of hydrographs are analyzed in such a way that during these
periods it is assumed that only groundwater storage contributes to the flow. A routine for the
estimation of the recession coefficient is included in the model. The method simultaneously
traverses the rainfall and the flow series to obtain the recession hydrographs. The procedure to
do this is:


61
1) Locate a point on hydrograph when the recession starts. A point (t
0
, q
to
) on the recession
curve will be chosen as starting point if there are n=3 consecutive dry decades before t
0
. A
dry decade may be defined as the decade with a rainfall less than a given limit, RLIM, that is
not able to generate flow of the river. For 10-day time step RLIM = 30 mm can be taken. It
has to be noted that this amount of rainfall is assumed to be evenly distributed over the ten
days period.

2) Check if the flow of the next decade is on recession hydrograph. The flows at time t
0
+1, is
selected if the decade is dry and the flow is less than q
t0
. If these conditions are not met go
to step 4.
3) Hold the pairs, time from the start of the recession point (t-t
0
) and the corresponding flows
q
t
. Repeat step 2.

4) Terminate a recession limb and start to look for the next recession limb until the flow series
is exhausted (go to Step1.)

This selection will result with a list of pairs of flows and time after the starting point t
0
. The
recession curves for different years will normally also be different. An example of the result of
the procedure for Mwambashi catchment is shown in Figure 3.14. To normalize the series each
series is divided by the corresponding initial q
o
. Taking the natural logarithm of the ratio we
have:

) ( ln
5 o
o
t
t
t t a
q
q
= (3.31)

The above equation simply shows that a
5
is the slope of the average line relating the logarithm
transformed ratio of flow plotted against the time delay (t-t
o
) (Figure 3.14). The unit of a
5
is
1/time step. Therefore the value of this parameter can be determined right before going to the
model to ease the competition of multiple parameters in representing the general hydrograph.
The flow cart that describes this procedure is given in Figure 3.15


62
0. 1
1
10
100
0 5 10 15 20
Time from recession point (10-day)
F
l
o
w

(
m
m
)
1969
' 70
' 71
' 72
' 73
' 74
, 76
' 77
Year

Figure 3. 13 Example of parallel recession curves (Mwambashi catchment, Zambia).

0. 01
0. 1
1
0 5 10 15 20 25
Time fr om r ecession point (10-day)
q
/
q
o
average recession line

Figure 3.14 Example of normalized recession hydrographs (Mwambashi catchment, Zambia)


63

Start
t=t +1
R
t-j
=Rlim
Q
t0
=q
t
t
0
=0
ic=1
Q
t
<q
t-1
q
t
=0.0
IC=IC+1
Y(IC)=ln(qt/qt0)
X(ic)=t-t0
t>n
Regress
X,Y-a5
t=t+1
t=t+1
Stop
No
Yes
No
Yes
Yes
NO

Figure 3.15 Flow-chart for recession parameter estimation


64

3.4 A parsimonious daily rainfall - runoff model

3.4.1 Background

The objective of this section is to extend the approach of conceptualization of the rainfall-runoff
processes at catchment scale considered in the previous sections to daily time step.

Conceptual rainfall-runoff models consist of two main components: the first represents the soil
water balance and the second represents the transfer of water to the closure section of the
catchment. The water balance component is the most important aspect that characterizes a
model. This component expresses the balance between the moisture content of the soil,
generally divided into several zones and the incoming (precipitation) and outgoing (evaporation
and runoff) quantities (Franchini and Pacciani, 1991). The earlier models described in this study
deal only with the first component. Due to the coarse time step of computation attention was
given only to the volume of water that could be proportioned to runoff and evaporation.
Diminishing the time step to daily requires to also routing the generated volume of runoffs from
each point in the catchment to the outlet in to account.

3.4.2 Model structure

The daily time step model proposed here (Figure 3.16) has the same structure as the 10-day time
step model at the water-soil level. In addition the fast flow and the interflow are subjected to a
transfer operation using linear reservoirs.


65

Rainfall
Infiltration
Soil moisture storage
Groundwater storage
Evapo-
transpiration
Fast flow
Percolation
Slow flow
Base flow
Total flow
h
f
h
s

Figure 3.16 Conceptual representation of the daily rainfall runoff model.

In daily or smaller time step the runoff generated at a point in the catchment should be subjected
to a transfer operation. Singh (1992) describes the runoff processes in two phases: (1) the
overland flow phase, or the land phase, and (2) the channel phase. The transformation
undergone by the inflow at a point until it reaches the outlet is due to a translation and a storage
effect consisting of a time lag, and a shape modification. Some examples of flow routings used
in conceptual-rainfall runoff models are described below.

In the SACRAMENTO (Burnash et al., 1973) model, four components of flow (direct runoff,
surface runoff, interflow, and base flow) are produced separately. The sum of the first three is
transferred to the outlet section of the catchment using a unit hydrograph. The base flow on the
other hand is added directly to the amount obtained from the transfer of the above mentioned
three components.

In the SSARR (Rockwood et al., 1972) model, each of the three components of flow (surface,
interflow and base flow) is transferred to the outlet section separately by means of an algorithm
based on the linear reservoir scheme.

In the XINANJIANG (Zhao et al., 1980) model, the runoff is divided into two parts
representing the surface component and the base flow component separately. The surface


66
component is transferred to the outlet section using a unit hydrograph. The base flow
component is transferred using a linear reservoir scheme.

The ARNO (Todini, 1995) model has peculiarity that the three components (the direct runoff,
the surface runoff, the interflow and the base flow) are added together and transferred to the
outlet section by a double application of the parabolic unitary hydrograph. This unit hydrograph
is obtained through analytical solution of the diffusive-convective flow equation (parabolic
partial differential equation with the boundary condition that the input is uniformly distributed
along the length of the drainage net (Franchini and Todini, 1988). Propagation towards the
closure section of a succeeding sub-basin is also performed by means of a parabolic unit
hydrograph (Todini and Bossi, 1986).

Among the various ways of routing, the Nash unit hydrograph based on a cascade of linear
reservoirs is the most commonly used in hydrological modelling. The details of the derivation of
the Nash unit hydrograph are available in a series of papers by Nash (1958,1959,1960) and
summarized by Singh (1992).

In the proposed model, the fast and the interflow components are transferred to the outlet using
a unit hydrograph described by a series of linear reservoirs. Similar to the SACRAMENTO
model, the base flow is directly added to the routed components since the base flow is already
described as a release from a linear reservoir considered in the water balance level, namely the
groundwater storage.

3.4.3 Properties of linear reservoirs

The continuity equation for a reservoir is given by
dt
dS
Q I + = (3.32)

where I is the rate of inflow, Q is the rate of out flow, and S is the storage. These quantities are
all functions of time. For the daily model (DRRM), the fast flow and interflow components,
which are generated by the water balance level of the model are considered as separate inputs.
The corresponding outflows are the delayed and routed response of each of the flow
components at the catchment outlet. The fast flow is expected to respond in short period of time


67
compared to the interflow and for this reason two separate unit hydrographs are implemented
which differ with their parameters. The highlight of the derivation of the unit hydrograph is
given below. Detailed derivations are available in any hydrology books.

Using a simple linear relationship between discharge and storage

kQ S = (3.33)

where k is the storage parameter, we have

dt
dQ
k Q I = (3.34)
By utilizing the operator notation D=d/dt,

) (
1
1
t I
kD
Q
+
= (3.35)

Mathematically this is equivalent to

dt I e e Q
k t k t
=
/ /
(3.36)

The operator 1/(1+kD) represents its effect on the inflow. The instantaneous unit hydrograph of
the linear reservoir obtained by replacing (I(t) by a unit impulse input (t) in equation (3.36) is
given as:

k
e
t h
k t /
) (
= (3.37)

3.4.4 The Nash-Unit Hydrograph

The Nash model is based on a cascade of identical linear reservoirs. When the instantaneous
unit effective rainfall, given by (t), is fed into the first reservoir, then equation 3.34 becomes.


68
) (t
dt
dh
k h = + (3.38)

Taking the Laplace transform of both sides of the above equation the outflow from the first
reservoir due to an instantaneous inflow is: (Singh, 1992)

)
1
(
1
) (
1
s
k
k
s h
+
= (3.39)

Where s is Laplace transform term that indicates the order of differences. Thus h
1
will be the
inflow to the second reservoir, and the continuity equation for the second reservoir is given as

1 2
h
dt
dh
k h = + (3.40)

Similarly Laplace transform gives

2
2
)
1
(
1
) (
s
k
k
s h
+
= (3.41)

Generally for the nth reservoir we can write

n n
s n
s
k
k
h
)
1
(
1
) (
+
= (3.42)

By taking the inverses of the Laplace transform the Nash model is expressed as

) (
1
/
1
) (
n
e
k
t
k
h
k t
n
t n
|
.
|
\
|
=
(3.43)

where is the gamma function of n. Equation (3.43) expresses the impulse.

=
0
1
) ( dx x e n
n x


69

With the flexibility obtained by allowing n to take fraction values, the impulse response of this
equation has the ability to represent adequately almost all impulse response of shapes
commonly encountered in the hydrological context.

If the input is represented as a series of pulses (block of uniform input over duration T) and the
output is similarly expressed as ordinate at intervals T, then the linear relationship is most
conveniently expressed through the pulse response. For the system whose impulse response is
given by a gamma function given as in equation (3.43), the corresponding pulse response can be
obtained numerically by

T
T t G t G
t T h
) ( ) (
) , (

= (3.44)

where
( ) dt k e
n k
dt t h t G
n
t
k
t
1
0
/
0
/
) (
1
) ( ) (
= =
(3.45)

and is a dummy variable given as t-T. From Equation 3.44 one can see that the ordinates of the
unit hydrographs can be obtained once the number of reservoirs, n and the storage parameter k,
are fixed. These parameters are obtained by optimization. The parameter pair, n and nk, should
however be chosen rather than n and k, because n is a shape parameter and the product nk is a
scale parameter. In this way the two parameters are likely to be more independent than would n
and k, both of which contribute to the scale.

Normally to set appropriate memory length one starts with a longer time and checks the
significance of the ordinates at the tail of the hydrograph. The memory length for the catchment
depends on the size and shape of the catchments. For the fast flow components a couple of
weeks are necessary while for the interflow component longer memory length from a month to
two month can be used.

3.4.5 Parameter sets


70
The set of parameters for the daily model can then sub-divided into two levels: The water
balance and the flow routing parameters. To keep the number of parameters to be optimized as
low as possible and for practical purposes, the parameter n which stands for the number of
reservoirs can be fixed to integer values 1, 2, 3 etc. For the fast flow component, a starting
value n=1 and for interflow n=2 are suggested. Thus the set of optimized parameters are (a
1
, a
2
,
a
3
, a
4
, a
5
, a
6
, a
7
). As previously described, a
5
can be determined from daily hydrograph analysis
hence the set of parameters to be optimized consists of (a
1
, a
2
, a
3
, a
4
, a
6
, a
7
).


71

Table 3. 4 Description of parameters of the daily rainfall-runoff model
Parameter Description Initial
value
Range

Water balance
a
1
Evaporation parameter 0.04 0.01-0.1
a
2
Percolation parameter 0.004 0.001-0.01
a
3
Inter flow Parameter 0.004 0.0001-0.01
a
4
Max. soil moisture storage 400 100 -1000
a
5*
Base flow recession parameter. 0.01 0.01-0.001
flow routing
a
6
Storage parameter for fast flow, k
f
1.0 1.0-30.0
a
7
Storage parameter for interflow k
s
2.0 1.0-100.0
fixed parameters
n
f
Number of storage for fast flow 1
n
s
Number of storage for interflow 2
m
f
Memory length for fast flow 1-30 days
m
s
Memory length for interflow 10-100
days

a
5
is possibly determined using hydrograph analysis.

3.5 Data requirement and scope of application of the developed models

The models developed here are intended to address the limited data availability in some regions
of the world and hence require the following hydrological inputs on a daily, 10-day and monthly
time steps. For calibrating these models a length of about 10 years of data is sufficient. For
daily time step model rather short period of data up to 3 years may be enough to obtain
optimum parameters.

Areal precipitation
Areal potential evaporation / mean temperature
River flow at outlet of a catchment
Catchment area

Since the models are strictly based on the water balance at a catchment scale any water taken
from or come in to a catchment from an adjacent catchment has to be taken care of separately.


72
The potential uses of these models are to extend flow data from precipitation and evaporation
data and filling missing flow data. They can also be used to separate fast and slow flow
components. Moreover the models also provide some intermediate variables such as the soil
moisture fluctuation during the season which can be used for observing soil moisture deficits for
agricultural purposes. The daily rainfall runoff model developed may also be used for flood
forecasting if coupled with a rainfall forecasting scheme.

Limitations

One of the limitations of the present models is the area of the catchment because of the
assumption of homogeneous distribution of rainfall and evaporation. If there is high spatial
variation of rainfall the model will not perform efficiently. On the other hand, too small
catchments will also give difficulties in that the groundwater and the topographic divides may
not coincide. Hence there would be a considerable loss of water to adjacent catchments and the
water balance would be violated.

The models are not applicable if there are large water bodies such as lakes in the catchment and
if significant frost and snow occur. For the latter problem, an extension of the model was
proposed by introducing one additional storage, which takes care of the snow storage (Ni-Lar -
Win, 1994). For such models temperature inputs are mandatory.


73

Chapter 4
4 STATISTICAL METHODS

4.1 Introduction...................................................................................................................................74
4.2 Data checking................................................................................................................................74
4.3 Parameter estimation.....................................................................................................................77
4.4 Model validation ...........................................................................................................................90

4.1 Introduction

With the increasing number of models in literature, efforts are devoted to developing a
statistical methodology to the evaluation of the quality of input data, estimation of initial
conditions, to parameter optimization techniques and to the validation of the performance of a
model. This chapter formulates the methodology that will be followed in this research.

4.2 Data checking

Advances in scientific hydrology and practice of engineering hydrology depend on good,
reliable and continuous measurements of hydrological variables. Model calibration more than
anything relies on the quality of data available. Hydrological data must be cleaned from random
and systematic error for erroneous data leads to either non-veritable rejection of a model or
wrong calibration that affects the usefulness of a model. Different types of error that
encountered in hydrological data are due to:

Instrument errors
Change of measuring sites and techniques
4. Statistical Methods

75
Data inputting and computation errors
Flow computation errors.

Instrument errors and change of measuring site

The first two types of errors can be identified using known methodologies such as mass curve
and double mass curve analysis. Inter comparison plots of the same variables measured at
adjacent stations can help to verify the consistency of the data series.

Another interesting procedure to check the stationarity of long records is to split the data set in
to two periods and to test the elementary statistics for the two periods. This test can be done if
there are long data series to observe and if there are general changes of the hydrological
variables due to the reasons listed above.

Let X
1j
, j=1, 2.n
1
and X
2j
, j=1,2n
2
be two time subseries of a very large hydrological data
series with the sample means
1
X and
2
X respectively. The t-statistic is mostly used for testing
whether the difference of the two means of the subseries is significant, that is

2 1
2 1
2 1
n n
n n
s
X X
t
+
= (4.1)
where

2
) ( ) (
2 1
2
2
2
1
2
1
1
1
+
+
=

= =
n n
X Xi X Xi
s
n
i
n
i
(4.2)

The variable t of Equation 4.1 is compared with the table value of the Student t-distribution with
(n
1
+ n
2
-2) degrees of freedom and 5% significance level. Helsel and Hirsh, (1992) state that the
above test is valid if the two subseries are independent and the series under study are normally
distributed.

Data inputting and computation errors


76

Data inputting and gross computational errors are those which generate random errors and it is
very difficult to detect them. Time series plot and visual inspection of the graphs may reveal
large errors such as decimal point shift and addition of digits. Quick annual water balance of
observed components may also reveal some gross errors.

Flow computation errors

The quality of discharge data plays a central role in the development of a model and also in
calibrating and applying it. However, the discharge records are often less accurate than
desirable. Discharges are computed from continuous stage records using an established rating
curve (Stage-Discharge relationship) at a gauging site on a river. Some of the causes of errors in
discharge computation include the unstability of the cross-section (the riverbanks and bed) so
that the discharge translation from stage can be corrupted. The complexity of a rating curve with
hysteresis also reduces the accuracy of flow. High flows are normally approximated by
extrapolation of rating curves and debris or sedimentation of a channel easily corrupts low
flows. Despite these and many other problems associated with obtaining an effective record of
stream discharge, the stream flow record is generally considered to be the most accurate element
of a catchment's hydrologic history (Burnash, 1995)

Most of the river gauging sites in the world are located on the natural cross-section of rivers.
Ideally, the sites (the banks and bed of a channel) should be stable and there should be a control
section downstream such as fall to avoid backwater effect. However, it is often impossible to
locate such a site which should also be accessible for daily reading and regular supervision. As a
result a reasonably good site is identified and instruments are installed. It is advisable in such
instances to investigate the cross-section of the gauging site and calculate indicative discharges
using Manning or Chezy formulas to avoid extrapolation errors that could occur by applying the
rating curves. An example of such problem is discussed in Section 5.2 for a river gauging
station in Ethiopia.

Since stream flow data are continuous in time and correlated in space it is possible for the
reliability of the observation to be checked by interpolation and statistical methods. WMO
(1994) outlines means of checking for internal consistency between observed discharges as
follows:


77

a) Qualitative assessment of the correspondence between discharges at adjacent stations;
b) Qualitative assessment of the relationship between the discharges and the value during the
previous measure;
c) Approximate check of the value of the discharge by seeing that it falls within the range of
the previous value for the given phase in the regime of the river; and
d) Approximate assessment of correspondence between the measured value and the regular
variation during the previous period.

4.3 Parameter estimation

Sorooshian, and Gupta (1995) classify parameters into two groups: physical and process
parameters. A physical parameter represents physically measurable properties of the watershed
(e.g. areas of the catchment, fraction of impervious area and surface area of water bodies,
surface slope etc). Process parameters represents properties of the watershed which are not
directly measurable e.g. average or effective depth of surface soil moisture storage, the effective
lateral inflow rate, the coefficient of non linearity controlling the rate of percolation to the
groundwater and so on. In fact the division between the two groups depends on the spatial
distribution and structure of a model. In a conceptual rainfall-runoff models almost all of the
parameters are of the second type where the optimum values are obtained by calibrating the
model using historical data.

Model calibration is one of the most important aspects of hydrologic modelling. Conceptually
realistic models can produce erroneous results if they are not properly calibrated. When
calibrating catchment models the following questions are frequently asked (Gan, 1988):

a) Which calibration approach should be used, manual or automatic?
b) For automatic calibration, how many parameters should be optimized and which ones?
c) When calibrating manually, what parameters should be changed and by how much?
d) How to determine if optimality has been reached and that any further adjustment would
only yield marginal returns?
e) What guidelines to use to determine if parameters are final or not?


78
In dealing with the questions above, the skill, experience and intuition of the hydrologist and
his/her knowledge of the hydrologic model play an important role. This is particularly true
when manual calibration is employed. It is recommended that a combination of manual and
automatic procedure be adopted for the model calibration (Gan, 1988). Manual calibration
alone is very tedious, time consuming, and requires the experience of the hydrologist. Because
of the time-consuming nature of the manual model calibration, there has been a number of
researches towards development of automated calibration methods. Automatic calibration on the
other hand relies heavily on the optimization algorithm and the specified objective function.
The following section summarizes some of the usual objective functions and techniques of
optimization algorithms.

4.3.1 Objective function

An objective function is an equation that is used to compute a numerical measure of the
deviation between the model calculated output and the observed catchment output. Typical
outputs used in hydrological modelling are stream flow hydrographs and groundwater levels.
Objective functions that have been shown to work well in practice include the least square and
maximum likelihood estimates.


79
Least square method

The least square method (LS) determines the parameter value for which the sum of squared
deviations between the observations and their expectations is as small as possible.

|
=
=
n
t
t t t
X d q w X F
1
2
) ( ) ( | (4.3)
where:
q
t
= observed streamflow value at time t;
d
t
(X) = model calculated flow value at time t;
X = vector of model parameters;
n = the number of data points.
w
t
= weight at time t

The method implicitly assumes that the differences of observed and calculated quantities have
equal weights throughout the data points. In practice, a weight factor can be introduced to
indicate the importance of particular hydrological values such as high flow regime. If all
weights are equal then one has the Simple Least Square (SLS), and if different weights are
introduced the Weighted Least Square (WLS). In case where the variance of the deviations
exhibits a large variation, a transformation of the observed and calculated quantities would be
necessary. Vandewiele et al. (1993) suggest square root transformation, which is believed to
have less computation difficulties than the logarithmic transformation. In the present study due
to large variation of flow, the square root transformation of the computed and observed flows is
used to compute the objective function.

Maximum Likelihood Methods

The maximum likelihood method (ML) determines the parameter value, for which the
probability of the deviation is as maximum as possible. Hence the method estimates the
parameters set that maximizes the probability of the observation. The Maximum Likelihood is
reduced to the SLS if the following two assumptions hold true:
the deviations are normally distributed with mean zero and variance
2

the deviations are independent of each other.


80
In reality, these assumptions are often violated and hence methods that account for non-
stationary variance in streamflow errors are developed (Sorooshian and Dracup, 1980;
Sorooshian, 1981). One of the forms of the maximum likelihood criteria called HMLE
(Heteroscedastic Maximum Likelihood Estimator) is defined as:

n
n
t
t
n
t
t t
X
w
w
n
HMLE
/ 1
1
1
2
,
1
max
(
=
=
=
(4.4)

where
t
=q
t
,
obs
-d
t
(X) is the model residual at time t; w
t
is the weight assigned to time t, computed
as:

) 1 ( 2
=
t t
f w (4.5)

where f
t
is the expected value of the simulated flow at time t. Recently Yapo et al. (1996)
suggest the use of the observed flow as a factor to obtain a stable estimator. Further they
confirmed that this estimator provides consistent performance across all flow ranges.

4.3.2 Optimization Algorithms

Once the objective function is identified, the next step is to obtain the parameter set, which
results in the extreme value of the function. The surface described by the set of parameter is
known as the response surface.
Figure 4.1 shows a surface described by one and two parameter models.


81

Figure 4.1 Response surface A: one parameter model; B: two-parameter model

The gradient G of the objective function is given as:

)
n
x
F(X)
,...,
x
F(X)
,
x
F(X)
( G(X)
=
2 1
(4.6)


82
One obtains the second derivative, which is often called Hessian matrix as:

2
2
2
2
1
2
2
1 2
2
2
2
2
1 2
2
1
2
2 1
2
2
1
2
) (
...
) ( ) (
... ... ...
) (
...
) ( ) (
) (
...
) ( ) (
) (
n n n
n
x
X F
x x
X F
x x
X F
x x
X F
x
X F
x x
X F
x x
X F
x x
X F
x
X F
x H
= (4.7)

Under general conditions this matrix is symmetrical i.e.

1 2
2
2 1
2
) ( ) (
x x
X F
x x
X F

(4.8)

Let ) ( ), ( ), (
0 0 0 0 0 0
X H H X G G X F F = = = then one can write the Taylor expansions as:

. . ) . ( . ) (
2
1
) . ( ) (
0 0 0 0 0 0
+ + + =
T T T
X X H X X X X G F X F
(4.9)

and for the gradient we have

.. ). ( ) (
0 0 0
+ + = H X X G X G (4.10)

If F is twice differentiable, the necessary condition for X
~
to be associated with a global or local
minimum value of F is that G( X
~
)=0, and that H( X
~
) be positive definite. Due to the non-
linearity nature of the system, the assumptions mentioned above are seldom realized. In reality
the objective function for the rainfall runoff models are not analytically differentiable. The
values are normally discrete values for each time step considered. The immediate solution at
this instance is using the differences. For instance if X is an n-parameters vector {a'
1,
a'
2
a'
n
},
the first element of the Hessian matrix, i.e. second derivative with respect of the same parameter
can be estimated as:


83
2
1
'
2
'
1 1
'
2
'
1
'
2
'
1
'
1
'
1
'
1
2
) (
) ' ... , ( ) ' ... , ( 2 ) ' ... , (
a
a a a a F a a a F a a a a F
a a
F
n n n
+ +
=

(4.11)

A similar formula can be used to estimate the diagonal of the matrix. For the other elements,

( )
( )
)
+ +
+ + +
=
) ' a ,..., a a , a a ( F ) ' a ,..., a a , a a ( F
) ' a ,..., a a , a a ( F ) a ,..., a a , a a ( F
a a 4
1
a a
F
n
'
2
'
2
'
1
'
1 n
'
2
'
2
'
1
'
1
n
'
2
'
2
'
1
'
1 n
'
2
'
2
'
1
'
1
'
2
'
1
'
2
'
1
2

(4.12)

In calculating the Hessian matrix it is better to use scaled parameters instead of the parameter
itself to a obtain well-conditioned matrix for inversion.

Different optimization algorithms are available for solving this problem with varying efficiency,
depending on the level of complexity of the problem such as number of parameters,
computation capacity etc. The major characteristics complicating the optimization problem in
conceptual rainfall-runoff modelling summarised by Duan et al. (1992).

a) regions of attraction, where more than one main convergence region exists;
b) minor local optima, where there are small peaks in each region;
c) roughness, when the response surface is rough with discontinuous derivatives;
d) sensitivity, poor and varying sensitivity of the response surface in the region of optimum
and non-linear parameter interaction; and
e) shape, resulting from a non convex response surface with long curved ridges.

Search algorithms are broadly classified into two groups, the local and global search methods.
The difference between these two methods is mainly the way they advance form one solution
point on the surface of the response to the next feasible point. A local search algorithm tries to
find the minimum point from a certain given point by investigating all possible ways and then
continues until it reaches a point where further move does not improve the solution. On the
other hand the global search algorithms attempt to exhaust all possible local minimum points by
starting from different initial points and then evolving following different of selection criteria
from the set of the points.


84
Local Search algorithms are further classified as direct search method and gradient methods.
Direct methods just use information of the function value to arrive at the optimal (evaluating
objective function in successive neighbor points). These algorithms select the best new direction
and best step to achieve improvement as fast as possible. Rosenbrock method (Rosenbrock,
1960), the Pattern Search Method (Hook and Jeeves, 1961) and the Simplex Method (Nelder
and Mead, 1965) are some examples of direct search method.

Gradient algorithms use the function value and additionally information coming from the
gradient in order to determine direction and step to reach the local minimum. Gradient methods
are based on a truncation of the Taylor Series expansion in the domain of the parameters
(Sorooshian 1995).

In this study two optimization methods are coupled with the developed models: the VA05A
method (Harwell subroutine library, 1974) which has been recommended as a better
optimization method for rainfall-runoff modelling by Demaree (1982) and the more recent
Shuffled Complex optimization method developed by Duan et al. (1992). The following two
sections briefly describe the two optimization methods.

4.3.3 The VAO5A algorithm

The VAO5A program is originally taken from the Harwell subroutine library (1974). The
method is based on the compromise between three different classical algorithms for minimizing
the sum of squares of errors, namely Newton Raphson, Steepest Descent and Marquardt. It
automatically obtains and improves an approximation of the Jacobian matrix G and the Hessian
Matrix H. From the inverse of the Hessian matrix H, the covariance matrix, , of the continuous
filter parameters is obtained by

(4.13)
1 2
H 2
=

where is the standard deviation which is obtained as:

n N
ssq
=
) min(
(4.14)


85

where N = number of data points
n = number of parameters to be optmized

From the covariance matrix, the correlation between two parameters a
i
and a
j
can be calculated
jj ij
ij
j i
i
a a Corr

= ) , ( (4.15)
Where
ij
is the (i,j)
th
element of the covariance matrix, . It follows that the standard deviation
of parameter a
i
is

ii
= (4.16)

The program is supplemented with subroutines that provide key statistics such as standard
deviation of a parameter and correlation matrix among the parameter. These statistics are not
given much attention in other methods such as shuffled complex algorithm.

The algorithm requires a predetermined order of magnitude of the parameters to scale them
down to values between 0.0 to 1.0. This is to allow using the same step of increment while
searching the direction. The other limitation of this method is that the parameters can not be
constrained. A user has to specifically define the parameter constraints in the structure of the
model to avoid unrealistic computation. For example for the models described in section 3.2, the
parameters a
2
and a
3
are bounded by zero and 1. The final values of parameters obtained from
this optimization method should be inspected, to see whether they fall within the specified
range. The definition of the constants and default setting used in the method are given Table 4.1.


86
Table 4.1 Inputs of the VA05A optimization method.
No. Symbol Definition
Recommended
value
1 DSTEP initial increment of parameters 0.01
2 ACC
stop criterion difference between consecutive
sum of squares of error.
0.001
3 MAXFUN
Maximum number of iterations to obtain
optimal vales.
10000

4.3.4 The shuffled complex evolution algorithm

The shuffled complex evolution algorithm (SCE-UA) is a global technique that can be applied
in a broad range of optimization problems. This method is classified as probabilistic because it
evaluates the objective function at randomly spaced points in the feasible parameter space. The
SCE-UA combines advantages and strategies of the Simplex procedure (Nelder and Mead,
1965), Controlled Random Search (Price, 1987), Competitive Evolution (Holland, 1975) and the
newly developed complex shuffling (Duan et.al, 1992). "The strategy in this optimization
method is analogous to giving the same difficult problem to 12 identically capable people and
asking them to solve it without conferring with each other. Clearly, a more efficient strategy
would be for them to spend some time working independently or in small groups and to get
together now and then to share information about their progress

SCE-UA treats the searching process as a natural biological evolution course. The
method begins with a population of points randomly chosen from the feasible space.
Then the population is partitioned into several communities, referred to as complexes,
each containing 2n+1 points; where n is the dimension of the problem (number of
optimized parameters). Each community is made to evolve based on a statistical
"reproduction" process that uses simplex geometric shape to direct the search in an
improvement direction. It ensures sharing information by periodically shuffling the
entire community (Duan et al., 1992).


87
The SCE-UA algorithm consists the following steps (Duan et al, 1992).

1. To initialize the process, select p1 and mn+1, where p is the number of
complexes, m is the number of points in each complex, and n is the
dimension of the problem. Compute the sample size s=pm

2. Then generate a sample as follows. Sample s points x
1
..x
s
in the feasible space
R
n
. Compute the function value F
i
at each point x
i
. In the absence of prior
information, use a uniform sampling distribution.

3. Rank the points as follows. Sort the s points in order of increasing function value.
Store them in an array D = {x
i
, F
i
, i=1... s}. So that i=1 represents the point with the
smallest function value.

4. Partition D into p complexes A
1
.. A
p
, each containing m points, such that A
k
= {x
k
j
,
F
k
j
/ x
k
j
=x
k+p(j-1)
, F
k
j
= F
k+p(j-1)
, j=1 .. m}.
5. Evolve each complex A
k
, k=1p, according to the complex evolution (CCE)
algorithm outlined separately.

6. Shuffle the complexes as follows. Replace A
1
.. A
p
into D, such that D={A
k
, k=1 ..
p}. Sort D in order of increasing function value

7. Check convergence. If the convergence criteria are satisfied or maximum number of
trials is exceeded. stop. Otherwise return to step 4. The convergence criteria is
expressed as the difference between the successive F values, and which should be
smaller than a given accuracy limit.

The competitive complex evolution (CCE) algorithm required for the evolution of each complex
in step 5 is outlined as follows.

1. To initialize the process, select q, , and , where 2q m, 1 and 1.

2. Assign weights as follows. Assign a trapezoidal probability distribution to A
k
, i.e.,


88

( )
( ) 1
1 2
+
+
=
m m
i m
i
, i= 1 .. m (4.17)
The point x
k
1
has the highest probability
1
=2/m+1. The point x
k
m
has the lowest
probability
m
=2/m(m+1).

3. Select parents randomly choosing q distinct points u
1
.. u
q
from A
k
according to the
probability distribution specified above (the q points define a subcomplex). Store
them in array B={u
i
, v
i
, i=1..q}, where v
i
is the function value associated with point
u
i
. Store in L the locations of A
k
, which are used to construct B.

4. Generate offsprings according to the following procedure:

a) Sort B and L so that the q points are arranged in order of increasing function value and
compute the centroid g using the expression:

=
1
1
1
1
q
j
j
u
q
g (4. 18)

b) Compute the new point r=2g-u
q
(reflection step).
c) If r is within the feasible space , compute the function value f
r
and go to step d; otherwise
compute the smallest hypercube HR
n
. that contains A
k
, randomly generate a point z within
H. Compute f
z
set r=z and set f
r
=f
z
(mutation step).
d) If f
r
<f
q
, replace u
q
by r, go to step f; otherwise compute c=(g+u
q
)/2 and f
c
(contraction step).
e) If f
c
<f
q
replace u
q
by c, go to step f; otherwise randomly generate a point z within H and
compute f
z
(mutation step). Replace u
q
by z.
f) Repeat steps a-e times, where 1 is a user-specified parameter. This value can be set to
unity.

5. Replace parents by offspring as fallow. Replace B into A
k
using the original
locations stored in L. Sort A
k
in order of increasing function value.


89
6. Iterate by repeating steps 2-5 times, where 1 is a user-specified parameter that
determines how many offspring should be generated (how far each complex should
evolve).

The FORTRAN coded program is available from the Authors. The user should provide an input
file that contains some of the constants and the range of parameters. Table 4.2 summarizes the
input required. The choice of the objective function is left for the user. This allows adapting the
optimization method to any kind of model involving parameter optimization.

The authors claim that the method developed is efficient in a wide range of problems. Recent
application of this method by Gan et al., (1996), and Franchini et al., (1998) also confirm the
method mostly guarantees the global optimization for conceptual rainfall-runoff models. The
main advantage of this method is that once the range of a parameter (lower bound and upper
bound) is identified from experience or from physical interpretation of a component of models,
it allows investigating a wider range of points in such a manner that the global optimum will be
obtained.

Unlike the first optimization method VA05A, this method does not require scaling of
parameters. This endows one of the advantages since in some cases prior identification of scales
of parameter can be difficult.

It is obvious from the description of the method, that it explores exhaustively in the space of the
feasible region bounded by the upper and lower limits of parameters. Hence it ensures the global
optimization. On the other hand the method does not give any statistics about the inter-
relationship that exists between parameters nor of sensitivity of the objective function. In the
present study, the parameters are optimized using either of the two methods. To obtain the
statistics related to the parameter optimization, the final solutions obtained by the Shuffled
complex algorithm are input to the VAO5A method. From the latter method the essential
statistics such as standard deviation of parameter, correlation matrix and sensitivity of the
objective function to individual parameter in a periphery of the optimal point are obtained.
Table 4.2 Inputs of the Shuffled complex optimization method.
No. Symbol Definition Recommended
value


90
1 MAXN Maximum number of trials 10000
2 KSTOP Number of Shuffling loops in which the criterion
value must change in a the specified percentage or
else optimization will be terminated
5
3 PCENTO Percentage by which criterion value must change
in the specified number of shuffling loops or else
optimization is terminated

4 NGS (p) Number of complexes used for optimization
search
2-20
5 ISEED Random seed in optimization search
6 NPG (m) Number of points in each complex 2nopt*+1
7 NPS (q) Number of points in each sub-complex Nopt+1
8 NSPL () Number of evolution steps taken by each complex
before next shuffling
2nopt+1
9 MINGS Minimum number of complexes required for
optimization search, if the number of complexes is
allowed to reduce as the optimization search
proceeds.
2-20
10** INIT Initial estimate of the parameter to be optimized
11 LOWER-B Lower bound of the parameter to be optimized
12 UPPER-B Upper bound of the parameter to be optimized

* Nopt= Number of parameters to be optimized; ** Inputs 10, 11, 12 are provided for each
parameter to be optimized.

4.4 Model validation


91
Selection of a 'good' model requires both subjective and objective judgements to determine
whether the results provide adequate information for answering the question facing decision-
makers. The results may be influenced by a variety of causes. Common problems in calibration
are among others;

errors in the data used in calibration
use of a period of record that does not contain enough of the physical processes needed to
calibrate key parameters

After the model has been chosen and parameters estimated, its merits have to be judged. This
testing or validation can be considered as an important stage of the model analysis. If the model
is good enough, then application proceeds. If not, the whole process starts again by changing
one or more working hypotheses. The criteria considered to judge on the performance of the
models developed in this study are as follows.

1) The significance of the parameters (confidence intervals)
2) Graphical comparison of simulated and observed flows
3) Comparison of monthly means of observed and computed flow
4) Model Efficiency: the percentage of explained variance
5) Model Quality factor (see Vandewiele et al,. 1992).
6) Residual Analysis (variance, seasonality and auto-correlation of residuals)


92

4.4.1 Checks on significance of a parameter

The question to be answered at this step is if all parameters are really necessary in the model to
describe the physical processes occurring. The hypothesis is tested that parameters a
i
are
significantly different from zero. This is examined by checking whether the zero value belongs
to the 95% confidence interval, CI.

) 96 . 1 , 96 . 1 (
i i i i
a a CI + = (4. 19)

and the half width of confidence interval (HWCI) is given as:

i i
a HWCI 96 . 1 ) ( = (4.20)

If the hypotheses a
i
= 0 is acceptable, then parameter a
i
can be set equal to zero without
diminishing the explanatory power of the model. The model then has to be recalibrated either
with by setting the value of this parameters to zero. Care must be taken at this stage that the
hypothesis suggests the superfluous parameter considering all the rest of the parameters at
optimum value.

4.4.2 Graphical comparison of simulated and observed flows

Graphical display of simulated and observed flows is very important because the traditional
method of evaluating model performance by statistical measures has limitations. Statistical
indices are not effective in communicating qualitative information such as trends, types of errors
and distribution patterns. In fact one should not depend on only single statistical measures of
model performance. Theses are sometimes misleading because of the high possibility of
compensation of errors from season to season or over years in long-term calibration. In this
research we propose that in cases of highly variable input cases the quality measures should also
account for this variation and hence seasonal means have to be compared.

4.4.3 Comparison of monthly means of observed and computed flow


93
The seasonal means are computed as follows. Let q
ij

denote the observed flow during year i and
season j, where j=1,2,..P and P is the number of seasons in the year. For example P=12 and
P=36 for monthly and decade time series respectively. The seasonal mean is given as:

q = q
j
(4.21)
Y
1
ij
Y
=1 i

where Y is the length of the data in years. Similarly long term seasonal means of all input and
output variables are computed using equation 4.21. Graphical or numerical display of these
values gives a summary of results on the water balance of a catchment. Moreover it depicts the
success or failure of the model in reproducing the expected values along the whole period of the
year.

4.4.4 Model Efficiency

The statistical index of modelling efficiency (E
f
) was defined by (Nash and Sutcliffe, 1970) as:
( ) ( )
( )

=
2
2 2
o o
o c o o
f
q q
q q q q
E (4.22)

where q
o
= observed flow,

o
q = mean observed flow and
q
c
= calculated flow.

The sum is taken over the whole period of the data used for calibration. The closer this value to
unity, the better the model explains the variance. A negative modelling efficiency means that the
model prediction is worse than simply using the mean of the observed flows. This measure is
highly affected by a few extreme errors and can be biased if a wide range of flow events is
experienced.

4.4.5 Model Quality factor


94
The model quality is a check on whether the model gives a computed output that is a
sufficiently close reproduction of the observed output. When the standard deviation, of the
random component is small, the unexplained part of runoff is small and the model fits observed
runoff very nearly. Therefore, is an inverse measure of quality. It is better to use the variation
coefficient vcf q
t
, which is a dimensionless quantity expressed as standard deviation divided by
expectation. Vandewiele et al., (1993) provides different expression for the vcf for transformed
flows. For square root transformed flows It can be shown that

t t
t
t
d d
d
q vcf

2
2 4

2
2
+
+
= (4.23)

where d is the computed flow. The approximated expressions are valid for small . The inverse
measure of the quality depends on d
t
, that is on time. Therefore the model variation coefficient
for d
t
equal to its mean is considered as;
t

(4.24)
d
d
d
d
mvcf
t
t

2
2 4

2
2
+
+
=

When the observed variation coefficient of runoff is small in a given basin, it can be expected
that the model variation coefficient will be small also and vice versa. This leads to the
comparison of these two variations by taking their quotient and to a definition of an 'absolute'
model quality measure, MQ.

mvcf
ovcf
runoff mean for vcf el
runoff of vcf observed
MQ = =
mod

(4.25)


95
For judging whether a difference in quality of two models is important, a confidence interval of
MQ is computed. Since MQ is approximately inversely proportional to , the HWCI of and
MQ are approximately proportional to and MQ respectively. With the expression of HWCI (
) leads to an approximate expression of the half width of a 95% confidence interval as;

n N
MQ
MQ HWCI
=
38 . 1
) ( (4. 26)

Where N is the number of data points and n is number of optimized parameters.

4.4.6 Residual Analysis

The computed runoff is only an approximation of the expected value of q
t
(Eq
t
). The difference
between the transformed observed and computed flows gives the residuals as:

t t t
d q u = (4.27)

Residual analysis is carried out to determine whether the residuals u
t
behave as required by the
model hypotheses, especially whether they are independent, homoscedastic and normally
distributed with zero expectation. Independence of residuals is checked by computing their
autocorrelations,
k
with time lag k;

( )( )
( )

1
2
1

=
+

=
N
t
t
K N
t
k t t
k
u u
u u u u
(4.28)
where
u
is the mean residual and N is the length of the time series (in case there are no gaps in
the runoff data series). The half width of 95% critical interval (5% significance level) is
approximately;

( )
=
+
1
1
2
2 1
96 . 1
=
k
i
i K
N
HWCI (4. 29)


96
The hypotheses
k
= 0 is true at 5% significance level when;

( )
k k
HWCI (4.30)

These expressions are bad estimates when k>N/5. It is sufficient to compute them for k = 24 up
to two years for monthly time step.

4.4.7 Sensitivity to calibration period

If the input and output data are assumed to be time homogeneous, the results of application of a
given model to two different calibration periods of the same catchment have to be the same. The
different calibration periods can be defined by splitting long observed time series. Let

) ,..., , (
2 1 n
a a a A & & &
&
= (4.31)
and

) ,..., , (
2 1 n
a a a A & & & & & &
& &
= (4.32)

be the estimates of vector A=(a
1
, a
2
,.., a
n
) of filter parameters obtained from two calibration
periods, where n is the number of parameters. Then and are stochastically independent and
approximately normally distributed. A level 5 % test on identity of the two models is
Vandewiele et al, 1993 proposes that at a level of 5%
A
&
A
& &

%) 5 ( ~ ) ( ) ' ' ' )( (
2 1
n
T
A A A A
& & & & & &
+
(4.33)

where ' and ' are the corresponding covariance matrices, T is transpose and is chi-
square distributed with n degrees of freedom. The (5%) is the 5% point of (see Table
4.3). Practically the individual confidence interval of parameters should be investigated to
contain the optimum parameters estimates obtained using two calibration periods.
'
2
n
2
n
2
n


97
Table 4.3 Some 5 % points of the Chi-square distribution with n degrees of freedom.

degrees of freedom, n 2 3 4 5 6 7
5% -value
2
n
5.991 7.815 9.488 11.070 12.592 14.067

Equality of model standard deviations & and & & can similarly be tested at the 5% level by
checking whether;

( ) ( )
96 . 1
' ' 2 ' 2
' ' '
2 ' ' 2 '
K N K N

(4.34)

where N' and N'' are the numbers of data in the two calibration periods. Besides these formal
tests, the general behaviour of residuals is compared.

Chapter 5
5 CASE STUDIES OF
MONTHLY WATER BALANCE MODELS FOR SEMI-
ARID AND ARID CATCHMENTS

5.1 Introduction...................................................................................................................................98
5.2 Data description.............................................................................................................................98
5.3 Monthly model application .........................................................................................................115
5.4 Conclusions .................................................................................................................................138

5.1 Introduction

This chapter begins with the description of the test catchments used in this study. The
hydrometeorololgical data are examined with regard to the quality and sufficiency for the
model calibration and verification. Then the conceptual monthly water balance models
developed in Section 3.2 are applied to the test catchments. Although originally developed for
humid regions, improvement of the performance of the models for arid areas is aimed at due
to incorporation concepts that account for the hydrological characteristics of the arid and
semiarid conditions.

5.2 Data description

5.2.1 Ethiopian catchments

5.Case studies of monthly water balance models

99
Four catchments located within the Upper Awash River Basin are used in this study. The
Awash River Basin is one of the major river basins in Ethiopia with a total catchment area of
112,697 km
2
. The Awash basin is narrow in the upper part and widens towards the north of
the catchment. The catchment is characterized by its closed river system (no outlet to the sea).

The catchments are located upstream of Koka dam (see Figure 5.1). The average annual areal
rainfall is 1000 mm for the upper part of the catchment and decreases to the 500 mm for the
lower basin. The major rainy season is between June and mid September with a short
monsoon rain in March to April. The annual minimum and maximum temperatures are 14
o
C
and 32
o
C respectively.

The topography of the catchment is characterized by highlands ranging from 2000- 3000m
above sea level for the upper part of the catchment and lowlands for the lower part of the
basin. There are two major soil types in the catchment; the deep red clay soil, Nitosol, and the
dark clay soil, Vertisol (Alem, 1989). The Nitosols are found in the upland areas, whereas the
Vertisols are found in lowland areas with slopes ranging from 2 to 8%. Agrarian farmers
practicing rain fed agriculture populate the upper part of the catchment (upstream of the Koka
Dam). Farmers practicing pastoral farming and mixed farming sparsely populate the lower
part, which is flat lowland. The existing water resources are mainly used for development of
hydropower and irrigation schemes with the water being regulated by the two dams, Koka and
Tendaho, on the main river. Figure 5.1 shows location of the rainfall and discharge gauging
stations in the Upper Awash River Basin.


100

N
Sudan
Ethiopia
Eritrea
Kenya
Awash River
Basin
Somalia
Djibouti
42
0
44
0
48
0

4
0

8
0

12
0
16
0
38
0

Figure 5.1 Location of Awash River Basin


101
Meteorological data

The rainfall and evaporation data are collected from the National Meteorological Agency.
Hard copies of daily data were edited in digital form. A total of seven stations within the
interest area are available (Table 5.1). From the location of the stations one can see that the
distribution of the raingauge stations is not uniform across the watershed but fair enough
compared to the data availability in the country.

Table 5.1 Description of daily rainfall data in upper Awash River Basin
STATION Lat. Long
Elve.
a.s.l (m)
Period
Annual
Rainfall (mm)
Addis Ababa 09
0
02' 38
0
43' 2408 1900-1996 1192
Addis Alem 09
0
03' 38
0
24' 2340 1980-1996 1117
Holeta 09
0
05' 38
0
30' 2000 1980-1995 1061
Sebeta * 08
0
55' 38
0
30' 2379 1980-1994 2347
Sendafa 09
0
10' 39
0
02' 2550 1980-1995 890
Debrezeit 08
0
26' 39
0
01' 1900 1980-1995 781
Modjo 08
0
37' 39
0
08' 1870 1980-1995 853
* Erroneous data

The stations at Sebeta showed extremely high rainfall records that do not accord with the
other stations. The annual rainfall record of this station reaches more than 3000 mm compared
to 1200 mm of the others. The station is located about 40 km from Addis Ababa station and
this amount of variation is not explained. Comparison of the 10 days moving average of data
of Sebeta and Addis Ababa shows the same trend but multiplied by different factor at different
times hence hinders correction of the data. Due to the possible erroneous data this station is
excluded and only the remaining 6 stations are considered.

The general quality of the other data of rainfall data is checked by inter comparison between
the stations. The correlation coefficient among the stations on monthly rainfall amount ranges
from 0.647 to 0.874 (Table 5.2). This shows fairly good agreement on the monthly rainfall
series. Daily rainfall data variation between the stations is pronounced due to localized storms.
Missing data of a few days at stations Debrezeit, Modjo and Holeta are encountered. The gaps
are filled using the data of the nearest station, which has a record for the missing periods.
Data on evaporation are only available for the station at Addis Ababa, for a continuous period
of two years (1988-89). To have the seasonal variation of evaporation, the monthly mean
evaporation rates (for the 12 months) are computed using these two years record (Table 5.3).


102
The data of this station is used for modelling for the four catchments in the Awash River
Basin.

0
10
20
30
40
50
60
70
1 31 61 91 121 151 181 211 241 271 301 331
Time series (from January 1)
1
0

d
a
y
s

M
o
v
i
n
g

A
v
e
r
a
g
e

r
a
i
n
f
a
l
l

(
m
m
)
Addis Ababa
Sebeta

Figure 5.2 10-days moving average of rainfall at Addis Ababa and Sebeta stations.

Table 5.2 Cross-correlation between monthly rainfall data of stations in the upper Awash
River Basin
Station Addis A. Addis
Alem
Holeta Sendafa Debrezeit Modjo
Addis A. 1.000
Addis Alem 0.875 1.000
Holeta 0.927 0.968 1.00
Sendafa 0.805 0.765 0.805 1.000
Debrezeit 0.828 0.812 0.884 0.753 1.000
Modjo 0.815 0.791 0.866 0.749 0.814 1.000


103
Table 5.3 Pan evaporation daily data
Addis Ababa : 1987-1988
Month Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Evap.
(mm)
158 150 155 104 152 106 75 77 94 164 189 114

Hydrological data

Daily stage readings of three rivers at four stations with corresponding rating curve equations
were obtained from the Hydrology Department of the Ministry of Water Resources of
Ethiopia (Table 5.4). Daily discharge values are computed using the rating curves, which were
developed for specific periods. The availability of the raw data enabled us to check the
reliability and quality of the data. Among the stations considered gauging site of Awash
Melka Kuntre and Modjo at Modjo are located in a very stable cross-section and hence have
unique rating curve for the entire record. Berga station exhibits morphological change. Three
stage-discharge relationships are developed which are valid in specific period (Table 5.5).

The gauging station at Hombole has also a stable channel but distinct relationships are
available flow depths less and greater than 0.72 m up to 1987. Afterwards it appears that a
single relationship is adequate for all stages.

Table 5.4 Daily River flow data

STATION
Area
(km
2
)
Period Rainfall Station
Berga -Addis Alem 4456 1986-1990 Addis Alem , Holeta
Modjo - Modjo 7656 1985-1989 Modjo, Sendafa
Awash - M. Kuntre 248 1985-1992 Holeta, Addis Alem
Awash - Hombole 1264 1963-1992 Addis Alem, Holeta, Addis A.,
Modjo Debrezeit, Sendafa

Table 5.5 Stage-discharge relationships rating curves for different periods. (h=Stage (m), and
Q= flow m
3
/s)
Valid during River
From To
Rating curve equations


104
Jan. 01, 1985 Aug. 02, 1986 ( )
77 . 1
44 . 0 56 . 9 = h Q
Aug. 03, 1986 Aug. 27, 1987 ( )
12 . 2
66 . 0 29 . 11 = h Q
Berga
Aug. 28, 1987 Dec. 31, 1990 ( )
98 . 1
52 . 0 64 . 9 = h Q

Awash
M.Kuntre
Jan. 01, 1980 Dec. 31, 1990 ( )
104 . 2
04 . 0 03 . 22 = h Q

Jan. 01, 1962 Dec. 31, 1987
( )
582 . 1
26 . 0 32 . 10 + = h Q
h0.72m
( )
754 . 11
04 . 0 67 . 19 = h Q
h0.72m
Awash
Hombole
Jan. 01, 1988 Dec. 31, 1990 ( )
73 . 2
78 . 0 195 . 3 + = h Q

Modjo Jan. 01, 1984 Dec. 31, 1990 ( )
168 . 5
1 . 0 636 . 4 = h Q

Data quality checking

The stages are checked through time series plots, which revealed a number of inconsistencies
due to data entry error. The flow data at a station are validated by simultaneously examining
the rainfall time series and the flow of the other rivers.

For example the consistency of the flow data of Awash River at two stations is verified using
correlation and time series plots. The cross correlation coefficients for these two stations are
0.92 and 0.89 for lag 0 and lag 1 day respectively. Figure 5.3 shows a typical hydrograph of
the two stations on the Awash River. Comparing the two hydrographs, one can see that the
contribution of the drainage area below the Melka Hombole is not only large but also highly
variable diurnally during the rainy season.

The major errors found during data checking deal with the high flows computed using the
rating curves. Several of these errors are observed. For example considering the Modjo station
the record shows that the stage readings of August 8 and August 9, 1985 are 2.08 m and 3.04
m respectively. Using the rating curve the corresponding discharges are 260 m
3
/s and 1715
m
3
/s. The later flow is extremely large, with an equivalent of 110 mm of rainfall over the
whole catchment area. The rainfall record at Modjo on that particular day is only 15 mm and
no rain on August 7 1985. Similarly over estimation is seen also on 15 August 1985.

This overestimation is due to (at least for this station) the nature of the equation of the rating
curve which is derived from pure regression analysis of the measured stage and discharge. For
example the rating curve equation for Modjo is:


105
( )
168 . 5
1 . 0 636 . 4 + = h Q

Where Q= discharge of the river in m
3
/s
h= Stage (water level) in m

From a theoretical point of view the value of the exponent should be in the range of 1.5 to 2.5
for wide trapezoidal cross-section. In this case the value of the exponent, as determined from
the curve fitting is 5.168, which explodes the computed discharge for a small increase of
stage. The limitation of the rating curves is mostly in the extrapolation of high flow regimes.
Hence due care must be taken in the use of the relationships for extreme stages. For this
catchment, the overestimated points are quite few. They are replaced by the mean of either
the previous or the following day. When the stage of the following day is smaller, the stage of
the previous day is taken. It is advisable in such instances to investigate the cross-section of
the gauging site and calculate the maximum capacity of the river channel discharges using the
Manning or Chezy formulas.

5.2.2 Chinese catchments

Two catchments from the Shanxi Province of China are used. The data are obtained from (Li,
1995). The region has semi-arid climate with an average annual precipitation between 500 and
600 mm. The maximum precipitation in a wet year reaches about 900-1300 mm and the
minimum precipitation in a dry year can be as low as 200-250 mm. Most of the precipitation
is received during the rainy season (July to September). The annual potential evaporation is
about 900 mm.


106

0
100
200
300
400
500
600
700
800
1 31 61 91 121 151 181 211 241 271 301 331 361
days (from January 1)
f
l
o
w

(
m
3
/
s
)
Melka Kuntre
Hombloe

Figure 5.3 Typical comparison of hydrograph of Awash River at Melka Kuntre and Hombole
stations (1988)

0.01
0. 1
1
10
100
1000
10000
0 50 100 150 200 250 300 350
Stage, h (cm)
D
i
s
c
h
a
r
g
e
(
m
3
/
s
)

August 9, 1985
h= 304 cm
Q= 1715 m3/s
August 8, 1985
h= 208 cm
Q= 260 m3/s

Figure 5.4 Over estimation of rating curves in extrapolation (Modjo River at Modjo)


107
5.2.3 West African Catchments

The two basins on Faleme River are tributaries of the Senegal River. The other catchments
Boa, N'zo, and Ferdougouba are tributary of Sassandra river. One catchment is the upper part
of the Niger River basin. The location of the basins is given in Figure 5.5. The detailed
description of the catchments can be found in Ni-Lar-Win (1994).

The upper Senegal basin is characterized by its tropical climate. The basin receives most of its
rainfall during July - October. The geology of the basin is composed of schist, micaschist,
quartzite cut by volcanic rocks (metaandesite and metabasite) and plutonic (granite and
granodiorite). To the south, the plutonic intrusion creates a succession of subtabular
morphologies covered by lateritic terrain within which rivers have dissected V shape valleys.
The vegetation cover of the two basins on Falem is savannah with trees and bushes, cultivated
area, savannah densely covered with trees and dry forest.

The Niger basin lies on the transition of the tropical region. The duration of the rainy season is
about four months (July - September) and the dry season is less severe than the in the tropical
region. The geology of the basin is composed of strongly metamorphosed and recrystalized
granites, that appears as Gnesis in many cases. The vegetation cover of the Niger basin is
densely covered with trees, dry forest, cultivation, and fallow land.

The geology of the Sassandra basin is somewhat similar to the upper basin of Niger. The
climate and vegetation covers of the four basins are different from one another. Boa and
Ferdougouba lie in the tropical transition region with one rainy season during the months of
July-September. The Boa basin is densely covered with Savannah trees while Ferdougouba is
covered with sparse trees and bushes. The climatic region of Bafing is tropical transition and
equatorial transition. It is covered by evergreen forest and secondary forest. The N'Zo basin
lies in the equatorial transition region characterized by scarce rain from march to June, and
much rain in September and October. It is covered with evergreen forest and secondary forest.


108
0 100
200 Km

7 Boa 8 N'Zo 9 Ferdougouba 10 Niger
11 Falem 12 Falem 13 Bafinga

Figure 5.5 Location of West African basins under study


109
5.2.4 Tanzanian catchments

Tanzania is a semi-arid country transversed on average by a line of 6 degree below the
equator. Hydrometeorological data of Little Ruaha catchment from Tanzania (Figure 5.6) at
the outlet of Makalala and Ihimbu station were available The area covered up to Makalala is
759 km
2
and for Ihimbu is 2480 km
2
. The catchment has on average an annual historical
rainfall of 1000 mm/yr. There are six rainfall-gauging stations in the catchment with
continuous data for 9 years (1966-1975). Discharges at Ihimbu and Makalala are available
from 1966 to 1975. The catchment has only one evaporation station at Madibira with data
from 1966 to 1975. The data for these catchments were obtained from University of Dar Es
Salaam, Tanzania.

5.2.5 Zambian catchments

Data of five catchments from Zambia were available for the study. Information on the
catchment was obtained from the country report prepared by Mott Mac Donald and Partners,
(1990). All the five catchments are located in the upper part of the Kafue River Basin (Figure
5.7). The basin encompasses a very large area (155 000 km
2
) which is about 20 % of Zambia's
total land area. The mean annual rainfall over the catchment ranges from 700 mm in the
south to over 1200 mm in the north and has an areal mean rainfall of 1050 mm. The mean
annual flow at the outlet at Kasaka is about 350 m
3
/s, equivalent to 71 mm. The average
runoff coefficient at this large basin scale is only 0.06, which implies that more than 90
percent of the rainfall disappears mainly as evaporation. The rainy season is well defined,
occurring in during the months of November to April.

According to Hughes, (1997) the Kafu river is underlain by shales, sandstones, dolmites and
quatzites of the Katanga system and some areas of basement complex comprising gneisses,
schists and micaceous quartzites. The associated soils vary from clay to sandy clay loams and
are generally quite deep. Vegetation and land use consists of natural woodland, forest
reserves, cultivated land and urban/industrial areas associated with mining.


110

Figure 5.6 Location of Ruaha Catchment (Tanzania)

Figure 5.7 Location of Kafu catchments. (Zambia)

Ruaha Catchment
Kafue
Catchments

111
The data available for the study in the basin are the monthly discharge, precipitation and pan
evaporation. For Mwambashi and Baluba daily data were available. These meteorological
and hydrological data are obtained from the Zambian Government Department of Water
Affairs and the Meteorological Department respectively.

5.2.6 Botswana catchment

Only data for one catchment was available from Botswana for this study. The brief description
for this catchment is taken from Hughes (1997). Tati river basin is located in the east part of
Botswana. The vegetation cover is savanna bush and trees, while the soils are predominantly
loamy sands to coarse sandy loams with 10 to 20 % clay contents and approximately 800 mm
in depth. In some areas of the upper catchments, the soils are sandy clay loams and sandy
clays that can be up to 2m deep. Six rainfall stations are used to calculate areal rainfall for
the catchment. Only mean monthly evaporation data are available for the study. The data
were obtained from Southern African " Friend" project.

Tati Catchment
Figure 5. 8 Location of Tati catchment, (Botswana).


112
5.2.7 Common characteristics of the test catchments

A total of 20 catchments from West Africa, North East and South Africa and China were used
in this study. The catchment areas range from 248 to 16820 km
2
. Seasonal variation of
rainfall, evaporation and flow characteristics are shown in Figure 5.10. Though for some of
the catchments the amount of rainfall seems high, its concentration within a short period of the
year causes low flow and in some years even zero river flows. On an annual basis, only up to
10-20% of the rainfall reaches the outlets as river flow for the East and West African
catchments and 40 % for the Chinese catchments.

5
3
1: Tanzania
2: Ethiopia
3: Ivory Coast
4: Guinea
5: Senegal
2
1
4
7
6
6: Botswana
7: Zambia

Figure 5. 9 Location of the study areas in Africa.

Table 5.6 Characteristics of the test catchments and available data on monthly time step.


113
Annual quantities (mm)
No River Station
Area
(km
2
) Rain Evap. Flow
Period
Tanzania
1. Little Ruaha Ihimbu 2480 1041 1446 232 1966-1975
2. Little Ruaha Makawali 759 1244 1449 175 1966-1975
Ethiopia
3. Berga Addis Alem 248 1089 1537 262 1986-1990
4. Modjo Modjo 1264 871 1537 98 1985-1989
5. Awash Melka Kuntre 4456 1123 1537 180 1986-1991
6 Awash Hombole 7656 980 1537 160 1963-1992
Ivory Coast
7. Boa Vialdougou 5770 1285 1458 144 1981-1988
8. N'zo Kahin 4300 1618 1309 364 1981-1988
9. Ferdougouba N'golod. 5020 1492 1361 330 1981-1988
Guinea
10. Niger Kouroussa 16280 1467 1518 264 1981-1988
Senegal
11. Faleme Fadougou 2370 995 1814 97 1981-1988
12. Faleme Moussala 5720 993 1814 84 1981-1988
13 Bafing Bafindala 6230 1450 1318 250 1981-1988

Botswana

14. Tati Tati 470 581 1591 87 1985-1991
Zambia
15. Mwambashi Mwambashi 809 1219 1554 273 1978-1989
16. Lufumpa Kasempa 1062 951 1564 134 1964-1987
17. Kafue Kipushi 440 1120 1734 233 1961-1991
18. Baluba Baluba 306 1501 1361 239 1975-1989
China
19. SX1 Qianghi Province 809 560 883 206 1957-1970
20. SX2 Shangxi Province 751 495 873 60 1956-1970


114

0
50
100
150
200
250
Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep
m
m
/
m
o
n
t
h
Ruha -Ihimbu
0
50
100
150
200
250
300
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
m
m
/
m
o
n
t
h
Rainfall
pot. l evap.
runoff
Awash-Hombole

0
50
100
150
200
250
300
m
m
/
m
o
n
t
h
Boa
0
20
40
60
80
100
120
140
160
180
m
m
/
m
o
n
t
h
Tati

0
20
40
60
80
100
120
140
160
180
m
m
/
m
o
n
t
h
Sx1
0
50
100
150
200
250
300
350
m
m
/
m
o
n
t
h
Mwambashi

Figure 5.10 Long term monthly pan evaporation, precipitation and flow for some of the study
catchments


115
5.3 Monthly model application

A total of 20 catchments are used to test the monthly water balance models developed in
Section 3.2.

5.3.1 Initial soil moisture index

For the models developed, it is shown that the starting moisture index should be
approximated. Since this is an average value that indicates the moisture within the catchment
in the form of unsaturated and saturated water in the soil, it is rather difficult to measure it. A
warming up period would be necessary before the calibration period so that the model results
would not be influenced by the approximated initial soil moisture index. It has been
suggested that 3 to 6 years may be required for the warming up period in the humid
catchments. However in regions where we have strong seasonal variation and extended dry
seasons 3 -6 years data of warming up period may not be necessary.

To investigate the warming up period required several catchments are tested by using 1, 2, 3
and 6 years of warming up periods. The calculated soil moisture storages just before the
calibration period begins is examined. Table 5.7 shows a procedure to compare the influence
of the length of warming up period.

Table 5.7 Representation of different warming up period
Warming up period Calibration Period
m
o
1 year m
01

m
o
2 years m
02

m
o
3 years m
03

m
o
6 years m
06

Table 5.8 indicates that the resulting initial soil moistures are almost the same, regardless of
the length of the warming up period. This is mainly due to the fact that during the extended
dry season the catchment will exhaust all its storage and consequently will have no carry over
for the following years. The implication of this result is that for such regions a rather short
warming up period is sufficient. In case of limited length of data one full hydrological year
would be sufficient as a warming up period. A hydrological year may be defined as a period


116
starting from one regime of flow to the next similar regime after 12 months, e.g. (from a low
flow regime to the next low flow regime). This definition allows to consider a complete
hydrological season for the whole year.

Another possible solution is to append long term mean quantities for a required year before
the calibration period. It is worth mentioning that at least one year of observed data should be
available as a warming up period to obtain realistic value of initial soil moisture storage.

Table 5.8 Influences of length of warming period on initial soil moisture storage using
MWBM-A (IR=2, b
1
=2, b
2
=2)

Catchment Starting
m
o
(mm)
m
01
(mm)
m
02
(mm)
m
03
(mm)
m
06
(mm)
Awash-Hombole 100 164.7 167.8 168.0 168.0
Baluba 500 593.1 592.1 594.0 597.8
Sx1 100 167.9 169.53 168.8 167.2

5.3.2 Parameter optimization

Two optimization methods, VAO5A and the Shuffled Complex Algorithms, are used to
estimate the parameters. For the monthly water balance models (with few parameters) the two
methods have given the same results. In general it may therefore be concluded that local
optima are not a problem with monthly models. The first method takes a relatively small
computation time. In fact computation time for calibration with either of algorithm is not a
constraint with the present personal computer capabilities. The order of magnitude of
calibration time for a model for a data length of 30 years is 1 to 3 minutes with Pentium-2
machines. Sometimes the Shuffled Complex algorithm was used when good scaling factors
are not obtained for the other method

5.3.3 Selection of the forms of the model

Two types of models are considered: The model without imposing an upper limit to the soil
moisture, (MWBM-A) and a model with an upper soil moisture limit, (MWBM-B). In each
case, because of the 2 evaporation equations (IR=1,2) and the 9 combinations of the two


117
discrete parameters (b
1
and b
2
), there are 18 possible forms of a model. Hence, for each
catchment, 18 forms of the two models had to be calibrated and analyzed. The details of this
analysis are given for the Boa catchment in Ivory Coast. The time series of monthly rainfall,
evaporation and flow are shown in Figure 5.11. Table 5.9 and Table 5.10 list the parameter
estimates and statistical summary of the results for the three of possible forms for MWBM-A
and MWBM-B. The estimated parameters are shown after scaling with appropriate factor as
described in Appendix B

0
50
100
150
200
250
300
350
400
0 12 24 36 48 60 72 84 96
months starting January 1981
(
m
m
)
Rain fall
Evaporation
Flow

Figure 5.11 Time series of Input data (Boa Catchment, Ivory Coast)

It appears that for most combination of the discrete parameters, the optimum parameters are
significant and uncorrelated (the absolute value of correlation coefficient between parameters
is not greater than 0.9). This justifies the structure of the model and means that each parameter
retains realistic information from the data used for calibration. Global statistics, mean and
standard deviation of residuals with the model performance measures are provided for each
calibration. Table 5.11 and Table 5.12 summarize the results of all combinations for the two
variants of the model.

Table 5.9 Example application of the MWBM-A
IR=2 IR=2 IR=2


118
b1=0.5, b2=0.5 b1=1, b2=1 b1=2, b2=2
Parameters
Evaporation a
1
0.149 0.07 0.789 0.38 0.257 0.07
Slow flow a
2
0.315 0.014 0.177 0.07 0.259 0.01
Fast flow a
3
0.169 0.002 0.671 0.01 0.086 0.03
Max. correlation 0.49 (a
1,
a
2
) 0.55 (a
1,
a
3
) 0.62(a
1,
a
3
)

Results
Model Quality 2.46 2.47 2.45
Efficiency 85% 95 95
Mean of Residuals -0.06 0.08 0.03
Standard Deviation 0.65 0.53 0.48

Table 5.10 Example application of the MWBM-B
IR=2,
b1=0.5, b2=0.5
IR=2,
b1=1,
b2=1
IR=2,
b1=2,
b2=2
Parameters
Evaporation a
1
0.140 0.090 0.717 0.30 0.2750.07
Slow flow a
2
0.315 0.015 0.170 0.07 0.2910.01
Fast flow a
3
0.118 0.002 0.671 0.01 0.0460.03
Upper storage a
4
0.948 0.30 0.757 0.10 0.5880.12
Max. correlation 0.54(a
1,
a
2
) 0.55.(a
1,
a
3
) 0.63(a
1,
a
3
)

Results
Model Quality 2.46 2.47 2.45
Efficiency 85% 95 95
Mean of Residuals -0.01 0.08 0.05
Standard Deviation 0.66 0.53 0.49

* Note the parameter and HWCI values are scaled (See Appendix B for appropriate scales)


119
Table 5.11 Summary of results for all 18 variants of the model WBAM-A. (Boa Catchment,
Ivory Coast)

Discrete parameters Discrete parameters
N0.
IR b
1
b
2

Model
Qual.
Eff.
%
N0.
IR b
1
b
2

Model
Qual.
Eff.
%
01 0.5 2.47 85 10 0.5 2.46 85
02 1.0 2.45 93 11 1.0 2.45 95
03
0.5
2.0 2.46 94 12
0.5
2.0 2.47 94
04 0.5 2.45 88 13 0.5 2.45 90
05 1.0 2.47 95 14 1.0 2.47 95
06
1.0
2.0 2.45 95 15
1.0
2.0 2.48 95
07 0.5 2.45 93 16 0.5 2.40 90
08 1.0 2.22 88 17 1.0 2.17 84
09
IR=1
2.0
2.0 2.45 95

18
IR=2
2.0

2.0 2.48 95

Table 5.12 Summary of results for all 18 variants of the WBAM-B. (Boa Catchment, Ivory
Coast)

Discrete parameters Discrete parameters
N0.
IR b
1
b
2

Model
Qual.
Eff.
%
N0.
IR b
1
b
2

Model
Qual.
Eff.
%
01 0.5 2.41 85 10 0.5 2.46 85
02 1.0 2.45 93 11 1.0 2.45 95
03
0.5
2.0 2.44 94 12
0.5
2.0 2.45 95
04 0.5 2.46 88 13 0.5 2.44 90
05 1.0 2.12 95 14 1.0 2.47 95
06
1.0
2.0 2.45 95 15
1.0
2.0 2.48 95
07 0.5 2.43 93 16 0.5 2.40 90
08 1.0 2.17 86 17 1.0 2.17 84
09
IR=1
2.0
2.0 2.45 95

18
IR=2
2.0

2.0 2.45 95


120
The task of selecting the best model is not easy since some of the quality criteria sets are often
not significantly different for different forms of a given model. The following priority list is
suggested to examine the model outputs

1) Success of the optimization method. One should at least get the minimum of the objective
function in the environment of the optimum values of parameters (Figure 5.12). It is noted that
for the model structures with an upper soil moisture limit, one obtains a flat limb of the
objective function which indicates that the parameters has no influence on the objective
function for values higher than the optimized value (See a
4
in Figure 5.12.)

2)The significance of the parameters (confidence intervals): before any analysis on results of
a model it is important to check if all the optimized parameters are statistically significant as
explained in section 4.4.

3) Graphical comparison of computed and observed flows: Visual inspection of the agreement
between computed and observed river flows allows the modeler to accept or reject a model.
Though this analysis relies on subjective analysis, it is found to be more practical in some
cases where other general statistics such as model quality and efficiency do not show variation
for the spectrum of model forms.


121
0
4 0
8 0
1 2 0
0 0 .1 0 .2 0 .3 0 .4 0 .5
Sca l ed Pa ra meter
a
1
S
S
Q
0
4 0
8 0
1 2 0
0 0 .2 0 .4 0 .6 0 .8
Sca l ed pa ra meter
a
2
S
S
Q
0
4 0
8 0
1 2 0
0 0 . 0 5 0 . 1 0 . 1 5 0 . 2
Sca led P a ra met er a
3
S
S
Q
0
4 0
8 0
1 2 0
0 . 3 0 . 4 0 . 5 0 . 6 0 . 7 0 . 8 0 . 9
Sca led P a ra met er a
4
S
S
Q

Figure 5.12 Plot of the parameter and the sum of squares of error in the neighborhood of the
optimum value. (Boa catchment -modeled with MWBM-B)


122
Figure 5.13 displays the time series of the observed and calculated flows using the two
variants of the model. It can be noted from the graph that both models resulted with good
agreement between computed and observed flows.

0
10
20
30
40
50
60
70
80
90
0 12 24 36 48 60 72 84 96
f
l
o
w

(
m
m
)
Obserrved
MWBM-A
MWBM-B

Figure 5.13 Comparison of observed and modeled flow with the two variants of monthly
water balance models (Boa catchment)

4) Comparison of monthly means of observed and computed flow: The seasonal statistics are
essential especially for regions with strong variation of flow in a year. This test reveals mainly
the general performance of the model over flow regimes.

5) Model efficiency and Quality Factor: These statistics define the overall variance explained
by the model. Perhaps the model efficiency is the most used in the hydrology community to
evaluate a model and compare its performance with the results of other models. Unlike the
previous tests there is no influence of subjectivity. Because of its singular value these tests
allow to draw straightforward conclusions. However precaution has to be given if the other
tests are also supporting the conclusions. For MWBM-A the model quality factor varies from
2.17-2.48 and the efficiency varies from 84-95% (Table 5.11). For the second variant of the
model, MWBM-B the model quality factor ranges from 2.12-2.48 and the efficiency ranges
from 85-95% (Table 5.12). From these tables we observe that for the catchment considered


123
the model quality factor and the model efficiency do not give clear indication of the best
model and hence selection of the best model should be supported with other criteria.

6) Residual analyses: Various analyses of residuals are implemented in the model as described
in Section 4.6. The influence of an input variable can be depicted by plots of residuals versus
the input variables. The point of attention is to observe whether the variance of the residual is
the same for lower and higher values of input variables. It should be noted that the residuals
analyzed are the differences between the square root of observed and square root of computed
flows. The residual series versus the rainfall, evaporation and computed flow for both models
have showed equal variance over the range of the water balance variables (see Figure 5.14
through Figure 5.19). The independence of the residual series is checked by the
autocorrelation function (Figure 5.20). A Time series plot of residuals (Figure 5.21) depicts
there is no trend and/or periodicity of residuals. The distribution of the residuals grouped by
season is also tested for normality as supposed by the model hypothesis.

In addition to the above tests the evolution of intermediate water balance variables such as soil
moisture and actual evaporation should be examined, to analyze whether the values are
realistic. For the example catchment considered, the soil moisture fluctuates rapidly. Figure
5.22 and Figure 5.23 show the evolution of the soil moisture and the actual evaporation
respectively. According to the model, actual evaporation equals the potential for a few months
during the rainy season and remains low for the dry periods. This observation can be
generalized for most of the test catchments

One of the features of the model is the separation of the total flow in to 'Fast flow' and 'Slow
Flow'. This definition corresponds to direct and base flow components based on classical flow
separation methods. The fast flow component comprises large proportion of the flow for the
catchment considered (Figure 5.23 and Figure 5.25). It is noticed that the two models show
almost a similar proportion of the flow components. However the MWBM-B tends to give a
slightly higher percentage of fast flow than MWBM-A.
.


124

-1.5
-1
-0.5
0
0.5
1
1.5
2
0 50 100 150 200 250 300 350 400
rainfall (mm)
r
e
s
i
d
u
a
l
s

Figure 5.14 Residual versus rainfall MWBM-A. (Boa Catchment, Ivory Coast)

-1.5
-1
-0.5
0
0.5
1
1.5
0 50 100 150 200 250 300 350 400
rainfall (mm)
r
e
s
i
d
u
a
l
s

Figure 5.15 Residual versus rainfall MWBM-B (Boa Catchment, Ivory Coast)


125

-1.5
-1
-0.5
0
0.5
1
1.5
0 10 20 30 40 50 60 70 8
computed flow (mm)
r
e
s
i
d
u
a
l
s
0

Figure 5.16 Residual versus computed flow MWBM-A. (Boa Catchment, Ivory Coast)

-1.5
-1
-0.5
0
0.5
1
1.5
0 10 20 30 40 50 60 70 8
computed flow (mm)
r
e
s
i
d
u
a
l
s
0

Figure 5.17 Residual versus computed flow MWBM-B. (Boa Catchment, Ivory Coast)


126
-1. 5
-1
-0. 5
0
0. 5
1
1. 5
70 80 90 100 110 120 130 140 150
potential evapotranspiration (mm)
r
e
s
i
d
u
a
l
s

Figure 5.18 Residual versus Potential evapotranspiration MWBM-A. (Boa Catchment, Ivory
Coast)

-1.5
-1
-0.5
0
0.5
1
1.5
70 80 90 100 110 120 130 140 150
r
e
s
i
d
u
a
l
s

Figure 5.19 Residual versus Potential evapotranspiration MWBM-B. (Boa Catchment, Ivory
Coast)


127
-1
-0. 8
-0. 6
-0. 4
-0. 2
0
0. 2
0. 4
0. 6
0. 8
1
0 3 6 9 12 15 18 21 24
lag (month)

C
o
r
r
e
l
a
t
i
o
n
lower Confidence limit
upper confidence limit

Figure 5.20 Auto correlation of residuals from MWBM-A. (Boa Catchment, Ivory Coast)

-1. 5
-1
-0. 5
0
0. 5
1
1. 5
0 12 24 36 48 60 72 84 96
r
e
s
i
d
u
a
l
s
MWBM-A
MWBM-B

Figure 5.21 Time series plot of residuals. (Boa Catchment, Ivory Coast)


128
0
100
200
300
400
500
600
700
0 12 24 36 48 60 72 84 96
s
o
i
l

m
o
i
s
t
u
r
e

s
t
o
r
a
g
e

(
m
m
)
MWBM-A
MWBM-B

Figure 5.22 Comparison soil moisture evolution flow with the two variants of monthly water
balance models (Boa Catchment, Ivory Coast)

40
60
80
100
120
140
160
0 12 24 36 48 60 72 84 96
e
v
a
p
o
r
a
t
i
o
n

(
m
m
)
PET
AE- MWBM-A
AE-MWBM-B

Figure 5.23 Comparison of the actual evaporation with the two variants of monthly water
balance models. (Boa Catchment, Ivory Coast)


129

0
20
40
60
80
100
120
140
0 12 24 36 48 60 72 84 96
f
l
o
w

(
m
m
)
FastFlow
Base flow

Figure 5. 24 Flow components using MWBM-A. (Boa Catchment, Ivory Coast).

0
10
20
30
40
50
60
70
0 12 24 36 48 60 72 84 96
f
l
o
w

(
m
m
)
FastFlow
Base flow

Figure 5.25 Flow components using MWBM-B. (Boa Catchment, Ivory Coast).

5.3.4 Summary of the results for the test catchments


130
The total number of model calibrations is 20 stations times 2 variants of the model times 18
possible combinations which equals to 360 calibrations

According to the list of tests described in the previous section for individual catchments it is
noted that the evaporation equation IR=2 yields often better results as compared to IR=1. It is
also observed that model forms associated to b
1
=0.5 and b
2
=0.5 result in a systematic
underestimation of both fast and slow flows. The model with the combination of IR=2, b
1
=2
and b
2
=2 results in a reasonable quality of agreement between simulated and observed flows
for all the sub-basins and may therefore be considered as the general model form. In instances
where the quality is not acceptable one can try other combinations of the discrete parameters.
It is possible in some cases that other forms of the models give a slightly better performance
in one of the criteria, however the results of only one combination are reported for even
comparison.

The estimates of model parameters with the associated model performance indicated by the
efficiency and quality factors are given in Table 5.13 and Table 5.14 for MWBM-A and
MWBM-B respectively. For the catchments considered, the model without imposing upper
soil moisture has efficiencies that vary from 60-97% and model quality factor ranging from
1.52-3.79. The model with upper soil moisture storage has efficiencies that range from 62-
95% and quality factors that vary between 1.57-4.0. The quality factor greater than 1.5 has
been classified as good and for ranges above 2.5 the model is considered to be very good for
monthly water balance models.

From Table 5.13 and Table 5.14 it is observed that the quality factor has similar magnitudes in a
given region. For Example for the Tanzanian catchments, the values are all around of 1.5
while for the west African catchment we have value in the range of 2.5-3.5. This is in
consistence with the assertion of Xu (1992) that the quality factor heavily depends on the
climatic and physical characteristics of the basin.


131
Table 5.13 Parameter estimates for without upper limit to storage -MWBM-A (IR=2, b
1
=b
2
=2.0)
parameter
Catchment
Evaporation
(a
1
)x10
-2

Slow flow
(a
2
) x10
-4

Fast flow
(a
3
) x10
-4

Eff.
%
MQ
1. Ruaha - Ihimbu .050 .159 .003 75 1.54
2. Ruaha - Mkalala .071 .208 .002 78 1.52
3. Berga .230 .571 .004 75 4.00
4. Modjo .120 .010 .050 62 3.15
5. Awash - Melka .145 .450 .090 97 3.19
6. Awash - Hombole .204 .120 .025 88 3.08
7. Boa .257 .259 .009 95 2.45
8. N'zo .716 .575 .020 79 3.79
9. Ferdougouba .278 .502 .012 82 3.09
10. Niger .431 .804 .010 77 3.04
11. Faleme - Fad. .809 .717 .038 80 2.92
12. Faleme - Mous. .855 .589 .030 80 2.64
13. Bafing .550 .458 .062 61 2.84
14. Tati .315 .080 .000 73 2.66
15. Mwambashi .171 .132 .014 79 3.22
16. Lufumpa .200 .467 .010 65 3.00
17. Kafu .100 .678 .020 60 2.42
18 Baluba .240 .794 .030 72 2.75
19. SX1 .254 .160 .600 82 3.19
20. SX2 .204 .255 .028 91 2.01

Table 5.14 Parameter estimates for with upper limit to storage MWBM-B (IR=2, b
1
=b
2
=2.0)
parameter
Catchment
Evaporation
(a
1
)x10
-2

Slow flow
(a
2
) x10
-4

Fast flow
(a
3
) x10
-4

Upper limit
(a
4
)
Eff.
%
MQ
1. Ruaha - Ihimbu .120 .613 .006 .775 68 1.57
2. Ruaha - Mkalala .130 .606 .002 .545 57 1.68
3. Berga .23 .429 .050 .414 79 3.93
4. Modjo .120 .010 .050 .471 67 3.14
5. Awash .231 .220 .020 .463 93 2.99
6. Awash .260 .821 .060 .263 83 3.28
7. Boa .275 .291 .046 .588 95 2.45
8. N'zo .201 .139 .009 .756 94 2.49
9. Ferdougouba .572 .575 .029 .554 68 4.00
10. Niger .330 .572 .019 .635 78 3.13
11. Faleme - Fad. .315 .579 .011 .000 78 3.03
12. Faleme -Mas. .442 .368 .038 .418 80 2.93
13. Bafing .212 .400 .070 .867 62 2.83
14. Tati .186 .100 .090 .280 71 2.65
15. Mwambashi .153 .731 .030 .657 70 2.17
16. Lufumpa .200 .385 .010 .638 65 2.97
17. Kafu .107 .120 .040 .733 73 2.54
18 Baluba .138 .177 .040 .818 75 2.47
19. SX1 .138 .544 .027 .470 88 2.82
20. SX2 .162 .159 .020 .451 78 2.07


132
Introducing the upper limit to the soil moisture marginally improved the model performance
only for a few catchments notably for the West African and Zambian catchments.

The variation of estimates of the parameters is shown in Figure 5.26. Because of the limited
information on catchment it has not been attempted to relate the parameter values with
catchment characteristics.

0. 0
0. 2
0. 4
0. 6
0. 8
1. 0
0 1 2 3
parmeters (a )
e
s
t
i
m
a
t
e
s

o
f

p
a
r
m
e
t
e
r
s
0. 0
0. 2
0. 4
0. 6
0. 8
1. 0
0 1 2 3 4
parmeters (a )
e
s
t
i
m
a
t
e
s

o
f

p
a
r
m
e
t
e
r
s

Figure 5.26 Ranges of estimates of parameter for MWBM-A (left) and MWBM-B (right)

5.3.5 Reproducing flow regimes

One of the requirements of a model is to reproduce the flow regimes. The models applied
generally conserve the long-term monthly mean flows for the high as well as for the low flow
regimes.

Figure 5.27 shows 6 examples of results obtained by using the two variants of the monthly
water balance models. The best agreement between the observed and simulated long-term
mean is shown for the Awash catchments in Ethiopia and Boa catchment in West Africa. For
some catchments as for example Mwambashi the models reproduce well the low flows and
medium flow regimes, but that high flows are not well reproduced. These differences may
probably be attributed to rainfall events concentrated in few days, which produce
exceptionally high flows.


133

0
5
10
15
20
25
30
35
40
m
m
/
m
o
n
t
h
Ruha-Ihimbu
0
10
20
30
40
50
60
70
80
m
m
/
m
o
n
t
h
Observed
MWBM-A
MWBM-B
Awash- Hombole

0
5
10
15
20
25
30
35
40
45
50
m
m
/
m
o
n
t
h
Boa
0
2
4
6
8
10
12
14
16
18
20
m
m
/
m
o
n
t
h
Tati

0
10
20
30
40
50
60
m
m
/
m
o
n
t
h
Sx1
0
10
20
30
40
50
60
m
m
/
m
o
n
t
h
Mwambashi

Figure 5.27 Comparison of seasonal observed and simulated monthly flows using MWBM-A
and MWBM-B.


134
As the model implicitly assumes a homogeneous rainfall distribution over the period of one
month, the effects of such short duration events are damped out. This phenomenon is linked to
the choice of the model time base and could be improved by reducing the time step of the
modelling. Similar results are observed for the other catchments.

5.3.6 Water balance variables

Besides establishing the rainfall-runoff relationship, the models compute the actual
evapotranspiration and from the simulations the water balance proportions can be obtained.
Considering rainfall as input to the system, the model results showed that from 70 to 90
percent of the water goes back to atmosphere through evaporation. Figure 5.28 through Figure
5.31 show the long-term average percentage of fast flow, base flow and actual evaporations
for the catchments. The percentage of fast flow and base flow varies greatly. The dominant
flow is fast flow for catchments from Ethiopia (Berga and Awash). For Tanzanian catchment
Ruaha the flow is dominated by base flow. The extreme case is that for Botswana catchment
Tati, where the flow is composed of only fast flow.

The possible reason for the differences of proportion is mainly due to the length of rainfall
season. With the exception of the Tati, catchments with longer rainy season have relatively a
higher proportion of slow flow components. Other possible reasons could be the type of soil
and geological formation of the catchments, this can only be verified by detailed studies of the
geology of the catchments.


135
Ruha -Ihimbu
77%
6%
17%
Ruha -Makawila
86%
3%
11%
Berga
76%
20%
4%
Awash -Melka
84%
14%
2%
Awash -Hombole
85%
13%
2%
Actual Evapor ation
Fast flow
Base flow

Figure 5.28 Proportion of water balances variables for East African catchments
Boa
89%
8%
3%
N' zo
12%
8%
80%
Ferdougouba
12%
78%
10%
Bafing
81%
10%
9%
Niger
83%
7%
10%
Faleme- Fadougou
6%
92%
2%

Figure 5. 29 Proportion of water balances for West African catchments


136
Lufumpa
86%
8%
6%
Tati
8%
92%
0%
Mwambashi
81%
5%
14%
Kafue
70%
5%
25%
Baluba
73%
5%
22%
Actual Evapor ation
Fast flow
Base flow

Figure 5. 30 Proportion of water balances for South African catchments

SX1
70%
14%
16%
Sx2
8%
88%
4%

Figure 5.31 Proportion of water balances for Chines catchments


137
5.3.7 Sensitivity of the parameters for the calibration period and extrapolation
test

Many researchers have advocated that model test and validation should be performed using
split-sample experiments. This is done by subdividing the input data into two parts. To
evaluate the sensitivity of the parameters to the calibration periods, the model is calibrated
separately using the entire data and two split parts. Consequently, the parameters obtained
from these three sets are compared. An example for a catchment is shown in Table 5.15. As -
in this case as well as for the other considered catchments - the confidence intervals for the
three sets overlap for all the parameters, it may be concluded that the three sets resulted in
comparable parameters and that the parameter values were not biased by the calibration
period.

Table 5.15 Comparison of parameter estimates and water balance variables for the three sets
of data (Awash catchment the discrete parameters: b
1
=2, b
2
=2)

Periods
Parameter

Scale
Of Par.
1963-1977 1978-1992 1963-1992
Evaporation Par, a
1
100 0.210 0.187 0.204 0.07
Fast Flow Par. a
2
10000 0.148 0.093 0.120 0.09
Slow Flow a
3
10000 0.045 0.053 0.041 0.02
Mean annual water Balance Variables (mm)
Rainfall 937 933.6 935
P. Evaporation 1794 1794 1794
Actual evaporation 758 777 770
Observed flow 183 165 175
Calculated flow 179 158 167

On the other hand extrapolation test can be performed by calibrating the model with the first
half of the data and simulate using the input of the second half. The performance of the model
in the two periods is compared. This shows the ability of the model to retain some of the
hydrological characteristic of the catchment, independent of the calibration period. The two
tests can be performed only if there is a sufficient length of data to obtain acceptable statistics.


138
0
20
40
60
80
100
120
140
160
0 12 24 36 48 60 72 84 96 108 120 132 144 156 168 180
months (1962 -1977)
f
l
o
w

(
m
m
)
Observed
MWBM-A

Figure 5.32 Comparison of observed and computed flow for the calibration period (Awash -
Hombole)

0
20
40
60
80
100
120
181 193 205 217 229 241 253 265 277 289 301 313 325 337 349
months (1978 - 992)
f
l
o
w

(
m
m
)
Observed
MWBM-A

Figure 5.33 Comparison of observed and computed flow for the verification period (Awash -
Hombole)
5.4 Conclusions

The conceptual monthly water balance models originally developed and tested for humid
regions are applied to arid and semiarid cases with slight modification of the structure of the


139
model apropos to the temporal variability of the hydrological processes in the latter regions.
More over it is shown that a particular representation of the evaporation term is justified and
some of the discrete parameters were narrowed to a definite range of values, indicating the
possible regionalization of the forms of the general equations governing the models.

The following conclusions may be specifically drawn from the present study:

For the catchments considered a warming up period of one to three year is sufficient to
avoid the influence of initial soil moisture assumption.
The model structure whereby the fast flow is expressed as a function of the average soil
moisture index resulted with good agreement of observed and calculated flow for most of
the catchment studied. Introducing an additional parameter such as upper limit of the soil
moisture index marginally improved the modelling in few catchments.
Evaporation Equation IR=2 is best suited for the catchments and also the discrete
parameters for slow flow and fast flow can be set to b
1
=2 and b
2
=2.
The models seem good enough to represent the rainfall runoff relationship for low flow
and medium flow but may over or under estimate the exceptional peaks.
The model parameters are not biased by the calibration period.
The seasonal mean flow statistics are well preserved.

The result of the modelling of the study catchments is encouraging for practical applications
such as extending and filling missing flow data for the periods where rainfall and evaporation
are available to obtain long-term flow data for water resource planning and management.

Chapter 6

6 CASE STUDIES OF
THE 10-DAY WATER BALANCE MODEL

6.1 Introduction.................................................................................................................................140
6.2 Data description...........................................................................................................................140
6.3 Detailed model tests with an example catchment.......................................................................143
6.4 Summary of the results for the test catchments...........................................................................154
6.5 Concluding remarks ....................................................................................................................158

6.1 Introduction

In this chapter, the conceptual 10-day water balance model (DWBM), described in section 3.3
is tested. Among the 20 catchments, 8 are selected due to availability of data on a 10-day time
step. The model results are compared to the once obtained by using the monthly model
structure of the second variant, MWBM-B with 10-day time steps. Since the structure of the
two models differs, a one-to-one comparison of the parameters will not be possible, however
general statistics of the results are compared.
6.2 Data description

Table 6.1 lists the 8 catchments used to test the models at 10-day time step. The characteristics
of the catchments were discussed in chapter 5. The 10-day rainfall and flow data for all test
6. Case studies of the 10-day water balance model.

141
catchments are obtained by aggregating the daily data. Figure 6.1 shows the mean 10-day
variation of rainfall and runoff in a year.

The monthly evaporation data were available for most of the catchments. To obtain the 10-day
time step values, the monthly evaporation amount is first divided in to three parts. Next, the
evaporation of the first decade of a month and the last decade are adjusted by taking the
average of the evaporation obtained by the month before and following. This gives a smooth
transition form one month to the next. Only for three catchments (Berga, Awash at Melka
Kuntre and Hombole) daily evaporation data at one station for period of two years (1988-
1989) were available. The 10-day mean for (36 values of the year) are calculated. These data
are used for all the catchments in the Awash River Basin.

Table 6.1 Test catchments for 10-day time step modelling.
No River Station Area Period
Tanzania
1. Little Ruaha Ihimbu 2480 1966-1975
2. Little Ruaha Makawali 759 1966-1975
Ethiopia
3. Berga Addis Alem 248 1986-1990
5. Awash Melka Kuntre 4456 1985-1991
6 Awash Hombole 7656 1985-1992

Botswana

14. Tati Tati 470 1985-1991
Zambia
15. Mwambashi Mwambashi 809 1978-1989
18 Baluba Baluba 306 1975-1989


142

0
20
40
60
80
100
r
a
i
n
f
a
l
l

(
m
m
)
0
5
10
15
20
f
l
o
w

(
m
m
)
rainfall
flow
Ruha -Ihimbu

0
20
40
60
80
100
120
r
a
i
n
f
a
l
l

(
m
m
)
0
5
10
15
20
25
30
f
l
o
w

(
m
m
)
rainfall
flow
Awash- Hombloe

0
20
40
60
80
100
120
140
160
r
a
i
n
f
a
l
l

(
m
m
)
0
5
10
15
20
25
f
l
o
w

(
m
m
) rainfall
flow
Mwambashi

0
10
20
30
40
50
r
a
i
n
f
a
l
l

(
m
m
)
0
1
2
3
4
5
6
7
f
l
o
w

(
m
m
)
rainfall
flow
Tati

Figure 6.1 10-day long-term mean of rainfall and flow for the study catchments


143
6.3 Detailed model tests with an example catchment

As in the monthly water balance models the output of the models, is too extensive to report for
all the studied catchments. Therefore we present the detailed result of only one catchment
(Awash at Hombole). Later, a summary of the results obtained for all the catchments and
conclusions derived are presented. Typical seasonal variation of the input variables is shown
in Figure 6.2. It is observed that the potential evaporation exceeds the rainfall for most of the
period, except during the rainy season. From the time series plot it is noted that peak flows
occur latter than the highest rainfall.

0
20
40
60
80
100
120
140
0 12 24 36 48 60 72 84 96 108 120 132 144 156 168 180
10-days starting January 1986
f
l
o
w

(
m
m
)
Rain
Pot. Evap.
flow

Figure 6.2 Time series of rainfall, evaporation and flow data for calibration (Awash at
Hombole)

6.3.1 Parameter optimization

In the process of optimization the limitation of the VAO5A procedure is clearly observed.
Specifically for the models studied in here, the problem is in selecting the right scale of
individual parameters. The procedure applied is first to use the Shuffled Complex Algorithm
by assigning a very wide range of parameters as lower and upper limit. Next, individual
parameters are scaled to intervals (0 - 1) and the scaling factor is retained for further call of
the VAO5A optimization method. The VAO5A algorithm then provides the necessary


144
statistics such as correlation between parameters, confidence intervals of individual
parameters.

The base flow parameter of the DWBM is estimated using the hydrograph analysis routine
incorporated in the model.

Regarding the parameter importance, when the time step of modeling is reduced from the
month to 10-day, it appears that the evaporation parameter value of a
1
has no influence on the
sum of squares of errors after a certain upper limit (See Figure 6.3). This phenomenon was
common for the two models. It has been said that in the second type of evaporation equation
the actual evaporation is computed as a function of available soil moisture and limited to the
potential evaporation. The physical explanation that follows is that at shorter time steps indeed
the potential evaporation is satisfied in most of the time during rainy days and hence the
influence of the evaporation parameter of a higher value is minimal. If one assumes the higher
value of the evaporation parameter, the model takes always the actual evaporation equal to the
potential.

The upper soil moisture limit, parameter a
4
, shows the same behavior as previously noted on
the monthly time step modelling. Once the critical maximum value corresponding to
minimum of sum of squares is achieved increasing this parameter will increase the sum of
errors of square slightly in the neighborhood of the optimum point and the effect will
disappear for the higher values. This is expected because for whatever values above the
critical point the soil moisture does never exceed this upper limit. In fact in the model
calibration process, the first trial is to use models without upper limit and investigate the
highest soil moisture attained. This gives an indication of the highest range of this parameter.
The other two parameters a
2
and a
3
are found to be fairly stable and significantly representing
the model structure.


145

Table 6.2 Parameter estimates of the MWBM-B and DWBM for Awash -Hombole)
MWBM-B DWBM
Parameter

estimates and
Confidence
interval
Parameters
estimates and
Confidence
interval
evaporation, a
1
x10
-2
0.78 0.37 evaporation, a
1
x10
-2
0.89 0.85
slow flow , a
2
x10
-2
0.76 0.42 percolation, a
2
x10
-1
0.46 0.28
fast flow, a
3
x10
-2
0.13 0.03 interflow, a
3
x10
-2
0.82 0.3
0.67 0.16 max. storage, a
4
x10
3
0.51 0.19
base flow, a
5
0.15
max. storage, a
4
10
3

Max. Correlation

0.85 (a
1
,a
2
)
Max. Correlation
0.677 (a
1
, a
3
)

6.3.2 Comparison of observed and computed flow

Generally speaking, the two models reproduce the flow series very well. From the comparison
of the observed and computed (Figure 6.4) except for the extreme peak flows the medium
flow and low flows are generally modeled quite satisfactorily. The model efficiencies for the
catchment considered are 84% and 86% for MWBM-B and DWBM respectively. It is noted
that as in the monthly water balance models modelling the peak flows remains the weak point
of the whole modelling with the time steps considered.


146
0
40
80
120
0. 3 0. 5 0. 7 0. 9
Parameter a
4
S
S
Q
0
40
80
120
0 0. 5 1 1.
Parameter a
1
S
S
Q
5
0
40
80
120
0 0. 5 1 1. 5 2
parameter a
2
S
S
Q
0
40
80
120
0 0. 05 0. 1 0. 15 0. 2
parameter a
3
S
S
Q

0
40
80
120
0. 3 0. 5 0. 7 0. 9
Parameter a
4
S
S
Q
0
40
80
120
0 0. 5 1 1.
Parameter a
1
S
S
Q
5
0
40
80
120
0 0. 5 1 1. 5 2
parameter a
2
S
S
Q
0
40
80
120
0 0. 05 0. 1 0. 15 0. 2
parameter a
3
S
S
Q

Figure 6.3 Plot of the parameter and the sum of squares of error in the neighborhood of
optimum value. (left: MWBM-B, and right DWBM).


147

0
10
20
30
40
50
60
70
80
0 12 24 36 48 60 72 84 96 108 120 132 144 156 168 180
10-day starting January 1986
f
l
o
w

(
m
m
)
Obserrved
MWBM-B
DWBM

Figure 6.4 Comparison of observed and modeled flow with MWBM and DWBM (Awash-
Hombole)

6.3.3 Flow components

The two models have different assumption with regard to the flow components: the MWBM-
B classifies the flows into two components (fast and slow flow) while the DWBM adds one
more component (interflow). The need for further classification is discussed in the model
development. Considering the example case, no fast flow component is generated. It means
that at the 10-day time step the intensity of the rainfall is always lower than the infiltration
capacity. It is observed that the base flow component is much lower than what is considered
as slow flow in the monthly model structure because the former explicitly emerges from the
groundwater storage. Therefore for the 10-day model structure the interflow component
includes almost all the flow generated except the base flow. The flow classification of the
DWBM appears more realistic for the fact that for e.g. the base flow contribution for this
catchment is very low (Figure 6. 6). For the MWBM, the flow separation flow the fast flow is
the flow that is generated immediately within the time step considered and the slow flow is the
flow that arrives with delay often are more than one time step.


148
0
5
10
15
20
25
30
35
40
45
0 12 24 36 48 60 72 84 96 108 120 132 144 156 168 180
f
l
o
w

(
m
m
)
FastFlow
Base flow

Figure 6. 5 Flow components using MWBM-B. (Awash-Hombole)

0
5
10
15
20
25
30
35
0 12 24 36 48 60 72 84 96 108 120 132 144 156 168 180
f
l
o
w

(
m
m
)
Inter flow
Base flow

Figure 6. 6 Flow components using DWBM (Awash-Hombole)


149
6.3.4 Intermediate water balance variables

Apparently, the soil moisture and actual evaporation evolution of the two models agree quite
reasonably (Figure 6.7 and Figure 6. 8). For the DWBM the two storages are added together for
comparison. Regarding the actual evaporation it is noted that during the rainy season the
actual evaporation strictly follows the potential evaporation while during dry season the actual
evaporation is controlled by the available water in the catchment as expressed by the soil
moisture index.

0
50
100
150
200
250
300
350
400
450
500
0 12 24 36 48 60 72 84 96 108 120 132 144 156 168 180
s
o
i
l

m
o
i
s
t
u
r
e

s
t
o
r
a
g
e

(
m
m
)
MWBM-B
DWBM

Figure 6.7 Comparison of soil moisture storage evolution with MWBM-B and DWBM. (Awash-
Hombole)
0
10
20
30
40
50
60
70
80
0 12 24 36 48 60 72 84 96 108 120 132 144 156 168 180
e
v
a
p
o
r
a
t
i
o
n

(
m
m
)
PET
AE- MWBM-B
AE-DWBM

Figure 6. 8 Comparison of 10-day actual evaporation with MWBM-B and DWBM. (Awash-Hombole)


150
6.3.5 Residual analyses

Residuals obtained from the two models are investigated to verify the effects of the input of
the model and also to check the versatility of the model in ranges of the input. Plots of
residuals versus rainfall are given in (Figure 6.9). In all the cases the residuals plotted are the
difference between the transformed observed and computed flows.

-1.5
-1
-0.5
0
0.5
1
1.5
0 20 40 60 80 100 120
rainfall (mm)
r
e
s
i
d
u
a
l
s

Figure 6.9 Residual versus rainfall using MWBM-B (Awash-Hombole).

-1.5
-1
-0.5
0
0.5
1
1.5
0 20 40 60 80 100 120 140
rainfall (mm)
r
e
s
i
d
u
a
l
s

Figure 6.10 Residual versus rainfall using DWBM. (Awash-Hombole)


151
Plots of residuals versus potential evaporation ( Figure 6.11) show that for MWBM-B the
residuals have a constant deviation for the range of potential evaporation. For DWBM (Figure
6.12) the residuals show a smaller deviation for the high amounts of evaporation.

-1.5
-1
-0.5
0
0.5
1
1.5
0 10 20 30 40 50 60 70 8
r
e
s
i
d
u
a
l
s
0

Figure 6.11 Residual versus potential evapotranspiration using MWBM-B. (Awash-
Hombole)

-1. 5
-1
-0. 5
0
0. 5
1
1. 5
0 10 20 30 40 50 60 70 8
r
e
s
i
d
u
a
l
s
0

Figure 6.12 Residual versus potential evapotranspiration DWBM. (Awash-Hombole)


152
Plots of residuals versus computed flow (Figure 6.13 and Figure 6.14) show that for both
models the residuals have a constant deviation for the range of computed flows. The scatter is
denser around the low flows for the fact that during large periods of time the flow is in the low
flow regimes.

-1.5
-1
-0.5
0
0.5
1
1.5
0 5 10 15 20 25 30 35 40 45 50
computed flow (mm)
r
e
s
i
d
u
a
l
s

Figure 6.13 Residual versus computed flow using MWBM-B. (Awash-Hombole)

-1.5
-1
-0.5
0
0.5
1
1.5
0 5 10 15 20 25 30 35 40 45 50
computed flow (mm)
r
e
s
i
d
u
a
l
s

Figure 6.14 Residual versus computed flow using DWBM. (Awash-Hombole).


153
From the time series plot of the residuals (Figure 6.15), the residuals from both models do not
show any trend. The autocorrelation coefficient of the residuals is slightly higher than the 95
confidence limit (Figure 6. 16) for only lag 1 (10-day). For the higher lags the correlation
coefficients are not significant. This can also be seen from the time series plot that, there is a
tendency of consistent over or under estimation for few periods.

-1. 5
-1
-0. 5
0
0. 5
1
1. 5
0 12 24 36 48 60 72 84 96 108 120 132 144 156 168 180
10-day starting January 1986
r
e
s
i
d
u
a
l
s
MWBM-B
DWBM

Figure 6.15 Time series plot of residuals. (Awash-Hombole)

-1
-0. 8
-0. 6
-0. 4
-0. 2
0
0. 2
0. 4
0. 6
0. 8
1
0 3 6 9 12 15 18 21 24
lag (10-day)

C
o
r
r
e
l
a
t
i
o
n
lower Confidence limit
upper confidence limit

Figure 6. 16 Autocorrelation of residuals for DWBM. (Awash-Hombole)


154
6.4 Summary of the results for the test catchments

6.4.1 General performance of the models

Table 6.3 lists the estimates of the parameters and the two measures of performance efficiency
and model factors for MWBM-B when applied to the test catchments at 10-day time step. The
model performed fairly well in reproducing the flows, the model efficiencies vary from 60 %
for the Ruaha-Ihimbu catchment to 84 % for the Awash-Hombole catchment. The Model
quality factors for the model vary between 0.98 for the Ruaha-Ihimbu catchment to 2.19 for
the Tati catchment.

For the DWBM the efficiencies vary from 60% for the Mwambashi catchment to 89 % for
The Awash-Hombole catchment. (Table 6. 4). The model quality factor on the other hand
varies from 1.54 for the Mwambashi catchment to 2.12 for the Tati catchment.

Considering the two model quality measures, especially the efficiency, the two models
explains more than 60 per cent of the variance which is fairly acceptable. Comparison of the
two models shows that the difference of their performance is marginal. Nevertheless the new
DWBM has not only performed as good as the monthly model structure, the conceptual
representation is quite realistic mainly in separating the catchment storage into groundwater
and soil moisture storages.

6.4.1 Reproducing flow regimes

In addition to the general performance of a model measured by the efficiency or model
quality, the versatility of the model in reproducing the flow regimes is checked by comparing
the long term seasonal flows. Figure 6.17 shows good agreement between the observed and
model computed long-term 10 days mean flows.


155

Table 6.3 Table Parameter estimates for the MWBM-B with 10-day time step.

Estimates of Parameters

Catchment
Evaporation
(a
1
)x10
-2

Slow flow
(a
2
)x10
-2

Fast flow
(a
3
) x10
-3

Upper limit
(a
4
) x10
3

Eff.% MQ
1. Ruaha -Mak .196 .169 .209 .811 75 1.00
2. Ruaha - Ih. .330 .170 .101 .414 60 0.98
3. Berga .020 .700 .198 .151 74 2.00
5. Awash-M .800 .630 .103 .572 79 1.89
6. Awash-H .780 .760 .130 .670 84 1.95
14. Tati .100 .000 .400 .697 64 2.19
15. Mwambashi .280 .200 .010 .583 70 1.56
18 Baluba .300 .200 .940 .900 70 1.53

Table 6. 4 Parameter estimates for the DWBM

Estimates of parameters

Catchment
Evaporation
(a
1
)x10
-2

Percolation
(a
2
)x10
-1

Interflow
(a
3
) x10
-2

Upper
limit
(a
4
) x10
3

Recession
a
5

Eff.
%
MQ
1. Ruaha -Mak .317 .300 .080 .612 .070 69 0.98
2. Little -Ih. .900 .200 .025 .792 .011 70 1.04
3. Berga .380 .060 .140 .429 .090 66 2.09
5. Awash-M .240 .540 .080 .679 .198 83 1.90
6. Awash-H .777 .461 .600 .510 .148 89 1.97
14. Tati .940 .200 .020 .306 .029 65 2.12
15. Mwambashi .600 .400 .020 .708 .118 60 1.54
18. Baluba .300 .030 .030 .790 .090 72 1.59


156
0
2
4
6
8
10
12
14
16
18
f
l
o
w

(
m
m
)
observed
MWBM-B
DWBM
Ruha Ihimbu
0
2
4
6
8
10
12
14
f
l
o
w

(
m
m
)
observed
MWBM-B
DWBM
Ruha Mak.

0
5
10
15
20
25
f
l
o
w

(
m
m
)
observed
MWBM-B
DWBM
Mwambashi
0
5
10
15
20
25
f
l
o
w

(
m
m
)
observed
MWBM-B
DWBM
Baluba

0
5
10
15
20
25
30
Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May
f
l
o
w

(
m
m
)
observed
MWBM-B
DWBM
Awash Hombole
0
1
2
3
4
5
6
7
f
l
o
w

(
m
m
)
observed
MWBM-B
DWBM
Tati

0
5
10
15
20
25
30
f
l
o
w

(
m
m
)
observed
MWBM-B
DWBM
Awash Melka.
0
10
20
30
40
50
60
f
l
o
w

(
m
m
)
observed
MWBM-B
DWBM
Ber ga

Figure 6.17 Comparison of long-term 10 day mean observed and computed flow using
MWBM-A and DWBAM.
6.4.2 Water Balance components


157
For all the considered catchments the fast flow component (infiltration excess) does not exist,
hence the total flow is only composed of interflow and base flow (Figure 6. 8. The reason for
this could be that with the 10 day time interval it is likely that the isolated intensive short
duration storms are smoothed, and excess runoff can not be generated. For e.g. the Awash
River, the base flow component is very low. Although isolated rain events occur during the
dry season, it appears that most of the rain is lost by evaporation and only part of it reaches the
river as interflow.

Ruha -Ihimbu
75%
0%
8%
17%
77
17
6
Ruha -Makawila
86%
0%
7%
7%
86
11
3
Berga
77%
0%
3%
20%
76
4
20
Awash -Melka
85%
1%
0%
14%
84
2
14
Awash -Hombole
85%
0%
1%
14%
85
2
13
Actual Evaporation
Fast flow
Inter flow
Tati
0%
0%
91%
9%
92
8
0
Mwambashi
77%
0%
15%
8%
81
14
5
Baluba
73%
0%
20%
7%
22
73
5
Base flow
86
values from
monthly model

Figure 6.18 Proportion of water balance s using DWBM.
For the Mwambashi, Ruaha and Baluba catchments, on the other hand, a substantial amount
of the flow comes as groundwater supply that maintains relatively high base flow. For Tati
catchment the flow is totally interflow. The water balance proportion obtained by the 10-day


158
model is almost the same as for the monthly model with regard to evaporation and flow
components.

6.5 Concluding remarks

The comparison between the two models suggests that the new 10-day model structure is as
good as the previous monthly one. It is also more realistic as it incorporates more concepts in
flow separation as well as flow recessions without increasing the number of parameters to be
optimized. The flows computed using the decade time step are aggregated to obtain monthly
time series. These time series are compared to the observed monthly flow and time series
obtained by using the monthly water balance models (Figure 6.19). The conclusion is that
regarding the monthly flow generation the three models (MWBM with monthly time step ,
MWBM with 10 day-time step and the DWBM) have nearly the same performance. The
model structure described for a monthly time step could therefore be used as the first trial in
modelling for the reason of simplicity. One can also see that none of the models has overcome
the problem of inaccurate estimation of the exceptional peaks.

The present study shows that a relatively small number of parameters are sufficient to
represent the rainfall-runoff relations in10-day time step. Further, it demonstrates that a
routine devised to compute the recession coefficient is useful to incorporate a priori
knowledge of a catchment in hydrological modelling. We believe that the approach of
competition of numerous parameters in minimization the objective function in standard


159
0
20
40
60
80
100
120
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59
m
m
/
m
o
n
t
h
observed
MWBM-B (Monthly)
MWBM-B (Decade)
DWBM

Figure 6.19 Comparison of monthly observed and computed flow using the MWBAM and the
DWBM. (Awash -Hombole)

Chapter 7
7 CASE STUDIES OF
DAILY RAINFALL RUNOFF MODELS

7.1 Introduction.................................................................................................................................160
7.2 Models from literature.................................................................................................................161
7.3 Application of the rainfall-runoff models....................................................................................164
7.4 Comparison of the PDRRM with the 10 days and monthly models............................................173
7.5 Concluding remarks ....................................................................................................................175

7.1 Introduction

The aim of this chapter is to test the parsimonious daily rainfall - runoff model proposed in
Section 3.4. The performance of the model is compared to three classical models, which have
been applied to a number of catchments in the world. The three candidate models are
XINANJIANG (XNJ) (Zhao et al., 1980), SMAR (OConnell et al., 1970) and NAM (Danish
Hydraulic Institute, 1982). The models are selected on the following merits:

1) Data requirement as of the available data, all of the models require time series of areal
rainfall and evaporation and flow data for calibration.
2) The models have equal complexity in conceptualizing rainfall-runoff processes
3) The three models have been applied in dry regions before. (see Gan and Biftu 1996;
Gan et al., 1997)
4) All the model parameters are obtained by automatic optimization procedure.

7. Case studies of daily rainfall runoff models

161
The chapter begins with a brief description of the conceptual representation and parameters of
the three models. The results of the three models and the newly developed model are then
discussed. Finally the summary of comparison of the results obtained using daily and coarse
time steps i.e. 10-day and monthly models are given. Four catchments (Awash - Melka,
Awash-Hombole, Modjo and Berga), from Awash River Basin in Ethiopia are used as a test
catchments. The discussion on the quality of data was given in Chapter 5.

7.2 Models from literature

7.2.1 The XNJ model

The XINANJIANG model (XNJ), developed in 1973, has been used to forecast floods flows
to the Xinanjiang reservoir (Zhao et al., 1980). Two distinct components are identified as for
most of conceptual rainfall runoff models: 1) soil-level water balance and 2) the transfer to the
closure section of catchments (Franchini and Pacciani, 1990). The XNJ model represents the
soil level water balance components by four interconnected storages, the upper zone tension,
the lower tension, deep zone tension and free water storages. Evaporation will take place at
potential rate from upper zone and with reduced rate from lower storages. The model takes
into account the variable area source concept through a probability distribution. The transfer
to the closure is represented by direct overland flow, saturation overland flow, interflow and
groundwater flow. A unit hydrograph is used for surface flow and linear reservoir for the base
flow routing. The model has 15 parameters that are described in Table 7.1.

7.2.2 The Soil Moisture Accounting and Routing (SMAR) model.

SMAR was developed at the University of Galway (OConnell et al., 1970). The model
assumes that the catchment is analogous to a vertical stack of horizontal soil layers which can
contain various amounts of water. Similar to the XNJ, model the evaporation from the top
layer occurs at potential rate. Evaporation from second layer occurs only on exhaustion of the


162
first layer, at the potential rate multiplied by a factor C (whose value is less than unity). On the
exhaustion of the second layer, evaporation from third layer occurs at potential rate multiplied
by C
2
. Thus a constant potential evaporation applied to the basin would reduce the soil
moisture in a roughly exponential manner. Three inter-connected storages are represented:
Surface storage, maximum of 5 layers of soil moisture storage and groundwater storage. The
model simulates four components of flow: direct overland flow, saturation overland flow,
interflow and groundwater flow. The model has 9 parameters that are described in Table 7.2.

Table 7.1 Description of XNJ model parameters (Zhao et al., 1980)
Soil moisture
phase
No. Parameter Description
Direct runoff 1 WM Mean areal soil moisture capacity (mm) 30 - 300
Upper zone 2 X Upper soil moisture storage capacity (mm) 0.0 - 0.5
Lower zone 3 Y Lower soil moisture storage capacity (mm) -
4 KE Evaporation coefficient 0.5 - 1.0
Upper 5 B Tension water distribution index 0.05 - 0.2
zone 6 SM Aerial free water capacity (mm) 0 - 40
7 EX Free water distribution index 0.0 - 2.0
8 CI Fraction of free water to interflow 0.0 - 1.0
9 CG Fraction of free water to ground water 0.0 - 1.0
10 IMP Impermeability coefficient 0.0 - 0.5
Lower 11 C Deep layer evaporation coefficient 0.0 - 0.3
zone 12 KI Interflow recession coefficient 0.5 - 0.9
13 KG Groundwater recession coefficient 0.95 - 1.00
14 N Parameter for linear reservoir 1 - 10
15 NK Number of time steps 1 - 10
Approximate
range


163

Table 7.2 Description of SMAR model parameters (O Connell et al., 1970)
Soil Moisture
Phase
No. Paramet
er
Description Range
Upper 1 C Evaporation 0.0 - 1.0
Zone 2 Z Total soil moisture capacity 50-500
3 Y Soil infiltration capacity 20-200
Direct 4 H Direct runoff area index 0.0 1.0
Runoff 5 T Potential evaporation factor 0.0 - 0.0
Groundwater 6 G Groundwater runoff coefficient 0.0-1.0
7 N Parameter for linear reservoir 0.5-5.0
8 NK Number of time step [Parameter nk] 0.5-40
9 KG Number of time step groundwater K 5-200

7.2.3 The NAM model

The "Nedbor-Afstromnings Model" NAM (Danish Hydraulic Institute, 1982) is a
deterministic, lumped conceptual rainfall-runoff model developed at the Technical University
of Denmark. The model is a soil moisture accounting one, utilizing five different, mutually
interrelated, moisture storages. The effects of snow are controlled by the temperature
conditions. Surface storage is composed of interception and depression storages. The root
zone is represented by the lower storage from which vegetation draws water for transpiration.

Rain and snow are subject to surface storage, evaporation and interflow. When maximum
surface storage is reached, some of the excess water enters the stream as overland flow and
the remainder infiltrates to the lower zone and groundwater. The groundwater recharge is
divided into two storages, upper and lower with different time constants. The groundwater
storage acts as a linear reservoir draining continuously to the stream as base flow. Overland
flow and interflow are routed through one linear reservoir before all the streamflow
components are added and routed through a final linear reservoir. The model has 15
parameters described in Table 7.3. For the present study the snow component is not relevant,
hence the parameter corresponding to this process is set to zero.


164
Table 7.3 Parameters of the NAM model (Ref. Danish Hydraulic Institute, 1982)

Soil moisture
phase
No. Parameter Description Range
Direct 1 UMAX Upper zone storage capacity (mm) 1 - 100
runoff 2 LMAX Lower zone storage capacity (mm) 20 - 500
3 CMELT
Snow melt coefficient (mm/
o
C/day)
-
Upper 4 CQOF Overland flow runoff coefficient 0.01 - 1.0
zone 5 CLOF Overland flow threshold coefficient 0.0 - 0.8
6 CQIF Interflow drainage coefficient 1E-5 - 0.05
7 CLIF Interflow threshold coefficient 0.0 - 0.95
Lower zone 8 CLG Recharge threshold coefficient 0.0 - 0.95
Ground- 9 CBFL Groundwater storage recharge 0.0 - 0.95
water 10 CK1 Linear reservoir (1) routing constant 6 - 120
11 CK2 Linear reservoir (2) routing constant 6 - 120
12 CKBFU Upper groundwater storage baseflow
routing constant
100 - 4000
13 CKBFL Lower groundwater storage baseflow
routing constant
200 - 9000
14 CAREA Permeability constant 1.0
15 CEVP Conversion factor from pan
evaporation data to potential
0.5 - 1.0

7.3 Application of the rainfall-runoff models
7.3.1 Calibration of the models

With exception of limiting the memory of the unit hydrographs for the flow routing
component, all the parameters are obtained by optimization. The five to six years of available
data are divided in to two parts to serve as calibration and verification periods as shown in
Table 7.4.

Table 7.4 Calibration periods
Catchment Calibration
Period 1
verification
Period 2
Berga 1986-1987 1988-1990
Modjo 1985-1987 1988-1989
Awash - Melka 1985-1988 1989-1991
Awash -Hombole 1985-1987 1988-1990

As described in the previous sections the number of model parameters are large for the three
models. The optimum values are obtained by automatic optimization of the sum of squares of


165
errors of square root transformed daily observed and model calculated flows. The optimum
parameters for the four models are given in Table 7.5 through 7.8.

Table 7.5 Model parameters: The XNJ Model
Catchments
No. Parameter Berga

Modjo Awash M.

Awash H.

1 WM 99.458 407.8 461.0 462.6
2 X 0.125 0.237 0.080 0.128
3 Y 0.331 0.930 0.904 0.640
4 KE 0.599 0.977 0.926 0.925
5 B 0.500 1.804 0.877 1.979
6 SM 20.438 2.229 0.000 7.917
7 EX 1.906 0.131 1.775 1.970
8 CI 0.080 0.246 0.909 0.045
9 CG 0.577 0.020 0.091 0.001
10 IMP 0.173 0.046 0.106 0.132
11 C 0.187 0.114 0.085 0.280
12 KI 0.900 0.509 0.506 0.500
13 KG 0.917 0.700 0.510 0.501
14 N 1.000 1.202 1.236 1.001
15 NK 7.775 6.094 9.137 6.294

Table 7. 6 The SMAR Model parameters:
Catchments
No. Parameter Berga

Modjo

Awash M.

Awash H.

1 C 0.512 0.934 0.917 0.999
2 Z 274.9 398.1 367.4 380.0
3 Y 167.221 105.4 135.1 101.0
4 H 0.551 0.216 0.587 0.629
5 T 0.695 0.714 0.850 0.638
6 G 0.441 0.609 0.812 0.006
7 N 0.502 0.775 0.718 0.702
8 NK 11.302 39.149 23.58 39.61
9 KG 31.707 199.99 156.11 141.73


166

Table 7. 7 Model parameters: The NAM Model
Catchments No. Parameter
Berga

Modjo Awash
M.
Awash H.

1 UMAX 3.120 55.2 30.8 8.947
2 LMAX 281.5 587.3 599.997 446.9
3 CMELT 2.000 2.000 2.000 2.000
4 CQOF 0.503 0.321 0.564 0.860
5 CLOF 0.391 0.000 0.000 0.015
6 CQIF 0.036 0.049 0.036 0.036
7 CLIF 0.432 0.730 0.797 0.803
8 CLG 0.000 0.326 0.137 0.108
9 CBFL 0.266 0.950 0.493 0.582
10 CK1 72.99 95.97 110.72 84.99
11 CK2 10.45 6.24 101.61 85.83
12 CKBFU 3645 100 407.3 147.7
13 CKBEL 5688.4 8999 8988. 8993
14 CAREA 1.0 1.0 1.0 1.0
0.874 0.816 1.000 15 CEVP 0.732

Table 7.8 Model parameters: The PDRRM
Catchments
No. Parameter
Berga Modjo Awash M. Awash H.
a
1
Evaporation 0.023 0.045 0.023 0.028
0.065 0.0042 0.037
3
Interflow 0.0011 0.0005 0.0012 0.0011
a
4
Upperlimit 402.2 896 764. 709.0
a
5
recession 0.069 0.012 0.02 0.010
a
6
k-fast flow UH 1.00 1.01 1.01 1.01
a
7
k-base flow UH 1.20 1.5 2.69 2.05
Max. correlation 0.86 (a
1,
a
2
) -0.82(a
2,
a
3
) -0.67 (a
2,
a
3
) -0.62(a
2,
a
3
)
a
2
Percolation 0.032
a

7.3.2 Comparison of performance of the models

To compare the performance of the models, the Root mean Squares Errors (RMSE) and Bias
of residuals are used in addition to the Nash-Sutcliff model efficiency. The two former indices
are widely used in comparison of daily rainfall runoff models. The RMSE is defined as:

n
q d
RMSE
n
t
t t
=
1
2
) (
(7.1)


167

and the Bias of residuals is calculated as:
100
) (
(%)
1
1
=
=
=
N
t
t
N
t
t t
d
q d
Bias (7.2)

The model parameters listed in Section 7.3.1 are used to estimate the flows during the
calibration and verification periods. Table 7.9 summarises the statistical indices of the four
models applied to four catchments. From Table 7.9 it is clearly seen that except for Modjo
catchment, all the models resulted with small biases ranging from 2%-7% for the calibration
period. Unfortunately this performance is not maintained when the model is used for the
period that has not be used for calibration. This can be seen from the scatter plot of observed
and estimated flow for the Awash- Hombole (Figure 7. 1).
where: q
t
= observed daily flow (mm)
d
t
= model computed flow (mm)
N = total number of data points used for calibration or verification.

Since the root mean square is function of the magnitude of the observed flows, only results
obtained by different models when applied to the same catchment and for the same period can
be compared. The requirement of a good model is to obtain a small root mean square error as
much as possible. The bias and efficiency indices allow comparison of model performance
when applied to various ranges of catchments. Positive values of the bias indicate over
estimation, whereas negative values indicate under estimation of flows.


168
Table 7.9 Comparison of statistics of the studied models

Calibration Verification
Catchment
Bias % RMSE (mm) Ef. % Ef .% Bias % RMSE (mm)

Berga -4 0.99 68 1.47 -39 25
Modjo -3 1.25 17 38 2.09 11
Awash M. -2 0.35 80 5.9 0.58 51
Awash H -2 0.33 82 5.8 0.49 77

0.91 69 -9 1.26 20
0.85 5 -39 1.58 5
Awash M. 7 0.31 83 17 0.49 59
Awash H -1.0 0.23 82 12 0.57 72
NAM
Berga -6 0.93 71 -20 1.55 16
Modjo 32 0.23 54 33 0.56 17
Awash M. 2 0.34 81 22 0.61 46
Awash H 2 0.31 76 84 -7 0.5

Berga -3 1.02 56 6 1.19 50
Modjo -17 1.68 21 25 12
Awash M. -6 0.47 74 12 0.53 58
Awash H -5 0.41 78 7.5 0.54 75
XNJ
SMAR
Berga 3
Modjo 28.
PDRRM

The scatter plots are fairly evenly distributed along the 45-degree line, which shows the fit of
the observed and estimated flows for the calibration period. However for the verification
period, the scatter plots are widely spread which shows poor representation of the hydrograph.

The above observation is confirmed by the reduction of the RMSE and the Efficiency of each
model from the calibration to verification periods. According to the three indices and to the
graphical comparisons of the flow hydrographs, the three models (XNJ, SMAR and NAM)
have well reproduced the flows for the calibration periods for the Awash River at Hombole
and Melka Kuntre. Berga River is also not badly modelled. The result of modelling the Modjo
River shows an extremely low performance for all the models.


169
N
A
M

X
N
J

S
M
A
R

P
D
R
R
M

CALIBRATION VERIFICATION
0
100
200
300
400
500
600
0 100 200 300 400 500 600
Observed flow, m
3
/ S
C
o
m
p
u
t
e
d

f
l
o
w
,

m
3
/
s
0
100
200
300
400
500
600
0 100 200 300 400 500 600
Observed flow, m
3
/ S
C
o
m
p
u
t
e
d

f
l
o
w

,

m
3
/
s
)
0
100
200
300
400
500
600
0 100 200 300 400 500 600
Observed flow (cms)
C
o
m
p
u
t
e
d

f
l
o
w
,

m
3
/
s
0
100
200
300
400
500
600
0 100 200 300 400 500 600
Observed flow (cms)
C
o
m
p
u
t
e
d

f
l
o
w
,
m
3
/
s
0
100
200
300
400
500
600
0 100 200 300 400 500 600
Observed flow, m
3
/ s
C
o
m
p
u
t
e
d

f
l
o
w
m
,

m
3
/
s
0
100
200
300
400
500
600
0 100 200 300 400 500 600
Observed flow, m
3
/ s
C
o
m
p
u
t
e
d

f
l
o
w
,

m
3
/
s
0
100
200
300
400
500
600
0 100 200 300 400 500 600
Observed flow, m
3
/ s
C
o
m
p
u
t
e
d

f
l
o
w
,

m
3
/
s
0
100
200
300
400
500
600
0 100 200 300 400 500 600
Observed flow, m
3
/ s
C
o
m
p
u
t
e
d

f
l
o
w
,

m
3
/
s

Figure 7. 1 A comparison of scatter plots of the observed and computed flow for the four
models (Awash River at Hombole).


170
The possible explanation for this catchment is the poor quality of data. It has been mentioned
in section 5.2 that the flow of Modjo River is questionable due to an inadequate rating curve.

Considering the statistic indices and the scatter plot of the residuals, and excluding the Modjo
catchment, the PDRRM shows a lesser efficiency for the calibration periods. Nevertheless the
difference is not that significant. For example the efficiency of XNJ model for the Awash-
Hombole is 82% while for PDRRM the efficiency is 78%. It can also be seen that the
PDRRM model has maintained the performance of the model for verification period, which is
not always the case for the other models.

For the purpose of investigating the efficiency of the model in reproducing various flow
regimes a typical hydrograph of Awash-Hombole for the year 1986 is compared with the
flows estimated by the four models (Figure 7.2). From this comparison, we observe that all the
models have attempted to follow the trend of the observed hydrograph. For this particular
hydrograph it appears that the rising limb of the hydrograph and the high flows are well
modelled. There is an over estimation of the low flows, because of some local peaks that
resulted from short storms during the dry season. The recession limb of the hydrograph is also
where one observes discrepancies. From calibration point of view and looking into the
statistics and the comparison of hydrographs the XNJ model is found to be slightly better than
the other models considered.

The PDRRM, in general was able to reproduce well the hydrographs in low and high flow
regimes. It should be mentioned that the model under estimates the peaks for almost all the
catchments and there is a trend of smoothening erratic daily flows. Figure 7.3 and Figure 7. 4
show a comparison of the estimated flow hydrographs for the Awash-Melka and Berga
catchments respectively. Like for the other three models, the Modjo catchment was not
modelled with acceptable efficiency. It should also be remembered that the model structure
used are first tested on a rather larger time steps, further application to other catchments and
verification are necessary.


171
0
50
100
150
200
250
1 31 61 91 121 151 181 211 241 271 301 331 361
day ( starting 1 January 1986)
f
l
o
w

m
3
/
s
observed XNJ

0
50
100
150
200
250
1 31 61 91 121 151 181 211 241 271 301 331 361
f
l
o
w

m
3
/
s
observed SMAR

0
50
100
150
200
250
1 31 61 91 121 151 181 211 241 271 301 331 361
f
l
o
w

m
3
/
s
observed NAM

0
50
100
150
200
250
1 31 61 91 121 151 181 211 241 271 301 331 361
f
l
o
w

m
3
/
s
observed PDRRM

Figure 7.2 Comparison of observed and computed hydrographs for the Awash-Hombole
catchment. (Typical year 1986.)


172

0
50
100
150
200
250
300
350
400
450
1 31 61 91 121 151 181 211 241 271 301 331 361
day starting 1 January 1988
f
l
o
w

m
3
/
s
Observed XNJ PDRRM

Figure 7.3 Comparison of observed and estimated flows for Awash-Melka catchment

0
5
10
15
20
25
1 31 61 91 121 151 181 211 241 271 301 331 361
day starting 1 January 1987
f
l
o
w

m
3
/
s
observed XNJ PDRRM

Figure 7. 4 Comparison of observed and estimated flows Berga catchment


173
7.4 Comparison of the PDRRM with the 10 days and monthly models

In chapter 5 and chapter 6, we have reported that simple models are capable of modelling
rainfall-runoff relationship at monthly and 10-day time steps. The question that may be asked
is that whether shorter time step modelling is beneficial in applications, which require 10-day
or monthly flows? An example of such applications is flow extension for reservoir design.
The argument is not to undermine the necessity of the short time step modelling, because in
applications such as flood forecasting one can not surpass the short time steps models. The
problem is mainly that the short time step requires detailed data and the quality of the data is
also very important to calibrate these models. For 10-day and monthly time step it is
commonly believed that the individual daily errors tend to compensate one another unless
there are systematic errors. Moreover some inputs such as potential evaporation are readilly
available on monthly basis.

Following the above argument, a catchment is modeled using the three time steps. The daily
flows estimated by the PDRRM is aggregated to give the 10-day flow to compare it with
flows that are estimated by the DWBM. Similarly the 10-day flows obtained by the DWBM
and those aggregated from daily model are again grouped on monthly basis to compare them
with the estimate of monthly water balance model.

These analyses show that the daily model is only slightly better than the 10-day time step
modelling in reproducing the long-term seasonal flows. Both models underestimate the high
flow regime and it is observed that the water balance is compensated by a slight over
estimation of the recession limb (Figure 7.5). Comparing the mean monthly flows obtained by
the PDRRM, the DWBM and MWBM, surprisingly, the MWBM shows the best fit to the
mean seasonal flows (Figure 7. 6). The model efficiencies on monthly flow representation for
the catchment considered is (90%, 89%, 90%) for MWBM, DWBM and PDRRM
respectively.


174
The implication of this finding is that, given the problem at hand one should start with the
simple and larger time step model. The shorter time step models are recommended for those
applications, which requires shorter time step and also in cases where reliable and good
quality data are available.
0
5
10
15
20
25
30
f
l
o
w

(
m
m
)
observed
PDRRM
DWBM
Awash Hombole

Figure 7.5 Comparison of the daily rainfall-runoff and the 10-day water balance models
(Awash-Hombole)
0
10
20
30
40
50
60
70
80
m
m
/
m
o
n
t
h
Observed
PDRRM
DWBM
MWBM-B
Awash-

Figure 7. 6 Comparison of the daily rainfall-runoff, the 10-day and monthly water balance
models (Awash-Hombole)


175
7.5 Concluding remarks

From the results of calibration of the models the following conclusions can be made:

1) Among the four considered models (XNJ, SMAR, NAM and PDRRM) tested on four
semi -arid catchments, the XNJ model performs the best for the calibration period
considered.

2) For all the models, it was observed that the efficiency reduces when the models are
applied to a verification period. During the verification period, all the models performed
equally well (or bad).
3) The results of the PDRRM show that, though the quality of the model is somewhat less
than for the other models tested (with more parameters), the general reproduction of the
hydrograph is promising. It is the author's opinion that with few refinements of the
structure of the model, the performance of the model can be improved.

4) Regarding the time step of modelling, it is found that, if the application requires seasonal
flows, the first model to try is the monthly water balance model.

Chapter 8
8 PRACTICAL APPLICATION OF
THE DEVELOPED MODEL

8.1 Introduction.................................................................................................................................1 76
79
86

8.2 Influences of the variability of rainfall on flow regimes .............................................................1
8.3 Low flow - duration - return period curves .................................................................................1
8.4 Reservoir capacity design............................................................................................................188

8.1 Introduction

In order to make decisions for planning, design and control of water resource systems long
runoff series are required. The latter are not often obtained with reasonable length. However,
rainfall observations are relatively long. In this chapter, the monthly water balance model is
used to extend flow data of a catchment to study long-term influences of variability of rainfall
on runoff, development of low flow duration curves and preliminary reservoir capacity design.
For the reason of availability of data, the Awash River basin is used. (See section 5.2 for
description of the catchment). The present study focuses only on the water head area of 7560
km
2
upstream of Hombole. In this sub catchment, a minimum human interference such as
dams and diversions on the river flow regime is anticipated. Five rainfall stations with
concurrent data length (1980 - 1995) were used to calculate the areal rainfall (Figure 8.1).
The station at Addis Ababa has a longer record of rainfall (1900-1995). To assess whether the
use of the longer data series of Addis Ababa station is justifiable, the correlation of the
calculated areal rainfall and main station is analyzed. Areal monthly rainfall computed using
the mean of the five stations and rainfall series at Addis Ababa are well correlated (Figure
8. Practical applications of the developed model

177
8.2). Therefore the rainfall station of Addis Ababa is considered as an index station for further
investigation of the temporal variability of the rainfall over the catchment. The linear equation
obtained is with a coefficient of determination R
AddisAbaba area
R R 794 . 0 =
2
=0.9. This equation
is used to extrapolate the areal rainfall from the observed data of Addis Ababa.

Rainfall Stations
Holeta
Sendafa
Debrezeit
Addis Alem
Addis Ababa
Areal Rainfall-observed
-Extrapolated
Flow
Awash at Hombole
Period 1900 1910 1960 1970 1980 1990
Avialble data Extrapolated Modeled

Figure 8.1 Schematic representation of extension of flow.

A monthly water balance model developed in section 3.2 is applied to reconstruct the long-
term record of flows from the observed rainfall series. The model requires monthly rainfall,
potential evaporation and river flow as inputs. The areal rainfall data obtained from the Addis
Ababa rainfall is used. Mean monthly evaporation data at Addis Ababa for the years (1988
and 1987) were available. Due to the fact that evaporation in such region is mainly controlled
by the available water, the long-term changes of the potential evaporation over the whole
period of simulation may not greatly influence the rainfall run-off processes. Hence, the
constant seasonal variation of evaporation is used throughout the simulation period.


178
0
50
100
150
200
250
300
350
400
450
0 50 100 150 200 250 300 350 400 450
Monthly Rainfall at Addis Ababa (mm)
M
o
n
t
h
l
y

A
r
e
a
l

R
a
i
n
f
a
l
l

(
m
m
)
Data points
Regression line
Upper confidence limit
Lower confidence limit
R
Area
=0.794R
Addis
R
2
=0.90

Figure 8.2 Relationship between areal and Addis Ababa monthly rainfall (R
2
is coefficient of
determination).

Monthly river flow data of the Awash River at the Hombole gauging site for the years 1963-
1992 are used to calibrate the model. The optimum parameters obtained previously in chapter
5 (see section 5.3.7) are used for reconstructing the flow series corresponding to the historical
rainfall series. It is further assumed that the catchment characteristics have not altered greatly.
Therefore, the parameters are assumed to be time invariant and, consequently, the monthly
flows series of the years (1900-1992) is simulated for the analysis in the following sections.
Figure 8.3 shows the comparison of the annual historical rainfall and reconstructed flows. One
can see good agreement for the segment of the time series where both calculated and observed
flows are available (1963-1992).


179
0
500
1000
1500
2000
2500
1900 1910 1920 1930 1940 1950 1960 1970 1980 1990
Year
R
a
i
n
f
a
l
l

(
m
m
)
0
100
200
300
400
500
600
F
l
o
w

(
m
m
)
Rainfall
Modeled flow
Observed flow

Figure 8.3 Time series comparison of long term rainfall series and modeled annual flows

8.2 Influences of the variability of rainfall on flow regimes
8.2.1 Introduction

Rainfall variability due to climatic change has remained a concern to hydrological science.
Globally one may see two types of rainfall variability: the continuous decline of the amount of
annual rainfall (e.g. West Africa since 1960s) and recursive anomalies of annual rainfall (e.g.
East and North East of Africa). From hydrological point of view each variability has a
different magnitude of influence on the water resource system of a region or a catchment. In
this study we investigate the link between the variability of annual runoff and the
corresponding annual rainfall and flow regimes by using a water head catchment of the Upper
Awash River Basin located in central Ethiopia.

From the annual rainfall record of this century, researchers have shown that there is a shift of
the mean annual rainfall in Northeastern Africa since the last half of this century. Seleshi


180
(1996) reported that there is a decline of mean summer rainfall over this region since the mid
of 1960s, which has also been observed for the Sahel regions. Comparing the data of the Blue
Nile catchment for two 20 years windows (1946-1965) and (1965-1984) Conway & Hulme
(1993) estimated that there is 18 percent of reduction of the mean annual flow corresponding
to 8 percent reduction of the annual mean rainfall in the second 20 year period. From analysis
of 5 and 10 years moving average of the flow of the Blue Nile at Dongola, Misganaw (1989)
asserted that there is a general trend of decline of the annual flows after mid 1960s.

Due to the high local variability of the rainfall and to the limited number of observations over
the large area of a basin like the Blue Nile, it is always difficult to precisely determine the
degree of influence of the anomaly of rainfall on the flow regimes. In the present study, a
medium size catchment is selected from the central part of Ethiopia with fairly distributed
rainfall stations. Because of the limited years of record of runoff (30 years), extension of flow
data is achieved through rainfall-runoff modelling of the catchment. It is evident that the
decline of the annual rainfall amounts is echoed to the annual flows. However it is important
to know the extent of the influences and to study the variability of flow regimes caused by the
temporal variability of rainfall on seasonal basis. Hence monthly rainfall and flow data are
analyzed to obtain a clearer link between the two variables.

8.2.2 Annual rainfall variability

From the rainfall record at Addis Ababa station, it can be stated (with the exception of the
decade of 1930s) that the first half of the twentieth-century has experienced higher rainfall
amounts than the second half. The lowest rainfall on the decade basis was observed during the
1950s. The lowest annual rainfall (904 mm) was observed in 1962. From Table 8.1, one can
see that there is a persistence of the lower annual mean values during the years after 1960. The
linear and polynomial trends fitted to the rainfall and runoff series resulted with a statistically
insignificant coefficient of determination (R=0.023 and 0.004) respectively. However the 5
and 10 years of moving average curves confirms the general decline of the amount of flow
and rainfall (Figure 8.4 and Figure 8.5).


181
Table 8.1 Rainfall Variability during the last 9 decades based on Addis Ababa observations.
(+) and (-) indicate values above and below long term mean respectively.

Period Annual
Rainfall
(mm)
Summer
Rainfall
(mm)
Period Annual
Rainfall
(mm)
Summer
Rainfall
(mm)
1901-1910 1213 (+) 886 (+) 1951-1960 1112 (-) 787 (-)
1911-1920 1327 (+) 981 (+) 1961-1970 1176 (-) 807 (-)
1921-1930 1317 (+) 951 (+) 1971-1980 1166 (-) 854 (-)
1931-1940 1113 (-) 803 (-) 1981-1990 1198 (-) 800 (-)
1941-1950 1231 (+) 933 (+)
Long term Mean (1900-1995) Annual = 1203 mm Summer= 864mm

Linear Trend
Rainfall = -1.114t + 1257.9
R
2
= 0.0235
500
700
900
1100
1300
1500
1700
1900
1900 1910 1920 1930 1940 1950 1960 1970 1980 1990
Year
R
a
i
n
f
a
l
l

(
m
m
)
Data Points
5-Years MA
10-Years MA
Linear trend

Figure 8.4 Annual variability of rainfall at Addis Ababa station.


182
Flow= -0.3129t + 156.95
R
2
= 0.004
50
100
150
200
250
300
350
1963 1968 1973 1978 1983 1988
Year
F
l
o
w

(
m
m
)
Data Points
Linear Trend
5 Years MA
10 Years MA

Figure 8.5 Annual variability of Awash River flow at Hombole

8.2.3 Influence of rainfall timing on flow regimes

The correlation coefficient between the annual areal rainfall and the reconstructed flow is 0.6,
which indicates that mere annual magnitude anomalies of rainfall can not fully explain the
effect on the total generated flow. Hence it is useful to dissociate the rainfalls in to seasons, to
study the influence of the timing and length of the rainy season on the flow regimes. It can be
argued that in regions with long dry periods, the first arrival of rain on bare soil will generate
high flows and leaves little or no water for groundwater supply. Consequently, smaller or no
base flow will occur during the non-rainy season. A more favorable condition for higher flow
is when there is intense rain during the last days of the rainy season when the upper soil is
saturated.

From the cross correlations of monthly rainfall and summer runoff (Table 8.2), it is observed
that the rainfall of June is more important than the July rain. The explanation is that the first
arrival of rain determines the losses. The most significant cross correlation of monthly rainfall
with the summer flow is obtained for the August rainfall. The interpretation is that in August,
there is a higher opportunity of flow generation (even for small storms) since the catchment is
already moist and the losses are minimal. The total summer rainfall is also well correlated


183
with the summer flow (correlation coefficient r =0.73). On the other hand, the individual
monthly rainfall and also the total summer (June-September) rainfall less influence the low
flow regime (November-February), due to fast response of the catchment.

Table 8.2 Cross correlation coefficients of rainfall and flow regimes: Underlined values show
significant correlation at 95 % confidence limit = 0.2)

Rainfall Flow regimes
June July Augus
t
September October Summer
Summer flow
(June-
September)
0.43 0.36 0.63 0.46 -0.05 0.73
Low flow
(Nov.
February)
-0.01 0.14 0.04 0.16 0.21 0.11

8.2.4 Influence of long -term rainfall variability on flow regimes

The reconstructed annual mean flows of the second half of the century are observed to be
below the long-term mean which is consistent with the rainfall variability. It is also noted that
there is an amplification of the variability of the flow, especially for the summer flows. For
example in the 1950s there was a departure of the flows of 25 % from the long term mean
(163 mm), corresponding to only 10 % departure of rainfall from the long term mean
(1200mm) (Figure 8.6 and Figure 8.7). On the other hand, during the same decade the base
flow component has shown only a departure of 10 % from long term mean of (25 mm) and the
annual low flow has only a departure of 7 % (Figure 8. 8). Moreover, the base flow and
annual low flow (November-February) does not show high variability during the whole period
of analysis.


184
-10
-5
0
5
10
15
D
e
p
a
r
t
u
r
e

f
r
o
m

l
o
n
g
-
t
e
r
m

m
e
a
n

(
%
)
1901-10 1911-20 1921-30 1931-40 1941-50 1951-60 1961-70 1971-80 1980-90
Decade
Mean Annual rainfall
Mean Summer rainfall

Figure 8.6 Departure of annual and summer rainfall from long term means on decade basis.

-30
-20
-10
0
10
20
30
40
D
e
p
a
r
t
u
r
e

f
r
o
m

l
o
n
g
-
t
e
r
m

m
e
a
n

(
%
)
1901-10 1911-20 1921-30 1931-40 1941-50 1951-60 1961-70 1971-80 1980-90
Decade
Mean Annual flow
Mean Summerflow

Figure 8.7 Departure of annual and summer flow from long-term means on decade basis.


185
-15
-10
-5
0
5
10
15
D
e
p
a
r
t
u
r
e

f
r
o
m

l
o
n
g
-
t
e
r
m

m
e
a
n

(
%
)
1901-10 1911-20 1921-30 1931-40 1941-50 1951-60 1961-70 1971-80 1980-90
Decade
Mean Annual base flowl
Mean low flow

Figure 8. 8 Departure of annual base flow and low flow from long term means on decade
basis.

8.2.5 Conclusions

The study has also shown that the runoff variability is amplified compared to the rainfall
variability. It is also demonstrated that the major flow regime (summer flow) is affected
greatly by the low anomalies of rainfall and that the low flow, which is normally small, does
not show large variation in this century. The analysis on the timing of rainfall showed that
June rainfall is significantly important for the total summer flow to the river, though August
rainfall is the most significant cause of high summer flows. It should be noted that this study
is limited on the causality of rainfall on the flow regimes. Further influence of land use and
vegetation cover changes on the flow regimes should be studied with more detail e.g. by
means of distributed hydrological models.
From simple correlation analysis of the available data in the catchment, it is shown that the
Addis Ababa rainfall station can be used as an index station to study the long-term temporal
variability of rainfall over that upper Awash catchment. It has been observed that the mean
annual and summer rainfall amounts have declined since the second half of this century and
that this low mean is maintained up to the beginning of the 1990s. However, the linear trend
of the decline is found to be not significant.


186

8.3 Low flow - duration - return period curves

Low flow frequency analyses use data that are independent and homogeneous and therefore
they can be used to determine the probability of occurrence of a flow event of specific
magnitude. To obtain an independent series the annual minimum series are used for the
analysis.
The procedure to obtain the curves is outlined below

1) From n years of monthly streamflow series, minimum flows with a given duration in each
year are obtained and ranked with the lowest flow being ranked, r = 1.
2) The return period of a flow lower or equal to the lowest minimum flow is then n years
and for minimum flow at rank r, it is equal to n/r years.
3) Flows and corresponding return periods then are plotted..
4) Similar curves can be plotted for other duration.

The low flow frequency analysis is performed on the observed flow series (1963-1992) and
the extended series (1900-1995) for Awash River at Hombole. The extension of the flow
series has allowed computation of higher return periods, which was not possible with the short
observed time series. From the analysis it can be concluded that the Awash River at Hombole
can cater a limited amount of supply without any regulation of the river. The river can supply
at least about 3 and 2 million m
3
per month, with a failure frequency once in 10 and 100 years
respectively (Figure 8.10 and Figure 8.9).


187
0. 1
1
10
100
1000
1 10
return period (years)
f
l
o
w

x
1
0
6

m
3
100
9 months
6
3
2
1

Figure 8.9 Low flow duration -return period relationships for Awash River at Hombole (1963-
1992)

0. 1
1
10
100
1000
1 10
return period (years)
f
l
o
w

x
1
0
6

m
3
100
9 months
6
3
2
1

Figure 8.10 Low flow duration -return period relationships for Awash River at Hombole.
(1900-1995)


188
8.4 Reservoir capacity design

McMahon and Mein (1978), give a comprehensive review and classification of most of the
currently used reservoir capacity estimation procedures. A critical period defined as a period
during which a reservoir goes from a full condition to an empty condition is used in this
analysis.

Let S
t
= the storage at the beginning of month t,
K= the capacity of the reservoir,
D
t
=the monthly demand and
Q
t
= monthly flow,
where all the volumes are in the same units, for example million m
3
. Then the storage at the
beginning of month t+1 is given as:

)
D
+ Q -
S
- K (0, + ,0)
D
- Q +
S
( =
S t
t
t t
t
t 1 + t
min max (8.1)

The inflows are monthly river flows obtained from the rainfall of 93 years using the runoff
model. In the above equation losses from the reservoir (such as evaporation) are assumed to
be incorporated within the demand. Using the above recursive equation the evolution of the
content of the reservoir can be computed for the flow series.

There is a shortage during month t whenever S
t+1
= 0. Let s be the number of years amongst
the total number of years N during which there is at least one month with shortage, then the
estimate of the corresponding return period is given as:
s
N
= R (8.2)
and the inverse F= 1/R is the probability of a shortage. The upper and lower boundaries of the
95% confidence interval are approximately (Vandewiele et al., 1993)


189
N
) (1.96
+ 1
N
4
) (1.96
+
N
F) - F(1
1.96
2N
) (1.96
+ F
2
2
2 2
(8.3)
1
10
100
20 60 100 140 180
reservoir capacity x 10
6
m
3
r
e
t
u
r
n

p
e
r
i
o
d

(
y
e
a
r
s
)
Deamand 20 Mm
3
/month
Deamand 30 Mm
3
/month
95 % Confidence limit

Figure 8.11 Return period versus reservoir capacity using the extended flow series for given
demands. (Awash River at Hombole).

Two constant demands, (20 and 30 millions of m
3
/month) are considered. Figure 8.11 shows
some examples of the reservoir capacities required for the two demands considered, with
corresponding return period of shortages for Awash River at Hombole. At this stage it may be
noted that, in these figures, the axis of the reservoir capacity can easily be transformed into
height of the dam required at a selected dam site which in turn is transformed to the
investment required. Hence this relationship is really a very important tool in making
decisions at the planning stage both in technical and economical aspects. Moreda, et al.,
(1998) give similar relationships using long runoff records generated by stochastic pure runoff
models.

Chapter 9

9 SUMMARY
CONCLUSIONS AND RECOMMENDATIONS

9.1 Summary

The main aim of this research was to contribute to the conceptual rainfall-runoff modelling in
regions with limited data. Three time steps (monthly, 10-dayly and daily) rainfall-runoff
models were developed systematically and applied to different catchments. The models were
coupled with two different optimization algorithms Shuffled Complex Evolution and VOA5A
sequentially to efficiently obtain the model parameters and their relevant statistics. The
conceptual monthly water balance models originally developed and tested for humid regions
were applied to arid and semiarid cases. Modification of the structure of the model was
necessary apropos to the temporal variability of the hydrological processes in the latter
regions. More over it was shown that a particular representation of the evaporation term is
justified and some of the discrete parameters were narrowed to a definite range of values,
indicating the possible regionalization of the forms of the general equations governing the
models. Based on the experiences on the monthly model, a 10-day water balance models was
developed and tested. The concept was extended to develop a daily time step model.
9. Summary, conclusions and recommendations

191
9.2 Conclusions

The following conclusions may be specifically drawn from the development and application
of the models:

9.2.1 Monthly water balance models

For semi-arid and arid catchments considered, a warming up period of one to three year is
sufficient to avoid the influence of initial soil moisture assumption.
The model structure whereby the fast flow is expressed as a function of the average soil
moisture index resulted with good agreement of observed and calculated flow for most of
the catchment studied. Introducing an additional parameter such as an upper limit of the
soil moisture index marginally improved the modelling in a few catchments only.
Evaporation Equation IR=2 is best suited for the catchments.

IR=2 ] ), 1 ( min[
1
t
a
t
e
t t
e e w r
=
1
0 a <1

where r
t
= actual evaporation e
t
= Potential evaporation
w
t
= available water a
1
= evaporation parameter
The discrete parameters for slow and fast flow equations can be set to b
1
=2 and b
2
=2.

fast flow equation:
t
b
t t
n m a f
1
3
=
Slow flow equation:
2
2
) (
b
t t
m a s =

where f
t
= f ast flow
s
t
= slow flow

t
m = soil moisture storage
a
1,
a
2,
= slow and fast flow parameters


192
No problems of local minima of the objective function were observed during the
calibration of the parameters.
The models well represent the rainfall runoff relationship for low flow and medium flow
but may over or under estimate the exceptional peaks.
The model parameters are not biased by the calibration period.
The seasonal mean flow statistics are well preserved.

The result of the modelling of the study catchments is encouraging for practical applications
such as extending and filling missing flow data for the periods where rainfall and evaporation
are available, to obtain long-term flow data for water resource planning and management.

9.2.2 10 days water balance models

The present study shows that a relatively small number of parameters are sufficient to
represent the rainfall-runoff relations in 10-day time steps. Further, it demonstrates that a
routine devised to compute the recession coefficient is useful to incorporate a priori
knowledge of a catchment in hydrological modelling. We believe that the approach of
competition of numerous parameters in minimization the objective function in standard

The comparison of performance the monthly water balance model (MWBM) with 10-day time
step and the DWBM (10-day water balance models) suggests that the new 10-day model
structure is as good as the previous monthly one. It is also more realistic as it incorporates
more concepts in flow separation as well as flow recessions without increasing the number of
parameters to be optimized.

9.2.3 Daily rainfall runoff models


193
The concept of a simple water balance model is extended to develop a parsimonious daily
rainfall runoff model, PDRRM. The proposed model and three classical daily rainfall runoff
models XNJ, (Zhao et al., 1980), SMAR (OConnell et al., 1970) and NAM (Danish
Hydraulic Institute, 1982) are applied to four semi-arid catchments from Upper Awash River
Basin, Ethiopia. Preliminary results comparing the proposed model and three classical daily
rainfall run-off models show that, though the quality of the PDRRM results are less than the
classical ones, the general reproduction of the hydrographs of the four test catchments is
promising. There is a quest for a parsimonious daily rainfall runoff model for flood
forecasting and reservoir regulations in small catchments

9.2.4 Practical application of the models

The practical application of the monthly water balance model is demonstrated by extending
the flow of the Awash River from the long meteorological records. The extended flows are
used to study the influence of temporal rainfall variability on flow regimes, low flow analysis
and reservoir capacity design. These analyses were not possible with the short river flow data.

The study on the influence of temporal rainfall variability on the flow regimes for the Awash
River Basin during the period of 1900-1995 shows that there is an amplified variability of the
annual flow compared to the rainfall variability. For example in the 1950s there was a
departure of the annual flows of 25 % from the long term mean (163 mm), corresponding to
only 10 % departure of rainfall from the long term mean (1200mm). The major flow regime
(summer flow) is affected greatly by the low anomalies of rainfall and that the low flow,
which is normally small, does not show large variation in this century. The influence of the on
the timing of the rainfall on the flow regimes showed the amount of rainfall that arrives early
in the rainy season (rainfall of the month of June), is significantly important for the total
summer flow to the river. The August rainfall is the most significant cause of high summer
flows.

From low flow analysis the Awash River at Hombole the flow - duration-return period curves
were established. From these curves we conclude that the Awash River at Hombole can cater
a limited amount of supply without any regulation of the river. The river can supply at least
about 3 and 2 million m
3
per month, with a failure frequency once in 10 and 100 years


194
respectively. The reservoir capacity curves, which are important at the feasibility study of
water resources projects, were also established.

9.3 Recommendations

For large catchment, there is a need for water balance modeling based on distributed input
variable. Recent work on grid inputs includes e.g. the work of Conway (1997) dealing
with a water balance of the Blue Nile.
The daily model proposed should be verified with more case studies.
For 10-day and daily water balance models, groundwater level data could be incorporated
to obtain optimum parameters by using a multi-objective global optimization procedure
(Yapo et al. 1998).
Good quality data collection should always be encouraged. No model can be calibrated or
even used without a good quality of data. The hydrology community should bridge the
gap that is existing in the advancement of model development and the data acquisition.
The latter is lagging considerably. In addition to the traditional ground observation of
hydrometeorological variables derived from satellite images and radar technology can
augment the data availability for water resources studies. Of course even theses recent
technologies can achieve certain goals if and only if they are well calibrated by ground
observations.


195

REFERENCES

Abayneh, B., 1992. Relationships between physical basin characteristics and parameters of a
monthly water balance model, MSc. Thesis. Laboratory of Hydrology, VUB,
Brussels, Belgium.
Abbott, M.B., Bathurst, J.C., Cunge, J. A., O'Connell, P.E. and Rasmussen, J., 1986a. An
introduction to the European Hydrological system - Systm Hydrologique
Europen- "SHE". 1:History and philosophy of a physically based distributed
modelling system. J. Hydrol, 87: 45-59.
Abbott, M.B., Bathurst, J.C., Cunge, J. A., O'Connell, P.E. and Rasmussen, J., 1986b. An
introduction to the European Hydrological system - Systm Hydrologique
Europen- "SHE" 2: Structure of a physically based distributed modelling system. J.
Hydrol, 87: 61-77.
Alem, G., 1989. Monthly pure runoff model for the upper Awash River catchment in Ethiopia.
MSc. Thesis, Laboratory of Hydrology, VUB, Brussels, Belgium.
Alley, W. M., 1984. On the treatment of evaporation, soil moisture accounting and aquifer
recharge in water balance models. Waer Resour. Res. 20(11), 37-1149.
Beven, K.J. and Kirkby, M.J., 1979. A physically based variable contributing area model of
basin hydrology. Hydrol. Sci. Bull., 24:43-69.
Beven, K.J., Lamb, R., Quinn, P., Romanwicz, R. and Free, J., 1995. TOPMODEL. In: V.P.
Sigh (Ed.). Computer Models of Watershed Hydrology. Water Resources
Publications, Highlands Ranch, CO, pp 627-668.
Blaney, H. F. and Crddidle, W.D., 1950. Determining water requirements in irrigated areas
from climatological and irrigation data. USDA Soil Conserv. Serv. TP - 96, 48pp.
Blschl, G. and Sivapalan, M., 1995. Scale issues in hydrological modelling: A review. In
Scale issues in Hydrological modelling (ed) Kalma, J.D. and Sivapalan, M. John
Wiley & Sons. p 9-48.
References

197
Box, G.E.P. and Jenkins, G.M., 1976. "Time series analysis. Forecasting and Cotrol" Holden-
day, San Francisco 543p.
Conway, D. and Hulme, M., 1993. Recent fluctuations in precipitation and runoff over the
Nile subbasins and their impact on main Nile discharge. Climate Change 25:127-
151.
Burnash, R.J.C., 1995. The NWS river forecast system-catchment modelling. In: V.P. Sigh
(Ed.). Computer Models of Watershed Hydrology. Water Resources Publications,
Highlands Ranch, CO, pp. 311-366.
Burnash, R.J.C., Ferral , R.L and McGuire , R.A., 1973. A general stream flow simulation
system - Conceptual modelling for digital computers. Report by the Joint Federal
State River Forecasts Senter, Sacramento, California.
Chow, V.T., 1964. Hydrology of arid and semiarid regions. In Handbook of Applied
Hydrology. 24-1 -24-42. McGrawhill Company. New York.
Chow, V.T., Maidment, D.R. and Mays, L.W., 1988. Applied Hydrology, McGraw-Hill, New
York.
Clarke, R.T., 1973. 'Mathematical models in hydrology.' Irrigation Drainage Paper 19. FAO,
Rome
Conway, D., 1997. A water Balance model of the Upper Blue Nile, Ethiopia. J. Hydrol Sci.
42(2) 265-286.
Crawford, N.H. and Linsley, R.K., 1966. Digital simulation in hydrology. Stanford
watershade model IV. Tech. Rep. No 39. Dep. Civ. Eng., Stanford University.
Danish Hydraulic Institute, 1982. NAM Model Documentation. 82-892, JCR/Skn.
Dawdy, D.R. and O' Donnel, T., 1965. Mathematical models of catchment behavior. J.
Hydraul. Di. Proc. ASCE, HY4

References

198
Demaree, G., 1982 Comparison of techniques for the optimization of conceptual hydrological
models. Mathematics and Computers in Simulation XXIV pp.122-130. North-Holland
Publishing Company.
Dias, P.P.G., 1993. Monthly water balance models for basins in Sir-Lanka and West Africa.
MSc. Thesis, Laboratory of Hydrology. VUB, Brussels, Belgium.
Dooge, J.C.I., 1957. The method for estimating flood peak. Engineering (London), 184:311-
374.
Dooge, J.C.I., 1973. The linear theory of the unit hydrologic systems. Tech. Bull. U.S. Dep.
Agric., No. 1468, U.S. Gov. Print. Off., Washington, D.C.
Duan, Q., Sorooshin, S. S. and Gupta, V.K., 1992. Effective and efficient global optimization
for conceptual rainfall runoff models. Water Resour. Res., 28(4): 1015-1031.
Dyck, S., 1983. Overview on the present status of concepts of water balance models, in: Van
der Beken, A. and Hermann, A. (eds), New Approaches in Water Balance
Computations (Proceeding of the Hamburg Workshop), IAHS publ. No.148, pp.3-19.
Franchini, M., Galeati, G. and Berra, S., 1998. Global optimization techniques for calibration
of conceptual rainfall-runoff models. Hydrol. Sci. J. 43(3), 443-457.
Franchini, M. and Pacciani, M., 1991. Comparative analysis of several conceptual rainfall-
runoff models. J.Hydrol., 122:161-219.
Franchini, M. and Todini, E., 1988. PABL: A parabolic and backwater scheme with lateral
inflow and outflow. 5
th
IAHR Int. Symp. Stochastic Hydraulics, Birmingham, UK.
Fuchs, M., 1973. The estimation of evapotranspiration. In B.Yaron, E. Danfors and Y. Vaadia
(Eds). Arid Zone Irrigation. New York. pp 241-247.
Gan T. Y., 1988. Application of scientific modelling of hydrological response from
hypothetical small catchments to assess a complex conceptual rainfall runoff model.
Water Resources Series Technical reports no. 111. Department of Civil Engineering,
University of Washington, Seattle, Washington

References

199
Gan, T. Y. and Biftu, G. F., 1996. Automatic calibration of conceptual rainfall-runoff models:
Optimization algorithms, catchment conditions, and model structure. Water Resour.
Res., 32(12), 3513-3524.
Gan, T. Y., Dlamini, E. M. and. Biftu, G.F., 1997. Effects of model complexity and structure,
data quality, and objective functions on hydrologic modeling, J. Hydrol., 192: 81-
103.
Grgens, A. H. M., 1983. Reliability of calibration of monthly rainfall-runoff model: the
semiarid case. Hydrolo. Sci. J. 28(4), 485-498
Haan, C.T., 1972. A water yield model for small watersheds. Water Resour. Res. 8(1), 28-69.
Harwell Subroutine Library, 1974. Subroutine VA05A, 4p.
Helsel, D.R. and Hirsch, R.M., 1992. Statistical Methods in Water Resources, Elsevier
Sciences Publishers Amsterdam, The Netherlands.
Hassan, A., 1984. Stochastic monthly water balance - Subcatchments of upper Awash Basin.
MSc. Thesis, Laboratory of Hydrology VUB, Brussels, Belgium.
Holland, J. H., 1975. Adaptation in natural and artificial systems. University of Michigan
Press. Ann Arbor, Michigan.
Hooke, R. and Jeeves, T.A., 1961. Direct search solutions of numerical and statistical
problems, Journ. Assoc. Comput. Mach., Vol. 8(2), 212-229.

References

200
Horton, R.E., 1939. The role of infiltration in the hydrologic cycle, Trans. Am. Geophys.
Union, vol. 14, pp.446-460.
Hsu, Kuo-Iin, Gupta, H.V. and Sorooshian, S., 1995. Artificial neural network modelling of
the rainfall-runoff process. Water Resour. Res. 31 (10), 2517-2530.
Hughes, D. A. and Metzler, 1998. Assessment of three monthly rainfall-runoff models for
estimating the water resources yield of semi-arid catchments in Namibia. Hydrol. Sci.
J. 43(2), 283-297.
Hughes, D. A., 1982. Conceptual catchment model parameter transfer studies using monthly
data from Southern Cape Coastal Lakes Region. Report 1/82. Hydrological Research
Unit, Rhodes University, Grahamstown.
Hughes, D. A., 1995. Monthly rainfall runoff models applied to arid and semi-arid catchments
for water resource estimation purposes. Hydrol. Sci. J. 40(6), 751-769.
Hughes, D. A., 1997. Southern African "FRIEND"- The application of rainfall-runoff models
in SADC Region. WRC Report No. 235/1/97. Institute for Water Research Rhodes
University.
Kalman, R.E., and Bucy, R., 1961. New results in linear filtering prediction J. Basic Eng.,
(ASME), 83D: p95-108.
Kalman. R.E., 1960. A new approach to linear filtering and prediction problems. J. Basic
Eng., (ASME), 82D: p35-45
Kohler, M.A. and Linsley, R.K., 1951. Predicting the runoff from storm rainfall. U.S. Weather
Bureau Research Paper 34.
Li, C., 1995. Further development of monthly water balance models in arid and semi-arid and
areas in China. MSc. Thesis, Laboratory of Hydrology. VUB, Brussels, Belgium.
Kuczera, G., 1983. Improved parameter inference in catchment models. 1, Evaluationg
parameter uncertainty, Water Resour. Res. 19(5), 1151-1162.

References

201
Makhlouf, Z. and Michel, C., 1994. A Two parameters monthly water balance model for
France watersheds, J. hydrol. 162: 299-318.
Mandeville, A.N., O'connell, P.E., Sutcliiffe, J.V. and Nash, J.E. 1970. River Forecasting
through conceptual models part III.- The Ray catchment at Grendon Underwood. J.
Hydrol. 11: 109-128.
McMahon, T.A. and Mein, R.G., 1978. Reservoir capacity and yield. Elsevier Scientific
publishing company, Amsterdam.
Misganaw D., 1989. Analysis of Drought in Ethiopia based on Nile River flow records: In the
Proceedings of the Sahel Forum - The state- of- the-art of Hydrology and
Hydrogeology in the Arid and Semiarid Areas of Africa Ouagadougou, Burkina Faso
18-23 February 1989.
Moore, R.J., 1989. Hydrological Modelling for water management in arid and semiarid areas
of Africa: In the Proceedings of the Sahel Forum - The state- of- the-art of
Hydrology and Hydrogeology in the Arid and Semi-arid Areas of Africa
Ouagadougou, Burkina Faso 18-23 February 1989.
Moreda, F. and Bauwens W., 1997. Conceptual monthly water balance models applied to arid
and semiarid catchments. Paper presented on 5
th
Scientific Assembly of IAHS in
workshop Flow forecasting under conditions of limited data. Rabat, Morocco. 12p.
Moreda, F. and Bauwens W., 1998. A conceptual 10-day water balance model for semi-arid
catchments. Proceedings of the 18th AGU Annual HYDROLOGY DAYS.
Hydrology days publications, 57 Sleby Lane, Antherton, CA. pp 211-221.
Moreda, F. Vandewiele, G., Demaree, G. and Bauwens, W., 1998. Application of pure run-off
models in low flow analysis of Awash River basin in Ethiopia. In Demare, G.
Alexander, J. and De Dapper, M. Proc. of International conference "Tropical
Climatology, Meteorology and hydrology." pp. 532-555.
Mott MacDonald and Partners, 1990. Country report: Zambia Sub Saharan Africa.
Hydrological Assessment SADC Countries, UK.

References

202
Nash, J.E., 1958. The form of instantaneous unit hydrograph. IUGG Gen. Assem. Toronto,
Vol. III, IAHS Publ.No. 45. p114-121.
Nash, J.E., 1959. Systematic determination of unit hydrograph parameters. Journal of
Geophysical Research 64(1): 111-115.
Nash, J.E. and Barsi, B.I., 1983. A hybrid model for flow forecasting catchments. J. Hydrol.
65(3):125-137.
O' Connell, P. E. and Todini, E., 1996. Modelling of rainfall, flows transport in hydrological
systems: an overview. J. Hydrol. 175: 3-16.
Pitman, W.V., 1973. A mathematical Model for Generating River flows from Meteorological
Data in South Africa, Report 2/73, Hydrological Research Unit, University of the
Witwaterstand, Johannesberg.
Nash, J.E., 1960 A unit hydrograph study, with particular reference to British catchments.
Proc. Inst. Civ. Eng. (London) 249-282.
Nash, J.E. and Sutcliffe, J., 1970. River flow forecasting through conceptual models, Part I A
discussion of principles. J. Hydrol. 10: 282-290.
Natale, L. and Todini, E., 1976. A stable estimator for linear models II: real hydrological
application. Water Resour. Res. 12(4) 672-676.
Nedler, J. A and Mead, R., 1965. A simplex method for function minimization. Computer J.
7: 308-313.
Ni-Lar-Win, 1994. Contribution to rainfall runoff modelling on basin scale. VUB Hydrologie-
(26), Brussels.
O'Connell, P.E., Nash, J.E. and Farrell, J.P., 1970. River flow forecasting through conceptual
models, Part II . J. hydrol. 10: 317-329.
Palmer, W.C., 1965. Meteorologic drought. Research Paper U.S. Weather Bur., 45, 58p.
Philip, J.R., 1957. Theory of infiltration: 1. The infiltration equation and its solution. Soil Sci.
83(5), 345-357.

References

203
Pitman, W.V., 1978. Flow generation by catchment models of different complexity - a
comparison of performance. J. Hydrol. 38: 59-70.
Price, W.L., 1987. Global optimization algorithim for a CAD workstation, Journal of
Optimization Theory and Applications, 55(1), 133-146.
Refsgaard, J. C. and Storm, B., 1995. MIKE SHE. In: V.P. Sigh (Ed.). Computer Models of
Watershed Hydrology. Water Resources Publications, Highlands Ranch, CO, pp 809-
846.
Roberts, P.J.T., 1978. A comparison of performance of selected conceptual models of the
rainfall runoff process in semi-arid catchments near Grahamstown, Report 1/78,
Hydrological Research Unit. Rhodes University, Grahamstown.
Roberts, P.J.T., 1979. Model FLEXIFIT: A conceptual rainfall-runoff model for extension of
monthly runoff records. Dept.of Environment Affairs, Hydrological Research
institute, Tech. Report TR98.
Rockwood , D. M., Davis, E. D. and Anderson, J.A., 1972. User manual for COSSARR
model. US Army Engineering Division, North Pacific, Portland, OR.
Rockwood, D. M. and Nelson, M.L., 1966. Computer application to transform flow synthesis
and reservoir regulation. IV Int. Cnof. Irrig. Drain.
Rosenbrock, H.H, 1960. An automatic method for finding the greatest or least value of a
function, Computer Journal, Vol. 3, 175-184.
Salas.J.D, Tabios III, G.Q. and Obeyeskera, J.T.B., 1986. Seasonal model for watershed
simulation, Users Manual, Dept. of Civil Engineering, Colorado State University.
Seleshi, Y., 1996. Stochastic Predictions of summer rainfall amounts over the northeast
African highlands and over India. Ph.D. Thesis, Laboratory of Hydrology Free
University Brussels, Belgium. 351p.
Servat, E. and Dezetter, A., 1993. Rainfall-runoff modelling and water resources assessment
in northwestern Ivory Coast. Tentative extension to ungauged catchments. J. Hydrol.,
148: 231-248.

References

204
Singh, V.P. and Xu, C.-Y., 1997. Evaluation and generalization of 13 mass-transfer equations
for determining free water evaporation. Hydrological processes. 11: 311-323.
Sorooshian, S., 1981. Parameter estimation of rainfall runoff Models with Hetroscedastic
streamflow errors: Non informative data case, In: Y. Haimes and J. Knolles (Eds.),
Water and related land resources systems, IFAC symposium, Cleveland, Ohio, May
28-31, 1980. Pergamon Press, New York N.Y., pp. 477- 485.
Sorooshian, S., and Gupata, V.K., 1995. Model calibration. In: V.P. Sigh (Ed.). Computer
Models of Watershed Hydrology. Water Resources Publications, Highlands Ranch,
CO, pp23-68.
Todd, D.K., 1980. Groundwater Hydrology. 2 & Sons, Inc, NewYork.
Singh, V.P., 1992. Elementary Hydrology. Prentice Hall, Englewood Cliffs.973p.
Sorooshian , S. and Dracup, J.A., 1980. Stochastic parameter estimation procedures for
hydrologic rainfall-runoff models: Correlated and hetroscedastic cases, Water Resor.
Res. 16(2), 430-442.
Sorooshian, S., Gupta, V.K, and Fulton J.L., 1983. Evaluation of maximum likelihood
parameter estimation techniques for conceptual rainfall-runoff models- influence of
calibration data variability and length on model credibility, Water Resour. Research,
19 (1), 251-259.
Starosolszky, O., 1987. Applied surface hydrology. Water Resources Publications, Littleton,
CO, p.526.
Szilagyi, J. and Parlange M. B., 1998. Base flow separation based on analytical solution of the
Boussinesq equation. J. Hydrol. 204: 251-260.
Thornthwaite, and C. W., Mather, J.R., 1955. The water balance. Publ. Climatol. Lab. Dresel
Inst. Technol. 8(1), 1-104
Thornthwaite, C. W. 1948. An approach toward a rational classification of climate. Geog. Rev.
38(1), 55-94.
nd
ed. John Willey

References

205
Todini, E. and Wallis, J. R., 1978. A real time rainfall-runoff model for an on-line flood
warning system. Proc. AGU Chapman Cof. Appl. Kalman Filtering Tech. Dep. Civ.
Eng. University of Pittsburgh, Pa.
Todini, E., 1978. Mutually interactive state parameters (MISP) estimation. Proc. AGU
Chapman Cof. Appl. Kalman Filtering Tech. Dep. Civ. Eng. University of
Pittsburgh, Pa.
Todini, E., 1988. Rainfall-runoff modelling- past, present and future. J.Hydrol. 100:341-352.
Todini, E., 1996. The ARNO rainfall-runoff model. J. Hydrol. 175:339-382.
Todini, E., and Bossi, A., 1986. PAB (Parabolic and backwater) an unconditionally stable
flow routing scheme particularly suited for real time forecasting and control. J.
Hydraul. Res. 24: 405-424.
Van der Beken, A., 1977. and J. Byloos, J., 1977. A monthly water balance model including
deep infiltration and canal losses. Hydrol. Sci. Bull. 22(3), 341-351.
UNESCO, 1979. Map of the world distribution of arid regions, MAB Technical note 7.
UNESCO.
Van der Beken, A., 1977. A monthly water balance model applied to two different watersheds
3rd International Symposium Fort Collins. Colorado. pp. 178-189.
Vandeweiele, G.L. and Atlabachew, E., 1995. Monthly water Balances of ungauged
catchments obtained by geographical regionalization. J. Hydrol. 170: 277-291.
Vandewiele, G.L. and Ni-Lar-Win, (1998). Monthly water balance models for 55 basins in 10
countries. Hydrol. Sci. J. 43(5), 689-699.
Vandewiele, G.L. Xu, C.Y and Ni-Lar-Win, 1992. Methodology and comparative study of
monthly water balance models in Belgium, China and Burma. J. Hydrol., 134:315-
347.

References

206
Vandewiele, G.L., Xu, C.Y and Ni-Lar-Win, 1993. Methodology for constructing monthly
water balance models on basin scale. VUB Hydrologie (20) Second ed. Brussels,
Belgium.
Vandewiele, G.L., Xu, C.Y and Huybrechts, W., 1991. Regionalization of physically based
water balance models in Belgium. Application to ungauged catchments. Water
Resource Mangement 5: 199-208.
Verma, R.D., 1979. A physical model of the rainfall-runoff relationships for semiarid lands. In
The hydrology of areas of low precipitation, (proceedings of the Canberra
Sympossium). IAHS Publ. no. 128.
Wang, Q.J., 1991. The genetic algorithm and its application to calibrating conceptual rainfall -
runoff models. Water Resour. Res, 27(9), 2467-2471.
WMO, 1975. Inter comparison of hydrological models. WMO publ., Geneva.
WMO, 1994. Data Review and coding. In Guide to Hydrological Practice. Fifth ed. WMO -
No 168. pp. 316.339.
Xu, C. Y. and Singh, V. P. 1998. A review on monthly water balance models for water
resources investigations. Water Resources Management 12: 31-50.
Xu, C. Y., 1988 Regional study of monthly rainfall runoff models. MSc. Thesis. Laboratory of
Hydrology, VUB, Brussels, Belgium.
Xu, C. Y., 1992. Monthly water balance models in different climatic regions. VUB
Hydrologie- (22), Brussels, Belgium.
Xu, C. Y., Seibert, J. and Halldin, S., 1995. Regional water balance modelling in the NOPEX
area: development and application of monthly water balance models. J. Hydrol. 180:
211-236.
Yapo, P.O., Gupta, H.V., and Sorooshian, S., 1996. Automatic calibration of conceptual
rainfall - runoff models: Sensitivity to calibration data. J. Hyrol. 181: 23-48.

References

207
Yapo, P.O., Gupta, H.V., and Sorooshian, S., 1998. Multi-objective global optimization for
hydrological models. J. Hyrol. 204: 83-97.
Zhao, R.J, Zhaung, Y.L, Fang, L.R., Lui, X.R. and Zhang, Q.S., 1980. The Xinanjiang model.
In Hydrological Forecasting, Proceedings of the Oxford Symposium, IAHS publ. No
129.

References

208

APPENDIX A The VUBMOD for Windows
APPENDIX B Orders of magnitude of the model parameters

Fekadu Moreda

Uploaded by

Copyright:

Available Formats

Fekadu Moreda

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fekadu Moreda

Uploaded by

Copyright:

Available Formats

Laboratory of Hydrology

Faculty of Applied Sciences, VUB

5.991 7.815 9.488 11.070 12.592 14.067

You might also like