Short-Term Load Forecasting Using Soft Computing Techniques: D. K. Chaturvedi, Sinha Anand Premdayal, Ashish Chandiok

Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Int. J.

Communications, Network and System Sciences, 2010, 3, 273-279


doi:10.4236/ijcns.2010.33035 Published Online March 2010 (http://www.SciRP.org/journal/ijcns/).

Short-Term Load Forecasting Using Soft


Computing Techniques
D. K. Chaturvedi1, Sinha Anand Premdayal1, Ashish Chandiok2
1
Department of Electrical Engineering, D. E. I., Deemed University, Agra, India
2
Department. of Electronics and Communication, B. M. A. S., Engineering College, Agra, India
Email: dkc_foe@ rediffmail.com
Received November 10, 2009; revised December 18, 2009; accepted January 21, 2010

Abstract
Electric load forecasting is essential for developing a power supply strategy to improve the reliability of the
ac power line data network and provide optimal load scheduling for developing countries where the demand
is increased with high growth rate. In this paper, a short-term load forecasting realized by a generalized neu-
ron–wavelet method is proposed. The proposed method consists of wavelet transform and soft computing
technique. The wavelet transform splits up load time series into coarse and detail components to be the fea-
tures for soft computing techniques using Generalized Neurons Network (GNN). The soft computing tech-
niques forecast each component separately. The modified GNN performs better than the traditional GNN. At
the end all forecasted components is summed up to produce final forecasting load.
Keywords: Wavelet Transform, Short Term Load Forecasting, Soft Computing Techniques

1. Introduction forecasting using GNN with wavelet for accurate load


forecasting results. This paper is organized as follows:
Short-term load forecasting (STLF) is an essential tech- Section 2 discusses various traditional and soft comput-
nique in power system planning, operation and control, ing based short term load forecasting approaches. Con-
load management and unit commitment. Accurate load cept of wavelet analysis required for prediction will be
forecasting will lead to appropriate scheduling and plan- discussed in Section 3 while elements of generalized
ning with much lower costs on the operation of power neural architecture needed will be described in Section 4.
systems [1–6]. Traditional load forecasting methods, A prediction procedure using wavelets and soft comput-
such as regression model [7] gray forecasting model [8,9] ing techniques and its application to time series of hourly
and time series [10,11] do not consider the influence of load forecasting consumption is discussed in Section 5.
all kind of random disturbances into account. At recent Section 6 includes discussion and concluding remarks.
years artificial intelligence are introduced for load fore-
casting [12–17]. Various types of artificial neural net- 2. Traditional and Soft Computing Techni-
work and fuzzy logic have been proposed for short term ques for Short Term Load Forecasting
load forecasting. They enhanced the forecasting accuracy
compared with the conventional time series method. The
ANN has the ability of self learning and non-linear ap- 2.1. Traditional Approaches
proximations, but it lacks the inference common in hu-
man beings and therefore requires massive amount of Time Series Methods
training data, which is an intensive time consuming proc- Traditional short term load forecasting relies on time
ess. On the other hand fuzzy logic can solve uncertainty, series analysis technique. In time series approach the
but traditional fuzzy system is largely dependent on the model is based on past load data, on the basis of this
knowledge and experiences of experts and operators, and model the forecasting of future load is done. The tech-
is difficult to obtain a satisfied forecasting result espe- niques used for the analysis of linear time series load
cially when the information is incomplete or insufficient. signal are:
This paper aims to find a solution to short term load 1) Kalman Filter Method

Copyright © 2010 SciRes. IJCNS


274 D. K. CHATURVEDI ET AL.

The kalman filter is considered as the optimal solution and current load data as input and output respectively.
to many data prediction and trend matching. The filter is The ANN learns from experience and generalizes from
constructed as a mean square minimization which re- previous examples to new ones. It is able to forecast
quires the estimation of the covariance matrix. The role more efficiently the load as the load pattern are non lin-
of the filter is to extract the features from the signal and ear and ANN is capable to catch trends more accurately
ignore the rest part. As load data are highly non linear than conventional methods.
and non stationary, it is difficult to estimate the covari- 2) Rule Based Expert Systems
ance matrix accurately [18]. An expert system is a logical program implemented on
2) Box Jenkins Method computer, to act as a knowledge expert. This means that
This model is called as autoregressive integrated mov- program has an ability to reason, explain and have its
ing average model. The Box Jenkins model can be used knowledge base improved as more information becomes
to represent the process as stationary or non stationary. A available to it. The load-forecast model can be built us-
stationary process is one whose statistical properties are ing the knowledge about the load forecast domain from
same over time, which means that they fluctuate over an expert in the field. The knowledge engineer extracts
fixed mean value. On other hand non stationary time this knowledge from the load domain. This knowledge is
series have changes in levels, trends or seasonal behavior. represented as facts and rules using the first predicate
In Box Jenkins model the current observation is weigh- logic to represent the facts and IF-THEN production
ted average of the previous observation plus an error rules. Some of the rules do not change over time, some
term. The portion of the model involving observation is changes very slowly; while others change continuously
known as autoregressive part of the model and error term and hence are to be updated from time to time [20].
is known as moving average term. A major obstacle here 3) Fuzzy Systems
is its slow performance [19]. Fuzzy sets are good in specialization, fuzzy sets are
3) Regression Model able to represent and manipulate electrical load pattern
The regression method is widely used statistical tech- which possesses non-statistical uncertainty. Fuzzy sets
nique for load forecasting. This model forms a relation- are a generalization of conventional set theory that was
ship between load consumptions done in past hour as a introduced as a new way to represent vagueness in the
linear combination to estimate the current load. A large data with the help of linguistic variable. It introduces
data is required to obtain correct results, but it requires vagueness (with the aim of reducing complexity) by
large computation time. eliminating the sharp boundary between the members of
4) Spectral Expansion Technique the class from nonmembers [21,22].
This method is based on Fourier series. The load data These approaches are based on specific problems and
is considered as a periodic signal. Periodic signal can be may represent randomness in convergence or even can
represented as harmonic summation of sinusoids. In the diverge. The above mentioned approaches use either reg-
same way electrical load signal is represented as summa- ression, frequency component or mean component or the
tion of sinusoids with different frequency. The drawback peak component to predict the load. The prediction of the
of this method is that electrical load is not perfect peri- load depends upon both time and frequency component
odic. It is a non stationary and non linear signal with which varies dynamically. In this paper, an attempt is
abrupt variations caused due to weather changes. This made to predict electrical load that combines the above
phenomenon results in the variation of high frequency mentioned features using generalized neurons and wave-
component which may not be represented as periodic let.
spectrum. This method is not suitable and also requires
complex equation and large computation time. 3. Elements of Wavelet Analysis

2.2. Soft Computing Approach Wavelet analysis is a refinement of Fourier analysis [9–
15,23–29] which has been used for prediction of time
Soft computing is based on approximate models working series of oil, meteorological pollution, wind speed, rain-
on approximate reasoning and functional approximation. fall etc. [28,29]. In this section some important vaults
The basic objective of this method is to exploit the tol- relevant to our work have been described. The underly-
erance for imprecision, uncertainty and partial truth to ing mathematical structure for wavelet bases of a func-
achieve tractability, robustness, low solution cost and tion space is a multi-scale decomposition of a signal,
best results for real time problems. known as multi-resolution or multi-scale analysis. It is
1) Artificial Neural Networks (ANN) called the heart of wavelet analysis. Let L2(R) be the
An artificial neural network is an efficient information space of all signals with finite energy. A family {Vj} of
processing system to perform non-linear modeling and subspaces of L2(R) is called a multi resolution analysis of
adaptation. It is based on training the system with past this space if

Copyright © 2010 SciRes. IJCNS


D. K. CHATURVEDI ET AL. 275

1) intersection of all Vj, j = 1, 2, 3, ...... be non-empty, It is quite clear that in the higher case the scaled, trans-
that is  V j   lated and normalized versions of  are denoted by
j

2)This family is dense in L2(R), that is, = L2(R)  j , k  t   2 j /2  2 j x  k  (6)


3) f (x)  Vj if and only if f (2x)  Vj + 1
4) V1  V2  ..... Vj  Vj + 1 With orthogonal wavelet  the set {j, k | j, k  Z}
There is a function preferably with compact support of is an orthogonal wavelet basis. A function f can be rep-
such that translates  (x – k) k  Z, span a space V0. A resented as
finer space Vj is spanned by the integer translates of the
scaled functions for the space Vj and we have scaling f =  c j ,k j ,k ( ,c j ,k  < f >) (7)
j  Z k Z
equation
The Discrete Wavelet Transform (DWT) corresponds
 ( x)  ak  (2 x  1) (1)
to the mapping f  cj,k. DWT provides a mechanism to
with appropriate coefficient ak, k  Z.  is called a scal- represent a data or time series f in terms of coefficients
ing function or father wavelet. The mother wavelet  that are associated with particular scales [24,26,27] and
is obtained by building linear combinations of  . Fur- therefore is regarded as a family of effective instrument
for signal analysis. The decomposition of a given signal f
ther more  and  should be orthogonal, that is,
into different scales of resolution is obtained by the ap-
 ( k ) , (l )  0,l,k  Z (2) plication of the DWT to f. In real application, we only
use a small number of levels j in our decomposition (for
These two conditions given by (1) and (2) leads to
instance j = 4 corresponds to a fairly good level wavelet
conditions on coefficients bk which characterize a mother
wavelet as a linear combination of the scaled and dilated decomposition of f).
father wavelets  : The first step of DWT corresponds to the mapping f to
its wavelet coefficients and from these coefficients two
 (x)=  bk  (2 x  k ) (3) components are received namely a smooth version, nam-
k z
ed approximation and a second component that corre-
Haar, Daubechies and Coefmann are some well known sponds to the deviations or the so-called details of the
wavelets. signal. A decomposition of f into a low frequency part a,
Haar wavelet (Haar mother wavalet) denoted by ψ is and a high frequency part d, is represented by f = a1 + d1.
given by The same procedure is performed on a1 in order to obtain
 1, 0  x 1 2 decomposition in finer scales: a1 = a2 + d2. A recursive
 decomposition for the low frequency parts follows the
 (x ) =  1, 1 2 < x 1 (4)
 0, x < 0, x > 1 directions that are illustrated in Figure 1.

The resulting low frequency parts a1, a2, ..... an are ap-
Can be obtained from the father wavelet proximations of f, and the high frequency parts d1, d2, .....
 1, 0  x 1 dn contain the details of f. This diagram illustrates a
 (x)=  (5) wavelet decomposition into N levels and corresponds to
0, x  0, x  1
f  d1  d 2  d3    d N 1  d N  aN (8)
In this case coefficients ak in (1) are a0 = a1 = 1 and ak
= 0 for k  0, 1. The Haar wavelets is defined as a linear In practical applications, such decomposition is ob-
combination of scaled father wavelets (x) =  (2x) – tained by using a specific wavelet. Several families of
 (2x – 1) which means that coefficients bk in (3) are b0 = wavelets have proven to be especially useful in various
1, b1 = –1 and bk = 0 otherwise, Haar wavelets can be applications. They differ with respect to orthogonality,
interpreted as Daubechie’s wavelet of order 1 with two smoothness and other related properties such as vanish-
coefficients. In general Daubechies’ wavelets of order N ing moments or size of the support.
are not given analytically but described by 2N coeffi-
cients. The higher N, the smoother the corresponding
Daubechies’ wavelets are (the smoothness is around 0-2N
for greater N). Daubechies’ wavelets are constructed in
a way such that they give rise to orthogonal wavelet
bases. It may be verified that orthogonality of translates
of  and  , requires that  ak = 2 and  bk = 2. Figure 1. Wavelet decomposition in form of coarse and de-
k k
tail coefficients.

Copyright © 2010 SciRes. IJCNS


276 D. K. CHATURVEDI ET AL.

4. Neuro Theory of Generalized Neuron type generalized neuron in the forward pass, as in the
Model feed-forward neural network, it is compared with the
desired output to find the error. Using back-propagation
algorithm the summation type GN is trained to minimize
The following steps are involved in the training of a the error. In this step, the output of the single flexible
summation type generalized neuron as shown in Figure 2. summation type generalized neuron is compared with the
desired output to get error for the ith set of inputs:
4.1. Forward Calculations
Error E i  (Y i  O i ) (12)
Step 1: The output of the  1 part of the summation type Then, the sum-squared error for convergence of all the
generalized neuron is patterns is

1 E p  0.5 E i 2 (13)
O  (9)
1  e  s*s _ net A multiplication factor of 0.5 has been taken to sim-
plify the calculations.
where s _ net   Wi X i  X o Step 5: Reverse pass for modifying the connection
Step 2: The output of the  part of the summation type strength.
generalized neuron is 1) Weight associated with the  1 and  2 part of the
2 summation type Generalized Neuron is:
O  e   p* pi _ net (10)
W (k )  W (k  1)  W (14)
where pi _ net   Wi X i * X o
where W   k (O  O ) X i  W (k  1)
Step 3: The output of the summation type generalized
neuron can be written as and  k   (Y i  O i )
O pk  O *(1  W )  O *W (11)
2) Weights associated with the inputs of the  1 part
of the summation type Generalized Neuron are:
4.2. Reverse Calculation
Wi (k )  Wi (k  1)  Wi (15)
Step 4: After calculating the output of the summation where Wi   j Xi  Wi (k  1)

s_bias
and  j    kW (1  O ) *O
3) Weights associated with the input of the  part of
the summation type generalized Neuron are:
Input, Xi Wi (k )  Wi (k  1)  Wi (16)
Output, Opk where Wi   j Xi  Wi (k  1)

and  j    k (1  W ) *(2* pi _ net ) *O


pi_bias
Momentum factor for better convergence.
Learning rate.
s_bias
Range of these factors is from 0 to 1 and is determined
by experience.

Input, 5. Generalized Neuron–Wavelet Approach


Xi Output, Opk
The Generalized Neuron–Wavelet approach has been
used to predict the electrical load. In this approach, Dau-
bechies wavelets Db8 have been applied in the decom-
pi_bias
position for the give data pattern. There are four wavelet
Figure 2. Learning algorithm of a summation type general- coefficients are used. All these wavelet coefficients are
ized neuron. time dependent (the first three wavelet coefficients from

Copyright © 2010 SciRes. IJCNS


D. K. CHATURVEDI ET AL. 277

d1 to d3 and the coarse approximation a3. These coeffi-


cients are illustrated in the Figure 3. We observe the
substantial difference of variability of the signals at dif-
ferent levels. The higher is the wavelet level, the lower
variation of the coefficients and easier prediction of them.
Our main idea is to substrate the prediction task of the
original time series of high variability by the prediction
of its wavelet coefficients on different levels of lower
variability’s, and then using Equation (4) for final pre-
diction of the power load at any time instant n. Since
most of the wavelet coefficients are of lower variability
we expect the increase of the total prediction accuracy.
The wavelet tool available in Matlab is used for the
process of wavelet decomposition of the time series rep-
resenting average of the power load data for 120 hours.
This step involves several different families of wavelets
and a detailed comparison of their performance. In our
case, The Daubechies wavelets of order 8 are performed.
Three level wavelet decomposition of the given time
series XN = f : is performed Figure 4. Mechanism for forecasting Procedure.
f  a3  d3  d 2  d1 (17)
The smooth part of f is stored in a3, and details on dif-
ferent levels are captured by d1, d2, d3. Consequently a
decomposition of the time series in three different scales
is obtained. Figure 4 illustrates the decomposition of the
original signals. The forecasting procedure methodology
explained in Section 4 is used to predict the next value.
The basic idea is to use the wavelet transforms and pre-
dict the data by soft computing techniques for individual
coefficients of wavelet transform represented by a3, d1,
d2, d3. The input to the architecture to predict the wavelet
coefficients is explained in Figure 5.

Figure 5. Actual and predicted training output using gener-


alized neuron model (GNN).

The total predicted power load at an instant (i) is given


by
F  i   f1  i   f 2  i   f 3  i   f 4  i  (18)

6. Results and Discussions

The electric load data have been collected for 120 hours
from Gujarat system and normalized them in the range
0–1. The Daubechies wavelet Db8 is used for decompo-
sition and the wavelet coefficients d1–d3 and a3 have
been calculated. The trend of coefficients has been used
for GN training and predicting the wavelet coefficients
for future loads. So wavelet is used to extract the feature
coefficients from data and then GN is implemented to
Figure 3. Wavelet decomposition of hour load data into predict the trend of the wavelet coefficient. The results of
wavelet coefficient. GN and actual load have been compared and shown in

Copyright © 2010 SciRes. IJCNS


278 D. K. CHATURVEDI ET AL.

Figures 5 and 6. The root means square error for training


and testing results are .1127 and .1152 mega watts (MW)
as in Table 1. When using generalized neuron and wave-
let conjunction model, training each coefficient and
combining to get the predicted output, a very high im-
provement is obtained in both training and testing results
as shown in Figures 7 and 8. The root means square er-
ror for training and testing data are .0554 and .0763 re-
spectively as in Table 1. The improvement in the results
shows that accuracy of forecasting increases in the com-
bined model and can give correct output for short term Figure 8. Actual and predicted testing output using gener-
alized neuron wavelet model (GNNW).
load forecasting.

7. References

[1] S. E. Papadakis, J. B. Theocharis, S. J. Kiartzis, and A. G.


Bakirtzis, “A novel approach to short-term load fore-
casting using fuzzy neuralnetworks,” IEEE Transactions
on Power Systems, Vol. 13, pp. 480–492, 1998.
[2] D. K. Chaturvedi, P. S. Satsangi, and P. K. Kalra, “Fuzz-
ified neural network approach for load forecasting
Figure 6. Actual and predicted testing output using gener- problems,” International Journal on Engineering Intelli-
alized neuron model (GNN). gent Systems, CRL Publishing, U. K., Vol. 9, No. 1, pp.
3–9, March 2001.
Table 1. Comparison between GNN and GNN Wavelet [3] A. A. Eidesouky and M. M. Eikateb, “Hybrid adaptive
technique. techniques for electric-load forecast using ANN and ARI-
MA,” IEE Proceedings–Generation, Transmission and
Min. Error Max. Error RMSE Distribution, Vol. 147, pp. 213–217, 2000.
Type
(Mw) (Mw) (Mw) [4] A. G. Bakirtzis, J. B. Theocharis, S. J. Kiartzis, and K. J.
Satsios, “Short term load forecasting using fuzzy neural
GNN (training) 0.0009521 0.3663 0.1127 networks,” IEEE Transactions on Power Systems, Vol.
10, pp. 1518–1524, 1995.
GNNW(training) 0.002184 0.1706 0.0554
[5] S. Rahman and O. Hazim, “A generalized knowledge-
GNN(testing) 0.1306 0.2094 0.1152 based short-term load forecasting technique,” IEEE Trans-
actions on Power Systems, Vol. 8, pp. 508–514, 1993.
GNNW(testing) 0.00462 0.1913 0.0673
[6] D. K. Chaturvedi, M. Mohan, R. K. Singh, and P. K.
Kalra, “Improved generalized neuron model for short
term load forecasting,” International Journal on Soft
Computing–A Fusion of Foundations, Methodologies and
Applications, Springer–Verlag, Heidelberg, Vol. 8, No. 1,
pp. 10–18, April 2004.
[7] G. Gross and F. D. Galiana, “Short term load fore-
casting,” Proceedings of the IEEE, Vol. 75, No. 1212, pp.
1558–1573, December 1987.
[8] D. K. Chaturvedi, P. S. Satsangi, and P. K. Kalra, “New
neuron models for simulating rotating electrical machines
and load forecasting problems,” International Journal on
Electric Power System Research, Elsevier Science, Ire-
land, Vol. 52, pp. 123–131, 1999.
[9] P. S. Addision, “The illustrated wavelet transform hand-
book: Introductory theory and applications in science,
engineering medicine and finance,” IOP Publishing LTD,
2002.
[10] Z. Can, Z. Aslan, and O. Oguz, “One dimensional wave-
Figure 7. Actual and predicted training output using gener-
let real analysis of gravity waves,” The Arabian Journal
alized neuron wavelet model (GNNW).

Copyright © 2010 SciRes. IJCNS


D. K. CHATURVEDI ET AL. 279

for Science and Engineering, Vol. 29, No. 2, pp. 33–42, ober 1989.
2004.
[21] C. H. Chen, “Fuzzy logic and neural network handbook,”
[11] Z. Can, Z. Aslan, O. Oguz, and A. H. Siddiqi, “Wavelet McGraw Hill Computer Engineering, 1996.
transform of meteorological parameters and gravity
[22] G. J. Klis and T. A. Folger, “Fuzzy sets uncertainty and
waves,” Annals Geophysics, Vol. 23, pp. 650–663, 2005.
information,” Prentice Hall of India Private Limited,
[12] I. Daubechies, “Ten lectures on wavelets,” SIAM, Phila- 1993.
delphia, 1992.
[23] P. Kumar and E. Foufoula–Georgiou, “Wavelet analysis
[13] A. F. Georgiou and P. Kumar, “Wavelet in geophysics,” for geophysical applications,” Reviews of Geophysics,
Academic Press, San Diago, 1994. Vol. 33, pp. 385–412, 1997.
[14] K. M. Furati, M. Z. Nashed, and A. H. Siddiqi, “Mathe- [24] S. Mallat, “A wavelet tour of signal processing,” Acade-
matical models and methods for real world and systems,” mic Press, NewYork, 1998.
Chapman and Hall/CRC, Taylor and Francis Group, Boca
[25] P. Manchanda, J. Kumar, and A. H. Siddiqi, “Mathema-
Raton, London, New York, Singapore, 2006.
tical methods for modeling price fluctuations of financial
[15] Z. Z. Hu and T. Nitta, “Wavelet analysis of summer rain- time series,” Journal of Franklin Institute, Vol. 344, pp.
fall over north China and India,” Journal Meteorological 613–636, 2007.
Society of Japan, Vol. 74, No. 6, 1996.
[26] A. H. Siddiqi, “Applied functional analysis,” Maral Dek-
[16] J. S. R. Jang and N. Gulley, “Fuzzy logic toolbox,” The kar, New York, 2004.
Mathworks Inc., 24 Prime Park Way, Natick, Mass, 1995.
[27] A. H. Siddiqi, G. Kovin, W. Freeden, U. Mosco, and S.
[17] J. S. R. Jang and C. T. Sun, “Neuro-fuzzy modeling and Stephan, “Theme issue on wavelet fractal in science and
control,” Proceedings of the IEEE, Vol. 83, No. 3, pp. engineering,” Arabian Journal for Science and Engi-
378–406, 1995. neering, KFUPM.
[18] A. K. Mahalanabis, D. P. Kothari, and S. I. Ahson, [28] O. Stanislaw and K. Garanty, “Forecasting of the daily
“Computer aided power system analysis and control,” meteorological pollution using wavelet and support
Tata McGraw Hill Publishing Company Limited, New vector machine,” Engineering Application of Artificial
Delhi, 1988. Intelligence, Vol. 20, pp. 745–755, 2007.
[19] G. E. Box and G. M. Jenkins, “Time series analysis: [29] S. Yousefi, I. Weinrich, and D. Reinarz–Chaos, “Wavelet
Forecasting and control,” Holden-Day, San Fransisco, -based prediction of oil prices,” Solution and Fractals,
1976. Vol. 25, pp. 265–275, 2005.
[20] I. Moghram and S. Rahman “Analysis and evaluation of [30] M. V. Wickerhauser, A. K. Peters, and Wellesley, “Adap-
five short-term load forecasting techniques,” IEEE Trans- ted analysis from theory to software,” 1994.
actions on Power Systems, Vol. 4, pp. 1484–1491, Oct-

Copyright © 2010 SciRes. IJCNS

You might also like