Journal of Hazardous Materials 192 (2011) 568–575

Contents lists available at ScienceDirect

Journal of Hazardous Materials

journal homepage: www.elsevier.com/locate/jhazmat

Modeling of membrane bioreactor treating hypersaline oily wastewater by

artificial neural network
Ali Reza Pendashteh a,b , A. Fakhru’l-Razi a,∗ , Naz Chaibakhsh c , Luqman Chuah Abdullah a ,
Sayed Siavash Madaeni d , Zurina Zainal Abidin a
a
Department of Chemical and Environmental Engineering, Faculty of Engineering, Universiti Putra Malaysia, 43400 UPM Serdang, Selangor D.E., Malaysia
b
Environmental Research Institute, Iranian Academic Center for Education, Culture & Research (ACECR), Rasht, Iran
c
Department of Chemistry, Faculty of Science, Universiti Putra Malaysia, 43400 UPM Serdang, Selangor D.E., Malaysia
d
Chemical Engineering Department, Razi University, Kermanshah, Iran

a r t i c l e i n f o a b s t r a c t

Article history: A membrane sequencing batch reactor (MSBR) treating hypersaline oily wastewater was modeled by
Received 14 November 2010 artificial neural network (ANN). The MSBR operated at different total dissolved solids (TDSs) (35,000;
Received in revised form 13 May 2011 50,000; 100,000; 150,000; 200,000; 250,000 mg/L), various organic loading rates (OLRs) (0.281, 0.563,
Accepted 17 May 2011
1.124, 2.248, and 3.372 kg COD/(m3 day)) and cyclic time (12, 24, and 48 h). A feed-forward neural network
Available online 23 May 2011
trained by batch back propagation algorithm was employed to model the MSBR. A set of 193 operational
data from the wastewater treatment with the MSBR was used to train the network. The training, vali-
Keywords:
dating and testing procedures for the effluent COD, total organic carbon (TOC) and oil and grease (O&G)
Membrane bioreactor
Artificial neural network
concentrations were successful and a good correlation was observed between the measured and pre-
Hypersaline oily wastewater dicted values. The results showed that at OLR of 2.44 kg COD/(m3 day), TDS of 78,000 mg/L and reaction
Halophilic microorganisms time (RT) of 40 h, the average removal rate of COD was 98%. In these conditions, the average effluent COD
Modeling concentration was less than 100 mg/L and met the discharge limits.
© 2011 Elsevier B.V. All rights reserved.

1. Introduction the Peoples Republic of China are 100 and 10 mg/L, respectively
[5].
Many industries generate billions of gallons of wastewaters con- Some physical and chemical methods such as photo-
taining organic matter and high concentrations of NaCl (>3.5%, electrocatalytic decontamination, hydrocyclones, coagulation
w/v). For example, large volumes of oil contaminated wastewa- and flocculation and membrane filtration have been investigated
ter are produced during extraction, transportation and refinery of to remove hydrocarbons from oily wastewaters [6]. Generally,
crude oil [1]. Oilfield wastewater or “produced water” is the largest biological treatment of wastewater is the most cost-effective alter-
high salinity and oily wastewater generated during oil production native when compared to other treatment technologies. However,
activities. The salt, oil and grease (O&G) and total organic carbon the salinity of hypersaline wastewaters affects the metabolism of
(TOC) concentration of oilfield produced water varies from a few microorganisms in activated sludge systems due to plasmolysis
parts per thousand to that of saturated brine [2]; 2–565 mg/L; and whereas halophilic microorganisms are usually able to survive
15–1500 mg/L, respectively [3]. in hypersaline environments [7]. SBR is a promising biological
Discharging untreated produced water can pollute surface and treatment system because of its flexibility and ease of operation [8].
underground water and soil. The permitted O&G limits for treated At high concentration of NaCl, microorganisms exhibit poor
produced water discharging from offshore facilities in the United settleability and cause high turbidity of the SBR effluents [9]. A
States are 42 mg/L daily maximum and 29 mg/L monthly average membrane coupled biological process as a separation step is able
[4]. To reduce the pollution load, many countries have implemented to retain microorganisms in the bioreactor [10]. Membrane biore-
more stringent discharge limits. The monthly average discharge actor (MBR) has many advantages including high effluent quality,
limits of COD and O&G of treated produced water prescribed by small footprint, high mixed liquor suspended solids (MLSS) concen-
tration, good disinfection capability, and high volumetric loading
[11].
Due to the rising concern about environmental issues, the
∗ Corresponding author. Tel.: +60 3 89466304; fax: +60 3 86567120. control and proper operation of wastewater treatment plants to
E-mail address: [email protected] (A. Fakhru’l-Razi). meet stringent effluent limitations have become very important.

0304-3894/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhazmat.2011.05.052
 A.R. Pendashteh et al. / Journal of Hazardous Materials 192 (2011) 568–575 569

Fig. 1. Schematic of experimental setup.

Treatment process models are essential tools to assure proper systems, UK) and recycled back to the bioreactor. The microorgan-
operation and better control of wastewater treatment plants [12]. isms separated by the membranes were returned to the bioreactor.
The ability of artificial neural networks (ANNs) in the model- Permeate flux was measured gravimetrically with an electronic
ing of complex systems that have nonlinear characteristics has balance (Tanita KD-200, Tanita Corporation, Tokyo, Japan). Table 1
made them the most popular tool for modeling of biological shows the UF characteristics and the membrane sequencing batch
processes [13]. reactor (MSBR) operating conditions.
In recent years, ANNs have been used for monitoring [14], con-
trolling [15], classification [16] and simulation [17–19] of activated 2.2. Synthetic wastewater preparation
sludge processes of wastewater treatment plants. In the literature
to date, a limited number of applications of ANNs have been made In order to determine the response of the MSBR system under
to MBRs for modeling of a plant operation [20,21]. Geissler et al. controlled conditions, synthetic wastewater was used during the
[20] used an ANN model to predict the filtration performance in whole study. Based on halophilic medium proposed by other
a submerged capillary hollow fiber membrane treating municipal researchers [7,22], produced water was simulated. The synthetic
wastewater. Cinar et al. [21] have also proposed an ANN model for produced water composition (TDS of 35,000 mg/L) in mg/L is shown
a submerged MBR treating cheese whey and evaluated its perfor- in Table 2.
mance at different sludge residence time (SRT). For TDS concentrations of 50,000, 100,000, 150,000, 200,000,
Up till now, there have been only few investigations on treating and 250,000 mg/L, NaCl was added at concentration of 46,000,
produced water by MBRs. Furthermore, no attempt has been made 96,000, 146,000, 196,000, and 246,000 mg/L, respectively. The com-
on the modeling of the produced water treatment systems. position of the wastewaters gave a C/N/P ratio of approximately
In this study, ANN was used to model the performance of 100/10/1 by adding NH4 Cl and KH2 PO4 . The pH was adjusted to
a membrane sequencing batch reactor treating hypersaline oily 7 using NaOH. All of the chemicals used in this study were of
wastewater at different organic loading rate (OLR), reaction time technical grade. Crude oil was collected from Malaysia oilfields
(RT) and total dissolved solid (TDS) in order to predict the effluent (Petronas BCOT, Sarawak). Synthetic produced water was pre-
characteristics to meet the effluent discharge standards. pared in a homogenizer (KIKA labortechnik, Staufen, Germany)

Table 1
2. Materials and methods
UF characteristics and the MSBR operating conditions.

2.1. Experimental setup Material Polyvinylidene difluoride (PVDF)

Tube diameter 1.25 cm
UF (FP200) Length 30 cm
A 5-L fermenter (Biostat-B.Braun Biotech International, Melsun- Molecular cut-off weight 200,000 Dalton
gen, Germany) was used as the SBR (Fig. 1). Dissolved oxygen (DO), (MWCO)
pH, temperature, aeration, and agitation were microprocessor con- Membrane area 0.012 m2
trolled. Aeration was provided by using an air compressor and a Agitation speed 300 rpm
sparger. The synthetic wastewater was fed by a peristaltic pump DO 3 mg/L
MSBR operating
(Peristaltic pump model: Watson-101U/R, Watson-Marlow, UK). conditions
Crossflow velocity 2 m/s
Mixed liquor was pumped through two tubular crossflow ultra- Temperature 30 ◦ C
Transmembrane pressure 2 bar
filtration (UF) membrane modules (MIC-RO 240, PCI membrane
570 A.R. Pendashteh et al. / Journal of Hazardous Materials 192 (2011) 568–575

Table 2
Chemical composition of the synthetic produced water.

Chemical NaCl CaCl2 ·2H2 Oa KCl MgCl2 ·6H2 Ob NaHCO3 NH4 Cl KH2 PO4
mg/L 31,173 60 2,000 50 800 860 99
a
Excluding the water of crystallization: 45 mg/L.
b
Excluding the water of crystallization: 23 mg/L.

by mixing salts and crude oil in a 5-L propylene container for

24 h (2400/min) to achieve equilibrium between the oil and water
phases [23]. Biochemical oxygen demand (BOD), COD and O&G of
synthetic produced water (1 mL oil/L) were 645, 2250 and 350 mg/L,
respectively.

2.3. Culture selection

Hypersaline soil from Morib onshore in Malaysia served as a

source of halophilic microorganisms. Isolation of microorganisms
capable of degrading crude oil in the hypersaline-produced water
began by placing approximately 6 g of soil into 200 mL of synthetic
produced water (1 mL oil/L and TDS of 35,000 mg/L). After 15 days of
mixing on a shaker table (150/min, 30 ◦ C), a 2 mL sample of the mix- Fig. 2. Structural organization of the neural network used for the estimation of
effluent characteristics.
ture was transferred to a fresh medium. After the three steps, the
resulting mixture was free of soil [7]. The culture was transferred
monthly to fresh medium for six months [22,24]. 2.7. Analytical methods

Since the chloride concentration was high, the COD of the sam-
2.4. Startup of the SBR
ples was determined according to the Freire and Sant’Anna [26]
method. MLSS, mixed liquor volatile suspended solids (MLVSS), and
The SBR was inoculated with the isolated microorganism’s
O&G were determined according to the standard methods [27]. The
culture. After inoculation, the bioreactor was operated with syn-
TOC was measured by a TOC analyzer (Shimadzu, Kyoto, Japan).
thetic produced water to increase the biomass concentration to
1000 mg/L. The reactor was operated at different operating condi-
tions and the temperature was kept constant at 30 ◦ C. The operation 2.8. Model development
mode was 12, 24, and 48 h cycles. The 24 h mode cycle consisted
of three stages: 1 h of feeding time, 20 h of reacting time, and 3 h A software package of NeuralPower version 2.5, CPC-X Software,
of decanting time. In the last stage, 2.5 L permeate was withdrawn USA, was applied in this study. A set of 193 operational data from
from the bioreactor. In this study, the membrane chemical cleaning the synthetic produced water treated with the MSBR was used to
(NaOCl 0.5%, w/v, and HCl 0.5%, w/v) and sonication was carried out train the network.
when flux declined to a value almost 40 L/(m2 h). After the cleaning, Multilayer normal feedforward neural network was used in
the decanting time was reduced to 1.5 h. order to predict the performance of the MSBR treating the syn-
thetic produced water. The networks were trained by different
learning algorithms (incremental back propagation, IBP; batch back
2.5. Effect of OLR, TDS and RT
propagation, BBP; Levenberg–Marquardt algorithm, LM; genetic
algorithm, GA; and quickprob, QP). The developed network con-
In this experiment, OLR was increased in a stepwise mode in
sisted of three layers including input layer that comprised four
different stages. Characteristics of the raw wastewater at differ-
nodes, which were experiment day, RT, OLR and TDS; one hid-
ent stages are presented in Table 3. The possible adverse effects
den layer consisting of several nodes, which were varied to obtain
of salt concentration on microbial activity were studied where
the best model and the output layer that had four output nodes
biological treatment of synthetic produced water was conducted
(which were TOC, COD, oil in sludge and MLSS). The structure of the
over different TDS levels (50,000, 100,000, 150,000, 200,000 and
proposed ANN used for prediction of the effluent characteristics is
250,000 mg/L). In order to study the effect of RT on the MSBR-
shown in Fig. 2.
product quality, the fermenter was started up again as in the
The transfer function determines the input–output behavior and
previous section and three cycle times of 12, 24, and 48 h (corre-
adds nonlinearity and stability to the network [28]. The transfer
sponding to hydraulic residence time (HRT) of 24 h, 48 h, 96 h) were
function of the hidden and output layers (sigmoid, hyperbolic tan-
studied.
gent function, gaussian, linear, threshold linear and bipolar linear)
was iteratively determined by developing several networks. The
2.6. Oil in mixed liquor best transfer function for the hidden layer was found to be hyper-
bolic tangent (tanh) function while the best transfer function for the
In order to measure the accumulation of undigested crude oil in output layer was a sigmoid one. Each network was trained until the
the activated sludge, extraction of hydrocarbons from bioreactor network average root mean squared error (RMSE) was minimum
mixed liquor samples were performed by shaking 5 mL sample and and coefficient of determination (R2 ) was equal to 1. Other param-
10 mL dichloromethane vigorously for 10 min. The extracts were eters for network were chosen as the default values of the software
filtered through anhydrous sodium sulfate to remove water. The (learning rate = 0.1 and momentum = 0.4).
samples were filtered through a 0.45-␮m pore size Teflon mem- The weights were initialized with random values and adjusted in
brane. Then the samples were dried and weighed on an analytical order to minimize the network error. A second set of validation data
balance [25]. was used to evaluate the quality of the network during training. In
 A.R. Pendashteh et al. / Journal of Hazardous Materials 192 (2011) 568–575 571

Table 3
Synthetic oily wastewater characteristics at different OLR, TDS and RT.

Time (day) mL oil/L Hydrocarbon (mg/L) COD (mg/L) RT OLR (kg COD/(m3 day)) O&G (mg/L) TOC (mg/L)a TDS (mg/L)

1–15 0.25 204.5 562.5 20 0.281 87.5 137 35,000

16–30 0.5 409 1,125 20 0.563 175 275 35,000
Effect of OLR 31–45 1 818 2,250 20 1.124 350 550 35,000
46–60 2 1,636 4,500 20 2.248 700 1,100 35,000
61–75 3 2,454 6,750 20 3.372 1,050 1,650 35,000

76–90 1 818 2,250 20 1.124 350 550 50,000

91–105 1 818 2,250 20 1.124 350 550 100,000
Effect of TDS 106–120 1 818 2,250 20 1.124 350 550 150,000
121–135 1 818 2,250 20 1.124 350 550 200,000
136–150 1 818 2,250 20 1.124 350 550 250,000

151–165 1 818 2,250 20 1.124 350 550 35,000

Effect of RT 176–180 1 818 2,250 8 1.124 350 550 35,000
181–195 1 818 2,250 44 1.124 350 550 35,000
a
Calculated.

addition, the performance of the trained network was estimated [31]. The difference between BBP and standard BP learning algo-
based on the accuracy of the network on the test dataset which rithm lies in timing of the weight update. The weight update of
was unseen by the developed network during training. For opti- the standard BP is performed after each single input data while for
mization of the conditions, GA was used. GA is an adaptive search the BBP, the update step with accumulated weight changes is per-
algorithm based on the principles of biological evolution, such as formed after full presentation of all training patterns [32]. In fact,
natural selection and genetic inheritance. In the GA, each solution the BBP is smoother in converging compared to that of the stan-
to a given problem is encoded as a chromosome, which evolves over dard BP and is best suited for nonlinear regression [33]. In order to
time towards a better solution. Some of the advantages of GA over minimize the total error of the network trained by BBP, the weights
the conventional optimization methods are short calculation time, are adjusted according to the following equation [34,35]:
flexibility, robustness and high convergence property [29]. The use
∂e
of GA requires the choice of a set of operational parameters such wji (n) = − × + a × wji (n − 1) (3)
∂wji
as population size, mutation rate and crossover rate. In this study,
the GA parameters were set at the default values of the software. where e is error function being minimized, wji is a generic weight
in the network, ˛ is a momentum factor,  is the learning rate or
2.9. Verification of the model step size parameter and n and n − 1 are two successive iterations.
Because gradient decent usually slows down near minima, so
The performances of the ANN models were measured by R2 the Levenberg–Marquardt (LM) method can be used to obtain faster
and RMSE between the predicted values of the network and the convergence. LM is a blend of simple gradient descent and the
experimental values, which were calculated by Eqs. (1) and (2), Gauss–Newton method. The algorithm for parameter updating is
respectively. presented by the following equation:
n ∗ (i) 2 −1 T
(yi − yp ) w = −[J T J + I] J ε (4)
R2 = 1 − i=1
n ∗ 2
(1)
i=1
(yi − ȳ) where ε = [e1 e2 . . . eP ]T is the error vector.  is a positive constant,
 I is the identity matrix and J is the Jacobian matrix given by:
 n ⎡ ⎤
 1  (i) ∂e1 ∂e1 ∂e1
RMSE = 
2
(yp − y∗ ) (2) ...
n i
⎢ ∂w1 ∂w2 ∂wN ⎥
i=1 ⎢ ⎥
⎢ ∂e2 ∂e2 ∂e2 ⎥
⎢ ... ⎥
J=⎢ ∂w ∂w2 ∂wN ⎥
(i)
where ȳ is the average of y over the n data, and yi∗ and yp are the (5)
⎢ .1 . ... . ⎥
ith target and predicted responses, respectively. ⎢ ⎥
⎢ . . ... . ⎥
⎣ ∂e ∂eP ∂eP
⎦
P
3. Results and discussion ...
∂w1 ∂w1 ∂wN
3.1. Artificial neural network modeling of the MSBR LM has found to be the fastest method for training moderate-
sized feedforward neural networks, where the training rate is 10 to
The goal of iterative neural network training is to update the 100 times faster than the usual gradient descent backpropagation
networks’ weights to minimize the difference between the net- method [36]. However, when the number of network weights is
work output and the desired output. Various feedforward neural large, the requirement for computation and memory becomes sig-
networks (FNNs) were trained using different learning algorithms nificant. Since in LM algorithm, inversion of square matrix JT J + I
for the estimation of the characteristics of the synthetic produced is involved thus a large memory space is required to store the Jaco-
water treated by the MSBR. The best algorithm was found to be BBP bian matrix and the Hessian matrix (JT J) along with inversion of
with an average R2 of 0.97339 and RMSE 89.385. The learning was approximated Hessian matrix in each iteration [34]. In this study,
completed after 10,000 iteration steps. LM showed considerable thrashing (rise and fall in error rate),
Back propagation (BP) is a commonly used algorithm that which slowed the conversion. This problem could be due to the
searches for the minimum of error function in weight space using large memory overheads. Therefore, BBP was selected as the best
the method of gradient descent [30]. Each iteration in BP involves algorithm for training the network.
two phases: forward activation with the computation of error, and Convergence rate and complexity of a model strongly depends
backward propagation of the computed error to modify the weights on the type of transfer function applied [37]. The best transfer
572 A.R. Pendashteh et al. / Journal of Hazardous Materials 192 (2011) 568–575

Fig. 3. Correlation between the experimental data and predicted values of the ANN model used for prediction of COD, TOC, MLSS, and oil in sludge.

function can be selected easily by trial and error. In this research, 3.2. Variation of effluent parameters at different stages
among all employed transfer functions for hidden layers, tanh
showed a better performance. The output layer contains sigmoid Fig. 4 shows the effluent characteristics variations at different
transfer function which produces a continuous value in the 0–1 experimental stages. In this study, the effect of OLR was investi-
range. gated at the cycle time of 24 h. During the experiment, OLR was
Finding the optimal number of neurons in hidden layer(s) is increased from 0.281 to 3.372 kg COD/(m3 day) corresponding to
crucial for processing capability of the network and accuracy of the influent COD and TOC concentrations from 562.5 to 6750 and
developed models [38]. Too few hidden neurons limit the ability of 137 to 1650 mg/L, respectively (Fig. 4a and b). At the lowest OLR
the network to model the process whereas excessive hidden neu- of 0.281 kg COD/(m3 day), minimum effluent COD and TOC concen-
rons lead to poor generalization for untrained data [31]. Although trations were observed. With an increase in the COD concentration
several approximation methods for determining the number of from 562.5 to 6750, the average COD concentration of the effluent
neurons in hidden layers are presented, the optimal number of increased from 37.9 to 184.3 mg/L. According to Fig. 4b, at the TDS
neurons is usually determined by trial and error [39]. In this study, concentration lower than l00,000 mg/L, the permeate COD concen-
the optimum number of neurons was chosen on the base of R2 and tration was less than 100 mg/L. At higher TDS, the COD increased
RMSE of the network. In order to find optimum number of neurons, and reached the maximum COD value of 238 mg/L at the highest
six different 4-x-4 architectures (x changes from 6 to 11) were used. TDS.
The optimum number of hidden neurons was found to be 9. Table 4 It was observed that MLSS concentration increased from
presents the corresponding average R2 and RMSE for the network 1560 mg/L to 7950 mg/L during 75 days and the oil concentration
trained with BBP with respect to the training data when the num- in the sludge increased from 571 mg/L to 6005 mg/L (Fig. 4c and d).
ber of neurons is varied. Correlation between the experimental data The decreasing trend of concentration of the oil trapped inside the
and the predicted values of the final trained ANN model with 9 hid- sludge flocs was observed when TDS was less than 100,000 mg/L,
den neurons is shown in Fig. 3. The scatter plots show that for all and it was increased at higher TDS. It may be concluded that salin-
the four parameters, the predictive capability was satisfactory and ity can reduce the biological degradation rate of hydrocarbons. The
the linear adjustment between actual and predicted values gives isolated microorganisms played significant role for biodegradation
almost a slope equal to 1. of hydrocarbons. In this study, bacteria of the genera Pseudomonas,
Ochrobactrum, Corynebacterium and Bulkhorderi were identified.
Identification of the isolated microorganisms was explained in
detail elsewhere [9].
Table 4
At RT of 44 h, the average UF-permeate COD and TOC concen-
Modeling error with respect to training data. trations were lower than 48 and 12 mg/L, respectively. It may be
concluded that, short contact time between microorganisms and
Model Average R2 Average RMSE
food affects biodegradation of organic matter. The results also
4-6-4 0.97024 113.450 showed that higher RT affects oil concentration in sludge. At RT of 8
4-7-4 0.97330 91.458
and 20 h, accumulation of crude oil was observed up to 2110 mg/L
4-8-4 0.97347 90.729
4-9-4 0.97339 89.385 and it was decreased to 1321 mg/L at the highest RT. This indi-
4-10-4 0.97299 93.859 cated that although the removal efficiency of the organic matters
4-11-4 0.96504 206.090 at lower RT was acceptable, the high concentration of oil trapped
 A.R. Pendashteh et al. / Journal of Hazardous Materials 192 (2011) 568–575 573

Fig. 4. Actual and predicted effluent variations: (a) TOC, (b) COD, (c) MLSS and (d) oil in sludge (predicted , actual ------ ).

inside sludge flocs might influence the treatment system at longer affected the effluent characteristics with an order of contribution
time, so a higher HRT could provide a sustained treatment system. OLR > RT > TDS. As shown in three dimensional plots obtained by
It should be mentioned that at RT of 8 h, the accumulation of ANN analysis, all the variables had significant effect on the treat-
370 mg of crude oil in the sludge was observed in comparison with ment efficiency (Fig. 5). Fig. 5a shows the effect of OLR and TDS
the total of 4090 mg of crude oil fed into the bioreactor per day. At RT and their interaction on COD of the effluent. As shown in the figure,
of 44 h, the decreasing trend proved the ability of the microorgan- minimum COD was achieved when TDS and OLR were at the lowest
isms to degrade crude oil. It was also inferred that no toxic or non- values. Similar trend was achieved for TOC (Fig. 5b).
biodegradable compounds had accumulated in the bioreactor [40]. Fig. 5c shows the effect of OLR and RT and their interaction
From Fig. 4, a good correlation can be observed between actual on the oil in sludge. Minimum accumulation of oil in sludge was
data and values predicted by the model. The network perfor- achieved when RT was 34 h. Therefore, for sustainable operating of
mance was also investigated with respect to validating, and testing the system and to prevent the accumulation of trapped hydrocar-
datasets. The R2 and RMSE for validating dataset were 0.99795 bons inside bioreactor sludge, the bioreactor RT should be longer
and 0.460326, respectively, while for the testing data set, R2 was than 34 h. In this research, it was observed that during whole exper-
0.99499 and RMSE was 0.622540. The results show that the devel- iment, O&G concentration was less than 5 mg/L, so based on EPA
oped BPP network is properly capable of learning the relationship limit [4], the treated wastewater can be discharged into the sea
between the input and output parameters and therefore could be and it also can be used to re-inject into oil-wells to enhance oil
employed in the further part of the study. recovery [41].

3.3. Effect of parameters 3.4. Optimization of the effluent COD for discharging to
environment
The importance of the parameters was obtained by summing
the absolute weights of the connections from the input neu- Based on the most stringent discharge limits, COD should be
rons to all neurons in the hidden layer. The operative variables less than 100 mg/L for discharging to sea [5]. Table 5 shows some
574 A.R. Pendashteh et al. / Journal of Hazardous Materials 192 (2011) 568–575

Table 5
Some optimum conditions for obtaining a COD less than 100 mg/L.

Input variable Output

OLR (kg COD/(m3 day)) TDS (mg/L) RT (h) Predicted COD (mg/L) Actual COD (mg/L)

2.44 78,000 40.4 100.2 95.3

0.9 164,000 8 98.2 92.4
0.281 56,500 11.6 42.2 48.5

optimum conditions for obtaining a COD less than 100 mg/L. It can is low and initial OLR is high. In addition, when initial OLR is low
be seen that the system is able to treat the wastewater at high OLR and initial TDS is more than 100,000 mg/L, the raw wastewater can
of 2.44 (kg COD/(m3 day) and TDS of 78,000 mg/L during 40.4 h. At be diluted to reduce inhibition effect of NaCl.
higher TDS, lower concentration of organic matter should be fed to
the bioreactor. In this part, a good correlation between the actual
4. Conclusions
and predicted values with a R2 value of 0.9822 was obtained.
Generally, based on the optimization results, in order to con-
A membrane sequencing batch reactor inoculated with isolated
trol the effluent characteristics to meet discharge standard limits,
halophilic microorganisms was used for the treatment of hyper-
influent wastewater can be pretreated chemically when initial TDS
saline oily wastewater. The training of an artificial neural network
with operational data from the MSBR has been successful. The
results of this study show that ANN-GA can easily be applied to
evaluate the performance of a membrane bioreactor even though it
involves the highly complex physical and biochemical mechanisms
associated with the membrane and the microorganisms.

Acknowledgement

The authors are grateful to the Universiti Putra Malaysia (UPM)

for supporting this work under the research grant RUGS No.
05/01/07/0158RU.

References

[1] M.P. Diaz, K.G. Boyd, S.J.W. Grigson, J.G. Burgess, Biodegradation of crude oil
across a wide range of salinities by an extremely halotolerant bacterial consor-
tium MPD-M, immobilized onto polypropylene fibers, Biotechnol. Bioeng. 79
(2002) 145–153.
[2] J.M. Neff, Bioaccumulation in Marine Organisms: Effects of Contaminants from
Oil Well Produced Water, Elsevier Science Publishers, Amsterdam, 2002.
[3] P.J.C. Tibbetts, I.T. Buchanan, L.J. Gawel, R. Large, A comprehensive determi-
nation of produced water composition, in: J.P. Ray, F.R. Engelhardt (Eds.),
Technological/Environmental Issues and Solutions, Plenum Publishing Corp,
New York, 1992, pp. 97–113.
[4] USEPA, http://www.epa.gov.
[5] G.T. Tellez, N. Nirmalakhandan, J.L. Gardea-Torresdey, Performance evalua-
tion of an activated sludge system for removing petroleum hydrocarbons from
oilfield produced water, Adv. Environ. Res. 6 (2002) 455–470.
[6] A. Fakhru’l-Razi, A.R. Pendashteh, A. Luqman Chuah, A.B. Dayang Radiah, S.S.
Madaeni, Z.A. Zurina, Review of technologies for oil and gas produced water
treatment, J. Hazard. Mater. 170 (2009) 530–551.
[7] C.R. Woolard, R.L. Irvine, Treatment of hypersaline wastewater in the sequenc-
ing batch reactor, Water Res. 29 (1995) 1159–1168.
[8] S. Venkata Mohan, N. Chandrashekara Rao, K. Krishna Prasad, B.T.V. Mad-
havi, P.N. Sharma, Treatment of complex chemical wastewater in a sequencing
batch reactor (SBR) with an aerobic suspended growth configuration, Process
Biochem. 40 (2005) 1501–1508.
[9] A.R. Pendashteh, A. Fakhru’l-Razi, A. Luqman Chuah, A.B. Dayang Radiah, S.S.
Madaeni, Z.A. Zainal, Biological treatment of produced water in a sequenc-
ing batch reactor by isolated halophilic microorganism consortium, J. Environ.
Technol. 31 (2010) 1229–1239.
[10] I.J. Kang, C.H. Lee, K.J. Kim, Characteristics of microfiltration membranes in a
membrane coupled sequencing batch reactor system, Water Res. 37 (2003)
1192–1197.
[11] P. Le-Clech, V. Chen, T.A.G. Fane, Fouling in membrane bioreactors used in
wastewater treatment, J. Membr. Sci. 284 (2006) 17–53.
[12] D. Guclu, S. Dursun, Artificial neural network modelling of a large-scale
wastewater treatment plant operation, Bioprocess Biosyst. Eng. 33 (2010)
1051–1058.
[13] J.S. Almeida, Predictive non-linear modeling of complex data by artificial neural
networks, Curr. Opin. Biotechnol. 13 (2002) 72–76.
[14] S.H. Hong, M.W. Lee, D.S. Lee, J.M. Park, Monitoring of sequencing batch reactor
for nitrogen and phosphorus removal using neural networks, Biochem. Eng. J.
35 (2007) 365–370.
[15] G.M. Zeng, X.S. Qin, L. He, G.H. Huang, H.L. Liu, Y.P. Lin, A neural network pre-
Fig. 5. Three dimensional plot showing the effect of individual parameters and their dictive control system for paper mill wastewater treatment, Eng. Appl. Artif.
mutual interactions on the characteristics of oil in sludge, COD and TOC. Intell. 16 (2003) 121–129.
 A.R. Pendashteh et al. / Journal of Hazardous Materials 192 (2011) 568–575 575

[16] O. Cinar, New tool for evaluation of performance of wastewater treatment [28] B.G. Liptak, Instrument Engineers’ Handbook: Process Control and Optimiza-
plant: artificial neural network, Process Biochem. 408 (2005) 2980–2984. tion, CRC Press, FL, 2006.
[17] J.C. Chen, N.B. Chang, W.K. Shieh, Assessing wastewater reclamation potential [29] E. Franco-Lara, H. Link, D. Weuster-Botz, Evaluation of artificial neural net-
by neural network model, Eng. Appl. Artif. Intell. 16 (2003) 149–157. works for modelling and optimization of medium composition with a genetic
[18] B. Raduly, K.V. Gernaey, A.G. Capodaglio, P.S. Mikkelsen, M. Henze, Artificial algorithm, Process Biochem. 41 (2006) 2200–2206.
neural-networks for rapid WWTP performance evaluation: methodology and [30] R. Rojas, Neural Networks: A Systematic Introduction, Springer-Verlag Berlin
case study, Environ. Modell. Softw. 22 (2007) 1208–1216. Heidelberg, New York, 1996.
[19] D. Hanbay, I. Turkoglu, Y. Demir, Prediction of chemical oxygen demand [31] I.A. Basheer, M. Hajmeer, Artificial neural networks: fundamentals, computing,
(COD) based on wavelet decomposition and neural networks, Clean 35 (2007) design, and application, J. Microbiol. Methods 43 (2000) 3–31.
250–254. [32] J. Mira, Artificial Neural Nets Problem Solving Methods, Springer-Verlag Berlin
[20] S. Geissler, T. Wintgens, T. Melin, K. Vossenkaul, C. Kullmann, Modelling Heidelberg, New York, 2003.
approaches for filtration processes with novel submerged capillary modules [33] T.W.S. Chow, S.Y. Cho, Neural Networks and Computing: Learning Algorithms
in membrane bioreactors for wastewater treatment, Desalination 178 (2005) and Applications, Imperial College Press, Singapore, 2007.
125–134. [34] N. Ye, The Handbook of Data Mining, Lawrence Erlbaum Associates, Inc., Pub-
[21] O. Cinar, H. Hasar, C. Kinaci, Modeling of submerged membrane bioreactor lishers, Mahwah, New Jersey, USA, 2003.
treating cheese whey wastewater by artificial neural network, J. Biotechnol. [35] Z. Zainuddin, N. Mahat, Y. Abu Hassan, Improving the convergence of the back-
123 (2006) 204–209. propagation algorithm using local adaptive techniques, Int. J. Comput. Intell. 1
[22] B. Peyton, T. Wilsona, D. Yonge, Kinetics of phenol biodegradation in high salt (2002) 79–82.
solutions, Water Res. 36 (2002) 4811–4820. [36] E. Al-Daoud, A comparison between three neural network models for classifi-
[23] B. Tansel, J. Regula, R. Shalewitz, Treatment of fuel oil and crude oil contami- cation problems, J. Artif. Intell. 2 (2009) 56–64.
nated waters by ultrafiltration membranes, Desalination 102 (1995) 301–311. [37] W. Duch, N. Jankowski, Survey of neural transfer functions, Neural Comput.
[24] G. Thouand, P. Bauda, J. Oudot, G. Kirsch, C. Sutton, J.F. Vidalie, Laboratory evalu- Surv. 2 (1999) 163–212.
ation of crude oil biodegradation with commercial or natural microbial inocula, [38] J. Kamruzzaman, R. Begg, R.A. Sarker, Artificial Neural Networks in Finance and
Can. J. Microbiol. 45 (1999) 106–115. Manufacturing, Idea Group Publishing, PA, 2006.
[25] J.C. Bertrand, M. Bianchi, M.A. Mallah, M. Acquaviva, G. Mille, Hydrocarbon [39] M. Islam, X. Yao, K. Murase, A constructive algorithm for training coop-
biodegradation and hydrocarbonoclastic bacterial communities composition erative neural network ensembles, IEEE Trans. Neural Netw. 14 (2003)
grown in seawater as a function of sodium chloride concentration, J. Esp. Mur. 820–834.
Biol. Ecol. 168 (1993) 125–138. [40] W. Scholz, W. Fuchs, Treatment of oil contaminated wastewater in a membrane
[26] D.D.C. Freire, G.L. Sant’Anna Jr., A proposed method modification for the deter- bioreactor, Water. Res. 34 (2000) 3621–3629.
mination of COD in saline waters, Environ. Technol. 19 (2002) 243–247. [41] J. Veil, M.G. Puder, D. Elcock, R.J.J. Redweik, A white paper describing pro-
[27] APHA, AWWA, WEF, Standard Methods for Examination of Water and Wastew- duced water from production of crude oil, natural gas, and coal bed methane,
ater, 21st ed., American Public Health Association, Washington, DC, 2005. www.netl.doe.gov/publications/oil pubs/prodwaterpaper.pdf, 2004.