Desalination 367 (2015) 240–247

Contents lists available at ScienceDirect

Desalination

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

Artificial neural network approach for predicting reverse osmosis

desalination plants performance in the Gaza Strip
Adnan M. Aish a, Hossam A. Zaqoot b,⁎, Samaher M. Abdeljawad c
a
Institute of Water and Environment, Department of Geology, Al Azhar University, P.O. Box 1277, Gaza Strip, Palestine
b
Environment Quality Authority (Palestinian Authority), Gaza Strip, Palestine
c
Ministry of National Economy (Palestinian Authority), Gaza Strip, Palestine

H I G H L I G H T S

• Small and large scale desalination plants' performance evaluation

• Prediction of desalination plants' performance using artificial neural network technology.
• Recommending possible enhancements for desalination plants' performance.

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

Article history: A rapidly growing technique for producing new water is desalination of seawater and brackish water. In the Gaza
Received 21 November 2014 Strip the maximum amount of the drinking water is produced through small private desalination facilities. The
Received in revised form 7 April 2015 present paper is concerned with using artificial neural network (ANN) technique to forecast reverse osmosis de-
Accepted 10 April 2015
salination plant's performance in the Gaza Strip through predicting the next week values of total dissolved solids
Available online 19 April 2015
(TDS) and permeate flowrate of the product water. Multilayer perceptron (MLP) and radial basis function (RBF)
Keywords:
neural networks were trained and developed with reference to feed water parameters including: pressure, pH
Artificial neural network and conductivity to predict permeate flowrate next week values. MLP and RBF neural networks were used for
MLP predicting the next week TDS concentrations. Both networks are trained and developed with reference to prod-
RBF uct water quality variables including: water temperature, pH, conductivity and pressure. The prediction results
Prediction showed that both types of neural networks are highly satisfactory for predicting TDS level in the product
Permeate flowrate water quality and satisfactory for predicting permeate flowrate. Results of both developed networks were com-
TDS pared with the statistical model and found that ANN predictions are better than the conventional methods.
Gaza
© 2015 Elsevier B.V. All rights reserved.

1. Introduction of salts and biological materials from seawater or brackish water.

Gaza Strip suffering from depletion of accessible fresh water because
of the groundwater over pumping which it is the only resource being
used for domestic purposes. The problem came to be more serious
with time as a result of seawater intrusion into the coastal aquifer, in ad-
dition to the infiltration of somewhat treated wastewater to the aquifer
[1,2]. Water is essential for socioeconomic growth and environmental
sustainability. The Gaza Strip is mostly in a catastrophic situation that
needs urgent and serious actions to improve the water status on condi-
tions of both quality and quantity. The desalination of brackish and sea-
water is at present a global motivation that has concerned overall
governmental and public awareness not only in arid areas but also in
other regions in the world. Desalination provides a mean of upgrading
brackish water or seawater to produce potable water by the removal
⁎ Corresponding author.
E-mail address: hanreen2@yahoo.com (H.A. Zaqoot).
There are a number of desalination techniques that are commercially
used such as vapor compression distillation, electro-dialysis, multi-
stage flash distillation; and reverse osmosis [3].
The groundwater in the Gaza coastal aquifer is approximately brack-
ish excluding some fresh water in the appearance of shallow lenses.
Thus, the quantity of fresh groundwater is almost negligible and exists
only in some areas in the Gaza Strip for example Beit Lahia. Desalination
of brackish and seawater in the Gaza Strip appears to be promising,
mainly in the absence of any other options. Though, utilizing desalina-
tion method as an alternative water supply denotes many challenges
such as energy cost and environmental characteristics [4]. On one
hand, depending on desalination as a source of water supply can solve
the increasing issue of water shortage in the vicinity and overcome
the worsening problem of water quality. However, the securing of pota-
ble water for drinking purpose to the community in the Gaza Strip is be-
coming an important goal to be implemented by the Palestinian Water
Authority (PWA) [5]. The Palestinian Water Authority identifies water

http://dx.doi.org/10.1016/j.desal.2015.04.008
0011-9164/© 2015 Elsevier B.V. All rights reserved.
 A.M. Aish et al. / Desalination 367 (2015) 240–247 241

desalination as the standard approach to solve the water shortage prob- topographical area is described as flat area gradually sloping with a
lem and provide people in the Gaza Strip with acceptable water quality range from 0 to 5%, westward toward the sea allowing for surface runoff.
for drinking and other purposes [6]. The only technology applied in Gaza The landscape is essentially a foreshore plain. A sandy beach stretches
for brackish water desalination purpose is Reverse Osmosis (RO). The along the coast, bound in the east by a bridge of sand dunes up to
first RO plant in the Gaza Strip was built in 1991 with a capacity of 40 m high. Land surface elevations range from mean sea level (MSL)
45 m3/h and recovery of 75% in Deir Al Balah town of the Middle area to about 110 MSL in the eastern parts. Gaza's water resources are essen-
governorate by EMS a subsidiary of Mekorot company. In the last two tially limited to that part of the coastal aquifer that underlies its area
decades seven RO desalination plants were built including: six brackish [19].
water desalination plants and one seawater desalination plant [7]. Five large and small scale brackish water desalination plants in the
Today, there are about 118 small private RO desalination units Gaza Strip were selected for this study. These plants were selected to de-
established and operated all over the Gaza Strip, and almost 30 units velop ANN models for predicting the performance of RO desalination
are licensed by PWA. plants in the Gaza Strip. The water quality data were generated from
Artificial neural networks (ANNs) are flexible tools which are being the selected five plants from southern, middle area, Gaza city and north-
used progressively to predict and forecast water resources variables. ern area. They are named Al-Salam plant (Rafah), Al-Sharqia plant
The neural network models must be developed according to the existing (Khanyounis), Al-Balad plant (Dier Al Balah), Hanneaf plant (Gaza)
data and information about water quality parameters for several years and Al-Radwan plant (Beit-Lahia). The selected desalination plants are
of a particular area [8]. It is assumed that nonlinear approaches such shown in Fig. 1.
as ANNs are adequate for predicting water quality parameters in the un-
derground and fresh water bodies. Earlier studies showed that ANNs
were successfully used to predict water quality parameters such as 2.2. Collection, processing and analysis of water samples
TDS, conductivity and flowrate in underground water, rivers and desali-
nation plants. The ANNs were able to capture nonlinear relationships A 500-ml and one liter polyethylene bottles were used for collecting
among input and output variables for modeling the performance of de- water samples. The water samples were collected once every week for a
salination plants [9–13]. period of six months from March to September 2013. The selected pa-
Cordoba [14], developed ANN models to evaluate and predict some rameters for this study include: temperature, pressure, flowrate, TDS,
of drinking water quality parameters within a water distribution sys- pH and electrical conductivity (EC). The samples were collected from
tem. [15] have successfully developed an artificial neural network the feeding wells and desalination plants' product water. Samples
model for predicting the two important parameters of RO desalination were analyzed in laboratories of Palestinian Ministry of National Econo-
plants including: salt rejection and permeate flowrate (flux). [16] used my and Water and Environment institute at Al-Azhar University-Gaza,
ANN technology to develop a model for predicting the performance of Palestine. Electrical conductivity and pH were measured directly in the
a reverse osmosis plant. The artificial neural network was fed with field using a portable instrument called Electrochemistry made by
three inputs including: feed pressure, temperature and salt concentra- CIBA-CORNING. The pressure was measured in (bars) with high preci-
tion to predict the water permeate flowrate. The developed network sion and accurate pressure gages, while temperature was measured in
learned the input–output data mapping with accuracy. [17] applied ar- degrees Celsius (°C) using a digital thermometer. The total dissolved
tificial neural network method for the prediction of RO desalination solids were measured by using the Oven method. The permeate rotame-
plants' performance. The permeate flowrate and salt rejection at differ- ter was used to measure the flowrate of permeate water (m3/h).
ent conditions of the process were predicted using ANN based on the
experimental water quality data. The model results were tested and
compared with the observed data and then error percentage was
calculated.
The main objective of this study is to predict the performance of RO
desalination plants in the Gaza Strip using ANN technology. For this pur-
pose the next one week total dissolved solids (TDS) content and the per-
meate flowrate are predicted using multilayer perceptron (MLP) and
radial basis function (RBF) neural networks. The conventional method
is used to compare the predicted results obtained from both developed
neural networks. Due to the tough and crucial circumstances in the Gaza
Strip, it was extremely challenging to generate and collect more exper-
imental data. However, this study used limited data sets for the devel-
opment of ANN models for predicting RO desalination plants'
performance. As a first case study for predicting the performance of de-
salination plants in the Gaza Strip, it is expected that the developed ap-
proach may prove helpful in improving plants' performance and
potentially increase their water production. In addition it may be used
as a new tool to improve current drinking water quality management
practices in Gaza.

2. Materials and methods

2.1. Study area

The Gaza Strip is a narrow strip of land on the eastern coast of the
Mediterranean sea, situated in the middle east at latitudes [31°16″ and
31°45″N] and longitudes [34°20″ and 34°25″E] bordered by the Medi-
terranean sea in the west and the Negev desert and Egyptian Sinai head-
land in the south with a total area of 365 km2 [18]. Most of the Gaza Strip Fig. 1. Map shows the selected desalination plants.
242 A.M. Aish et al. / Desalination 367 (2015) 240–247

Table 1 85
Summary of data statistical analysis used for TDS predictive model development.

Recovery Rate (%)

Parameters Data set Range Mean StDev
75
Water temp (°C) Training 21.8–26.7 24.32 1.04
Testing 22.8–25.90 24.49 0.95
pH Training 5.12–6.87 05.89 0.43 65
Testing 5.21–6.54 05.89 0.38
EC (μs/cm) Training 16.07–643 235.51 206.18
Testing 18.96–548 171.05 188.27
55
TDS (mg/l) Training 10–430 146.01 127.81
Testing 11.80–340 106.04 91.91
Pressure (bars) Training 9.5–17 12.80 2.04
Testing 10–15.5 13.36 1.91 45
1 2 3 4 5
Desalination Plants
2.3. Statistical analysis
Fig. 2. Recovery rate of all plants.
The water quality data is generated and being used to develop ANN
models to predict the water quality for understanding desalination
plants' performance in the Gaza Strip. The generated data were entered possible. There are a number of concerns complicated in setting-up
as Microsoft Excel sheets, uploaded to Statistical Package for Social and training a multilayer perceptron network. These concerns include:
(SPSS) and analyzed using minimum, maximum, mean, standard devi- selecting how many hidden layers to use in the network, deciding
ation (StDev) statistics. In addition the Pearson correlation coefficient how many neurons to use in each hidden layer, finding a comprehen-
(a measure of linear association) is used to study the relationship sively optimal solution that avoids local minima, converging to an opti-
among the selected parameters in both training and testing data sets. mal solution in a reasonable period of time and validating the neural
Training and testing of the developed ANN models were carried out network for testing overfitting in case it occurred during the training
using neural network toolbox in the language of technical computing process [21]. Training of multilayered perceptron network is normally
(MATLAB). Two types of feedforward networks are used. They are mul- performed in a supervised manner. One assumes that a training set is
tilayer perceptron and radial basis function neural networks. In addition available, which contains both input patterns and the corresponding
multiple linear regression (MLR) model is used to compare the predic- wanted output patterns (also called target patterns). The training is nor-
tion results obtained from both developed MLP and RBF neural mally based on the minimization of some error extent between the
networks. network's outputs and the wanted outputs. It involves a backward prop-
agation through a network similar to the one being trained. For this rea-
son the training algorithm is normally called backpropagation [22]. In
2.4. Artificial neural networks (ANNs) general, MLP nueral network model has performed well in a number
of hydrologic, water resource management and desalination plants' per-
The artificial neural network is made with a methodical step-by-step formance applications, such as [9,11–13,23–32].
method to enhance a performance norm or to monitor some indirect in- Radial basis function neural network is a type of feedforward neural
ternal limitation, which is generally signified as the learning imperative network that learns by using a supervised training technique.
or process. The learning process comprises updating network architec- Broomhead and Lowe were the first researchers to use radial basis func-
ture and connection weights so that a network can proficiently achieve tions in the design of neural networks. One specific feature of RBF is that
a specific recognition or classification task. In artificial neural networks, response decreases or increases (monotonically) with distance from a
the designer selects network topology, performance function, learning center point. Moreover it has been perceived that RBF networks are
rule and training algorithms, and criterion to stop the training step, able to approximate any practical continuous function mapping with a
but the system certainly adjusts the parameters [20]. Many ANN archi- satisfactory degree of accuracy [33]. The RBF network, which has three
tectures are available, but multilayer networks are commonly used for layers, can appear as a special class of multilayer feedforward networks.
forecasting. An ANN adapts to learn the relationship or mapping be- Each neuron in the hidden layer employs a radial basis function, such as
tween input and outputs during the training process [9]. the Gaussian Kernel, as an activation function. The output neurons im-
Multilayered perceptron network is the most popular class of multi- plement a weighted sum of hidden neuron outputs. The network is cen-
layer feedforward networks. The MLP is divided into three layers: input tered at the point specified by the weight vector associated with the
layer, hidden layer and output layer, where each layer in this order gives unit. Both the positions and the widths of these functions are learnt
the input to the next. The extra layers give the structure needed to rec-
ognize non-linearly separable classes. The aim of the training process in
98
MLP is to find the set of weight values that will mark the output from the
96
TDS Removal (%)

neural network to contest with the real target values as accurately as

94
92
Table 2 90
Summary of data statistical analysis used for flowrate predictive model development.
88
Parameters Data set Range Mean StDev 86
pH Training 6.96–7.84 7.29 0.277 84
Testing 7–7.83 7.32 0.285 82
EC (μs/cm) Training 861–6900 3912.74 1757.63 80
Testing 873–6480 3708.97 1836.34
Pressure (bars) Training 11–19.5 14.70 1.98
1 2 3 4 5
Testing 11.5–19.5 15.33 1.89 Desalination Plants
Flowrate (m3/h) Training 5.4–60 26.55 17.89
Testing 8.1–60 27.51 19.85
Fig. 3. TDS removal of all plants.
 A.M. Aish et al. / Desalination 367 (2015) 240–247
Fig. 4. MLP training performance for predicting TDS concentration.
from training patterns. Each output unit implements a linear combina-
tion of these radial basis functions. The method for training radial
basis function networks can be made in two steps. The first step in-
cludes the determination of a proper set of centers and widths. The sec-
ond step involves with the determination of the connection weights
from the hidden layer to the output layer [21]. Certainly, selection of
RBF network centers is the most critical issue in designing the network.
These should be placed according to the demands of the system to be
modeled (data to be predicted). A number of different methods have
been proposed for the selection of appropriate RBF centers [34].
2.5. Data processing and training
At the initial stage of the prediction of desalination plants' perfor-
mance, real water quality data of the selected study area in the Gaza
Strip, over a period of six months beginning from March to September
(2013) was generated. A total of five sampling locations in the Gaza
Strip are considered. These made up a total of 120 readings that are
combined in one data set. The main selected water quality parameters
including: water temperature, pressure, flowrate, pH, EC and TDS. Be-
cause the input and output variables have very different orders of mag-
nitude it is recommended to rescale the data. In this way, more reliable
predictions can be made. The normalization of data is usually done with
{0,1} [35]. Though, during this study the variables are rescaled to be in-
cluded within the interval {0,1} which could cover all variations of the
data sets used for the development of ANN prediction models.
Fig. 5. RBF training performance for predicting TDS concentration.
243
In this study, training and testing of developed ANN models were
carried out using neural network toolbox in the MATLAB. To develop
MLP network several algorithms are used during the training session in-
cluding: Resilient backpropagation, Levenberg Marquardt, Variable
learning rate backpropagation, Broyden Fletcher Goldfarb Shanno
(BFGS) Qusai-Newton, Bayesian rule and Gradient descent. The MLP
network had achieved very good results when trained using the
backpropagation incorporated with Levenberg Marquardt (LM) algo-
rithm. The tangent hyperbolic function is used as the activation function
in the hidden layer neurons. The linear activation function is used in the
output layer neurons [21]. The RBF network is trained using the
backpropagation incorporated with the orthogonal least squares (OLS)
algorithm and the Gaussian radial basis function is used as the activation
function in the hidden layer. The linear activation function is used in the
output layer [34]. Before running ANN models the data set was divided
into two data sets 70% of the data used for training purpose and 30% of
the data used for testing the networks performance and then data was
normalized to be included within the interval {0,1} [35]. The MLP net-
work training procedure started with utilizing two neurons in the hid-
den layer, then gradually the number of neurons increased and at 6
neurons the performance of the developed network was very good.
The architecture of the developed MLP neural network for predicting
TDS contains three layers, four neurons in the input, six neurons in the
hidden layer and one neuron in the output layer. The input neurons
made from four important parameters including: water temperature,
pressure, pH and conductivity. The structure of MLP neural network to
predict permeate flowrate consists of three layers, three neurons in
244 A.M. Aish et al. / Desalination 367 (2015) 240–247

Table 3 3. Results and discussion

Summary of developed ANN models and MLR results for predicting TDS.

Models Data set MSE R StDev MAE S.E. Error Range (mg/l) 3.1. Statistical analysis of data set
MLP Training 0.0227 1 127.01 0.0994 13.70 0.0010–0.5285
Testing 0.0233 1 91.92 0.1028 16 0.0026–0.5079 The statistical analysis of training and testing data sets showed that
RBF Training 0.0072 1 127.81 0.0585 13.70 0.0008–0.2467 TDS values are ranged from 10 mg/l to 430 mg/l for training data set and
Testing 0.0810 1 91.89 0.1761 15.99 0.0057–0.9346 from 11.8 mg/l to 340 mg/l for testing data set. The mean value among
MLR Training 0.0260 1 127.81 0.1153 13.70 0.0051–0.5860
all desalination plants for both training and testing data sets is
Testing 1.2810 0.9999 91.88 1.0885 15.99 0.0090–1.5871
146.01 mg/l and 106.04 mg/l respectively (see Table 1). The statistical
analysis showed a positive very strong correlation between product
the input, six neurons in the hidden layer and one neuron in the output TDS and EC for both training and testing data sets whereas r values
layer. The input neurons made from three important parameters in- are found to be 0.99 for both of them. A negative strong correlation is
clude: feed pressure, feed pH and feed conductivity. For training RBF found to be between TDS and feed pressure for both training and testing
neural network same input neurons was used as utilized for MLP net- data sets whereas the r values are found to be −0.77 and −0.71 respec-
work. During training process the RBF neural network performed tively. A positive moderate correlation is found to be between TDS and
good at 87 neurons in the hidden layer. According to Rounds [36] linear pH for both training and testing data sets whereas the r values are
regression can be regarded as a different example of ANN model, which 0.59 and 0.42 respectively. The correlation between TDS and tempera-
uses linear transfer functions and no hidden layers. If the linear model ture is found to be positive and moderate for both training and testing
accomplishes as well as other complex ANN, then using the nonlinear data sets whereas the r values are 0.34 and 0.56 respectively.
neural networks may not be reasonable and so the linear models are The statistical analysis of training and testing data sets of permeate
suitable as a basis for comparison. Though, in this work the linear re- flowrate values are found to be ranged between 5.4 and 60 m3/h for
gression is used to compare the prediction results achieved from both the training data set and ranged between 8.1 and 60 m3/h for the testing
developed MLP and RBF neural networks. Generally, there are three sta- data set. The mean value among all desalination plants for both training
tistical criteria used to evaluate the performance of ANN and MLR and testing data sets is found to be 26.55 m3/h and 27.51 m3/h respec-
methods. These criteria were coefficient of correlation (r) mean square tively (see Table 2). The statistical analysis reveals a positive strong cor-
error (MSE) and mean absolute error (MAE). Correlation of coefficient relation between permeate flowrate and feed EC for both training and
(r) is a common criterion for goodness of fit for regression models testing data sets whereas r values are found to be 0.80 and 0.78 respec-
[37]. In this study three different prediction reliability measures of the tively. A negative correlation is found to be between permeate flowrate
mean square error, the mean absolute error and correlation coefficient and feed pressure for both training and testing data sets whereas the r
were used to evaluate models results and to compare ANN and MLR: values are found to be −0.36 and −0.42 respectively. A positive corre-
lation is found to be between flowrate and pH for both training and test-
1 Xn 2
ing data sets whereas the r values are 0.36 and 0.59 respectively.
mse ¼ ðt −ak Þ
k¼1 k
ð1Þ Reliable evaluation of all selected desalination plants was made
n
based on recovery rate and TDs removal. As regards the recovery rate
Al Sharqia plant-Khanyounis(2) had the highest recovery rate of
where tk and ak are the observed and predicted values.
(78%), while Al Radwan plant-North Gaza (5) had the lowest recovery
(52%) as shown in Fig. 2. As shown in Fig. 3 the highest and lowest
Xn TDS removal was for Al Salam plant(1)-Rafah (97%) and Al Sharqia
jobserved−predictedj plant (85%) respectively. Meaning that the best performing plant is Al
MAE ¼ i¼1
ð2Þ
n Salam plant.

X X X 3.2. TDS predictive model

n xi yi − xi yi
r xy ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi ð3Þ
X 2 X 2 X 2 X 2
n xi − xi n yi − yi To predict the future values (one week ahead) of TDS concentrations
in the desalination plants of the Gaza Strip, feedforward MLP and RBF
neural networks are employed. The MLP and RBF neural network
where xi and yi are the ith observed and estimated values, respectively prediction results were compared with the multiple linear regression
and n is the total number of data. (MLR) model. The trained MLP and RBF networks performance is

TDS predicted values-MLR Model Testing Data

TDS predicted values-RBF Model Testing Data
TDS predicted values-MLP Model Testing Data
TDS observed values
400
350
TDS values

300
250
200
150
100
50
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33

Fig. 6. Comparison between ANN and MLR prediction results.

 A.M. Aish et al. / Desalination 367 (2015) 240–247 245

Fig. 7. MLP training performance for predicting permeate flowrate.

presented in Figs. 4 and 5 respectively. There are many statistical tools
for model validation, but the primary tools for most process modeling
applications include mean square error, correlation coefficient (r),
mean absolute error, standard error (SE), standard deviation and error
range (mg/l). Summary of these statistical tools used in the developed
models result for training and testing data sets as well as multiple re-
gression model are given in Table 3. The results obtained from MLP
used several different algorithms revealed that the created MLP net-
work which trained with backpropagation incorporated with LM algo-
rithm is the most fitting model for predicting product TDS in the
desalination plants of Gaza Strip. It can be seen from figures that the cre-
ated RBF network performance is mostly similar to the MLP network. It
is also understood from the results presented in Table 3 which shows
the coefficient correlations between the observed and predicted values
of TDS using MLP, RBF and MLR for training and testing the developed
models.
The correlations between the predicted TDS and actual values for
MLP, RBF model training and testing are found to be very strong and
slightly better than MLR model whereas coefficient correlation
values are [1–1], [1–1] and [1–0.9999] respectively. The MLP and
RBF networks and MLR performances have been tested with differ-
ent data sets and the obtained results show very good performance.
The results obtained prove that the developed RBF, MLP (neural
network models) and MLR have high capability and accuracy in
predicting TDS concentrations in the water quality of Gaza Strip de-
salination plants as shown in Table 3.
Fig. 6 shows comparisons of MLP and RBF model prediction results
with the conventional method predictions. From the figure it can be un-
derstood that the performances of ANN models and conventional
models for predicting TDS concentrations are almost the same. This
good prediction result obtained from the ANN and MLR models is due
to the strong correlation between the selected input and output data.
The model gave close approximations of the actual observations, sug-
gesting that these approaches are applicable for modeling the TDS
dataset. As a first case study for desalination plants TDS content predic-
tion, the simulation results prove that the artificial neural networks and
multiple linear regression models are suitable and robust for predicting
the performance of the desalination plants in the Gaza Strip.
3.3. Permeate flowrate predictive model
In this section the MLP and RBF neural networks are employed to
predict the future values (one week ahead) of permeate flowrate in
the desalination plants of Gaza Strip for the purpose of performance as-
sessment. The neural network prediction results were compared with
MLR model. The trained MLP and RBF networks performance is present-
ed in Figs. 7 & 8 respectively. The results of developed models during
training and testing data sets as well as multiple regression model are
given in Table 4. The results obtained from MLP using six different
algorithms showed that the developed MLP network trained with
backpropagation incorporated with LM algorithm is the most appropri-
ate model for predicting permeate flowrate in the desalination plants of

Fig. 8. RBF training performance for predicting permeate flowrate.

246 A.M. Aish et al. / Desalination 367 (2015) 240–247

Table 4 Table 5
Summary of developed ANN models and MLR results for predicting PFR. The predictive ANN models verification prediction results.

Models Data set MSE R StDev MAE S.E. Error Range (m3/h) ANN model Observed Predicted ANN Model Observed Predicted

MLP Training 3.7004 0.9941 17.78 1.3644 1.90 0.001–6.53 (m3/h) (mg/l)
Testing 9.5218 0.9904 18.60 1.9145 3.23 0.061–8.91
Permeate Flowrate 20.0000 23.4028 TDS 145.700 134.452
RBF Training 07.9713 0.9873 17.66 1.9170 1.89 0.017–8.85
08.4000 09.2535 11.7430 10.6547
Testing 12.6450 0.9853 18.59 2.4382 3.23 0.026–9.81
08.3000 09.2712 11.4510 10.4897
MLR Training 60.8297 0.9081 15.99 6.1412 1.71 0.08–23.05
50.0000 49.9817 64.7000 59.8663
Testing 82.6492 0.8976 18.03 6.6217 3.13 0.04–22.68

Gaza Strip. The performance of the developed ANN models obtained
after training the data sets was tested using an unknown data set.
MSE, MAE, correlation coefficient, standard error and standard devia-
tion tools are used for the models validation.
It can be seen from Figs. 7 & 8 that the MLP network performance is
slightly better than the RBF network. Also this result illustrated in
Table 4 shows the coefficient correlations between the predicted values
of permeate flowrate using MLP, RBF and MLR for testing the developed
model. The correlations between the predicted flowrate and actual
values for MLP, RBF model testing is found to be strong and better
than the MLR model whereas coefficient correlation values are 0.9904,
0.9853 and 0.8976 respectively. The results of the developed MLP and
RBF neural network models indicate permeate flowrate can be predicted
by considering other water quality parameters. In other words, artificial
neural network through exploring the relationship between the input pa-
rameters is able to predict the parameter permeate flowrate. Since the
comparison of ANN prediction results with measured values of flowrate,
artificial neural network is showing good accuracy. Therefore, it can be
stated that the performance of an artificial neural network is suitable for
predicting output parameter. Ultimately, by considering the MLP and
RBF ANN model structures and the achieved prediction results presented
in Table 4 it can be said that both MLP and RBF models are more accurate
than the MLR model for predicting permeate flowrate in the desalination
plants of Gaza Strip. Therefore, the methodology used for the develop-
ment of MLP and RBF predictive models may be extended to other
water quality parameters belonging to underground water.
In Fig. 9, it is clear from the comparison between ANN model perfor-
mance with the statistical method of the testing data set that MLP model
performance is better than both RBF neural network and MLR method.
The percentage of ANN model accuracy is found to be nearly between
90% and 95% as it can be observed from both the training and the testing
of data sets. It is clear from the prediction results that the developed
MLP is better than RBF neural network. Also, both neural networks are
found to be more accurate than the multiple regression method to pre-
dict flowrate in the desalination plants of Gaza Strip. The method used
for the development of ANN as a predictive model can be further ex-
tended for optimizing the operation of RO desalination plants in the
Gaza Strip.
3.4. Predictive models verification
The performance of the developed ANN predictive models obtained
after training and testing is required to be verified by using other un-
known data set. However, the required data for developed models ver-
ification purpose is collected from four different desalination plants
located in Gaza Strip governorates. For verification purposes the MLP
parameters (i.e. weights and biases) of the developed predictive ANN
models are used to predict water quality parameters including: perme-
ate flowrate and TDS for testing developed models performance. The
ANN developed model verification prediction results are presented in
Table 5. Results showed that the ANN developed models have good ca-
pability and accuracy in predicting the performance of desalination
plants in the Gaza Strip.
4. Conclusion
Generally all plants showed normal performance but the need to im-
prove and increase their production rate without increasing their water
resource abstraction is necessary to meet the water demand of the Gaza
Strip population. Recovery rate and TDS removal was made to evaluate
all selected desalination plants' performance.
MLP and RBF neural networks were successfully developed to pre-
dict one week ahead values of TDS concentrations and permeate
flowrate in the Gaza Strip desalination plants. It was found that the
MLP predictive model was slightly better than RBF predictive model.
The models were developed based on the data collected from five desa-
lination plants in the Gaza Strip. The developed models were compared
with the statistical model and it was found that the prediction results of
ANN were better than the conventional methods. Further research ef-
fort is suggested to be directed toward simulation of the desalination
plant's performance through membrane separation. This may be made
by using ANN to predict the performance of a single unit then predicting
a whole desalination plant. As a first case study for Gaza desalination
plants' performance modeling, the prediction results prove that the ar-
tificial neural network is suitable and capable for modeling the water
quality in the desalination plants. The developed ANN models in this

Predicted values-MLR Testing Data Predicted values-RBF Testing Data

Predicted values-MLP Testing Data Permeate Flow Rate Observed Values

70
Permeate flow rate values

60

50

40

30

20

10

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33

Fig. 9. Comparison between ANN and MLR prediction results.

 A.M. Aish et al. / Desalination 367 (2015) 240–247
study may be used as a new predictive tool to improve the performance
and management practices of desalination plants in the Gaza Strip.
Acknowledgment
The authors would like to thank the Middle East Desalination Re-
search Centre (MEDRC) and Palestinian Water Authority (PWA) Project
No: 12-AS-021 for their support and funding of the research.
References
[1] B. Shomar, Groundwater contaminations and health perspectives in developing
world case study: Gaza Strip, Environ. Geochem. Health 33 (2) (2011) 189–202.
[2] K. Qahman, A. Larabi, D. Oauzar, A. Naji, A. Cheng, Optimal extraction of groundwa-
ter in Gaza coastal aquifer, J. Water Resour. Prot. 1 (4) (2009) 249–259.
[3] H. Baalousha, Desalination status in the Gaza Strip and its environmental impact,
Desalination 196 (2006) 1–12.
[4] H. Hamdan, Artificial recharge of groundwater with storm water as a new water
resource-case study of the Gaza Strip, Palestine(PhD Thesis) Institute of Applied
Geosciences, Berlin University of Technology, Germany, 2012.
[5] Palestinian Water Authority (PWA), Water supply to Gaza preparatory studies for
seawater desalination plant, Final Technical Report (project information memoran-
dum) prepared by TECC (a local consultant) for PWA, 2012.
[6] R. El Sheikh, M. Ahmed, S. Hamdan, Strategy of water desalination in the Gaza Strip,
Desalination 156 (2003) 39–42.
[7] M.R. Ahmed, Water resources and management issues, Transboundary International
Conference, Aquifers: Challenges and New Direction (ISARM), 6–8 December, 2010.
[8] H.R. Maier, G.C. Dandy, Neural Networks for the prediction and forecasting of water
resources variables: a review of modelling issues and applications, Environ. Model.
Softw. 15 (2000) 101–124.
[9] P. Mageshkumar, T. Pradeep, M.R. Stalin John, S. Amal Raj, S. Anandakumar, Neural
network modelling of TDS concentrations in Cauvery River Water, Tamilnadu, India,
J. Appl. Sci. Eng. Res. 1 (6) (2012) 739–746.
[10] R.D. Yang, S.I. Kim, J.J. Jean, H.J. Kim, S.J. Lee, G.Y. Lee, S.Y. Lee, Artificial neural net-
work for optimizing operation of seawater reverse osmosis desalination plant, Desa-
lination 249 (2009) 180–189.
[11] D. Libotean, J. Giralt, F. Giralt, R. Rallo, T. Wolfe, Y. Cohen, Neural network approach
for modelling the performance of reverse osmosis membrane desalting, J. Membr.
Sci. 326 (2) (2009) 408–419.
[12] S. Nemati, L. Naghipour, M.H. Fazeli Fard, Artificial neural network modeling of total
dissolved solid in the Simineh River, Iran, J. Civ. Eng. Urban. 4 (1) (2014) 08–14.
[13] M. Khayet, C. Cojocaru, Artificial neural network model for desalination by sweeping
gas membrane distillation, Desalination 308 (2013) 102–110.
[14] G.A.C. Cordoba, Using of artificial neural network for evaluation and prediction of
some drinking water quality parameters within a water distribution system,
JUNIORSTAV, 3. Water Management and Water Structures2011.
[15] R. Righton, Development of an artificial neural network model predicting the perfor-
mance of a reverse osmosis (RO) unit(Master thesis) Curtin University of Technolo-
gy, 2009.
[16] A. Abbas, N. Al-Bastaki, Modelling of an reverse osmosis water desalination unit
using neural networks, Chem. Eng. J. 114 (1–3) (2005) 139–143.
[17] Z.V.P. Murthy, M.M. Vora, Prediction of reverse osmosis performance using artificial
neural network, Indian J. Chem. Technol. 11 (2004) 108–115.
[18] A. Aish, Drinking water quality assessment of the Middle Governorate in the Gaza
Strip, Palestine, Water Resour. Ind. 4 (2013) (2013) 13–20.
247
[19] M. Al-Talmas, Y. Mogheir, Risk Assessment and Mitigation of the Seawater Intrusion
Using Modelling Approach, Gaza Strip, Islamic University of Gaza, Palestine, 2012.
[20] O.A. Adeoti, P.A. Osanaiye, Effect of training algorithms on the performance of ANN
for pattern recognition of bivariate process, Int. J. Comput. Appl. 69 (2) (2013)
(0975–8887).
[21] S. Haykin, Neural Networks — A comprehensive Foundation, 3rd ed. Prentice-Hall,
Upper Saddle River, NJ, 2007.
[22] E. Quilty, T. Farahmand, P. Hudson, Validation and correction of high frequency
water quality data, 57th Canadian Water Resources Association Annual Congress,
2004.
[23] F. Anetil, A. Rat, Evaluation of neural network stream flow forecasting on 47 water-
sheds, J. Hydrol. Eng. 10 (1) (2005) 85–88.
[24] V. Nourani, A.H. Baghanam, M. Gebremichael, Investigating the ability of artificial
neural network (ANN) models to estimate missing rain-gauge data, J. Environ. In-
form. 19 (1) (2012) 38–50.
[25] P.A. Maedeh, N. Mehrdadi, G.R. Nabi Bidhendi, H. Zare Abyaneh, Application of arti-
ficial neural network to predict total dissolved solids variations in groundwater of
Tehran Plain, Iran, Int. J. Environ. Sustain. 2 (1) (2013) 10–20.
[26] A.H. Zare, V.M. Bayat, A.P. Daneshkare, Forecasting nitrate concentration in ground-
water using artificial neural network and linear regression models, Int. Agrophys. 25
(2011) 187–192.
[27] J.S. Alagha, Md Azlin Md Said, Y. Mogheir, M. Seyam, Modelling of chloride concen-
tration in coastal aquifers using artificial neural networks — a case study:
Khanyounis Governorate Gaza Strip–Palestine, Caspian J. Appl. Sci. Res. 2
(AICCE'12 & GIZ' 12) (2013) 158–165.
[28] K. Al-Mahallawi, J. Mania, A. Hani, I. Shahrour, Using of neural networks for the pre-
diction of nitrate groundwater contamination in rural and agricultural areas, Envi-
ron. Earth Sci. 65 (2012) 917–928.
[29] M. Seyam, Y. Mogheir, Application of artificial neural networks model as analytical
tool for groundwater salinity, J. Environ. Prot. 2 (2011) 56–71.
[30] H. Sirhan, M. Koch, Prediction of dynamic groundwater levels using an artificial neu-
ral network (ANN) approach in the Gaza Coastal Aquifer, South Palestine, First Inter-
national Colloquium REZAS12: Water Resources in the Arid and Semi-arid
Regions — Challenges and Prospects, Case of the African Continent. Beni Mellal,
Morocco, November 14–16, 2012.
[31] K. Al-Shayji, Y.A. Liu, Predictive modelling of large-scale commercial water desalina-
tion plants: data-based neural network and model-based process simulation, Ind.
Eng. Chem. Res. 41 (2002) 6460–6474.
[32] I.S. Al-Mutaz, B.A. Al-Sultan, Prediction performance of RO desalination plants, Desa-
lination 120 (1998) 153–160 (Presented at The Third Gulf Water Conference,
Muscat, Sultanate of Oman, 8–13 March 1997).
[33] D.S. Broomhead, D. Lowe, Multi-variable functional interpolation and adaptive net-
work, Complex Syst. 2 (1988) 321–355.
[34] S. Chen, C.F.N. Cowan, P.M. Grant, Orthogonal least squares learning algorithm for
radial basis function networks, IEEE Trans. Neural Netw. 2 (2) (1991) 302–309.
[35] F.R. Saen, The use of artificial neural networks for technology selection in the presence
of both continuous and categorical data, World Appl. Sci. J. 6 (2009) 1177–1189.
[36] S.A. Rounds, Development of a neural network model for dissolved oxygen in the
Tualatin River, Oregon, Presented at the Proceedings of the Second Federal Inter-
agency Hydrologic Modeling Conference (Sub-Committee on Hydrology of the In-
teragency Advisory Committee on Water Information), Las Vegas, Nevada, July 29
to August 1, 2002.
[37] M.A. Razi, K. Athappilly, A comparative predictive analysis of neural networks
(NNs), nonlinear regression and classification and regression tree (CART) models,
Expert Syst. Appl. 29 (2005) 65–74.