Ain Shams Engineering Journal (2017) 8, 549–557

Ain Shams University

Ain Shams Engineering Journal

www.elsevier.com/locate/asej
www.sciencedirect.com

CIVIL ENGINEERING

Data-driven modeling for water quality prediction

case study: The drains system associated with
Manzala Lake, Egypt
Mosaad Khadr *, Mohamed Elshemy

Irrigation and Hydraulics Engineering Department, Faculty of Engineering, Tanta University, 31734 Tanta, Egypt

Received 10 October 2015; revised 25 June 2016; accepted 11 August 2016

Available online 12 September 2016

KEYWORDS Abstract Manzala Lake, the largest of the Egyptian lakes, is affected qualitatively and quantita-
Data-driven modeling; tively by drainage water that flows into the lake. This study investigated the capabilities of adaptive
Water quality parameters; neuro-fuzzy inference system (ANFIS) to predict water quality parameters of drains associated with
Manzala Lake; Manzala Lake, with emphasis on total phosphorus and total nitrogen. A combination of data sets
Egypt was considered as input data for ANFIS models, including discharge, pH, total suspended solids,
electrical conductivity, total dissolved solids, water temperature, dissolved oxygen and turbidity.
The models were calibrated and validated against the measured data for the period from year
2001 to 2010. The performance of the models was measured using various prediction skill criteria.
Results show that ANFIS models are capable of simulating the water quality parameters and pro-
vided reliable prediction of total phosphorus and total nitrogen, thus suggesting the suitability of
the proposed model as a tool for onsite water quality evaluation.
Ó 2016 Ain Shams University. Production and hosting by Elsevier B.V. This is an open access article under
the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction the classification by considering the physical, chemical and

biological characteristics according to the water usage range
The quality and quantity of water resource worldwide is a sub- [5]. In water quality modeling, the mathematical modeling usu-
ject of ongoing concern [1]. Assessment and management of ally involves several parameters that cannot be measured or
long-term water quality of water resources is also a challenging involve considerable expense [6,7]. A deterministic model
problem [2–4]. The determination of the water quality refers to may also have inevitably errors originated from model struc-
tures or other causes. Water quality models are still therefore
simplified approximations of reality, and they inevitably con-
* Corresponding author.
tain certain kinds of errors that result in uncertainty in the
E-mail addresses: [email protected] (M. Khadr),
[email protected] (M. Elshemy).
model results [8]. Therefore the researchers tend to rely on con-
Peer review under responsibility of Ain Shams University.
ceptual or empirical models in practical applications to reduce
this uncertainty. A new modeling paradigm such as data-
driven modeling or data mining has recently been a
considerable growth in the development and application of
Production and hosting by Elsevier

http://dx.doi.org/10.1016/j.asej.2016.08.004
2090-4479 Ó 2016 Ain Shams University. Production and hosting by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
550 M. Khadr, M. Elshemy

computational intelligence and computer tools with respect to kilometers to the south of Port Said City. To the west, the
water-related problems [9,10]. Damietta branch of the Nile River borders the lake and the
These techniques are an approach to estimate the water southern side of the lake is bordered by cultivated land [30].
quality parameters based on the field data sets and to map The Lake is exposed to high inputs of pollutants from indus-
the relationship between the water quality parameter accord- trial, domestic, and agricultural sources. The southern region
ing to the temporal and spatial variation [11,12]. Data-driven of the lake is characterized by lower values of salinities and
models refer to a wide range of models that simulate a system high concentration of nutrients and heavy metals as a result
by the data experienced in the real life of that system. Data- of receiving high volumes of low salinity drainage water
driven modeling (DDM) is based on analyzing the data char- through different drains. The Lake is enriched by drainage
acterizing the system under study; in particular, a model can water transplanted by the drains which are connected to the
be defined on the basis of finding connections between the sys- Lake at the South and South Eastern Borders. Six major
tem state variables (input, internal and output variables) with- drains contribute by a flow rate of about 4170 million cubic
out explicit knowledge of the physical behavior [13]. DDM meters annually [32]. The main two drains flow into Manzala
includes different categories generally divided into statistical Lake, which were considered in this study, are Bahr El-
and artificial-intelligent models which include neural networks, Baker Drain system and Bahr Hadous Drain system. Bahr
fuzzy systems and evolutionary computing as well as other areas El Baqar drain, which is heavily polluted and anoxic over its
within artificial intelligence and machine learning [14–17]. entire length, transports untreated and poorly treated wastew-
The use of ANNs and fuzzy logic has many successful ater to Lake Manzala over a distance of 170 km. Water quality
applications in hydrology; in modeling rainfall-runoff pro- data, that were used to develop the ANFIS models, are mea-
cesses [18–21]; replicating the behavior of hydrodynamic/ sured values at outfall measuring stations and have a record
hydrological models of a river basin where ANNs are used length of 10 years covering between 2001 and 2010. The data
to provide optimal control of a reservoir [22]; modeling set includes discharge (Q), pH, total suspended solids (TSS),
stage-discharge relationships [23]; simulation of multipurpose electrical conductivity (EC), total dissolved solids (TDS),
reservoir operation [24–26]; and deriving a rule base for reser- water temperature, dissolved oxygen (DO) and turbidity
voir operation from observed data. The development and cur- (TU). The output of the model is two water quality parame-
rent progress in the integration of various artificial intelligence ters; total phosphorus (TP) and total nitrogen (TN). Both
techniques (knowledge-based system, genetic algorithm, artifi- parameters were chosen to be the model objective output
cial neural network, and fuzzy inference system) in water qual- due to their main impacts on water quality status and control.
ity modeling, sediment transportation and DO concentration Table 1 summarizes the statistical properties of input and out-
have been studied by many researchers [27–29]. Egyptian put data used in the simulations.
northern lakes have been regarded highly as a fishery; there-
fore, monitoring of water quality of the drainage water input 2.2. Adaptive neuro-fuzzy inference system – ANFIS
to the northern lakes is a major task for maintaining their ecol-
ogy. In this study, adaptive neuro-fuzzy inference system
An adaptive neuro-fuzzy inference system (ANFIS) is a fuzzy
(ANFIS) models were developed for prediction and simulation
inference system formulated as a feed-forward neural network.
of water quality parameters in drains systems associated with
Hence, the advantages of a fuzzy system can be combined with
Manzala Lake, the most important among all Egyptian Lakes,
a learning algorithm [33,34]. ANFIS was introduced as an
with emphasis on total phosphorus (TP) and total nitrogen
effective tool to represent simple and highly complex functions
(TN). TN and TP are considered as the most essential param-
more powerfully than conventional statistical methods. Neuro-
eters to assess and control the water quality and trophic status
fuzzy modeling is a technique for describing the behavior of a
of water bodies. In order to measure these two parameters,
system using fuzzy inference rules within a Neural Network
laboratory examinations should be done using water samples,
(NN) structure. Using a given input/output data set, adaptive
which is costly and time consuming process. To the best of our
neuro-fuzzy inference system (ANFIS) constructs a FIS whose
knowledge, the issue on predicting of water quality in the study
member ship function parameters are tuned using a back prop-
area using ANFIS model so far has not been addressed. It is
agation algorithm [34]. So, the FIS could learn from the train-
hoped that the proposed approach and our findings obtained
ing data. In this study, the ANFIS models were developed in
in this study are useful and valuable to assist in reporting the
the MATLAB environment.
status of water quality in the study area.
ANFIS was used to extract the relation of the total phos-
phorus (TP), total nitrogen (TN), discharge (Q), pH, total sus-
2. Method and materials pended solids (TSS), electrical conductivity (EC), total
dissolved solids (TDS), water temperature, dissolved oxygen
2.1. Study area and water quality data (DO) and turbidity (TU). The consequent part is total phos-
phorus (TP) or total nitrogen (TN). The structure of the
Manzala Lake, which is located at the northern edge of the ANFIS model consists of a Sugeno type fuzzy system with gen-
Nile Delta, is the largest of the Egyptian lakes along the eralized bell input membership functions, which provided the
Mediterranean coast (Fig. 1) [30,31]. The lake is bordered at best results in this study, and a linear output membership func-
the north by a sandy margin which separates the lake from tion. The Sugeno model makes use of ‘‘if then” rules to pro-
the Mediterranean Sea except at three outlets where exchange duce an output for each rule. It is similar to the Mamdani
of water occurs. These outlets are El-Gamil, El-Boughdady, method in many respects. The first two parts of the fuzzy infer-
and the new El-Gamil [31]. The eastern side of the lake is con- ence process, fuzzifying the inputs and applying the fuzzy
nected with the Suez Canal through El-Raswa Canal, a few operator, are exactly the same. The main difference between
Data-driven modeling for water quality prediction case study 551

Figure 1 Layout of El-Manzala Lake and main canals and drainage system associated with it.

Table 1 Statistical measures of water quality parameters at the outlet of Bahr El Baqar and Bahr Hadous drains.
Bahr El-Baqar drain Bahr Hadous drain
Min. Max. Mean Std. Min. Max. Mean Std.
3
Discharge (m /s) 42 59.9 50.38 4.93 2.34 11.65 5.44 2.77
PH 6.6 8.45 7.48 0.28 6.82 8.41 7.52 0.27
TSS (lg/L) 0 197 77.26 42.09 0.00 499.00 38.54 51.81
EC (lg/L) 1.15 6.57 4.36 0.94 0.50 2.28 1.41 0.27
TDS (lg/L) 291 4468 2779.1 645.21 342.0 1420.00 967.11 183.36
Temperature (°C) 13 31 23.05 5.31 11.00 32.00 21.90 5.35
DO (lg/L) 0.36 5.6 2.27 1.08 0.08 7.80 1.58 1.52
TU 39 200 108.73 44.7 14.00 73.00 39.59 15.41
TP (lg/L) 0.14 1.87 0.92 0.3 0.05 2.08 0.79 0.36
TN (lg/L) 0.7 55.6 15.41 13.44 0.31 80.77 10.29 13.17

Layer 1 Layer 2 Layer 3 Layer 4 Layer 5

A1 w1 w1 w 1 f1
X
A2
F
B1
Y w 2 w 2 w f2 2

B2

Figure 2 An ANFIS architecture for a two rule Sugeno system.

Mamdani and Sugeno is that in the Sugeno type rule outputs 1. A rule base containing a number of if-then rules.
consist of the linear combination of the input variables plus 2. A database which defines the membership function.
a constant term; the final output is the weighted average of 3. A decision making interface that operates the given rules.
each rule’s output. Adaptive neuro-fuzzy inference system 4. A fuzzification interface that converts the crisp inputs into
mimics the operation of a Takagi–Sugeno–Kang (TSK) fuzzy ‘‘degree of match” with the linguistic values such as high or
system. Fig. 2 presents the typical architecture of ANFIS with low.
a multilayer feed-forward network, which is linked with a 5. A defuzzification interface that reconverts to a crisp output.
fuzzy system for two inputs (x and y). Fuzzy inference systems
are composed of five functional blocks and the ANFIS model The rule base in the Sugeno model has the following form:
contains the following [35]: If x is A1 and y is B1 then f1 ¼ p1 x þ q1 y þ r1 ð1Þ
552 M. Khadr, M. Elshemy

If x is A2 and y is B2 then f2 ¼ p2 x þ q2 y þ r2 ð2Þ employed. The MAD measures the average magnitude of the
errors in a set of prediction, without considering their direc-
where x and y are predefined membership functions, Ai and Bi
tion. It measures accuracy for continuous variables. Expressed
are membership values, pi, qi, and ri are the consequent param-
MAD is calculated as follows:
eters that are updated in the forward pass in the learning algo-
Pn
rithm, and fi is the output within the fuzzy region specified by jWQoi  WQfi j
the fuzzy rule. MAD ¼ i¼1 ð8Þ
n
Let the membership functions of fuzzy sets Ai and Bj, be lAi
where WQo is the observed value, WQf is the predicted value
and lBi respectively. The five layers that integrate ANFIS are
and n is the number of data points.
as follows: In statistics, the coefficient of determination, R2 is used in
Let the output of the ith node in layer l is denoted as O1,i, the context of statistical models whose main purpose is the pre-
then, diction of future outcomes on the basis of other related infor-
mation. The absolute fraction of variance, R2, is calculated as
Layer 1: Every node i in this layer is an adaptive node with follows:
node function; Pn
ðWQoi  WQfi Þ2
Q1;i ¼ lAi ðxÞ for i ¼ 1; 2; or Q1;i ¼ lBi2 ðyÞ R2 ¼ 1  i¼1Pn 2
ð9Þ
i¼1 ðWQoi Þ
for i ¼ 3; 4 ð3Þ
with the variables already having been defined. The RMSE is
where x (or y) is the input to the ith node and Ai (or Bi2) is
the square root of the variance of the residuals. It indicates
linguistic labels. the absolute fit of the model to the data–how close the
Layer 2: This layer consists of the nodes labeled which mul-
observed data points are to the model’s predicted values.
tiply incoming signals and send the product out. Each node
Whereas R-squared is a relative measure of fit, RMSE is an
output represents the firing strength of a rule;
absolute measure of fit and lower values of RMSE indicate
O2;i ¼ wi ¼ lAi ðxÞ lBi ðyÞ for i ¼ 1; 2 ð4Þ better fit. RMSE is calculated as follows:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn 2
i¼1 ðWQoi  WQfi Þ
Layer 3: In this layer, the nodes labeled N act to scale the
firing strengths to provide normalized firing strengths; RMSE ¼ ð10Þ
n
wi The correlation coefficient is a concept from statistics, and
i ¼
O3i ¼ w ; i ¼ 1; 2 ð5Þ
w1 þ w2 it is a measure of how well trends in the predicted values follow
Layer 4: The output of layer 4 is comprised of linear com- trends in past actual values (historical releases). The correla-
bination of inputs multiplied by normalized firing strengths. tion coefficient is calculated as follows:
P
This layer’s nodes are adaptive with node functions; Pn ð WQoi ÞðWQfi Þ
i¼1 WQoi WQfi 
Cr ¼ s ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

n
ffi
Pn 2  Pn 2 
O4i ¼ wi fi ¼ wi ðpi x þ qi y þ ri Þ ð6Þ Pn ð i¼1 oi Þ WQ P n ð i¼1 WQfi Þ
i¼1 WQoi  i¼1 WQfi 
2 2
n n
where, wi is the output of layer 3, and {pi, qi, ri} are the param-
eter set. Parameters of this layer are referred to as consequent ð11Þ
parameters.
Layer 5: This layer consists of a single node and computes The efficiency E proposed by Nash and Sutcliffe [37] is
the final output as the summation of all incoming signals; defined as one minus the sum of the absolute squared differ-
ences between the predicted and observed values normalized
X P by the variance of the observed values during the period under
wi f
O5i ¼ i fi ¼ Pi¼1 i
w ð7Þ investigation. It is calculated as follows:
i¼1 wi
i¼1 "Pn 2
#
Layers represented by squares are adaptive and their values i¼1 ðWQoi  WQfi Þ
E ¼ 1  Pn 2
ð12Þ
are adjusted when carrying out the system training. Layers rep- i¼1 ðWQoi  WQo Þ
resented by circles remain invariable before, during and after
the training [36]. Fig. 3 illustrates the network used in this where WQo is the average of the considered parameter.
paper and consists of eight inputs, and one output membership
function (TP or TN). 3. Results and discussion

2.3. Performance measures The adaptive neuro-fuzzy inference system (ANFIS) was used
to derive and to develop models for prediction of water quality
Several measures of goodness of fit were used to evaluate the parameters in Bahr El-Baker Drain system and Bahr Hadous
prediction performance of all the aforementioned ANFIS Drain system. To simulate and predict the behavior of water
models. The measures that were used include Mean Absolute quality parameters in the two drains systems, a time series of
Deviations (MAD), the coefficient of determination (R2), Root the ten previously noted parameters in a 10-year (120-
Mean Square Error (RMSE), correlation coefficient (Cr) and month) period was used. For ANFIS models construction,
Nash–Sutcliffe coefficient (E). To investigate whether there is monthly data set has been randomly partitioned into two parts
a significant difference between the mean from the observed for the training and testing processes by considering 70% and
and predicted data, the two-sample t-test for the means was 30% respectively, which are common divisional percentages in
Data-driven modeling for water quality prediction case study 553

Figure 3 An ANFIS architecture for TP and TN prediction.

Table 2 Performance of particular ANFIS models according to the RMSE, with respect to the selection of input–output membership
functions.
Bahr El-Baqar drain Bahr Hadous drain
gaussmf dsigmf trapmf gbellmf trimf gaussmf dsigmf trapmf gbellmf trimf
Training TP 0.073 0.092 0.068 0.029 0.061 0.151 0.182 0.142 0.102 0.132
TN 1.352 1.295 1.086 0.756 0.954 0.092 0.087 0.093 0.018 0.076
Testing TP 0.084 0.094 0.073 0.023 0.062 0.334 0.297 0.161 0.122 0.143
TN 1.272 1.383 1.317 1.109 1.253 0.705 0.833 0.917 0.478 0.624

data-driven models. Accordingly, data set was divided into

Load historical data (TP, TN, Q, pH, TSS, EC, TDS, temperature, DO
used 7 and 3-year periods, respectively, for the training and and FTU) from i=1: i=N
testing data sets. Fuzzy inference system structure of a
Sugeno-type was then generated using subtractive clustering
and the separate sets of input and output data as input argu- Check data and filling missing values
ments and this was applied to TP and TN as well. The aim
of this step was to determine the number of rules and antece- Selection of desired output (TP or TN)
dent membership functions and then used linear least squares
estimation to determine each rule’s consequent equations to Generate Fuzzy Inference System (FIS) from i=1: i=t
cover the feature space.
A hybrid learning algorithm was used to identify parame- Optimization of FIS Parameters
ters of Sugeno-type fuzzy inference systems by applying a com-
bination of the least-squares method and the back propagation
gradient descent method for training FIS membership function ANFIS Training from i=1: i=t
parameters to emulate a given training data set. The network
was trained to obtain the nearest output to the target. In order ANFIS validation from i=t+1: i=N
to find the best model, an optimization model was developed
to find the ANFIS parameters that give best performance.
Calculation of Output
The performance function that was used for feed-forward is
the root of mean square error between the network outputs
and the target output. More than 119 models were tested to Model performance measurement
select the best model which fits data space with best perfor-
mance. According to the least RMSE from Table 2 it is very Figure 4 Flowchart of the water quality parameters simulation
obvious that the best ANFIS model, used for training and test- using ANFIS.
ing, is the one using generalized bell-shaped membership func-
tions per each input and a linear output membership function. out and compared with actual measured values to validate
Once the best working model was selected through ANFIS and examine the reliability of the developed ANFIS model
training, the TP and TN predicted values would be worked as shown in Fig. 4.
554 M. Khadr, M. Elshemy

Figure 5 Observation and estimation of water quality parameters using ANFIS model in Bahr El-Baker Drain. (a) TP training period;
(b) TP testing period; (c) TN training period; (d) TN testing period.

Figs. 5 and 6 present the monthly values of total phospho- values is similar except few records which are more deviated
rus (TP) and total nitrogen (TN) estimated by ANFIS versus from actual measured values.
the corresponding measured values for the training and the test The training, testing and validation results for both the
data set for Bahr El-Baker Drain system and Bahr Hadous total phosphorus (TP) and total nitrogen (TN) prediction
Drain system respectively. It is shown in Figs. 5 and 6 that models are summarized in Table 3. It is clearly seen from
the two curves of observed and estimated data almost overlap Table 3 that the ANFIS performs satisfactory and the overall
each other and the trend between the measured and estimated prediction results are fairly good from the RMSE, MAD and
Data-driven modeling for water quality prediction case study 555

Figure 6 Observation and estimation of water quality parameters using ANFIS model in Bahr Hadous Drain. (a) TP training period; (b)
TP testing period; (c) TN training period; (d) TN testing period.

Table 3 Performance measures for comparison of observed and predicted water quality parameters.
Bahr El-Baqar drain Bahr Hadous drain
MAD R2 RMSE Cr E MAD R2 RMSE Cr E
Training TP 0.016 0.98 0.029 0.976 0.992 0.071 0.98 0.102 0.971 0.942
TN 0.457 0.99 0.756 0.945 0.983 0.009 0.97 0.018 0.981 0.932
Testing TP 0.015 0.94 0.023 0.901 0.994 0.036 0.97 0.122 0.802 0.953
TN 0.682 0.92 1.109 0.857 0.971 0.209 0.91 0.478 0.829 0.928
556 M. Khadr, M. Elshemy

R2 viewpoints. All values of RMSE shown in Table 3 are very [5] Hambright KD, Parparov A, Berman T. Indices of water quality
small compared to mean values of both TP and TN during for sustainable management and conservation of an arid region
training and testing periods, which is an indicator for the high lake, L. Kinneret (Sea of Gali-lee), Kinneret. Aquat Conservation,
efficiency of the data-driven model. Table 3 also indicates that Mar Freshwater Ecosyst 2000;10(6):393–406.
[6] Rode Michael, Arhonditsis George, Balin Daniela, Kebede
the RMSE, for all models, is always closer to or equal to the
Tesfaye, Krysanova Valentina, Van Der Zee ATM. New chal-
MAD which indicates that all the errors are of the same mag- lenges in integrated water quality modelling. Hydrol Process
nitude. A significant positive correlation was obtained between 2010;24:3447–61.
the two groups of data for both TP and TN in case of Bahr [7] Maia Celsemy Eleutério, Rodrigues Kelly Kaliane Rego da Paz.
El-Baker Drain with values of 0.901 and 0.857 respectively; Proposal for an index to classify irrigation water quality: a case
however, in Bahr Hadous Drain Cr equals 0.802 and 0.929 study in northeastern Brazil. R Bras Ci Solo 2012;36:823–30.
respectively. The large values of R2 are indicative of a perfect [8] Ghajarnia N, Haddad OB, Mariño MA. Performance of a novel
relationship between the observed and predicted values. hybrid algorithm in the design of water networks. Proc Inst Civ
According to Eq. (12), the range of E lies between 1.0 (per- Eng Water Manage 2011;164(4):173–219.
fect fit) and 1; high value of E is indicative of a more effi- [9] Orouji H, Orouji H, Bozorg Haddad O, Fallah-Mehdipour E.
Modeling of water quality parameters using data-driven models. J
cient data-driven model. Values of E in the range (1, 0)
Environ Eng Ó ASCE 2013:947.
occur when the mean observed value is a better estimation [10] Daniel Edsel B, Camp Janey V, Le Boeuf Eugene J. Watershed
than the model prediction or simulation value, which indicates modeling and its applications: a state-of-the-art review. Open
unacceptable performance. Values of E shown in Table 3 are Hydrol J 2011;5:26–50.
greater than 0.9, which indicates that proposed models have [11] Liu Z-J, Weller DE. A stream network model for integrated
perfect fit for all of the quality parameters in both training watershed modeling. Environ Modell Assess 2007;13(2):291–303.
and testing data sets. The two-sample t-test failed to reject [12] Wu CL, Chau KW. Data-driven models for monthly streamflow
the null hypothesis at the 5% significance level (h = 0) that time series prediction. Eng Appl Artif Intell December
both predicted and measured data come from independent 2010:1350–67.
random samples from normal distributions with equal means. [13] Abrahart R et al. Data-driven modelling: concepts, approaches
and experiences, practical hydroinformatics. Water science and
technology library. Berlin Heidelberg: Springer; 2008. p. 17–30.
4. Conclusions [14] Kholghi M, Hosseini SM. Comparison of groundwater level
estimation using neuro-fuzzy and ordinary kriging. Environ
Model Assess 2009;14(6):729–37.
In this study, the capabilities of data-driven models to predict
[15] Ghavidel SZZ, Montaseri M. Application of different data-driven
water quality parameters were investigated. This study methods for the prediction of total dissolved solids in the
adopted ANFIS models to achieve easier and faster water Zarinehroud basin. Stoch Environ Res Risk Assess 2014;28
quality parameter predictions with emphasis on total phospho- (8):2101–18.
rus (TP) and total nitrogen (TN) in drain systems associated [16] Sanikhani H, Kisi O. River flow estimation and forecasting by
with Manzala Lake, the most important among all Egyptian using two different adaptive neuro-fuzzy approaches. Water
Lakes. Two main ANFIS models, for each drain system, were Resour Manage 2012;26(6):1715–29.
constructed for both TP and TN. The performance of the [17] Sanikhani H, Kisi O, Kiafar H, Ghavidel SZZ. Comparison of
developed models was measured on a 10-year database of different data-driven approaches for modeling lake level fluctua-
tions: the case of Manyas and Tuz Lakes (Turkey). Water Resour
ten water quality parameters. Comparison between predicted
Manage 2015;29(5):1557–74.
and measured data, using several evaluation criteria, showed
[18] Dawson CW, W.R.. An artificial neural network approach to
the efficiency of the applied models and confirmed the accu- rainfall-runoff modelling. Hydrol Sci J 1998;43(1).
racy of the developed ANFIS models. Validation statistics also [19] Dibike YSD, Abbott MB. On the encapsulation of numerical-
indicate that the correlation between predicted and actual mea- hydraulic models in artificial neural network. J Hydraul Res
sured values was fairly good. With reference to our findings, 1999;37(2).
we propose the developed models as a simple tool for predict- [20] Abrahart RJ, S.L.. Comparing neural network and autoregressive
ing water quality parameters and for onsite water quality moving average techniques for the provision of continuous river
parameters evaluation. flow forecast in two contrasting catchments. Hydrol Process
2000;14.
[21] El-Shafie A, Jaafer O, Seyed A. Adaptive neuro-fuzzy inference
References system based model for rainfall forecasting in Klang River,
Malaysia. Int J Phys Sci 2011;6(12):2875–88.
[1] Diamantopoulou MJ, Antonopoulos VZ, Papamichail DM. The [22] Solomatine DP, TL. Neural network approximation of a hydro-
use of a neural network technique for the prediction of water dynamic model in optimizing reservoir operation. In: Proc. 2nd
quality parameters of Axios River in Northern Greece. European int conference on hydroinformatics, Balkema, Rotterdam; 1996.
Water 2005;11/12:55–62, 2005. [23] Sudheer KP, J.S.. Radial basis function neural network for
[2] Koschel R, Adams DD. An approach to under- standing a modeling rating curves. ASCE J Hydrol Eng 2003;8(3).
temperate oligotrophic Lowland Lake (Lake Stechlin, Germany). [24] Fontane DG, Gates TK, Moncada E. Planning reservoir opera-
Archiv fur Hydrobiol, Spec Issues Adv Limnol 2003;58(2003):1–9. tions with imprecise objectives. J Water Resour Plan Manage,
[3] Parparov A, Hambright KD, Hakanson L, Ostapenia AP. Water ASCE 1997;123(3):154–62.
quality quantification: basics and implementation. Hydrobiologia [25] Shrestha BP, Duckstein LE, Stokhin Z. Fuzzy rule-based mod-
2006;560(1):227–37, 2006. eling of reservoir operation. J Water Resour Plan Manage
[4] Parparov Arkadi, Gal Gideon, Hamilton David, Kasprzak Peter, 1996;122(4):262–9.
Ostapenia Alexandr. Water quality assessment, trophic classifica- [26] Khadr M, Schlenkhof A. Integration of data-driven modeling and
tion and water resources management. J Water Resour Prot stochastic modeling for multi-purpose reservoir simulation. In:
2010;2:907–15. 11th international conference on hydroscience & engineering
Data-driven modeling for water quality prediction case study 557

(ICHE2014), Hamburg, Germany, 28 September to 2 October [37] Nash JE, Sutcliffe JV. River flow forecasting through conceptual
2014. models part 1 - A discussion of principles. J Hydrol 1970;10
[27] Broomhead D, Lowe D. Multivariable functional interpolation (3):282–90.
and adaptive networks. Complex Syst 1998;2(6):321–55.
[28] Dogan E, Sengorur B, Koklu R. Modeling biological oxygen
demand of the Melen River in Turkey using an artificial neural Mosaad Khadr is an Assistant Professor in
network technique. J Environ Manage 2009;90:1229–35. Irrigation and Hydraulics Engineering
[29] Ranković V, Radulović J, Radojević I, Ostojić A, Ćomić L. Department, Faculty of Engineering, Tanta
Neural network modeling of dissolved oxygen in the Gruza University, Egypt. In 1997, he did B.Sc. in
reservoir, Serbia. Ecol Model 2010;221:1239–44. Civil Eng. (Structural Engineering), and in
[30] Ali Mohamed HH. Assessment of some water quality character- 2002, M.Sc. in Civil Engineering, Irrigation
istics and determination of some heavy metals in Lake Manzala, and Hydraulic Engineering, Faculty of Engi-
Egypt. J Aquat Biol Fish 2008;12(2):133–54. neering, Tanta University, Egypt. In 2011, he
[31] Abdel Mola Hesham R, Abd El – Rashid Mohamed. Effect of did Ph.D. in Water Resources Management,
drains on the distribution of zooplankton at the southeastern part University of Wuppertal, Germany.
of Lake Manzala, Egypt. J Aquat Biol Fish 2012;16:57–68, ISSN
1110–1131.
[32] Laila Shakweer. Ecological and fisheries development of lake
Manzala, Egypt. Egypt J Aquat Res 2005;31(1). Mohamed Elshemy is an Assistant Professor in
[33] Zadeh LA. Outline of a new approach to analysis of complex Irrigation and Hydraulics Engineering
systems and decision processes. IEEE Trans Syst Man Cybernet- Department, Faculty of Engineering, Tanta
ics, Smc 1973;3(1):28–44. University, Egypt. In 1997, he did B.Sc. in
[34] Labani MM, Salahshoor K. Estimation of NMR log parameters Civil Eng. (Structural Engineering), and in
from conventional well log data using a committee machine with 2002, M.Sc. in Civil Engineering, Irrigation
intelligent systems: a case study from the Iranian part of the South and Hydraulic Engineering, Faculty of Engi-
Pars gas field, Persian Gulf Basin. J Petrol Sci Eng neering, Tanta University, Egypt. In 2010, he
2010;72:175–85. did Ph.D. in Civil Engineering (Environmen-
[35] Venugopal C, Devi SP, Rao KS. Predicting ERP user satisfaction- tal Engineering), University of Braunschweig,
an adaptive neuro fuzzy inference system (ANFIS). Approach Germany.
Intell Inf Manage 2010;2:422–30.
[36] Kablan A. Adaptive neuro-fuzzy inference system for financial
trading using intraday seasonality observation model. World
Acad Sci, Eng Technol 2009;58:479–88.