Journal of Environmental Science and Health, Part A

Toxic/Hazardous Substances and Environmental Engineering

ISSN: 1093-4529 (Print) 1532-4117 (Online) Journal homepage: https://www.tandfonline.com/loi/lesa20

A novel framework of multivariate modeling of

3
water distribution network through 3 factorial
design and artificial neural network

Partha S. Ghosal, Ashwini Javaregowda, Ashok K. Gupta & Dineshwar P.

Singh

To cite this article: Partha S. Ghosal, Ashwini Javaregowda, Ashok K. Gupta & Dineshwar P.
3
Singh (2019) A novel framework of multivariate modeling of water distribution network through 3
factorial design and artificial neural network, Journal of Environmental Science and Health, Part A,
54:6, 551-562, DOI: 10.1080/10934529.2019.1571308

To link to this article: https://doi.org/10.1080/10934529.2019.1571308

Published online: 22 Feb 2019.

Submit your article to this journal

Article views: 167

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=lesa20
JOURNAL OF ENVIRONMENTAL SCIENCE AND HEALTH, PART A
2019, VOL. 54, NO. 6, 551–562
https://doi.org/10.1080/10934529.2019.1571308

A novel framework of multivariate modeling of water distribution network

through 33 factorial design and artificial neural network
Partha S. Ghosala, Ashwini Javaregowdaa, Ashok K. Guptaa, and Dineshwar P. Singhb

a
Environmental Engineering Division, Department of Civil Engineering, Indian Institute of Technology Kharagpur, Kharagpur, India; bDwarka,
New Delhi, India

ABSTRACT ARTICLE HISTORY

The water distribution network is largely affected by the change in the influencing factors, such as Received 1 August 2018
input pressure, demand and supply duration. The change in each parameter requires the extensive Accepted 8 January 2019
design of the network and the interactive effect of the influencing parameters are hardly explored.
KEYWORDS
The main hurdles for the water providers lie in the absence of a prediction model, which can be
Water distribution networks;
used as a decision tool to assess the effect of the change in parameter and estimating the cost 33 factorial design; artificial
for the changed scenario. The present study developed a novel framework based on the artificial neural network;
neural network for multivariate prediction modeling taking the response as the cost of the pipe hydraulic design
network. The application of the 33 factorial design was used for the selection of the influencing
parameters and outcome was taken as the input to the neural network model. The adequacy of
the model was tested through error functions and analysis of variance. The low values of the error
functions (0.0004–0.228) and high F value (162,442) and R2 (0.999) established the significance of
the model. The model can be used for predicting the cost of the changed scenarios and assess-
ment of the optimal solution for the system variables.

Introduction water supply are the main factors significantly influencing

Water distribution network (WDN) plays an important role
in the water supply system. Designing WDN for continuous
water supply (CWS) and operating in the intermittent water
supply (IWS) create substantial pressure variation and ham-
per equitable water supply.[1,2] Nevertheless, the intermit-
tency in a continuous water supply system instigates the
pressure and flow distributions to be inadequate and not
homogeneous.[3] Short duration supply in IWS leads to large
peak flow, increased pressure-loss, tail end consumer suffers
low pressure and insufficient demand.[4] Supply duration
less than 12 h per day significantly increase the chance of
hydraulic losses and energy loss in the water distribution
network.[5] Life of the distribution system reduces due to
frequent leakage caused by the transient pressure in the
intermittent system. In the IWS system, network experiences
regular flow restarting, loss of pressure due to high peak
flow, draining as supply stops and ingression of contami-
nated water as pressure falls below positive value. These
processes influence the mechanism affecting water quality
throughout the distribution system.[6] However, the WDN
in most of the developing countries follows the IWS with a
wide variation of the supply hour.
Nodal pressures, delivered flows, line velocities, nodal ele-
vation differences, size of supplied areas and duration of
the equity of water supply in WDN.[7] The difference in
hydraulic gradient and nodal elevation defines the pressure
at the junction point; pressure tends to reduce as the eleva-
tion rises. Pressure head at the source and head loss in the
network define the hydraulic gradient line (HGL) through-
out the network and directly affects the nodal pressure
hence influences demand. Flow rate influences the velocity
and pressure of the network. The increase in flow rate due
to increasing demand in an existing network leads to mal-
function of the system.
The major issue with the water supply system in the
developing countries is the increasing demand due to
growth/migration of population, pilferage of water and
unmetered consumption. The unaccounted for flow
increases with time and additional capacity for supply is
often required. Although the treatment capacity of the sys-
tem can be increased by proving additional treatment units,
the increase in the capacity of the distribution line is often
difficult. Moreover, the difficulties in decision making for
augmenting the capacity or checking the adequacy of the
system is often observed due to nonavailability of predictive
model considering the several variables influencing the sys-
tem. The details of a distribution system have to be assessed
through a comprehensive hydraulic design, which requires

CONTACT Ashok K. Gupta agupta@civil.iitkgp.ac.in Environmental Engineering Division, Department of Civil Engineering, Indian Institute of Technology,
Kharagpur 721 302, India.
Color versions of one or more figures in the article can be found online at www.tandfonline.com/lesa.
Supplemental data for this article is available online at https://doi.org/10.1080/10934529.2019.1571308.
*Retired
ß 2019 Taylor & Francis Group, LLC
552 P. S. GHOSAL ET AL.
effort and time. The associated cost of modification is also
challenging to arrive without performing the detailed
hydraulic analysis of the network. The optimized solution
for the modification of a distribution network can only be
achieved by developing a multivariate prediction model for
the WDN.
The modeling of water supply system parameters associ-
ated with the WDN can be conducted by the univariate
approach by performing the hydraulic design of WDN with
the variation of a single parameter. However, a univariate
study rarely produces the interactive effect of the major sys-
tem parameters. In several areas of environmental modeling,
the applications of multivariate modeling have become a
popular choice to overcome limitations of the univariate
study. Among the several tools of multivariate modeling, the
artificial neural network (ANN) has become a preferred
choice. The vast application of artificial neural network
(ANN) on nonlinear multivariate modeling in environmen-
tal engineering for complex problems appeared promis-
ing.[8–15] Specifically, some inherent advantages, such as
universal approximation and mapping capacity, adaptability
to changes in the dataset, robustness and fault tolerance,
ability to process data in parallel (fast processing) and dis-
tributed manner are associated with the adequate perform-
ance of ANN model.[16,17] Especially, ANN can be
considered as a suitable tool to model the processes, which
are not described by an appropriate mathematical relation-
ship. Although, the efficiency of ANN model is largely
dependent on a large number of input data, it was observed
that a systematic input data through the statistical design of
experiment might drastically reduce the requirement of data
as well as enhance the performance of ANN.[18] Even, an
ANN model can perform better than the regression model if
the training of the network is conducted with suitable
experimental data.[19] Although, various techniques of opti-
mization and modeling were attempted for designing, solv-
ing, operational optimization, quality assessment and so
forth of the distribution network,[20–30] the multivariate
modeling for appraising the effect of major influencing par-
ameter on the distribution network was rarely conducted,
which is a significant gap area in the field of planning and
management of water supply system. Furthermore, the
application of ANN on WDN was rarely conducted,
although, it can be an efficient tool in this context.
In this article, the analysis of WDN at various conditions
of essential factors of the system was analyzed, and a multi-
variate prediction model was developed through ANN. The
systematic dataset with a 33 factorial design was framed con-
sidering pressure head at the source, network demand and
duration of supply as independent variables and the cost of
the system as a response. The number of neurons in the
hidden layer, training algorithms, transfer functions and so
forth, were optimized to develop the neural network. The
prediction capacity of the model was tested with a validation
study conducted by 32 factorial model for 8 h supply. The
efficiency of the model was tested by simulating the network
concerning the actual values of response.
Materials and methods
Selection of major influencing parameters
A water distribution network is characterized by several
physical and hydraulic parameter. Each parameter influences
the network capacity in different ways. The hydraulic
parameters of the water distribution systems are inter-
dependent and have a nonlinear interactive effect on one
another. Hence, each parameter has to be carefully studied
before considering for optimization of the network. The
three major influencing parameters of water distribution
network design are selected and investigated through several
simulations to model the optimized distribution network
design. Input pressure (pump head) plays a vital role in
determining the nodal pressure; the pressure head at source
defines hydraulic gradient and pressure at all nodes and cor-
responding flow at the respective nodes. Head at source also
defines pipe diameter to maintain required pressure
throughout the network. Nodal pressure defines the water
supply at the nodes.
Network demand is always an uncertain parameter.
Predicting future demand of water distribution system is the
biggest hurdle.[31] A sudden increase in demand within the
network or due to the nontechnical expansion of the net-
work decreases network capacity. The demand of the water
distribution network regulates the flow, velocity and head
loss in the system. Duration of supply in the network affects
the other hydraulic parameter of the network like flow, vel-
ocity, head loss and pressure in the network. The demand
increases with the decrease in supply duration. Thus, the
duration of supply, pressure head at the source and network
demand are considered as the significant factors for multi-
variate modeling of the network.
Multivariate modeling
The variation of parameters mentioned above has a signifi-
cant impact on the water distribution network. The specific
inter-relationships among the parameters are not available.
Moreover, the parameters nonlinearly affect the distribution
system. The assessment of the variation of the distribution
network is reflected by the variation of the pipe diameter,
which affects the overall cost of the network. Hence, the
evaluation of the overall effect of those parameters on the
distribution system may be carried out by the cost of the
pipe network.
In the present study, the artificial neural network (ANN)
was considered as the modeling tool, as there is no specific
mathematical relationship is available. The input of the
ANN should comprise of the dataset of variation of the cost
of the pipeline (response) concerning the variation of input
pressure, hours of supply and network demand (factors).
Eventually, the multivariate modeling through ANN requires
the generation of the different set of data showing the vari-
ation of response concerning the factors. The dataset was
prepared by simulation of a hypothetical network for a dif-
ferent set of factors and respective outcomes. In this context,
the generation of the input data was conducted through

Figure 1. Algorithm for multivariate modeling with 33 factorial design and ANN.
statistical design to produce the systematic data as input to
ANN instead of providing a large set of informal data
as input.
33 Factorial design
Design of Experiments (DoE) has pivotal importance in gen-
erating the dataset in a systematic way of achieving the opti-
mum model performance with a minimum number of
input. Among the several techniques, such as Box–Behnken
design (BBD), and central composite design (CCD), 33 fac-
torial design exhibits higher accuracy for prediction.[32] In a
3k factorial design, all possible combinations of the k factors
on three levels are investigated comprising a total number of
experimental run required is N3k ¼ 3k þ Cr; where, k is the
total number of factors and Cr is the central run replicates.
The total 27 number of hydraulic simulation was conducted
by varying the factors according to the 33 factorial model. In
this case, the central run is not required as the output is
generated from a hydraulic simulation software which is free
from pure error. The design run of 33 factorial design is
used as input to the ANN model.
JOURNAL OF ENVIRONMENTAL SCIENCE AND HEALTH, PART A 553
ANN model
Among the various techniques of multivariate modeling,
ANN is an efficient tool that can be applied to any informal
dataset to model both the linear or nonlinear relationships
of the independent and dependent variables.[33] The major
advantage of ANN is considered as their ability to learn
from the rule without depending on any mathematical rela-
tionship.[19] The most frequently used ANN architecture
consists of the input layer, the output layer and one or more
hidden layer forming a multilayer feed forward-back propa-
gation training network.[34] Several factors, like the complex-
ity of the problem, characteristics of input data, the error
goal and so forth influence the proper choice of the training
algorithm to obtain the optimum performance of the ANN.
In the present study, the number of perceptron in the hid-
den layer, training algorithms and transfer functions were
varied to develop the optimized ANN architecture.
Levenberg–Marquardt, resilient BP, BFGS quasi-Newton,
scaled conjugate gradient, Fletcher–Powell conjugate gradi-
ent, variable learning rate backpropagation, Polak–Ribiere
conjugate gradient and one step secant were chosen as the
554 P. S. GHOSAL ET AL.
different training algorithms. The input to the hidden layer
U is computed from input variables as follows:
fU g ¼ hWifI g (1)
where, W is the weight and I is the input. Each term of U
matrix can be presented as follows:
X
n
uj ¼ wi ii h (2)
i¼1
where h is the associated bias.
The linear (purelin) and the nonlinear functions, viz.,
log-sigmoid (logsig) and hyperbolic tangent sigmoid (tansig),
transfer functions were used in the hidden layer and are
presented in Equations (3)–(5), respectively:
f ðU Þ ¼ u (3)
f ðU Þ ¼
1
(4)
ð1 þ eðuÞ Þ
2 11,250 and 15,000 with 150 L per capita demand.
f ðU Þ ¼ (5)
ð1 þ eð2uÞ Þ 1
The mean square errors (MSE) and correlation coefficient
(R) were evaluated to appraise the performance of the ANN
model. The mean square errors (MSE) is represented as fol-
lows:
1X n
 2
MSE ¼ jYpred  Y exp j (6)
n i¼1
where Ypred and Yexp are the predicted and experimental
value of the dependent variable, respectively. The ANN ana-
lysis was performed using MATLAB, 2010a environment
(The MathWorks, INC.). The number of epochs, minimum
gradient and Mu was taken as 1000, 110 and 0.01, respect-
ively. The ANN architecture with the arrangement of the
input data, the hidden layer and the output layer has been
presented in Figure 1. The data of 33 factorial design was
randomly split into three portions, that is, 70% for training,
15% for validation and 15% for testing for framing the
ANN network.
Analysis of WDN
Hardy–Cross method, Newton–Raphson method, linear the-
ory method and linear graph theory method are few meth-
ods of solving WDN. However, it is quite cumbersome to
use these methods for solving large real WDN manually.
Hence, based on these analytical methods, several simulation
software packages have been developed to design and ana-
lyze complex functions of a distribution network facilities
using advanced computing. In the present work, hydraulic
simulator – WaterGEMS V8i (Bentley, USA) has been used
for design and analysis of water distribution network. In this
study, a hypothetical network was framed and used to assess
and compare the cost of the network by varying the influ-
encing factors. The WDN is presented in Figure S1 in
Supporting Information and was used for hydraulic analysis.
Table 1. Experimental range and levels of factors used in 33 factorial design.
Level
Factors –1 0 þ1
Input pressure (m H2O) 10 15 20
Network demand (L/s) 13.02 19.53 26.04
Supply hour (h) 4 14 24
Water distribution network with 55 pipeline and 31 junc-
tions with source reservoir were built in WaterGEMS simula-
tion software. The total length of the distribution network
pipeline was 9,688 m and minimum pipe diameter considered
was 100 mm. The design domain of the different independent
variables as per the 33 factorial design is presented in Table 1.
The network was optimized for every simulation using
Darwin designer tool in WaterGEMS. Pressure constraints
and velocity constraints were provided for every simulation.
Minimum and maximum pressure head considered for the
design was 7 m and 22 m, respectively.[35] Maximum velocity
constraint was restricted to 1.5 m/s.
The demand arrived for a residential population of 7500,
Accordingly, the discharge required at the source was calcu-
lated to be 13.02, 19.53 and 26.04 L/s. The system was ana-
lyzed for different pressure head at the source. Elevation of
the source was considered as 55.5 m, and 10 m, 15 m and
20 m pressure head at reservoir were considered for the ana-
lysis. Hence, the hydraulic gradient line was initiated with
65.5 m, 70.5 m and 75.5 m. Under each pressure head the net-
work was optimized and analyzed for three different supply
duration viz. 4 h, 14 h and 24 h. Fixed demand pattern was
considered for all the supply hours. Each experimental run
for different values of the simulation parameter is listed in
Table 2 concerning the 33 factorial design. The network was
optimized using Darwin designer by minimization of cost
subject to pressure and velocity constraints. The cost of the
pipe for different diameter was given in Table S1 in
Supporting Information. The Darwin designer performs with
genetic algorithm (GA) for the selected method of cost mini-
mization. In options, GA parameters were chosen as per the
recommended values of WaterGEMS to get the better results.
Results and discussion
Hydraulic design of different networks concerning 33
factorial design
The high requirement of the resources is associated with the
planning, designing, maintenance, operation and manage-
ment of the WDN. Any changes in factors affecting the
WDN involved a thorough hydraulic analysis of the net-
work. In those consequences, the major hindrance for the
decision makers is the unavailability of data for predicting
the system behavior and the financial involvement, which is
closely related to the physical parameters, such as the diam-
eter of the pipe of the network. The prediction of the net-
work adequacy is the involvement of many parameters and
their complex relationship. Suitable multivariate modeling
with an adequate prediction capacity may be the solution
for this concern.
 JOURNAL OF ENVIRONMENTAL SCIENCE AND HEALTH, PART A 555

Table 2. 33 factorial design cost as ANN input and the predicted cost from ANN model.
Std. order Input pressure (m) Demand (L/s) Duration (H) Actual cost of pipe (Million INRa) Predicted cost of pipe (Million INR a)
1 10 13.02 4 10.78582 10.78054799
2 15 13.02 4 9.8481 9.811654087
3 20 13.02 4 9.97938 9.980842496
4 10 19.53 4 11.7447 11.74732765
5 15 19.53 4 10.64111 10.63621111
6 20 19.53 4 11.21668 11.20864866
7 10 26.04 4 12.52188 12.47920058
8 15 26.04 4 11.22488 11.20886973
9 20 26.04 4 12.07116 12.07011329
10 10 13.02 14 9.37148 9.351650369
11 15 13.02 14 9.33968 9.331180613
12 20 13.02 14 9.33968 9.346317004
13 10 19.53 14 9.47262 9.476737023
14 15 19.53 14 9.41212 9.41344767
15 20 19.53 14 9.41212 9.414209771
16 10 26.04 14 9.57978 9.582185618
17 15 26.04 14 9.47792 9.480885333
18 20 26.04 14 9.47792 9.496223862
19 10 13.02 24 9.32888 9.331297426
20 15 13.02 24 9.32888 9.330934847
21 20 13.02 24 9.32888 9.33184668
22 10 19.53 24 9.33968 9.345947309
23 15 19.53 24 9.32888 9.333738007
24 20 19.53 24 9.32888 9.332367233
25 10 26.04 24 9.36128 9.357466229
26 15 26.04 24 9.37148 9.358441212
27 20 26.04 24 9.33968 9.335677044
a
Million INR¼ 14381.00 US $.

Figure 2. Variation of nodal pressure and pipe diameter at a. input pressure 10 m, supply hour 4 h, b. input pressure 10 m, supply hour 14 h, c. input pressure 10
m, supply hour 24 h, d. input pressure 15 m, supply hour 4 h, e. input pressure 15 m, supply hour 14 h, f. input pressure 15 m, supply hour 24 h, g. input pressure
20 m, supply hour 4 h, h. input pressure 20 m, supply hour 14, i. input pressure 20 m, supply hour 24 h for the flow rate of A. 13.02, B. 19.53 and C. 26.04 L/s.
556 P. S. GHOSAL ET AL.
Figure 2. Continued
To model the optimum distribution network with
selected hydraulic parameters, 27 scenarios of the water dis-
tribution network were simulated with a different combin-
ation of demand, input pressure and supply duration. The
scenarios created were optimized using Darwin designer tool
in the WaterGEMS. The new optimized network in the new
scenario was run, and the results were noted. In a few opti-
mized networks, the chronology of the pipe was missing and
they were manually corrected, and the results were checked
for satisfying the constraints. After confirming the chron-
ology of the pipe, the pipe table from flex table was copied,
and the cost of the network was calculated and are pre-
sented in Table 2. The pipe network concerning the vari-
ation of pipe diameter and nodal pressure were presented
for 13.02, 19.53 and 26.04 L/s flow rate at different supply
hour and input pressure as per the 33 factorial design run in
Figure 2A–C.
Variation of hydraulic parameters and pipe cost
The hydraulic design of the network conducted by restrict-
ing the design parameters, that is, velocity, pressure head
within the specified values as mentioned in the earlier sec-
tion. Each network was optimized for a minimum cost by
genetic algorithm in WaterGEMS. Hence, the major distin-
guishing parameter showing the hydraulic properties of the
network may be considered as pipe diameter and nodal
pressure. The increasing trend of nodal pressure concerning
to increase in input pressure and supply hours was estab-
lished from Figure 2A–C. However, the flow rate has an
antagonistic effect on the nodal pressure. The diameter of
the pipe in the network was decreased with increasing input
pressure and supply duration. On the other hand, the flow
rate has a synergistic effect on the diameter.
The cost of the water supply distribution network mainly
depends on the total cost of the pipe used in the network. It
is evident from Table 2 that the increase in supply hour
decreases the design cost of the water distribution system.
Supply hour also affects the flow rate of pipes. Increase in
supply duration decreases the flow rate of the network.
Increase in demand increases flow rates in the network,
meanwhile increase the diameter of the pipe to maintain the
velocity of flow and increases the cost of the system
(Fig. 2A–C).
The pressure in the network at any point is the difference
in hydraulic grade line and elevation of that point. The
hydraulic grade is the difference between inlet pressure and
head loss of the pipe. As input pressure increase, the inlet
pressure increases and hydraulic grade increase accordingly
the pressure in the system increases. Hence, head loss in the
system has to be decreased to obtain the required pressure.
Increasing the size of the network pipe decreases head loss

Figure 2. Continued
in the system eventually enhance the design cost of the dis-
tribution network. The outcome of the factorial design was
used as input to ANN.
ANN study
The multivariate modeling with the neural network was car-
ried out with the 27 numbers 33 factorial data points as the
input to the network. The proposed formulation was pre-
sented in Figure 1, and the design run was presented in
Table 2. The variation in the number of neurons in the hid-
den layer was first conducted to optimize the hidden layer.
Subsequently, the different training functions and transfer
functions were studied for the optimization of the ANN
model. In this study, eight training function and three trans-
fer function pertinent to this modeling were attempted to
optimize the network. The model outcome for these 24 cases
of network variation was presented in Table 3. The opti-
mized ANN architecture was chosen from the performance
indicator as MSE and R.
Validation of model
The modeling of the water distribution has to be validated.
For making comprehensive use of the validation experiment
of the model, a 32 factorial design with nine sets of scenario,
JOURNAL OF ENVIRONMENTAL SCIENCE AND HEALTH, PART A 557
with the combination of demand, input head and supply
duration was created in the WaterGEMS and were opti-
mized using Darwin designer tool. The simulation order of
the ANN model for the validation experiment is presented
in Table 4. Typically, the intermittent supply duration in
different cities of India was found to target 8 h. Hence, the
same was considered for validation of the model.
Optimization of number of neurons in hidden layer
The performance of the ANN is reliant on the computation
carried out in the hidden layer, which in turn dependent on
the number of perceptrons adopted. The inappropriate
choice of the number of neurons creates either a complex or
an under-performing network. The optimization of the
number of the neuron for the hidden layer configuration is
an integral part of the efficient ANN model. The numbers
of neurons were varied from 4 to 20, and the corresponding
mean square of error (MSE) and R were represented in
Figure 3. The network with 10 neurons in hidden layer
exhibited the optimum performance with an MSE of
1.744  104, and corresponding R is above 0.999 (Fig. 3).
The correlation statistics of the optimum number of neurons
for training, testing, validation and overall ANN model are
0.999, 0.997, 0.999 and 0.999, respectively (Fig. 3).
558 P. S. GHOSAL ET AL.

Table 3. Performance study of training algorithm and transfer function of ANN model.
Sl. No. Training algorithm Transfer function MSEa RMSEb Rc
4
1 Levenberg–Marquardt tansig 1.744  10 0.0132 0.999
logsig 0.0106 0.1031 0.996
purelin 0.0801 0.2830 0.967
2 BFGS quasi Newton tansig 0.0149 0.1219 0.992
logsig 0.0118 0.1085 0.994
purelin 0.0592 0.2434 0.967
3 Scaled conjugate gradient tansig 0.0022 0.0467 0.999
logsig 0.0120 0.1094 0.994
purelin 0.0602 0.2454 0.964
4 Resilient BP tansig 0.0133 0.1061 0.994
logsig 0.0183 0.1353 0.990
purelin 0.0885 0.2974 0.961
5 Polak–Ribiere conjugate gradient tansig 0.0026 0.0506 0.999
logsig 0.0071 0.0842 0.996
purelin 0.0707 0.2659 0.963
6 Fletcher–Powell conjugate gradient tansig 0.0116 0.1076 0.994
logsig 0.0040 0.0632 0.998
purelin 0.0618 0.2487 0.968
7 Variable learning rate backpropagation tansig 0.0104 0.1022 0.995
logsig 0.0059 0.0769 0.997
purelin 0.0685 0.2618 0.965
8 One step secant tansig 0.0094 0.0969 0.996
logsig 0.0169 0.1301 0.992
purelin 0.0627 0.2503 0.963
a
MSE is mean square errors.
b
RMSE is root mean square errors.
c
R is corellation coefficient.

Table 4. Validation set of 32 factorial design results.

Std. order Input pressure (m) Demand (L/s) Duration (h) Actual cost of pipe (Million INRa) Predicted cost of pipe (Million INRa)
1 10 13.02 8 9.57178 9.794936
2 15 13.02 8 9.42652 9.397797
3 20 13.02 8 9.46712 9.446128
4 10 19.53 8 9.7178 10.35198
5 15 19.53 8 10.07302 9.774478
6 20 19.53 8 9.5701 9.859356
7 10 26.04 8 10.72252 11.69009
8 15 26.04 8 9.8216 9.701655
9 20 26.04 8 9.97938 10.10908
a
Million INR¼ 14381.00 US $.

Selection of training algorithm and transfer function
The performance of the ANN is largely dependent on the
training algorithm, and the transfer function chosen for the
network analysis. In this study, a set of 24 network run was
conducted with eight training algorithm and corresponding
three transfer functions. The performance indicator as MSE
and R was presented for each training algorithm, such as
Levenberg–Marquardt, resilient BP, BFGS quasi-Newton,
scaled conjugate gradient, Fletcher–Powell conjugate gradi-
ent, variable learning rate backpropagation, Polak–Ribiere
conjugate gradient and one step secant and corresponding
log-sigmoid, hyperbolic tangent sigmoid transfer function
and linear transformation. The R2 and error functions were
computed for the models generated from 33 factorial data
(Table 3). The Levenberg–Marquardt training algorithm and
tansig transfer function exhibited the optimum results.
the actual values showing very good predictability of the
ANN model. The validation experiment for the 32 has also
been simulated by ANN model, and the predicted values
and actual values are shown in Figure 4B. The predicted
points are coming to the close proximity of the actual value
which shows the good predictability of the ANN model.
Nevertheless, the predicted values for the original ANN
model is obviously higher than that of the validation set.
The performance of the ANN model was further assessed
with some error functions and analysis of variance and
regression statistics. The error functions used in this study
are as follows:
The Sum of the square of the error[36]
X
n
SSE ¼ ðYpred  Y exp Þi
2
(7)
i¼1

The sum of absolute error[36]

Performance of ANN model
X
n
The predicted values of the ANN model for 3 model was 3 SAE ¼ jYpred  Y exp ji (8)
i¼1
computed. The actual values and the predicted values are
shown in Figure 4A. The predicted values are very close to
 JOURNAL OF ENVIRONMENTAL SCIENCE AND HEALTH, PART A 559

Figure 3. (a) Variation of MSE and R with the number of neurons in the hidden layer and (b) R of optimum number of neurons for training, testing, validation and
overall ANN model.

The average relative error[36] The chi-square test statistic[37]

" #
100 X
n
Ypred Y exp X n
ðYpred  Y exp Þ2
ARE ¼ j ji (9) v ¼
2
(12)
n i¼1 Y exp Ypred
i¼1
The hybrid fractional error function[36]
" # where, n – p is the degree of freedom; n is the number of
100 X n
ðYpred  Y exp Þ2 the data point and p is the number of parameters. The per-
HYBRID ¼ (10)
n  p i¼1 Y exp formance of ANN model for both the cases was further ana-
i
lyzed by computing the error functions, and conducting an
The Marquardt’s percent standard deviation[36] analysis of variance (ANOVA); the results were presented in
Table 5 and Table S2 in Supporting Information, respect-
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
" #2ffi ively. The SSE, SAE, ARE, HYBRID and MPSD error and v2
u
u 1 X n
ð Y  Y Þ values were also appeared significantly lower for the model
MPSD ¼ 100t
pred exp
(11) from the 33 factorial set data than that of the validation set.
n  p i¼1 Y exp
i The v2 value for both the cases was appeared the lowest as
560 P. S. GHOSAL ET AL.

Figure 4. Performance of ANN model for (a) 33 factorial design, (b) validation experiments (The cost is in Million INR; 1 Million INR¼ 14381.00 US $.).

Table 5. Comparison of experimental data with predicted values of ANN models generated from 33 factorial data
and validation data with 32 factorial design.
Test statistic ANN from 33 factorial data ANN for validation set
The Sum of the square of the error 0.005 1.593
The sum of absolute error 0.228 2.712
The average relative error 0.083 2.991
The hybrid fractional error function 0.002 1.935
The Marquardt’s percent standard deviation 0.036 4.340
The v2 test statistic 0.0004 0.145

0.0004 and 0.145, respectively. Whereas, the sum of absolute
error was the highest for 33 factorial model as 0.228 and the
Marquardt’s percent standard deviation was the highest for
the validation set. Furthermore, the F value of the model
from 33 factorial data was significantly high as 162,442, and
the P value was less than 0.001, which in turn established
the significance of the ANN model. However, the same test
statistics was 19.480 and 0.004 for the model for validation
sets. The higher R2, adjusted R2, F value and the lesser P
value and error functions in the ANN model from 33 factor-
ial set data attributed to the better performance level com-
pared to the validation set of 32 factorial design.
In the present study, the method of multivariate model-
ing through ANN can be used as a predictive and estimating

tool for WDN. In most of the cases, the WDN suffers from
a change in demand, supply duration and so forth for which
the thorough hydraulic simulation and redesign of the dis-
tribution network is required. Eventually, the authorities
face the hurdle to predict the cost requirement of the system
as well as the suitable solution to handle the changed situ-
ation, such as imposing additional head and changing sup-
ply duration In this context, the multivariate model can
perform as a predictive tool for the decision makers as well
as for the assessment of the interactive behavior of the sys-
tem variables.
Conclusions
An established prediction model on the design of a distribu-
tion system was rarely developed, and the assessment of the
interactive effect of the system parameters like pressure and
flow was also seldom explored. Subsequently, the multivari-
ate approach in the present study was conducted for the first
time to model the influential system parameters concerning
the cost of the pipe network. Generally, vivid network ana-
lysis is the solution to assess the cost of the system in differ-
ent scenarios. Whereas, the water supply authorities/agencies
often require to find out the cost of the system for the vari-
ous combinations of the process variables, for example, the
changed demand, extension or reduction of supply hour and
so forth.
Apart from that, the individual influence of each param-
eter can be assessed by conducting the univariate study.
However, the actual behavior of the system with a variation
of all the important parameters was rarely evaluated.
Nevertheless, the required process dynamics lies in the
assessment of the effect of all parameters at a time. The
approach of the present study explored this gap area. The
performance analysis of the multivariate model through the
ANN and 33 factorial design exhibited that it can serve as a
prediction model. The model can assess the combined effect
of the different parameters as well as it can serve as a deci-
sion tool for the authorities to assess the effect of changed
scenario without performing the vigorous hydraulic analysis
of the system.
Disclosure statement
No potential conflict of interest was reported by the authors.
References
[1] Vairavamoorthy, K.; Gorantiwar, S. D.; Pathirana, A. Managing
Urban Water Supplies in Developing Countries—Climate
Change and Water Scarcity Scenarios. Phys. Chem. Earth 2008,
33, 330–339. doi:10.1016/j.pce.2008.02.008.
[2] Ilaya-Ayza, A. E.; Campbell, E.; P
erez-Garc
ıa, R.; Izquierdo, J.
Network Capacity Assessment and Increase in Systems with
Intermittent Water Supply. Water (Switzerland) 2016, 8, 126.
doi:10.3390/w8040126.
[3] Fontanazza, C. M.; Freni, G.; Loggia, G. L. Analysis of intermit-
tent supply systems in water scarcity conditions and evaluation
of the resource distribution equity indices. WIT Trans. Ecol.
Environ. 2007, 103, 635–644.
JOURNAL OF ENVIRONMENTAL SCIENCE AND HEALTH, PART A 561
[4] Totsuka, N.; Trifunovic, N.; Vairavamoorthy, K. Intermittent
Urban Water Supply Under Water Starving Situations.
Presented at the 30th WEDC International Conference,
Vientiane, Lao PDR, 2004; pp. 505–512.
[5] Abu-Madi, M.; Trifunovic, N. Impacts of Supply Duration on
the Design and Performance of Intermittent Water Distribution
Systems in the West Bank. Water Int. 2013, 38, 263–282. doi:
10.1080/02508060.2013.794404.
[6] Kumpel, E.; Nelson, K. L. Intermittent Water Supply:
Prevalence, Practice, and Microbial Water Quality. Environ. Sci.
Technol. 2016, 50, 542–553. doi:10.1021/acs.est.5b03973.
[7] Ilaya-Ayza, A. E.; Martins, C.; Campbell, E.; Izquierdo, J.
Gradual Transition from Intermittent to Continuous Water
Supply Based on Multi-Criteria Optimization for Network
Sector Selection. J. Comput. Appl. Math. 2018, 330, 1016–1029.
doi:10.1016/j.cam.2017.04.025.
[8] Yabalak, E. Degradation of Ticarcillin by Subcritical Water
Oxidation Method: Application of Response Surface
Methodology and Artificial Neural Network Modeling. J.
Environ. Sci. Health A 2018, 53, 1–11.
[9] Ma, Y.; Huang, M.; Wan, J.; Hu, K.; Wang, Y.; Zhang, H.
Hybrid Artificial Neural Network Genetic Algorithm Technique
for Modeling Chemical Oxygen Demand Removal in Anoxic/
Oxic Process. J. Environ. Sci. Health A 2011, 46, 574–580. doi:
10.1080/10934529.2011.562821.
[10] Ghosal, P. S.; Gupta, A. K. Sorptive Equilibrium Profile of
Fluoride onto Aluminum Olivine [(FexMg1-x)2SiO4] Composite
(AOC): Physicochemical Insights and Isotherm Modeling by
Non-Linear Least Squares Regression and a Novel Neural-
Network-Based Method. J. Environ. Sci. Health A, 53, 2018,
1–13. doi:10.1080/10934529.2018.1474590.
[11] Chronopoulos, K. I.; Tsiros, I. X.; Dimopoulos, I. F.; Alvertos,
N. An Application of Artificial Neural Network Models to
Estimate Air Temperature Data in Areas with Sparse Network
of Meteorological Stations. J. Environ. Sci. Health A 2008, 43,
1752–1757. doi:10.1080/10934520802507621.
[12] Flores-Asis, R.; M
endez-Contreras, J. M.; Ju
arez-Mart
ınez, U.;
Alvarado-Lassman, A.; Villanueva-V
asquez, D.; Aguilar-Lasserre,
A. A. Use of Artificial Neuronal Networks for Prediction of the
Control Parameters in the Process of Anaerobic Digestion with
Thermal Pretreatment. J. Environ. Sci. Health A 2018, 53,
883–890. doi:10.1080/10934529.2018.1459070.
[13] Hu, K.; Wan, J. Q.; Ma, Y. W.; Wang, Y.; Huang, M. Z. A
Fuzzy Neural Network Model for Monitoring A2/O Process
Using On-Line Monitoring Parameters. J. Environ. Sci. Health
A 2012, 47, 744–754. doi:10.1080/10934529.2012.660102.
[14] Kim, M.-Y.; Kim, M.-K. Dynamics of Surface Runoff and Its
Influence on the Water Quality Using Competitive Algorithms
in Artificial Neural Networks. J. Environ. Sci. Health A 2007,
42, 1057–1064. doi:10.1080/10934520701418490.
[15] € Eco-friendly Approach to Mineralise 2-
Yabalak, E.; Yilmaz, O.
Nitroaniline Using Subcritical Water Oxidation Method: Use of
ANN and RSM in the Optimisation and Modeling of the
Process. J. Iran. Chem. Soc. 2018, 6, 1–10.
[16] _ Modeling and Optimization II: Comparison
Baş, D.; Boyacı, IH.
of Estimation Capabilities of Response Surface Methodology with
Artificial Neural Networks in a Biochemical Reaction. J. Food
Eng. 2007, 78, 846–854. doi:10.1016/j.jfoodeng.2005.11.025.
[17] Desai, K. M.; Survase, S. A.; Saudagar, P. S.; Lele, S. S.; Singhal,
R. S. Comparison of Artificial Neural Network (ANN) and
Response Surface Methodology (RSM) in Fermentation Media
Optimization: Case Study of Fermentative Production of
Scleroglucan. Biochem. Eng. J. 2008, 41, 266–273. doi:10.1016/
j.bej.2008.05.009.
[18] Gupta, A. K.; Ghosal, P. S.; Srivastava, S. K. Modeling and
Optimization of Defluoridation by Calcined Ca-Al-(NO3)-LDH
Using Response Surface Methodology and Artificial Neural
Network Combined with Experimental Design. J. Hazard.
Toxic. Radioact. Waste 2017, 21, 04016024. doi:10.1061/
(ASCE)HZ.2153-5515.0000343.
562 P. S. GHOSAL ET AL.
[19] Abbasi, B.; Mahlooji, H. Improving Response Surface
Methodology by Using Artificial Neural Network and Simulated
Annealing. Expert Syst. Appl. 2012, 39, 3461–3468. doi:10.1016/
j.eswa.2011.09.036.
[20] Ostfeld, A.; Oliker, N.; Salomons, E. Multiobjective
Optimization for Least Cost Design and Resiliency of Water
Distribution Systems. J. Water Resour. Plann. Manage. 2014,
140, 04014037. doi:10.1061/(ASCE)WR.1943-5452.0000407.
[21] Yazdani, A.; Jeffrey, P. Applying Network Theory to Quantify
the Redundancy and Structural Robustness of Water
Distribution Systems. J. Water Resour. Plann. Manage. 2012,
138, 153–161. doi:10.1061/(ASCE)WR.1943-5452.0000159.
[22] Aksela, K.; Aksela, M.; Vahala, R. Leakage Detection in a Real
Distribution Network Using a SOM. Urban Water J. 2009, 6,
279–289. doi:10.1080/15730620802673079.
[23] Rathi, S.; Gupta, R. Optimal Sensor Locations for
Contamination Detection in Pressure-Deficient Water
Distribution Networks Using Genetic Algorithm. Urban Water
J. 2017, 14, 160–172. doi:10.1080/1573062X.2015.1080736.
[24] Tabesh, M.; Shirzad, A.; Arefkhani, V.; Mani, A. A
Comparative Study Between the Modified and Available
Demand Driven Based Models for Head Driven Analysis of
Water Distribution Networks. Urban Water J. 2014, 11,
221–230. doi:10.1080/1573062X.2013.783084.
[25] Song, I.; Romero-Gomez, P.; Andrade, M. A.; Mondaca, M.;
Choi, C. Y. Mixing at Junctions in Water Distribution Systems:
An Experimental Study. Urban Water J. 2018, 15, 32–38. doi:
10.1080/1573062X.2017.1364395.
[26] Wang, M.; Barkdoll, B. D. A Sensitivity Analysis Method
for Water Distribution System Tank Siting for Energy
Savings. Urban Water J. 2017, 14, 713–719. doi:10.1080/
1573062X.2016.1241285.
[27] Sadiq, M.; Zaidi, T. H.; Muhanna, H.; Al.; Mian, A. A. Effect of
Distribution Network Pipe Material on Drinking Water Quality.
J. Environ. Sci. Health A Environ. Sci. Eng. Toxicol. 1997, 32,
445–454. doi:10.1080/10934529709376553.
[28] Clark, R. M. Water Quality Modeling in Distribution Systems.
J. Environ. Sci. Health A Environ. Sci. Eng. Toxicol. 1992, 27,
1329–1366.
[29] Wu, Z. Y. Optimal Calibration Method for Water Distribution
Water Quality Model. J. Environ. Sci. Health A 2006, 41,
1363–1378. doi:10.1080/10934520600657115.
[30] Abbaszadegan, M.; Yi, M.; Alum, A. Stimulation of 2-
Methylisoborneol (MIB) Production by Actinomycetes After
Cyclic Chlorination in Drinking Water Distribution Systems. J.
Environ. Sci. Health A 2015, 50, 365–371. doi:10.1080/
10934529.2015.987526.
[31] Vasan, A.; Simonovic, S. P. Optimization of Water Distribution
Network Design Using Differential Evolution. J. Water Resour.
Plann. Manage. 2010, 136, 279–287. doi:10.1061/(ASCE)0733-
9496(2010)136:2(279).
[32] Ghosal, P. S.; Gupta, A. K.; Sulaiman, A. Multivariate
Optimization of Process Parameters in the Synthesis of
Calcined CaAl (NO3) LDH for Defluoridation Using 33
Factorial, Central Composite and Box-Behnken Design. J.
Environ. Sci. Health. A Tox. Hazard. Subst. Environ. Eng. 2016,
51, 86–96. doi:10.1080/10934529.2015.1086212.
[33] Agami, N.; Atiya, A.; Saleh, M.; El-Shishiny, H. A Neural
Network Based Dynamic Forecasting Model for Trend Impact
Analysis. Technol. Forecast. Soc. Change 2009, 76, 952–962.
doi:10.1016/j.techfore.2008.12.004.
[34] Uddameri, V. Using Statistical and Artificial Neural Network
Models to Forecast Potentiometric Levels at a Deep Well in
South Texas. Environ. Geol. 2007, 51, 885–895. doi:10.1007/
s00254-006-0452-5.
[35] CPHEEO. Manual on Water Supply and Treatment. 1999.
Ministry of Urban Development: New Delhi, India.
[36] Kundu, S.; Gupta, A. K. Arsenic Adsorption Onto Iron Oxide-
Coated Cement (IOCC): Regression Analysis of Equilibrium
Data with Several Isotherm Models and Their Optimization.
Chem. Eng. J. 2006, 122, 93–106. doi:10.1016/j.cej.2006.06.002.
[37] Ayoob, S.; Gupta, A. K. Insights into Isotherm Making in the
Sorptive Removal of Fluoride from Drinking Water. J. Hazard.
Mater. 2008, 152, 976–985. doi:10.1016/j.jhazmat.2007.07.072.