Desalination and Water Treatment 104 (2018) 99–110

February
www.deswater.com
doi: 10.5004/dwt.2018.21911

Optimal design of water distribution networks using simple modified particle

swarm optimization approach

Alireza Moghaddama,*, Amin Alizadehb, Alireza Faridhosseinib, Ali Naghi Ziaeib,

Danial Fallah Heravic
a
Department of Water Engineering, University of Urmia, Urmia, Iran, email: [email protected]
b
Department of Water Engineering, Ferdowsi University of Mashhad, Mashhad, Iran, emails: [email protected] (A. Alizadeh),
[email protected] (A. Faridhosseini), [email protected] (A.N. Ziaei)
c
Department of Mechanical Engineering, Ferdowsi University of Mashhad, Mashhad, Iran, emails: [email protected]

Received 17 February 2017; Accepted 15 January 2018

abstract
Water distribution networks design belongs to a class of large combinatorial nonlinear optimization
problems, involving complex implicit constraints, such as conservation of mass and energy equations,
which are commonly satisfied by using hydraulic simulation solvers. Recently, many researchers have
shifted the focus of traditional optimization methods to the use of metaheuristic approaches for han-
dling this complexity. Particle swarm optimization (PSO) is one of the evolutionary algorithms which
was developed for optimization problems with continuous variables. Also it has been adapted success-
fully in other problems contexts with discrete variables. In this research, a simple modified particle
swarm optimization (SMPSO) was applied to minimize water distribution networks cost. The SMPSO
was used as novel factor to decrease inertia weight linearly with time for each iteration to facilitate
the balance of global and local researches. The SMPSO algorithm was linked to a hydraulic simulator,
EPANET 2.0. This approach was applied to three benchmark in water distribution network optimi-
zation problems. The results indicate that a significant improvement in performance of PSO could be
achieved by decreasing inertia weight over the iterations.
Keywords: SMPSO; Inertia weight; EPANET 2.0; Water distribution networks

1. Introduction In the last decade, several new nontraditional optimiza-

tion methods for such non-deterministic polynomial-time
The efficiency is one of the main elements in designing
(NP)-hard problems, which contain nonlinear, constrained,
new water distribution networks (WDNs). The optimal
non-smooth, non-convex, and multimodal functions, have
design of WDNs has been studied comprehensively over
been explored [1]. At first, most of the optimization tech-
the past few decades due to its computational and engi-
niques made some initial solution, then using deterministic
neering complexity. Most of these studies focused on
search methods, until no more reduction in cost took place.
the least-cost optimization. However, it is necessary to
Therefore, the final solution depended on the initial solution.
investigate the reliability of network design to ensure the
Yates et al. [2] stated that either explicit enumeration or
sufficient head. Moreover the existence nonlinear relation
an implicit enumeration technique such as dynamic pro-
between flow, head-loss, and discrete variables such as
gramming could guarantee the optimal solution to NP-hard
pipe diameter in optimal design of WDNs is a highly
WDN design problem. For example, in a WDN with 20 num-
challenging problem.
ber of pipes and 10 commercially available pipe sizes, the
total number of solutions is 1020, giving a very wide search
* Corresponding author. space. Therefore, the complete enumeration method for real

1944-3994/1944-3986 © 2018 Desalination Publications. All rights reserved.

100 A. Moghaddam et al. / Desalination and Water Treatment 104 (2018) 99–110

size WDN problems becomes stubborn. Gessler [3] proposed 2. Optimal design of a water distribution network
a partial enumeration technique in which certain inferior
The optimal design of a WDN is often noticed as a least-
solutions were rejected from being evaluated by the hydrau-
cost optimization problem. The decision variables are the
lic simulation model. In addition, partial enumeration tech-
diameters of each pipe in WDN. The optimal solution is
niques are difficult to use for the optimal design in large size
obtained by minimizing the total cost. For a given layout,
realistic WDNs.
the source head, elevation and demand values for nodes,
Many studies in literature have focused on minimizing
pipe lengths and pipe roughness are known in advance. The
the costs of the objective function in optimization of WDNs
objective is to find a combination of different sizes of pipe
such as linear programming, nonlinear programming (NLP),
that can satisfy the nodal head constraints at the lowest cost.
enumeration techniques, heuristic methods, and evolution-
In order to facilitate the comparison of results obtained by
ary techniques. Alperovits and Shamir [4] applied the linear
other authors, the following objective function was used to
programming gradient method that is a linearization model.
minimize the cost for a WDN by Eq. (1):
Many researchers [5–7] have developed approaches for
obtaining the global optimum and dominated by lineariza-
Fobj = ∑ i =1 C Di Li 
n
tion technique limitations. Moreover, many researchers [8–13] (1)
have applied the NLP optimization approach to pipe network
problems due to the nonlinear nature of these problems. The
NLP techniques do not guarantee the identification of global where Di, Li, and CDi are the diameter, the length and the unit
optimal solution because they depend on the initial solution cost of the i-th pipe, respectively, and n is the total number
and practice of discrete variables, such as commercial pipe of pipes in the network. Typically, the constraints of WDNs
diameters, reduces the quality of optimal solution [14]. optimization include: flow continuity at each node, energy
Additionally, the research applied the stochastic optimi- conservation in each primary loop, and the minimum allow-
zation models such as genetic algorithms (GAs), simulated able head requirement at each node. These constraints can be
annealing (SA), harmony search optimization (HS), shuf- mathematically expressed as:
fled frog leaping algorithm (SFLA), ant colony optimization
(ACO), differential evolution (DE), and shuffled complex q inj − q out
j
− qj = 0 j = 1, 2 ,… , d j ;  (2)
evolution (SCE) in optimal design of WDNs. The search strat-
egy in most of these models is based on the objective function
values to move to a better solution in successive iterations
and efficient in handling discrete variables. Some research-
(∑ npl
i =1
HLi ) L
=0 L = 1, 2 ,… , nL;  (3)

ers [15–24] applied GAs for solving network design prob-

lems. Simulated annealing used by Loganathan et al. [25]
H j ≥ H jmin j = 1, 2 ,… , nd;  (4)
and Cunha and Sousa [26]. Geem et al. [27], Geem [28], Yazdi
[29], and Eusuff and Lansey [30] developed the HS method
and the SFLA. Maier et al. [31] used the ACO approach and Dmin ≤ D ≤ Dmax (5)
outperformed GAs both in terms of computational efficiency
and their ability to find near global optimal solutions. Shie- in
where q j is the flow entering at node j, q j is the flow leav-
out
Yui Liong [32] and Vasan and Simonovic [14] applied the SCE
and DE approach, respectively. ing at node j towards the downstream nodes, qj is the demand
Particle swarm optimization (PSO) is one of the evo- at node j, HLi is the head loss in pipe i, npl is the number of
lutionary algorithms, which has proven its possibility and pipes in a loop, nL is the number of loops in the WDN, Hj
min
performance in solving various optimization problems is the hydraulic head available at node j, H j is the mini-
[33–37]. The PSO algorithm was developed by Kennedy and mum hydraulic head required at node j, dj is the number of
Eberhart [38] and inspired by the social behavior of a group demand nodes, and in this context, Dmin and Dmax are the min-
of migrating birds trying to reach an unknown destination. imum and maximum allowable pipe sizes, respectively. The
This algorithm with certain modifications was used in this loop refers to the closed circuit formed by the pipes. Eq. (2)
research to find solutions for the optimal design in water is referred to as the nodal mass balance equation; Eq. (3) is
supply networks. referred to as the loop energy balance equation; Eq. (4) is the
In this study, a simple modified PSO (SMPSO) is intro- minimum hydraulic head requirement constraint and Eq. (5)
duced initially. This model uses a new constant factor to is the constraint for the pipe diameters.
decrease inertia weight linearly using time for each itera- The head loss in each pipe is the head difference between
tion. This strategy can significantly facilitate the balance of connected nodes, and can be computed using the Hazen–
global and local searches. In the following steps, the per- Williams equation:
formance of this algorithm is evaluated for two standard
benchmark networks, and the results are compared with the Li
previous studies. Also, the sensitive analysis was performed HL = ω qiα  (6)
C Diβ
α
i
for determining the best parameter values of SMPSO algo-
rithm on the network. The optimization problems addressed
herein linked with hydraulic simulator of EPANET 2.0 [39]. where ω is a numerical conversion constant (dependent on
The goal is to minimize the cost, with pipe diameter as deci- units); Ci is roughness coefficient of pipe i (dependent on
sion variables. material); α and β are regression coefficients.
 A. Moghaddam et al. / Desalination and Water Treatment 104 (2018) 99–110 101

Researchers have used different values for the numerical and velocity Vit = ( vi 1 , vi 2 ,..., vis ) at iteration t. Each parti-
conversion constant ω and regression coefficients α and β. cle keeps tracking of its position vector pbest, which has
The higher constant ω, the greater the head loss is and vice achieved the best fitness function so far. The position vector
versa. Thus, an optimal solution with higher value of ω will gbest, which is the best value of fitness function, obtained by
be costlier than solution with lower value of ω [28]. Savic and any particle so far that is also remembered. The values of the
Walters [18] reported the smallest and largest value of ω used fitness function for these are stored. The PSO concept con-
in the literature as 10.5088 (α = 1.85, β = 4.87) and 10.9031 sists of changing the velocity of each particle toward its pbest
(α = 1.852, β = 4.87). Cunha and Sousa [26] and Geem et al. and gbest. Once the velocities are determined, then position
[27] used ω = 10.5088 (α = 1.85, β = 4.87). Eusuff and Lansey vectors of the particles will be updated. At these updated
[30] and Shie-Yui Liong [32] coupled their algorithms with positions, the fitness function is recalculated and the position
EPANET 2.0 in which ω = 10.667 (α = 1.852, β = 4.871). vectors pbest and gbest are updated. This process continues
To solve the problem mentioned above, the constrained until the given iterations are over. The following equations
model is converted into an unconstrained one by adding were used, which iteratively modify the particle velocities Vijt
the amount of constraint violations to the objective function and positions Xijt at iteration number t: [37,40]
as penalties. Although the conservation of mass and energy
constraints are satisfied externally via EPANET 2.0 [32], the
pressure constraint is required to be considered in the penalty ( ) ( )
Vijt +1 = wVijt + c1r1t pbest ( ij ) − Xijt + c2 r2t gbest ( j) − Xijt  (9)
costs. Thus, the total cost of the network is considered as the
sum of the network cost and a penalty cost is defined as Eq. (7):
Xijt+ 1 = Xijt + Vijt  (10)
Fobj = ∑ i =1 C Di Li + PF 
n
(7)
where i = [1, 2,..., P] and j = [1, 2,..., n]. c1 and c2 are accelera-
tion constants and r1, r2 are random numbers between [0,1].
The penalty function PF only applies when the pressure The position vector gbest (global best position) and pbest
in any node is less than a predetermined minimal value. For (particle best position) are modified during the iteration.
nodes with pressure larger than this minimal value, the asso- Proper fine-tuning of the parameters c1 and c2 in Eq. (9) may
ciated individual penalties are vanished, and 1 is used as the result in faster convergence of the algorithm, and allevia-
usual Heaviside step function ϑHeaviside in the explicit expres- tion of the problem of local minima. To control the changes
sion for PF as Eq. (8): in velocity, Clerc [41] introduced the constriction factor into
the standard PSO algorithm to ensure the convergence of
( ) ( )
PF = ∑ j =1 ϑHeaviside H jmin − H j .a. H jmin − H j 
nd
(8) the search. The role of inertial weight w in Eq. (9) is con-
trolling the impact of previous velocities on the current one.
A large inertial weight facilitates global exploration (search-
where a is the penalty multiplier that is defined by user and ing new areas), while a small weight tends to facilitate local
in this study is assumed to be 9 × 109. exploration. Hence, selection of a suitable value for the iner-
tial weight w usually helps in reduction of the number of
3. Hydraulic simulator: EPANET 2.0 iterations that required to locate the optimum solution [42].
Shi and Eberhart [43], Shi and Eberhart [44] suggested that
EPANET 2.0 is a robust model which is used by a large the allowable of w changes between 0.4 and 0.9, in standard
community of users in the world in order to run the hydrau- PSO algorithm.
lic simulations of the WDNs [39]. It combines all the main In this research, a SMPSO is present by using a reduction
infrastructures of supplying systems, such as gravity and factor, wdamp, to adjust the convergence speed of an algorithm
pump systems, valves (e.g., relief, pressure reducing, regu- to find the optimal solution. It is important to determine the
lating, control and isolation valves), reservoirs (of fixed or appropriate value of wdamp, as it reduces w following a linear
variable level), by which it is possible to make operating con- form in each iteration:
ditions. EPANET 2.0 calculates flow in each pipe, pressure in
each node, water level in each reservoir and concentration of
chemicals during the simulation period. It considers the bal- w t + 1 = w t .wdamp  (11)
ance conditions, for a set of equations, using the method of
gradient and runs static and quasi-steady simulations of the
hydraulic and water quality situation of pipe network [27]. To manage any changes in the particle velocities, the
relevant upper and lower limits were defined as follows:

4. Simple modified particle swarm optimization algorithm

Vmin ≤ V ≤ Vmax (12)
PSO is a promising new optimization technique devel-
oped by Kennedy and Eberhart [38] which models a set of The standard PSO algorithm is applicable to continu-
potential solutions as a swarm of particles moving about in a ous problems and cannot use for discrete problems. Various
virtual search space. The method was inspired by the move- approaches were put forward to tackle discrete problems with
ment and interaction of flocking birds with their neighbors in PSO [34,35]. Essentially, this algorithm only takes integer parts
the group. A swarm of P particles optimizes in n-dimension of flying velocity vector components into account. Following
search space. Each particle i has position Xit = ( xi 1 , xi 2 ,..., xis ) the new velocity Vijt +1, that is an integer, the new position
102 A. Moghaddam et al. / Desalination and Water Treatment 104 (2018) 99–110

vector components also will be integer (Eq. (13)). As a result, 5.1. Two-loop network
the initial position vectors are generated with integer values.
The two-loop network, which is shown in Fig. 1, was orig-
inally presented by Alperovits and Shamir [4]. The network
( ( ) (
Viji +1 = round wVijt + c1r1t pbest ( ij ) − Xijt + c2 r2t gbest ( j) − Xijt )) has seven nodes and eight pipes with two loops and is fed
by gravity from a reservoir with a 210 m fixed head. Nodal
 (13)
demands and elevations are given in Table 1. The pipes are all
1,000 m length with the assumed Hazen–Williams coefficient
For discrete variables, round() is a function that takes
of 130. The required minimum head of other nodes is 30 m
the integer part of its argument. The particle velocity is com-
above ground level. There are 14 commercial diameters for
puted by Eq. (13) and follows exactly the limits which are
selection and costs for each pipe size are given in Table 2.
established by Eq. (12). Vmax is calculated by Eq. (14):
Thus, the problem search space consists of 148 different net-
work designs, which made this illustrative example difficult
Vmax = 0.5. ( Xmax − Xmin )  (14) to solve [18].
The results of this study were compared with the pre-
vious researches, which solved this problem by different
where Xmax and Xmin are maximum and minimum diameters
evolutionary algorithms such as GA, SA, SFLA, HS, and SS
that can be considered for each network.
(Table 3). According to Table 3, the cost obtained due to the
optimization with PSO algorithm is $419,000, which is the
5. Testing the benchmark problems minimum cost reported for this network so far and obtained
after 3,100 times of number of function evaluation (NFE)
The performance of developed SMPSO-based model for
(Fig. 2), whereas other methods reached this cost after at
optimization of WDN design problem is evaluated through
least 3,215 times of NFE. The constant of Hazen–Williams
three well-known benchmark case studies: the two-loop net-
equation (ω, Eq. (6)) was considered 10.667, 10.5088, 10.5088,
work, the Hanoi network, and the Kadu network. For each
and 10.55879 for PSO, GA, SA, and HS algorithms, respec-
case study, a preliminary sensitivity analysis was performed
tively. The increase in ω coefficient will increase the head
to determine the effective parameter values of the SMPSO
loss of the pipes, so, larger pipe diameters should be selected.
algorithm on the basis of the range that was suggested by
Clerc and Kennedy [45].
Table 2
Pipe sizes and costs for two-loop network

Pipe number Diameter (mm) Cost ($/m)

1 25.4 2
2 50.8 5
3 76.2 8
4 101.6 11
5 152.4 16
6 203.2 23
7 254.0 32
8 304.8 50
9 355.6 60
10 406.4 90
11 457.2 130
12 508.0 170
13 558.8 300
Fig. 1. Layout of the two-loop network.
14 609.6 550
Table 1
Node demands and elevations for two-loop network Table 3
Results for design of two-loop network by various researchers
Node number Elevation (m) Demand (m3/h)
Model Cost ($) NFE ω
2 180 100
3 190 100 GA [18] 419,000 250,000 10.5088
4 185 120 SA [26] 419,000 25,000 10.5088
5 180 270 SFLA [30] 419,000 11,323 10.667
6 195 330 HS [28] 419,000 5,000 10.5879
7 190 200 SS [46] 419,000 3,215 10.667
Reservoir 1 210 –1,120 SMPSO (This work) 419,000 3,100 10.667
 A. Moghaddam et al. / Desalination and Water Treatment 104 (2018) 99–110 103

The obtained diameters of pipes and pressure of nodes for had the optimal solution ($419,000) in 13 and 15 iterations,
this network are shown in Table 4. respectively. Compared with the previous state, the number
The most important step in starting the optimization is of iterations had a noticeable reduction.
determining the best values for the algorithm parameters. As one can see in Figs. 6 and 7, the investigation was
For this purpose, in the two-loop network, SMPSO parame- done on c1 and c2 within their authorized ranges (4–2) [43].
ters change within their standard range and their sensitivity These two parameters are very sensitive and determining
is investigated. their exact values is time consuming. In Fig. 6, the values
At the first step, the initial population sizes of 20, 60, 100, of 2.05, 2.3, and 2.45 offer the minimum cost for the param-
140, and 180 for the two-loop network was introduced to eter c1, whereas, according to Fig. 7, the values of 2.05 and
SMPSO algorithm. The lowest cost of optimization (equal to
$420,000) was obtained in the population size of 100 (Fig. 3).
After 22 iterations, no change was seen in the cost reduction
process in all charts in Fig. 3. While determining the popula-
tion size, w = 0.9, wdamp = 1 and c1 = c2 = 2 were selected as default.
The behavior of parameter w was studied by increasing
from 0.4 to 0.9 and assuming a population size of 100 (Fig. 4).
In w = 0.4, minimum network cost ($419,000) was obtained in
iteration 23.
After determining the value of w, the optimum value of
wdamp should be determined. After performing successive iter-
ations of the algorithm, the appropriate range of wdamp (0.9–1)
is recommended for this network. wdamp with its effect on w,
caused the optimal solution to be found in fewer iterations.
According to Fig. 5, the values of wdamp equal to 0.92 and 0.98

Fig. 3. Changes of population size for the two-loop network.

Fig. 2. Convergence chart of SMPSO algorithm in two-loop Fig. 4. Changes of w for two-loop network.
network optimization.

Table 4
Pipe diameters and node pressures due to optimization with
SMPSO method for two-loop network

Pipe number Diameters Node number Pressure

(mm) (mH2O)
1 457.2 2 53.24
2 254 3 30.46
3 406.4 4 43.44
4 101.6 5 33.80
5 406.4 6 30.44
6 254 7 30.55
7 254
8 25.4
Fig. 5. Changes of wdamp for two-loop network.
104 A. Moghaddam et al. / Desalination and Water Treatment 104 (2018) 99–110

2.5 are recommended as the best values for parameter c2. 5.2. Hanoi network
Finally, after examining the combined changes of these two
The Hanoi network in Vietnam (Fig. 8), first presented
parameters for the two-loop network, it was concluded that
by Fujiwara and Khang [7], is a new design as all new pipes
c1 = c2 = 2.05, should be selected.
are to be selected. The network consists of 32 nodes and 34
Finally, the best selected SMPSO input parameters are as
pipes organized in three loops. The system is gravity fed by a
follows: population size = 100; w = 0.4; wdamp = 0.98; c1 = c2 = 2.05
single reservoir and network details that are given in Table 5.
and the maximum number of iterations = 30.
The minimum required head pressure for all nodes is 30 m.
There are six available pipe diameters to be selected for each
new pipe; thus, the total search space consists of 634 possible
designs. Table 6 lists the pipe cost per meter for the six avail-
able pipe diameters.
Table 7 fully shows the results obtained from the SMPSO
algorithm in comparison with other algorithms. In SA and
HS algorithms, the minimum cost is reported as $6.056 × 106,

Fig. 6. Changes of c1 for the two-loop network.

Fig. 8. Layout of Hanoi network.

Table 6
Pipe sizes and costs for Hanoi network

Pipe Diameter (in) Cost ($/m)

1 12 45.726
2 16 70.400
3 20 98.378
4 24 129.333
5 30 180.748
Fig. 7. Changes of c2 for the two-loop network. 6 40 278.280

Table 5
Network data for the Hanoi problem

Pipe Pipe number 1 2 3 4 5 6 7 8 9 10 11

data Length (m) 100 1,350 900 1,150 1,450 450 850 850 800 950 1,200
Pipe number 12 13 14 15 16 17 18 19 20 21 22
Length (m) 3,500 800 500 550 2,730 1,750 800 400 2,200 1,500 500
Pipe number 23 24 25 26 27 28 29 30 31 32 33 34
Length (m) 2,650 1,230 1,300 850 300 750 1,500 2,000 1,600 150 860 950

Node Node number 1 2 3 4 5 6 7 8 9 10 11

data Demand (m /h)3
–19,940 890 850 130 725 1,005 1,350 550 525 525 500
Node number 12 13 14 15 16 17 18 19 20 21 22
Demand (m3/h) 560 940 615 280 310 865 1,345 60 1,275 930 485
Node number 23 24 25 26 27 28 29 30 31 32
Demand (m3/h) 1,045 820 170 900 370 290 360 360 105 805
 A. Moghaddam et al. / Desalination and Water Treatment 104 (2018) 99–110 105

Table 7
Solutions for Hanoi network obtained by different techniques

Pipe diameters (in)

Pipe GA [18] ACO [48] SCE [32] SA [26] HS [27] SMPSO (This work)
1 40 40 40 40 40 40
2 40 40 40 40 40 40
3 40 40 40 40 40 40
4 40 40 40 40 40 40
5 40 40 40 40 40 40
6 40 40 40 40 40 40
7 40 40 40 40 40 40
8 40 40 30 40 40 40
9 40 40 30 40 40 40
10 30 30 30 30 30 30
11 24 24 30 24 24 24
12 24 24 24 24 24 24
13 20 20 16 20 20 20
14 16 12 12 16 16 16
15 12 12 12 12 12 12
16 12 12 24 12 12 12
17 16 20 30 16 16 16
18 20 24 30 20 20 24
19 20 20 30 20 20 20
20 40 40 40 40 40 40
21 20 20 20 20 20 20
22 12 12 12 12 12 12
23 40 40 30 40 40 40
24 30 30 30 30 30 30
25 30 30 24 30 30 30
26 20 20 12 20 20 20
27 12 12 20 12 12 12
28 12 12 24 12 12 12
29 16 16 16 16 16 16
30 16 16 16 12 12 16
31 12 12 12 12 12 12
32 12 12 16 16 16 16
33 16 16 20 16 16 16
34 20 20 24 24 24 20
Cost ($ millions) 6.195 6.134 6.220 6.056 6.056 6.097
NFE 1,000,000 85,571 25,402 53,000 200,000 30,300
ω 10.9031 10.667 10.667 10.5088 10.5088 10.667

whereas, ω = 10.5088 and if the solution proposed by these evaluated. According to Fig. 10, the population size equal to
algorithms is simulated by EPANET 2.0 software (ω = 10.667); 300 had a better convergence process. Based on the results of
the least minimum pressure of 30 m was not respected in all previous example, c1 = c2 = 2.05 were set in this network first,
nodes [47]. In this network, the cost obtained due to optimi- and then the analysis was performed on other parameters of
zation with SMPSO algorithm is $6.097 × 106, obtained after SMPSO algorithm.
30,300 times of NFE (Fig. 9). The pressure at each node is After several times of Hanoi network assessment, it was
shown in Table 8. concluded that the network has many local minimums and
Since in Hanoi network the number of network pipes solutions, and should be investigated with greater caution by
(decision variables) is higher than that of the two-loop avoiding premature convergence of the algorithm; and wdamp
network, the initial population size should also increase. values with very little changes in the scope of (0.990–1). As
Therefore, after reviewing different population sizes in this can be seen in Figs. 11 and 12, the best value for w and wdamp,
network, the values of 100, 140, 180, 220, 260, and 300 were were obtained as 0.6 and 0.998, respectively.
106 A. Moghaddam et al. / Desalination and Water Treatment 104 (2018) 99–110

After sensitivity analysis, the best SMPSO parameters In addition, the minimum allowable pressure for network
were chosen as follows: the population size = 300, w = 0.6, nodes are different values, which leads to more complexity of
wdamp = 0.998, c1 = c2 = 2.05, and the maximum number of iter- the problem compared with two previous networks.
ations = 100. This network has already been optimized by GA and
GA-ILP (integer-linear programming) algorithms and the min-
5.3. Kadu network imum cost offered in these methods are 131,312 and 815 rupees
that have been obtained after 4,440 times of NFE. In this study,
A two-reservoir network with 26 nodes, 34 links, and 9 however, SMPSO algorithm obtained 130,666,043 rupees
loops is shown in Fig. 13, which was initially introduced and cost after 45,150 times of NFE (Table 11). The maximum NFE
optimized by Kadu et al. [49]. Two reservoirs with heads of increased in comparison with the two previous methods,
100 and 95 m feed the network through nodes 1 and 2, respec-
tively. Number of nodes, pipes, and demand of each node
are shown in cubic meters per minute in Fig. 13; and Hazen–
Williams’s coefficient is 130 for all pipes. Other information
such as the length of pipes and demand of nodes are provided
in Table 9. There are 10 commercial diameters, which can be
selected to optimize the network that is expressed in Table 10
along with cost per length unit of them. As a result, in this
problem, 1438 is different states for possible designing, which
should be evaluated in the absence of optimization method.

Fig. 10. Changes of population size for Hanoi network.

Fig. 9. Convergence chart of SMPSO algorithm in Hanoi network

optimization. Fig. 11. Changes of w for Hanoi network.

Table 8
Node pressures due to optimization with SMPSO method for Hanoi network

Node number Pressure (mH2O) Node number Pressure (mH2O) Node number Pressure (mH2O)
2 97.14 13 30.07 24 39.26
3 61.67 14 35.65 25 36.02
4 56.92 15 33.91 26 32.29
5 51.05 16 31.65 27 31.24
6 44.85 17 33.60 28 35.80
7 43.40 18 49.97 29 31.11
8 41.67 19 55.11 30 30.15
9 40.29 20 50.57 31 30.62
10 39.27 21 41.22 32 31.89
11 37.71 22 36.05
12 34.28 23 44.44
 A. Moghaddam et al. / Desalination and Water Treatment 104 (2018) 99–110 107

but the final cost for the network decreased. Fig. 14 shows that as w = 0.4, c1 = c2 = 2.05, and wdamp = 0.998. Moreover, the
the minimum cost was fixed after 2,200 times of NFE and the optimization started based on the initial population of 150
algorithm performance is evident on accelerating convergence. and finished after 300 iterations.
The minimum required pressure and calculated pres-
sure for each node of Kadu network, after optimization with 6. Summary and conclusions
SMPSO, are provided in Table 12. The pressure at each node
was higher than the minimum allowable pressure. Designing an optimal WDN is a complex task. Being
The analysis results of the SMPSO parameters on Kadu nonlinear, nonconvex, and discrete in nature make this
network are shown in Figs. 15–17. problem difficult to solve. In general, two major aims are
According to Fig. 15, cost of 180,611,053 rupees were followed in optimizing water pipe networks: (1) obtaining
obtained for the network in the population size of 150 and the global solution and (2) developing a computationally
in iteration 270, which is better than other populations. After efficient procedure [50]. Various deterministic and heuristic
applying different w, the value w = 0.4 was to reduce network algorithms have been proposed and attempted for solving
cost to 139,444,688 rupees and 130 iterations (Fig. 16). Finally, this problem.
the minimum cost of 130,666,043 rupees was obtained for Kadu
network in iteration 120 by applying wdamp = 0.998 (Fig. 17).
After optimization of this network with SMPSO
algorithm, the final values for its parameters were selected

Fig. 12. Changes of wdamp for Hanoi network. Fig. 13. Layout of Kadu network.

Table 9
Network data for the Kadu network

Pipe Pipe number 1 2 3 4 5 6 7 8 9 10 11

data Length (m) 300 820 940 730 1,620 600 800 1,400 1,175 750 210
Pipe number 12 13 14 15 16 17 18 19 20 21 22
length (m) 700 310 500 1,960 900 850 650 760 1,100 660 1,170
Pipe number 23 24 25 26 27 28 29 30 31 32 33 34
Length (m) 980 670 1,080 750 900 650 1,540 730 1,170 1,650 1,320 3,250

Node Node number 1 2 3 4 5 6 7 8 9 10 11 12 13

data Demand (m3/min) – – 18.4 4.5 6.5 4.2 3.1 6.2 8.5 11.5 8.2 13.6 14.8
Node number 14 15 16 17 18 19 20 21 22 23 24 25 26
Demand (m3/min) 10.6 10.5 9 6.8 3.4 4.6 10.6 12.6 5.4 2 4.5 3.5 2.2

Table 10
Pipe sizes and costs for Kadu network

Pipe number Diameter (mm) Cost (rupees/m) Pipe number Diameter (mm) Cost (rupees/m)
1 150 1,115 6 400 4,255
2 200 1,600 7 450 5,172
3 250 2,154 8 500 6,092
4 300 2,780 9 600 8,189
5 350 3,475 10 700 10,670
108 A. Moghaddam et al. / Desalination and Water Treatment 104 (2018) 99–110

In this research, a simple modified PSO algorithm, In the second study case, the Hanoi problem, comparison of
SMPSO, applied for the optimal design of WDN. The min- the results showed that SMPSO was able to find the best solu-
imum cost obtained by SMPSO programming linked via tion in fewer NFEs than other best-performing algorithms,
water network hydraulic solver EPANET 2.0. The perfor- such as SCE and ACO.
mance of the proposed SMPSO algorithm studied on three In the third example, which was a network offered by
benchmark networks and the results were compared with the Kadu, the complexity of the problem increased due to the
previous studies. In a two-loops network, SMPSO obtained
the optimal solution in fewer NFE than other stochastic opti-
mization algorithms, including GA, SA, SFLA, HS, and SS.

Table 11
Solutions for the Kadu network obtained by different techniques

Pipe diameters (mm)

Pipe number GA-ILP [49] GA [48] SMPSO
(This work)
1 1,000 1,000 900
2 900 900 900
3 400 400 500
4 350 350 250
5 150 150 150
6 250 250 200
7 800 800 900
8 150 150 150
9 400 400 600 Fig. 14. Convergence chart of SMPSO algorithm in Kadu network.
10 500 500 700
11 1,000 1,000 900
12 700 700 700
13 800 800 500
14 400 400 450
15 150 150 150
16 500 500 450
17 350 350 300
18 350 350 450
19 150 150 500
20 150 200 150
21 700 700 600
22 150 150 150
23 450 400 150
Fig. 15. Changes of population size in Kadu network.
24 400 400 400
25 700 700 500
26 250 250 150
27 250 250 350
28 200 200 350
29 300 300 150
30 300 300 300
31 200 200 200
32 150 150 150
33 200 250 200
34 150 150 150
Cost (rupees) 131,312,815 131,678,935 130,666,043
NFE 4,440 360,000 45,150
ω 10.667 10.667 10.667
Fig. 16. Changes of w for Kadu network.
 A. Moghaddam et al. / Desalination and Water Treatment 104 (2018) 99–110 109

Table 12
The node pressures due to optimization with SMPSO method for Kadu network

Node Minimum Pressure Node Minimum Pressure Node Minimum Pressure

number allowable (mH2O) number allowable (mH2O) number allowable (mH2O)
pressure (mH2O) pressure (mH2O) pressure (m)
3 85 98.08 11 85 87.86 19 82 87.77
4 85 94.83 12 85 85.88 20 82 84.38
5 85 93.37 13 82 86.41 21 82 84.31
6 85 87.54 14 82 93.97 22 80 85.88
7 82 85.28 15 85 88.43 23 82 82.25
8 82 91.50 16 82 82.39 24 80 85
9 85 92.64 17 82 90.37 25 80 83.32
10 85 89.64 18 85 85.62 26 80 80.46

prior speed on the current speed. This factor was effec-

tive in increasing the speed of convergence, and far away
from trapping in a local minimum. However, it is better
to increase wdamp from 0.980 to 0.998 as the network size
increases.

Finally, it seems that the SMPSO algorithm can solve opti-

mization problems of WDNs with a few parameters and easy
implementation. It also can find the optimal solution in less
duration than other algorithms.

References
[1] R. Baños, C. Gil, J.I. Agulleiro, J. Reca, A Memetic Algorithm
for Water Distribution Network Design. Soft Computing in
Industrial Applications, Springer, 2007, pp. 279–289.
[2] D.F. Yates, A.B. Templeman, T.B. Boffey, The computational
complexity of the problem of determining least capital cost
Fig. 17. Changes of wdamp for Kadu network. designs for water supply networks, Eng. Optim., 7 (1984)
143–155.
difference in the minimum allowable pressure of the nodes. [3] J. Gessler, Pipe Network Optimization by Enumeration, Part of
The results showed that NFE increased by the proposed Computer Applications in Water Resources, 1985, pp. 572–581.
SMPSO algorithm, while the least new cost was less than the [4] E. Alperovits, U. Shamir, Design of optimal water distribution
other previous works. systems, Water Resour. Res., 13 (1977) 885–900.
[5] G.E. Quindry, E.D. Brill, J.C. Liebman, Optimization of looped
After the sensitive analysis of the parameters of SMPSO water distribution systems, J. Environ. Eng. Div., 107 (1981)
algorithm for all three benchmark networks, the following 665–679.
results were achieved: [6] R.E. Featherstone, K.K. El-Jumaily, Optimal diameter selection
for pipe networks, J. Hydraul. Eng., 109 (1983) 221–234.
1. According to the standard range of w, when it is closer [7] O. Fujiwara, D.B. Khang, A two-phase decomposition method
to 0.4, the convergence was accelerated, but if a problem for optimal design of looped water distribution networks,
Water Resour. Res., 26 (1990) 539–549.
has numerous local minimums such as Hanoi network, it [8] U. Shamir, Optimal design and operation of water distribution
is likely for the algorithm to fall into local minimum trap. systems, Water Resour. Res., 10 (1974) 27–36.
Conversely, by increasing the value of w to 0.9, algorithm [9] A.N. El-Bahrawy, A.A. Smith, A methodology for optimal
reviews the solutions with more caution, but the number design of pipe distribution networks, Can. J. Civ. Eng., 14 (1987)
of iterations will be increased. 207–215.
[10] Y.C. Su, L.W. Mays, N. Duan, K.E. Lansey, Reliability‐based
2. With increasing the number of decision variables of optimization model for water distribution systems, J. Hydraul.
a problem (i.e., diameter of network pipes), the size of Eng., 113 (1987) 1539–1556.
initial population should also increase (e.g., Hanoi and [11] N. Duan, L.W. Mays, K.E. Lansey, Optimal reliability-based
Kadu networks, where the initial population size was design of pumping and distribution systems, J. Hydraul. Eng.,
116 (1990) 249–268.
more than that of the two-loop network). [12] K.E. Lansey, L.W. Mays, Optimization Models for Design of
3. In the two-loop and Hanoi and Kadu networks, c1 and Water Distribution Systems, in Reliability Analysis of Water
c2 were changed to the amounts suggested by Eberhart Distribution Systems, Part 1: State-of-the-Art, 1989.
(between 2 and 3). After examining all different scenar- [13] C. Ciaponi, E. Creaco, L. Franchioli, S. Papiri, The importance of
ios, it was concluded that c1 = c2 = 2.05 are the best values the minimum path criterion in the design of water distribution
networks, Water Sci. Technol. Water Supply, 17 (2017) 1558–1567.
for all the networks in this study. [14] A. Vasan, S.P. Simonovic, Optimization of water distribution
4. In this research, a new factor called wdamp was used to network design using differential evolution, J. Water Resour.
decrease w in each iteration and damp the impact of Plann. Manage., 136 (2010) 279–287.
110 A. Moghaddam et al. / Desalination and Water Treatment 104 (2018) 99–110

[15] L.J. Murphy, A.R. Simpson, G.C. Dandy, Pipe Network [34] Y.-X. Jin, H.-Z, Cheng, J.-y. YanL Zhang, New discrete method for
Optimization Using an Improved Genetic Algorithm, particle swarm optimization and its application in transmission
Department of Civil and Environmental Engineering, network expansion planning, Electr. Power Syst. Res., 77 (2007)
University of Adelaide, Adelaide, Australia, 1993. 227–233.
[16] A.R. Simpson, G.C. Dandy, L.J. Murphy, Genetic algorithms [35] C.-J. Liao, P. Luarn, A discrete version of particle swarm
compared to other techniques for pipe optimization, J. Water optimization for flowshop scheduling problems, Comput.
Resour. Plann. Manage., 120 (1994) 423–443. Oper. Res., 34 (2007) 3099–3111.
[17] G.C. Dandy, A.R. Simpson, L.J. Murphy, An improved genetic [36] S. Janson, D. Merkle, M. Middendorf, Molecular docking
algorithm for pipe network optimization, Water Resour. Res., with multi-objective Particle Swarm Optimization, Appl. Soft
32 (1996) 449–458. Comput., 8 (2008) 666–675.
[18] D.A. Savic, G.A. Walters, Genetic algorithms for least-cost [37] A. Moghaddam, J. Behmanesh, A. Farsijani, Parameters
design of water distribution networks, J. Water Resour. Plann. estimation for the new four-parameter nonlinear muskingum
Manage., 123 (1997) 67–77. model using the particle swarm optimization, Water Resour.
[19] I. Gupta, A. Gupta, P. Khanna, Genetic algorithm for Manage., 30 (2016) 2143–2160.
optimization of water distribution systems, Environ. Modell. [38] J. Kennedy, R. Eberhart, Particle Swarm Optimization, in Neural
Software., 14 (1999) 437–446. networks, Proc. IEEE International Conference, Vol. 4, 1995, pp.
[20] I. Lippai, J.P. Heaney, M. Laguna, Robust water system design 1942–1948.
with commercial intelligent search optimizers, J. Comput. Civil [39] La. Rossman, EPANET 2: Users Manual. Cincinnati US
Eng., 13 (1999) 135–143. Environmental Protection Agency National Risk Management
[21] K. Vairavamoorthy, M. Ali, Optimal design of water distribution Research Laboratory, Vol. 38, 2000, p. 200.
systems using genetic algorithms. Comput. Aided Civ. [40] AAde. Meneses, M.D. Machado, R. Schirru, Particle swarm
Infrastruct. Eng., 15 (2000) 374–382. optimization applied to the nuclear reload problem of a
[22] E. Creaco, G. Pezzinga, Embedding linear programming in pressurized water reactor, Prog. Nucl. Energy, 51 (2009)
multi objective genetic algorithms for reducing the size of the 319–326.
search space with application to leakage minimization in water [41] M. Clerc, The Swarm and the Queen: Towards a Deterministic
distribution networks, Environ. Modell. Software, 69 (2015) and Adaptive Particle Swarm Optimization, in Evolutionary
308–318. Computation, 1999. CEC 99, Proceedings of the 1999 Congress
[23] S.R. Dongre, R. Gupta, Optimal design of water distribution on 1999, IEEE.
network under hydraulic uncertainties, ASCE-ASME J. Risk [42] K.E. Parsopoulos, M.N. Vrahatis, Recent approaches to global
Uncert. Eng. Syst., Part A Civ. Eng., 3 (2017) G4017001. optimization problems through particle swarm optimization,
[24] J.P. Matos, A.J. Monteiro, N.M. Matias, A.J. Schleiss, Redesigning Nat. Comput., 1 (2002) 235–306.
water distribution networks using a guided evolutionary [43] Y. Shi, R.C. Eberhart, Parameter selection in particle swarm
approach, J. Water Resour. Plann. Manage., 142 (2015) C4015004. optimization, International Conference on Evolutionary
[25] G.V. Loganathan, J.J. Greene, T.J. Ahn, Design heuristic for Programming VII SE – 57.1447, 1998, pp. 591–600.
globally minimum cost water-distribution systems, J. Water [44] Y. Shi, R. Eberhart, A Modified Particle Swarm Optimizer, in
Resour. Plann. Manage., 121 (1995) 182–192. Evolutionary Computation Proceedings, 1998 IEEE World
[26] M. Cunha, J. Sousa, Water distribution network design Congress on Computational Intelligence, The 1998 IEEE
optimization: simulated annealing approach, J. Water Resour. International Conference, 1998, pp. 69–73.
Plann. Manage., 125 (1999) 215–221. [45] M. Clerc, J. Kennedy, The particle swarm - explosion, stability,
[27] Z.W. Geem, J.H. Kim, G.V. Loganathan, Harmony search and convergence in a multidimensional complex space, IEEE
optimization: application to pipe network design, Int. J. Model. Transactions Evolutionary Computation, Vol. 6, 2002, pp. 58–73.
Simul., 22 (2002) 125–133. [46] M.-d. Lin, Y.-h. Liu, G.-f. Liu, C.-w. Chu, Scatter search heuristic
[28] Z.W. Geem, Optimal cost design of water distribution networks for least-cost design of water distribution networks, Eng.
using harmony search, Eng. Optim., 38 (2006) 259–277. Optim., 39 (2007) 857–876.
[29] J. Yazdi, Decomposition based multi objective evolutionary [47] H. Wang, S. Liu, F. Meng, M. Li, Gene expression programming
algorithms for design of large-scale water distribution networks, algorithms for optimization of water distribution networks,
Water Resour. Manage., 30 (2016) 2749–2766. Procedia Eng., 37 (2012) 359–364.
[30] M.M. Eusuff, K.E. Lansey, Optimization of water distribution [48] A.C. Zecchin, A.R. Simpson, H.R. Maier, J.B. Nixon, Parametric
network design using the shuffled frog leaping algorithm, J. Study for an Ant Algorithm Applied to Water Distribution
Water Resour. Plann. Manage., 129 (2003) 210–225. System Optimization, IEEE Transactions Evolutionary
[31] H.R. Maier, A.R. Simpson, A.C. Zecchin, W.K. Foong, K.Y. Computation, Vol. 9, 2005, pp. 175–191.
Phang, H.Y. Seah, et al., Ant colony optimization for design of [49] M.S. Kadu, R. Gupta, P.R. Bhave, Optimal design of water
water distribution systems, J. Water Resour. Plann. Manage., networks using a modified genetic algorithm with reduction
129 (2003) 200–209. in search space, J. Water Resour. Plann. Manage., 134 (2008)
[32] Shie-Yui Liong, M. Atiquzzaman, Optimal design of water 147–160.
distribution network using shuffled complex evolution, J. Inst. [50] A. Haghighi, H. Samani, Z. Samani, GA-ILP Method for
Eng. Singapore, 44 (2004) 93–107. optimization of water distribution networks, Water Resour.
[33] Y. Dong, J. Tang, B. Xu, D. Wang, An application of swarm Manage., 25 (2011) 1791–1808.
optimization to nonlinear programming, Comput. Math. Appl.,
49 (2005) 1655–1668.