See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/225778485

An Optimization Strategy for Water Distribution Networks

Article in Water Resources Management · January 2008

DOI: 10.1007/s11269-008-9270-8

CITATIONS READS

32 933

2 authors:

Önder Ekinci Haluk Konak

Kocaeli University, Asım Kocabiyik Vocational School, Turkey Kocaeli University
5 PUBLICATIONS 32 CITATIONS 32 PUBLICATIONS 86 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Haluk Konak on 03 June 2015.

The user has requested enhancement of the downloaded file.

Water Resour Manage (2009) 23:169–185
DOI 10.1007/s11269-008-9270-8

An Optimization Strategy for Water

Distribution Networks

Önder Ekinci & Haluk Konak

Received: 6 March 2007 / Accepted: 6 March 2008 /

Published online: 22 April 2008
# Springer Science + Business Media B.V. 2008

Abstract An optimization strategy based on head losses minimization is developed for the
least cost design of water distribution networks. A new weighting approach is suggested for
calculating the initial flow distribution and optimum pipe diameters of the weighted flow
distribution is presented by using least square method. In the mean time homogenous and
isotropous head losses are maintained with implications of head loss path choice. The
model is employed for designing and/or modifying pipe sizes while the classical Hardy-
Cross network solver is used to balance the flows. The whole algorithm is programmed and
applied to a two-looped network selected from the literature and the results are presented on
a comparative basis. A FORTRAN software with the necessary steps in the flow chart is
written for the optimization calculations in this paper.

Keywords Algorithms . Optimization . Water distribution . Network . Cost

1 Introduction

There are many models that have been developed to solve water distribution network
optimization non-linear problem using different techniques. Among them are the
deterministic, dynamic, linear, non-linear, enumeration etc. approaches, in addition to
heuristic techniques such as genetic algorithm (Savic and Walters 1997), simulated
annealing (Cunha and Sousa 1999, 2001), harmonic search (Geem et al. 2000, 2002, Geem
and Tseng 2002), ant colony optimization algorithm (Maier et al. 2003) and shuffled frog
leaping algorithm (Eusuff and Lansey 2003).
The objective functions focus usually on the cost minimization while the reliability or
precision of the optimal solutions are not assessed sufficiently. Reliability is the ability of a
system or component to perform its required functions under stated conditions for a
specified period of time whereas the precision is the state or quality of being exact.

Ö. Ekinci (*)
Asım Kocabıyık Vocational School, University of Kocaeli, Hereke, 41800 Kocaeli, Turkey
e-mail: oekinci@kou.edu.tr

H. Konak
Faculty of Engineering, Deparment of Geodesy and Photogrammetry Engineeering,
Information Technologies Research Center, University of Kocaeli, 41100 Kocaeli, Turkey
e-mail: hkonak@kou.edu.tr
170 Ö. Ekinci, H. Konak

Recently, several papers addressed the issue of reliability (or redundancy) of a water
distribution networks, together with cost minimization and adopting a multi-objective
approach to solve the problem (Tolson et al. 2004, Prasad and Park 2004, Farmani et al.
2005, Dandy and Engelhardt 2006).
Among the main difficulties in use of such models are insufficient solutions for the
application problems (Walski 2001) and inability to provide better alternatives as compared to
the traditional methods (Goulter 1992). The main precision problem is how to evaluate the
initial parameters, flows and diameters. If decision makers use different initial value sets, they
can find occasionally not global but local optimum. The solution of non-linear equations could
be performed by consecutive approaches in such a manner that the results in the first stage
should not be far from the target values. Hence, estimation of the initial parameters is one of the
main problems, although it does not seem to be an important process at the beginning.
In any water distribution network system design, one of the major objectives is to
minimize the head losses along the pipelines. The most important constraint is to get a
solution with the minimum cost. Therefore, a minimum head loss optimization strategy,
which could provide a rapid, realistic and practicable solution, is developed based on a
precision process. Such a model provides a reliable solution for the investigation of the
optimal flow distributions and corresponding diameters.
In this paper a weighted approach is introduced to determine the initial pipe flows as the
most suitable and the most approximate calculation for the real case. Flows and diameters
are estimated initially while the heads are taken as variables. The proposed approach is
applied successfully to the Alperovits and Shamir (1977) two-looped network system and
data. The weighted flow rate pre-balance by Hardy–Cross solver is imposed among the
initial values. Optimum diameter determination according to the initial flow rates is
performed by a weighted least square method.

2 Model Formulation

2.1 The Determination of the Objective Functions for the Weighting Optimization

The Darcy–Weisbach functional relation gives the head loss, hi, at i-th pipe as,
8 Li
hi ¼ 2 fi 5 Q2i ð1Þ
p g Di
where g is the acceleration of gravity; fi is the friction factor of pipe, which can be
calculated using the Colebrook–White formula Li is the pipe length, Di is the pipe diameter
and Qi is the discharge in pipe i.

2.1.1 Homogeneously Distributed Head Losses

Equation 1 is difficult to use in the case of pipe networks, and therefore, it is convenient to
write it similar to the Chezy–Manning formula as,
hi ¼ Ki Q2i ð2Þ
where Ki is a constant. If the corrections are aimed to be in the normal distribution, the
functional model would be arranged by its unknowns as,
1 pffiffiffiffi
pffiffiffiffiffi hi ¼ Qi ð3Þ
Ki
An optimization strategy for water distribution networks 171

Perturbation of variables hi and Qi in Eq. 3 after the ignorance of non-linear terms leads to
the following expression where hi is taken as the average values of the unknowns.
1 pffiffiffiffi 1
pffiffiffiffiffi hi þ pffiffiffiffiffiffiffiffi dhi ¼ Qi þ ni ð4Þ
Ki 2 K i hi
Herein, vi indicates the correction amount (error due to perturbation) in the discharge.
Multiplication of both sides in Eq. 2 by hi yields Ki hi ¼ Ki2 Q2i , acceptance of hi0 ffi 0 and
their substitution into Eq. 4 leads to,
1
dhi ¼ Qi þ ni ð5Þ
2Ki Qi
with the correction expression as,
1
ni ¼ dhi  Qi ð6Þ
2Ki Qi
In order to determine the most suitable estimation of the unknowns and the most probable
corrections of the measurements, it is necessary to have the sum of the squares of the
corrections as minimum,

vT v ) min ð7Þ
where

vT ¼ ½n 1 ; n2 ; . . . ; nn 
is the correction vector. Equation 7 yields a square matrix with diagonal elements
 
A ¼ a1;1 ; a2;2 ; . . . ; an;n

and coefficients as
1
ai;i ¼
2Ki Qi
The collection of head losses in pipes appears in the form of an unknown vector as,

xT ¼ ½dh1 ; dh2 ; . . . ; dhn 

Additionally by considering the flow measurement vector as

‘T ¼ ½Q1 ; Q2 . . . ; Qn 
the normal equations can be established through the error propagation law as,

AT Ax  AT ‘ ¼ 0 ð8Þ
or
 1
x ¼ AT A AT ‘ ð9Þ

Constraint functions are established among the unknowns (head losses) for calculating the
flow corrections (vi). As a restriction in each loop as a condition the total head loss should
be equal to zero. From the aforementioned concepts such a restriction requires that
h1  h2 ¼ K1 ðQ1 þ $QÞn K2 ðQ2  $QÞn ¼ 0 ð10Þ
172 Ö. Ekinci, H. Konak

Equations 8 and 10 indicate that the head loss values are directly dependent on the
unknown coefficients. If the flow distribution and the corresponding pipe diameters are
chosen appropriately, then the head losses will be homogeneous accordingly. In an ideal
situation the head losses are expected to be almost equal.
Eigenvalues, li, are the most important criterion of precision in a loop optimization. The
coefficient matrix, AT A, is a diagonal matrix, and therefore its eigenvalues are equal to its
diagonal elements, which mean that
1
li ¼ ð11Þ
4Ki2 Q2i
This expression can be rewritten by considering Eq. 2 as

li ¼ Ci =Ki Q2i ¼ Ci =hki ð12Þ

where Ci ¼ 1=4Ki . According to this relation, there is only a simple linear conversion
between the eigenvalue and the head loss. A favorable objective function can be chosen by
considering the precision requirements as one of the following alternatives.
1) The maximum head loss should be minimum,
maxfhi g ) min ð13Þ
2) The head losses should have the same values (isotropous),
maxfhi g ¼ minfhi g ð14Þ
3) The distribution of head losses should be symmetrical (homogeneous):
maxfhi g
" ð15Þ
minfhi g

where ɛ is an acceptable ration value. The head losses are dependent on the pipe diameters
rather than the initial flow distribution, which can be expressed by considering Eqs. 1 and 2 as,

Ki ¼ fi Li c D5i ð16Þ
 2
where c ¼ 8 p g is a constant. This last expression shows that if the pipe diameter increases,
the head loss decreases. Accordingly, the cost increases, if a cost function is constituted with
the commercial pipe diameters, then the cost increases also with the quality.

2.1.2 An Isotropous Head Loss Distribution for Whole Network

In many optimization procedures, the precision criteria with global objectives representing
whole network are preferred instead of those with scalar objectives. A homogeneous and
isotropous matrix can be chosen as the objective function for an ideal water distribution
network. For this purpose, first an isotropous criterion matrix is established as the objective
function as,
Q x ¼ ce E ð17Þ

where ce is a constant and E is a unit matrix, and the functional model can be established as,
 1 :
AT A ¼Qx ð18Þ
An optimization strategy for water distribution networks 173

where ¼˙ is the inconsistency sign. If the problem is converted to an inverse of criterion

matrix, then the basic model could be obtained with direct approach (Konak 1994,
Grafarend et al. 1979).
:
AT A ¼ Q1
x
ð19Þ

which can be written explicitly as,

2 3
1 0 0  0
2 3 6 0 1 0  07
1
4K12 Q21
0 0  0 6 7
6 7 60 0 1  07
6 1
0  0 7 6 .. 7
6 4K22 Q22 7 : 6 7
6 1
 0 7 ¼66
. 7
7 ð20Þ
6 4K32 Q23 7 6 7
4 5 6 7
.. 1 6 7
. 4Kn2 Q2n 4 5
0 0 0  1
This expression indicates clearly that the unknowns are Ki coefficients. By definition as,

pi ¼ 1 Ki2 ð21Þ

for weights on ai which are the diagonal elements of coefficients matrix one can write that
:
BT PB ¼ E ð22Þ
with

1 1 1
diagonal ðBi Þ ¼ ; ;...; ð23Þ
2Q1 2Q2 2Qn
On the other hand, provided that
  :
BT ΘBT p ¼ q ð24Þ

then the weights can be calculated as,

:  1
p ¼ BT ΘBT q ð25Þ

where p is the vector (P), q is the vector (E), Θ is the Khatri–Rao multiplication.
Equation 22 can be solved by the least square method (LSM). If (*) is Hadamard
multiplication, then the LSM solution will be as,
   
BBT *BBT p  BT ΘBT q ¼ 0 ð26Þ

The residue matrix is defined as,

 1
D ¼ BT B Q ð27Þ

with residue vector as while d=vector (D)

d T d ) min ð28Þ
174 Ö. Ekinci, H. Konak

By dealing with the basic Eq. 22, it is easily seen that the criterion matrix is the solution of
a simple normal equation system with (n×n) dimensions when it is a diagonal matrix. If the
system is solved for each pipe, than the following equations can be obtained as,

1
pi ¼ qi ð29Þ
4Q2i
and
pi ¼ 4Q2i qi ð30Þ
From Eq. 21, the pi and Ki coefficients are obtained as,

1
pi ¼ ð31Þ
Ki2
and
1
Ki ¼ pffiffiffiffi ð32Þ
pi

respectively, Ki are the most suitable coefficients for the isotropous structured head loss
distribution. Hence, if the commercial diameters are to be found, then weight coefficients
for the maximum pipe diameter D0 should be calculated for each pipe by Eqs. 33 and 34 as,
c
K0 ¼ f0 L0 ð33Þ
D50
and
c
Ki ¼ fi Li ð34Þ
D5i
where f0 and fi are friction factors. The ratio of these two last expressions leads to,

Ki fi Li cD50
¼ ð35Þ
K0 f0 L0 cD5i
where ff0i is a favorable coefficient and it can be taken as almost equal to 1 ( ff0i ffi 1). Hence,
the optimum diameters could be obtained from Eq. 36 as,

K0 fi Li 5
D5i ¼ D ð36Þ
Ki f0 L0 0
If the head loss distribution is calculated by balancing the head losses with the
determined commercial pipe diameters then a homogeneous (symmetrical) head loss
distribution can be obtained insofar. In addition, the optimum pipe diameters for the initial
flow distribution could be determined, while the head loss distributions are controlled in the
optimization procedure.

3 A Minimum Head Loss Optimization Strategy

The minimum head loss optimization strategy includes three steps as the optimization of
precision, reliability and cost. Herein, the cost and reliability optimizations are especially
An optimization strategy for water distribution networks 175

important. However, the optimization of precision seems to be necessary for all the stages.
Therefore, cost and reliability optimization processes should be performed simultaneously.
In this paper, a FORTRAN software with flow chart in Fig. 1 is developed for the
optimization strategy.

START

Calculate the weighted initial flows

Determine the diameter distribution for the initial flows

Balance initial flows at once

Calculate the optimum diameter distribution

System Controller
Network Solver
Diameter optimization by
Hardy Cross Calculate the head loss distribution Least Squares Method

No
Objective Function
max{hi} ⇒ min;
Constraint: Hmin ≤Hij ≤Hmax

Obtain Datum Yes

(Main Pipe Diameter, D0)
Cost objective function
min M 0 = ∑ M iLi
i

Reduce all the diameters to the next lower

commercial diameters

Calculate the optimum diameter distribution

Calculate the head loss distribution

Yes
Objective Function
max{hi} ⇒ min;
No
Calculate an alternative cost
Constraint: Hmin ≤Hij ≤Hmax

Decision

Obtain Datum
(Main Pipe Diameter, D0)
STOP

Fig. 1 The flow chart for the minimum head-loss optimization strategy
176 Ö. Ekinci, H. Konak

The optimization of precision procedure is applied to the network by choosing the

minimization of the maximum head loss as the priority objective function as

max hfi ) min

then the network is taken to its optimum condition with respect to the precision. Under
these conditions, the objective function is performed as the minimization of the total pipe
cost
X
Mi ðD; LÞ ) min

Here, the most important constraint is the pressure constraint which is taken as 30≤Hi ≤
80 m.

4 Weighted Flow Distribution

If a decision-maker chooses initial flow distribution based on his/her expert view then it is
highly probable that the solution algorithm converges to a local optimum. In this case, the
number of iterations increases and with the change of the initial flow distribution, another
local optimum can be obtained. The significance of initial flow distribution determination,
and its effects on the solution can be seen in the numerical applications clearly (Goulter
et al. 1986). Morgan and Goulter (1985) proposed the first weighting algorithm for flow
distribution. Their algorithm balanced by Hardy–Cross method, the ratio between the flow
transferred by a pipe to the node and the total flow at the node is accepted as the weight of
the pipe related to the node. In other studies, the initial flow distribution is chosen
according to the practice of decision maker, but the criteria for the selection of initial flow
distribution are not explained clearly.
According to the weighting algorithm based on a loop approach in this study, the sum of
the flows drawn from the beginning and the end of the link (pipelines) is rated to the sum of
the flows drawn from the total network then the weighted flow is obtained. If this pre-
weight is taken independently, a deceptive situation that is not in accordance with the reality
may appear such that the weight of the pipes following the main pipe may be less than that
of the pipes at the end of the network. In this paper, the sum of the pipe weight and the
weights of the consecutive pipes fed by this pipe in the network expresses the real weight of
this pipe in the network. The main steps for this algorithm are as follows
1. The path choice number, s, is constituted
2. The weights of each pipe are calculated for the path choice. If qi is the discharge flow
at the node i, and qsij is the average discharge flow at the pipe, which connects i, j nodes
then,
 
qsij ¼ qi þ qj 2 ð37Þ

and the weighted flow of the connection pipe from i to j nodes asij can be calculated
as,

qsij
asij ¼ P s ð38Þ
qij
An optimization strategy for water distribution networks 177

with the control condition that

X
asij ¼ 1 ð39Þ

3. The weighted flows, Qij, of the successive pipes for the path choice can be calculated
as,
X
Qijr;kþ1 ¼ Qijr;k  qi arij ð40Þ

where

Qij ¼ Qijr;kþ1 ð41Þ

Herein, r is the number of successive pipes at the path, and k is the number of the
calculation steps.
4. The steps 2 and 3 are repeated for each path choice.

5 The Determination of the Optimum Path

The determination of the consecutive pipes fed by a pipe and the calculation of the heads at
the nodes require the determination of optimum path in the network. After the
determination of weighted flow distribution, the flows are obtained in a consistent manner
related to the network conditions. The Hardy–Cross network solver balances this pre-initial
flow distribution to get the optimum initial flow distribution. In this manner, the first step
for optimum solution with a few iterations can be obtained by initial flow distribution
determination. This approach forms a basis for the choice of optimum path, which is then
determined by the application of the minimum head loss optimization strategy including
head loss distribution, distribution of diameters for the main pipe and the other pipes and
the trials for the determination of the path with the minimum cost.

5.1 Optimization of Diameters

Since the pipe diameters are one of the input parameters in a solution strategy, the initial
once-balanced flows are not sufficient alone for the solution. The optimum initial diameter
determination according to the initial flows is a significant problem. It is essential that a
solution should be found for the pipe diameters’ determination that are consistent with the
network conditions, according to the distribution of optimum initial once-balanced flows.
The proposed methodology in this paper includes a weighting optimization process by
the LSM with a criterion matrix in a completely isotropic structure, which represents the
whole network and it is used as the objective function. Since the initial value (input
parameter) is the diameter of the main pipe (datum) in the diameter optimization process,
the optimum diameter for the system could be obtained after few iterations starting from the
maximum commercial diameter possible in the system. Thus, an initial diameter, D0, of
main pipe could be obtained as the most approximate value.
By the application of diameter optimization for searching the initial, optimized and
alternative solutions, both the optimum diameters could be determined, and the control of
178 Ö. Ekinci, H. Konak

Fig. 2 The two-looped network

of Alperovits and Shamir (1977)

the distribution of head losses become possible in the system. In this manner, the decision-
maker could obtain the optimum solution with the least cost rapidly by reducing the number
of iterations, for example, from 122 to 6.

5.2 Balancing the Distribution of Head Losses

If the head loss is aimed to be minimum in each pipe, increasing the diameter of the pipe at
the maximum head loss location with the following upper commercial diameter in mind
until the pressure constraint is satisfied, should reduce the head loss and consequently the
distribution of head losses should be homogenized. In this manner, the pressure constraint
is satisfied by changing the pipe diameters in the algorithm. Provided that that the
pressure is not satisfied at any node in the network, the pipe diameter under the maximum
head loss is increased to the next upper diameter, and this process continues until the
pressure constraint is satisfied at each node. The head losses distribution is homogenized
for approximating the ideal solution. It should be noted that the cost function tends
downward when the optimization of diameters is performed.

5.3 Searching the Alternative Solutions

Since there are many local optimums in the solution space of non-linear problems, the
probability of the presence of more optimal solutions should be concerned also. Although
the decision-maker could make some changes interactively for searching such solutions, it
is difficult to reach an optimum solution converging to the real value by this way, when the
errors caused by the increase in the number of iterations and choices of the decision-maker
are taken into account. Therefore, the changes should not be local and they should include
all elements in the system. In the approach proposed here, search for reducing all the
diameters performs a better optimal solution determination as optimum in the first run to the
An optimization strategy for water distribution networks 179

next lower commercial diameters. Hence, it could be said that the system solution is leaped
from the local optimum. It is observed in the numerical applications, in general, that the
alternative solution does not differ from the optimum solution.

6 Numerical Applications

For the comparison of the proposed model with widely accepted ones in the literature, it is
necessary to consider objectivity, consistency, sufficiency and reliability in the final testing.
In this study, the proposed model is compared with the Alperovits and Shamir (1977) two-
looped network which provides enough data for testing (see Fig. 2).
In a water distribution network, the aim is to get the head losses through the pipelines
connecting the pairs of the nodes as minimum and equal as possible. In addition, the cost of
design is required to be minimum with respect to the commercial pipe diameters. On the
other hand, the design solution has to satisfy the constraints of pressure, velocity, etc. For a
minimum cost optimization solution procedure should follow a strategy including the
following four steps depending on and controlling each other.

1. Optimization of the initial flows (weighted flow distribution),

2. Determination of the initial pipe diameters: diameter optimization (datum optimization),
3. Determination of the optimum path choice according to the first two steps, and
4. The head loss optimization to control the successive steps above.

6.1 Testing of the Model and Reliability

Alperovits and Shamir (1977) presented a stable network model with equal pipe lengths
that are really connected at right angles and hence it is possible to observe the head loss
distribution through the network in a reliable manner.
The diameter optimization procedure on the initial flows that are obtained by the initial
data group is formed by the weighted flow distribution approach. The flow distribution
from the commercial diameter optimization procedure may not be effective to minimize the
inconsistency effect.
In the numerical application, the proposed model is applied to the path choices
determination according to nodes of 5 (the first path) and 7 (the second path) and the datum
are observed as 20 and 18 in. respectively. The results are clearly compatible with the
results of Alperovits and Shamir (1977) as shown in Table 1.
For models given in the literature, it is possible to see the pipe diameters as low as 1 in.,
which is not used in practical use. The proposed strategy produces the minimum pipe
diameter as 4 in. which can be used in practice and it provides a homogenous head loss
distribution with plausible commercial pipe diameters in addition to solutions with less cost
as in Tables 2 and 3.

6.2 The Comparisons of the Results

The optimum design cost solutions of the two-looped test network in the literature gives the
unit costs as 479.525 (Alperovits and Shamir 1977), 435.015 (Goulter et al. 1986), 417.500
(Kessler and Shamir 1989), 415.271 (Fujiwara and Khang 1990), 402.352 (Eiger et al.
 180

Table 1 Datum and diameter optimization results

Node 5 Node 7

Initial flows of Alperovits and Shamir (1977) Weighted initial flows Initial flows of Alperovits and Shamir (1977) Weighted initial flows

Optimum solution Alternative solution Optimum solution Alternative solution Optimum solution Alternative solution Optimum solution Alternative solution

No D0 D1 Di D0 D1 Di D0 D1 Di D0 D1 Di D0 D1 Di D0 D1 Di D0 D1 Di D0 D1 Di

1–2 355.6 355.6 508.0 457.2 355.6 508.0 406.4 406.4 508.0 457.2 406.4 457.2 406.4 406.4 457.2 406.4 406.4 457.2 406.4 406.4 457.2 406.4 406.4 457.2
2–3 254.0 254.0 304.8 254.0 254.0 304.8 304.8 304.8 304.8 254.0 304.8 304.8 304.8 304.8 304.8 254.0 304.8 304.8 355.6 355.6 355.6 304.8 355.6 355.6
2–4 355.6 304.8 406.4 355.6 304.8 406.4 355.6 355.6 406.4 355.6 355.6 355.6 355.6 355.6 406.4 355.6 355.6 406.4 355.6 355.6 355.6 304.8 355.6 355.6
4–5 152.4 203.2 203.2 152.4 203.2 203.2 254.0 254.0 254.0 203.2 254.0 254.0 203.2 152.4 152.4 76.2 152.4 152.4 254.0 203.2 203.2 152.4 203.2 203.2
4–6 304.8 304.8 355.6 304.8 304.8 355.6 355.6 355.6 355.6 304.8 355.6 355.6 355.6 355.6 355.6 304.8 355.6 355.6 304.8 304.8 304.8 254.0 304.8 304.8
6–7 254.0 254.0 254.0 203.2 254.0 254.0 304.8 254.0 254.0 203.2 254.0 254.0 304.8 304.8 304.8 254.0 304.8 304.8 152.4 203.2 203.2 152.4 152.4 152.4
3–5 203.2 254.0 254.0 203.2 254.0 254.0 254.0 304.8 304.8 254.0 304.8 304.8 254.0 304.8 304.8 254.0 304.8 304.8 355.6 304.8 304.8 254.0 304.8 304.8
5–7 203.2 152.4 152.4 152.4 152.4 152.4 152.4 101.6 203.2 152.4 254.0 254.0 254.0 304.8 304.8 254.0 304.8 304.8 254.0 254.0 254.0 203.2 254.0 254.0
Cost 296 295 473 345 295 473 390 397 507 366 406 476 397 421 491 336 426 496 400 388 428 309 376 416
Approximate 355.6 406.4 406.4 406.4
datum
Optimum 508.0 457.2 457.2 457.2
datum
Ö. Ekinci, H. Konak
Table 2 The trend of cost function

Approximate datum 508.0 457.2 406.4 355.6 304.8 254.0

Optimum datum 508.0 508.0 457.2 457.2 508.0 508.0

Path Initial Optimum Alternative Optimum Alternative Optimum Alternative solution Optimum Alternative Optimum Alternative solution Optimum Alternative
flows solution solution solution solution solution solution solution solution solution solution
An optimization strategy for water distribution networks

Node 5 1 682 690 544 514 476 476 473 473 480 480 491 491
2 722 690 544 514 507 476 482 482 480 480 491 491
Node 7 1 723 713 553 572 491 496 507 507 500 500 529 529
2 676 675 535 505 428 416 435 435 435 435 440 440

1 Initial flows of Alperovits and Shamir (1977), 2 weighted initial flows

181
 182

Table 3 Minimum head loss optimization strategy

Optimization of head loss distributions Optimization of head loss distributions for alternative solution

1 2 3 4 5 6 7

3 3 3
No Q0 m /h D0 mm Q1 m /h hi m Hi m Di mm Qi m /h hi m Hi m Di mm Qi hi m Hi m Dt mm Qi hi m Hi m Di mm Qi hi m Hi m Di mm Qi m3/h hi m Hi m
m3/h m3/h m3/h

1–2 −1120 406.4 −1120 11.3 48.7 406.4 −1120 11.3 48.7 457.2 −1120 6.09 53.9 406.4 −1120 11.3 48.7 406.4 −1120 11.3 48.7 457.2 −1120 6.09 53.9
2–3 −554.6 355.6 −457.7 3.78 34.9 355.6 −436.3 3.43 35.3 355.6 −436.3 3.43 40.5 304.8 −436.3 7.67 31.1 355.6 −432.2 3.36 35.4 355.6 −432.2 3.37 40.5
2–4 465.9 355.6 561.7 5.69 38.1 355.6 584.1 6.14 37.6 355.6 584.1 6.14 42.8 304.8 584.1 13.7 30.0 355.6 587.2 6.21 37.5 355.6 587.2 6.22 42.7
4–5 −76.5 254.0 61.2 0.39 38.9 203.2 74.4 1.84 36.8 203.2 74.4 1.82 41.9 152.4 61.2 5.59 18.3 203.2 69.3 1.60 36.7 203.2 69.3 1.60 41.9
4–6 345.6 304.8 304.8 3.74 29.3 304.8 313.9 3.97 28.6 304.8 313.9 3.97 33.8 254.0 327.2 11.2 13.9 304.8 322.1 4.18 28.3 304.8 322.1 4.19 33.5
6–7 15.3 152.4 −25.5 0.98 26.7 203.2 −16.3 0.09 25.7 203.2 −16.3 0.08 30.9 152.4 −3.1 0.02 14.2 101.6 −8.2 0.75 25.2 101.6 −8.2 0.75 30.3
3–5 −454.7 254.0 −357.8 2.31 42.7 304.8 −336.4 4.55 40.8 304.8 −336.4 4.45 45.9 254.0 −336.4 11.8 29.3 304.8 −332.3 4.45 40.9 304.8 −332.3 4.45 46.1
5–7 −110.1 254.0 −149.9 2.36 26.9 254.0 −140.7 2.08 26.6 254.0 −140.7 2.07 31.7 203.2 −128.4 5.50 8.4 254.0 −132.5 1.83 26.5 254.0 −132.5 1.83 31.7

Cost 400 388 428 309 376 416

Iteration number 3 5 1 5 4 1
Precision criterions
hmax (m) 11.25 11.25 6.14 13.72 11.25 6.22
hmin (m) 0.392 0.087 0.85 0.015 0.75 0.75
hmax =hmin P 28.71 128.94
P P7.256 928.6
P 15.04
P P8.28
hi ¼ 30:50m hi ¼ 33:34 hi ¼ 28:16 hi ¼ 66:73 hi ¼ 33:66 hi ¼ 28:50
Ö. Ekinci, H. Konak
An optimization strategy for water distribution networks 183

Table 4 Solutions for the Alperovits and Shamir (two-loop) network

Literature Solutions Cost (units) Pipe diameter (mm)

Min Max

Alperovits and Shamir 1977 Linear programming 497.525 101.6 508.0

Goulter et al. 1986 Linear programming 435.015 25.4 508.0
Kessler and Shamir 1989 Linear programming 417.500 50.8 457.2
Fujiawara and Khang 1990 Linear programming 415.271 25.4 457.2
Eiger et al. 1994 Linear programming 402.352 25.4 457.2
Loganathan et al. 1995 Simulated annealing 412.931 25.4 457.2
Savic and Walters 1997 Genetic algorithm 419.000 25.4 457.2
Sherali and Smith 1997 Linear programming 436.684 76.2 457.2
Sherali et al. 1998 Linear programming 436.915 25.4 457.2
Abebe and Solomatine 1998 Genetic algorithm 419.000 25.4 457.2
Cunha and Sousa 1999 Simulated annealing 419.000 25.4 457.2
Todini 2000 Resilience index 419.000 25.4 457.2
Geem et al. 2002 Harmonic search 419.000 25.4 457.2
Eusuff and Lansey 2003 Shuffled frog leaping algorithm 419.000 25.4 457.2
Prasad and Park 2004 Genetic algorithm 419.000 25.4 457.2
Present model Minimum head loss strategy 416.000 101.6 457.2

1994), 405.381 (Loganathan et al. 1995), 436.684 (Sherali and Smith 1997), 436.915
(Sherali et al. 1998), and 419.000 (Savic and Walters 1997, Abebe and Solomatine 1998,
Cunha and Sousa 1999, Todini 2000, Geem et al. 2002, Eusuff and Lansey 2003, Prasad
and Park 2004). On the other hand, the proposed model produces the cost value of 416.000
units with less number of iteration (Ekinci 2003, Ekinci and Konak 2005; Table 4).
The minimum head loss strategy is not aimed only at the minimum costs due to its objective
functions, which means that the main object of the optimization is to provide expectations of its
objective functions sufficiently while the design cost reaches to an optimum value. The design
cost is one of the most less cost value than the others. Another important result from the
proposed model is the minimum pipe diameter value as 101.6 mm (4 in.) which is obtained
from the model solution that can be used in practice. This situation shows that the model yields
also a homogenous pipe diameter distribution as in Table 1 (Ekinci 2003, Ekinci and Konak
2005). Among others only Alperovits and Shamir (1977) study gave 4 in. as the minimum
pipe diameter where design cost is the largest in the literature.
If the problem is taken from the objective functions and constraints points of view then
the precision degree of the proposed model is acceptable within the expectations of decision
makers (Tables 2 and 3). This situation shows the model’s interior consistency.

7 Conclusions

In design of the water distribution network, the head losses through the pipelines should be
minimum and nearly equal as possible as they can be. In this study, the initial flows and pipe
diameters are determined by a weighting optimization process to get a reliable solution
satisfying these conditions in a concrete manner. A datum optimization procedure is realized to
get a less cost design alternative giving the optimum conditions for the system. The proposed
model is shown to perform and guarantee a proper solution with a minimum head loss
optimization strategy and the main problems included in the three important stages are
controlled simultaneously. A more effective decision support system could be developed in the
future by paying attention to this model providing consistent, objective and reliable solutions.
184 Ö. Ekinci, H. Konak

Acknowledgements The writers are grateful to Dr. Aykan Karademir for his valuable contributions and
comments.

References

Abebe AJ, Solomatine DP (1998) Application of global optimization to the design of pipe networks. Proc.,
Hydroinformatics’ 98, Copenhagen Balkema Rotterdam, pp. 986–996
Alperovits E, Shamir U (1977) Design of optimal water distribution systems. Water Resour Res 13(6):885–
458
Cunha MC, Sousa J (1999) Water distribution network design optimization: simulated annealing approach. J
Water Resour Plan Manage 125(4):215–221
Cunha MC, Sousa J (2001) Hydraulic infrastructures design using simulated annealing. J Infrastruct Syst 7
(1):32–39
Dandy GC, Engelhardt MO (2006) Multi-objective trade-offs between cost and reliability in the replacement
of water mains. J Water Resour Plan Manage 132(2):79–88
Eiger G, Shamir U, Ben-Tal A (1994) Optimal design of water distribution networks. Water Resour Res 30
(9):2637–2646
Ekinci Ö (2003) Su dağıtım şebekeleri için bir optimizasyon modeli (An optimization model for water
distribution networks). Dissertation, Institute of Natural Sciences, University of Kocaeli (in Turkish)
Ekinci Ö, Konak H (2005) Su dağıtım şebekeleri için minimum yük kayıplı bir optimizasyon stratejisi (A
minimum head loss optimization strategy for water distribution networks). Bull Chamb Survey Cadastre
Eng Turkey 2005(92):44–54 (in Turkish)
Eusuff MM, Lansey KE (2003) Optimization of water distribution network design using the shuffled frog
leaping algorithm. J Water Resour Plan Manage 129(3):210–225
Farmani R, Walters GA, Savic DA (2005) Trade-off between total cost and reliablity for anytown water
distribution network. J Water Resour Plan Manage 131(3):161–171
Fujiwara O, Khang DB (1990) A two-phase decomposition method for optimal design of looped water
distribution networks. Water Resour Res 26(4):539–549
Geem ZW, Kim JH, Loganathan GV (2002) Harmony search optimization: Application to pipe network
design. Int J Model Simul 22(2)
Geem ZW, Tseng CL (2002) Engineering applications of harmony search. Proc., Genetic and Evolutionary
Computation, New York City, pp. 69–173
Geem ZW, Kim TG, Kim JH (2000) Optimal layout of pipe networks using harmony search. Proc.,
Hydroscience and Engineering, Seul
Goulter IC (1992) System analysis in water-distribution network design: from theory to practice. J Water
Resour Plan Manage 118(3):238–248
Goulter IC, Lussier BM, Morgan RD (1986) Implications of head loss path choice in the optimization of
water distribution networks. Water Resour Res 22(5):819–822
Grafarend E, Heister H, Kelm R, Kropff H, Schaffrin L (1979) Optimierung Geodaetischer Messoperationen.
Herbert Wichmann Verlag, Karlsruhe
Kessler A, Shamir U (1989) Analysis of the linear programming gradient method for optimal design of water
supply networks. Water Resour Res 25(7):1469–1480
Konak H (1994) Yüzey Ağların Optimizasyonu (Surface Network Optimization). Dissertation, Institute of
Natural Sciences, Karadeniz Technical University (in Turkish).
Loganathan GV, Greene JJ, Ahn TJ (1995) Design heuristic for globally minimum cost water-distribution
systems. J Water Resour Plan Manage 121(2):182–192
Maier HR, Simpson AR, Zecchin AC, Foong WK, Phang KY, Seah HY, Tan CL (2003) Ant colony
optimization for design of water distribution systems. J Water Resour Plan Manage 129(3):200–209
Morgan DR, Goulter IC (1985) Optimal urban water distribution design. Water Resour Research 21(5):642–
652
Prasad TD, Park NS (2004) Multiobjective genetic algorithms for design of water distribution networks. J
Water Resour Plan Manage 130(1):73–82
Savic DA, Walters GA (1997) Genetic algorithms for least cost design of water distribution networks. J
Water Resour Plan Manage 123(2):67–77
An optimization strategy for water distribution networks 185

Sherali HD, Smith EP (1997) A global optimization approach to a water distribution network design
problem. J Glob Optim 11:107–132
Sherali HD, Totlani R, Loganathan GV (1998) Enhanced lower bounds for the global optimization of water
distribution networks. Water Resour Res 34(7):1831–1841
Todini E (2000) Looped water distribution networks design using a resilience index based heuristic approach.
Urban Water 2:115–122
Tolson BA, Maier HR, Simpson AR (2004) Genetic algorithms for reliability-based optimization of water
distribution systems. J Water Resour Plan Manage 130(1):63–72
Walski TM (2001) The wrong paradigm-why water distribution optimization doesn’t work. Editorial. J Water
Resour Plan Manage 127(4):203–205

View publication stats