An Iterative Penalty Method for the Optimal Design of Pipe

Networks

M. H. Afshar1,*, A. Afshar2, M. A. Mariño, Hon. M. ASCE3

Received: May 2008, Accepted: March 2009

Abstract: This paper presents the application of an iterative penalty method for the design of water distribution pipe
networks. The optimal design of pipe networks is first recasted into an unconstrained minimization problem via the use
of the penalty method, which is then solved by a global mathematical optimization tool. The difficulty of using a trial
and error procedure to select the proper value of the penalty parameter is overcome by an iterative use of the penalty
parameter. The proposed method reduces the original problem with a priori unknown penalty parameter to a series of
similar optimization problems with known and increasing value of the penalty parameters. An iterative use of the
penalty parameter is then implemented and its effect on the final solution is investigated. Two different methods of
fitting, namely least squares and cubic splines, are used to continuously approximate the discrete pipe cost function
and are tested by numerical examples. The method is applied to some benchmark examples and the results are
compared with other global optimization approaches. The proposed method is shown to be comparable to existing
global optimization methods.

Keywords: Pipe Networks; Water distribution; Design; Optimization.

1. Introduction constrained minimization where the hydraulic

requirements constitute some of the constraints
A typical water distribution network is a of the problem. Various investigators have
collection of pipes, reservoirs, pumps, and addressed this problem in a number of different
different kinds of valves connected to each other ways during the past decades. Enumeration
in order to meet specified demand at nodes. techniques, though reliable, suffer from limited
Basically, the optimal design of pipe networks is practical application due to an extraordinary wide
a multi-objective task involving hydraulics, search space and consequently an enormous
reliability, and water quality. The multi-objective computational time required when applied to
design of water distribution networks real-world size networks, where optimization is
necessitates the development of proper mostly needed (Yates et al. 1984). The class of
algorithms before it can be effectively applied in constrained minimization, in particular
practice. However, investigations indicate that decomposition methods, has become popular in
much can be gained by handling the matter as a recent years. These algorithms can be divided
single objective problem of component design in into two main groups, namely linear and
which only the size optimization of different nonlinear programming methods. The first linear
components such as pipes, tanks, etc with a least decomposition method, called linear
cost objective is considered. programming gradient (LPG) suggested by
Optimal design of pipe networks, when Alperovits and Shamir (1977) and later extended
formulated mathematically, is clearly one of by Kessler and Shamir (1989), assumes the
length of the pipes in each arc to be the decision
variable for a given flow distribution and solves
* Corresponding Author: E-mail: mhafshar@iust.ac.ir
1 Associate Prof., Dept. of Civil Engineering, Iran Univ.
the linear programming problem so constructed.
of Science and Tech., Tehran, Iran. Modifications and comments to the original LPG
2 Prof., Dept. of Civil Engineering, Iran Univ. of method were made by Quindry et al. (1979),
Science and Tech., Tehran, Iran. E-mail: Saphir (1983), and Fujiwara et al. (1987).
a_afshar@iust.ac.ir Quindry et al. (1981) presented an analogous
3 Prof., Dept. of Land, Air and Water Resources and
Dept. of Civil and Environmental Engineering, Univ.
approach to the LPG method in which the
of California, Davis, CA. E-mail: problem is solved for a given set of hydraulic
MAMarino@ucdavis.edu heads. The first of the nonlinear decomposition

International Journal of Civil Engineerng. Vol. 7, No. 2, June 2009 109

models was devised by Mahjoub (1983) who layout in order to provide the required pressure
started with an assumed initial link flow rate and and quantity of water at every demand node. A
then solved for link head losses. The obtained penalty method is used to formulate the optimal
optimal link head losses were then fixed and the design of the pipe network as an unconstrained
cost minimized for link flows. The whole optimization problem, which is then solved by a
procedure was repeated until no improvement general purpose optimization package called
could be attained. An improvement to the above DOT (Vanderplaats, Miura and Associates 1994).
method was suggested by Fujiwara and Khang The formulation of the problem is described next.
(1990), who used the Lagrange multipliers of the The effect of the penalty parameter and the cost
optimal solution obtained in the first phase to function approximation on the final solution is
modify the flow distribution, achieving a then investigated and finally some verifying
reduction in system cost before the second phase numerical examples are presented.
is started. Recently, genetic algorithms (GA)
have been applied to the design of pipe networks 2. Problem Formulation
(Simpson and Goldberg 1993; Simpson et al.
1994; Savic and Waters 1997, Wu and Simpson The optimal design of a pipe network with a
2002; Wu et al. 2002). Some problems associated pre-specified layout in its standard form is
with GA are the uncertainty about the termination described as follows:
of the search and, as in all random search m
methods, the absence of a guarantee for the Min ∑L C
l =1
l l (1)
global optimum.
Global (unconstrained) optimization methods Subject to:
are among the first proposed for pipe network
design (Abebe and Solomatine, 1999). . These 2.1. Hydraulic constraints:
methods convert the original constrained
problem to an unconstrained problem via the use ∑q
l∈ k
l
= Qk k = 1, 2,..., n (2)
of a penalty function or a Lagrange multiplier
method. The solution of the resulting problem is
therefore much easier but the final solution is ∑J l =0 p = 1, 2,..., P (3)
l∈ p
inferior to methods so far described. However, a
global optimization formulation of the pipe
network design problem has the advantage of q l = Kchl d αl ( J l / Ll ) β (4)
simplicity and practicality for engineering use.
Furthermore, all the random search and evolution
methods such as GA employ some sort of global 2.2. Head and flow constraints:
optimization formulation of the pipe network
optimization. Abebe and Solomatine (1999) used H k ≥ H min k = 1,2,..., n (5)
the penalty method to define the pipe design
problem as an unconstrained optimization
problem, which was then solved by a global ql ≥ qmin l = 1, 2,..., m (6)
optimization package, GLOBE (Solomatine
1998), incorporating four different algorithms.
In this paper, the optimal design of pipe 2.3. Pipe size constraints:
network is addressed from a hydraulic point of
view. Only steady state condition is considered d min ≤ d l ≤ d max l = 1,2,..., m (7)
here; however, the extension to the case of
dynamic condition is straightforward. The paper where Ll = length of the lth pipe ; Cl = per unit
addresses the determination of optimal diameters cost of the lth pipe ; dl = diameter of the lth pipe
of pipes in a network with a pre-determined ; ql = flow in the lth pipe; Jl = head loss in the lth

110 International Journal of Civil Engineerng. Vol. 7, No. 2, June 2009

pipe ; Hk = nodal head at the kth node; Qk = 3. Penalty prameter treatment
consumption at node k ; Hmin = minimum
allowable hydraulic head; qmin = minimum It is a common practice in the penalty
allowable pipe flow; dmin and dmax = minimum formulation of the pipe network design to assume
and maximum allowable pipe diameter, a large number for the penalty parameter and
respectively; and n, p, m = total number of minimize the penalized objective function
nodes, loops, and links in the network, (Abebe and Solomatine 1999). The first difficulty
respectively; K = Hazen-Williams constant with this approach is that the numerical value of
whose value depends on the system of units used; the penalty coefficient is not known a priori and
ch = Hazen-Williams roughness coefficient hence a trial and error procedure should be used
whose value depends on the pipe characteristics; to get the proper value, ensuring the enforcement
α = 2.63 and β = 0.54. The first set of constraints of the constraints included in the objective
(2)-(4) describes the flow continuity at nodes, function. Secondly, this large value of the penalty
head loss balance in loops, and the Hazen- coefficient could change from one application to
Williams equation. The second set (5)-(6) refers another, requiring the proper setting of the
to the minimum nodal head and pipe flow penalty parameter for each application. From a
requirements while the last constraint (7) requires mathematical point of view, the penalty method
the optimal pipe diameters to be between the is an iterative approach, which asymptotically
maximum and minimum available pipe converges to the solution of the original
diameters, respectively. Equation (1) describes constrained problem as the value of the penalty
the total cost of the pipes in the network. coefficient increases. Iterative nature of the
In this work, hydraulic constraints are satisfied method eliminates the need for choosing the
via the use of an element by element simulation numerical value of the penalty coefficient via a
program, which explicitly solves the set of trial and error process. Furthermore, iterative
hydraulic constraints for nodal head and pipe implementation of the penalty method can
flows (Afshar 2001). The second set of improve the effectiveness of the method as will
constraints are, therefore, included in the be shown later when numerical examples are
objective function via the use of an exterior considered. In this paper, the following strategies
penalty method resulting in the following are used to set the value of the penalty parameter:
penalized problem:
m m n Set the initial design
Min ∑ Cl Ll + ∑ α l (ql − qmin ) + ∑ α k (H k − H min )
2 2
(8) d l = d max l = 1,2,..., m
l =1 l =1 k =1 or
d l = d min l = 1,2,..., m

Where α l, α k = penalty parameters with large

values when corresponding constraints are Set penalty parameters α l and α k = 1
violated, and zero when the constraints are
satisfied. For each network with the known pipe
Minimize the penalized objective
size diameters obtained, the distribution of the function defined by Eq. (8)
nodal heads and pipe flows are calculated by the
simulation program and used to calculate the Check if the nodal pressures and pipe
Yes Stop
flow, or velocity, are feasible.
penalty terms in Eq. (8). The pipe size constraints
No
are handled by the optimization package, DOT,
as box constraints when the pipe diameters are For pressure constraint violation, set α k = 10α k
For flow, or velocity, constraint violation, set α l = 10α l
taken as decision variables. Next, we consider the
numerical treatment of the penalty term and the
type of the analytical function used to
Set the optimized solution as the
approximate the cost per unit length of the pipes initial design for the next iteration

and present the results obtained.

Fig. 1 Iterative penalty method flow diagram

M. H. Afshar, A. Afshar and M. A. Mariño, Hon. M. Asce 111

 Table 1 Cost data for the two-loop network
Diameter(cm) 2.54 5.08 7.62 10.16 15.24 20.32 25.4 30.48 35.56 40.64 45.72 50.8 55.88

Cost (units/m) 2 5 8 11 16 23 32 50 60 90 130 170 300

(1) Choose a very large value for the penalty to a set of N discrete data can therefore be easily
parameter (herein, 10 Cmax, where Cmax is the cost constructed by defining a different cubic spline
of the most expensive design); and (2) use an for each N-1 reach such that their first and
iterative approach in which the value of the second derivative at N discrete points are
penalty parameter is increased by an order of continuous. This leads to 2(N-1) equations
magnitude starting from unity, until the stemming from the fact that each of the N-1
constraints are exactly satisfied. The iterative piecewise functions at their two end points
procedure employed here is depicted in Fig. 1. should equate the corresponding discrete data
plus 2(N-2) equations stemming from the
4. Pipe cost function continuity of the first and second derivative at N-
2 internal points, summing up to 4(N-1) - 2
An analytical function of the form a d b is equations in terms of 4(N-1) unknowns defining
usually used to approximate the cost per unit N-1 piecewise third-order functions. Two
length of commercial pipes when a continuous boundary conditions regarding the curvature of
method of optimization is employed for the the spline at the two extreme points of the set are
optimum design of the networks. Unit cost used to close the system of simultaneous linear
parameters, a and b, are usually obtained via a equations. The size of the system of equations
least- squares fit of the aforementioned function can be reduced by a proper formulation of the
to the discrete cost data. It is obvious that this fit problem in terms of the value of the second
would not be exact in terms of both function derivatives at the N discrete points. The solution
value and its gradient. Most of the common of this system yields the value of the second
mathematical search methods use function value derivatives required to define each of the cubic
or its derivative to find the optimum value of the splines at N-1 reach. This function can then be
objective function. Thus, it is expected that used to approximate the value of the discrete
improving the accuracy of the approximate cost function at any arbitrary points, including the
function would improve the optimization results. discrete data points. It is obvious that cubic spline
In this work, a piecewise cubic spline is used to approximation would always yield exact values
approximate the cost per unit length of the pipes at the discrete data points and therefore is a more
and the results are compared with the usual least- accurate representation of the approximated
squares form. Cubic splines are, by definition, function compared to a least-squares fit. Figure 2
third-order functions enjoying first and second shows the least squares and piecewise cubic
derivative continuity. A piecewise cubic spline fit spline approximations of the discrete cost
function used in the literature for the two-loop
network (Table 1).

5. Test problems

The first problem addresses a two-loop

network with 8 pipes, 7 nodes, and one reservoir
as shown in Fig. 3 (Abebe and Solomatine 1999).
All the pipes are 1,000 m long and the Hazen-
Williams coefficient is assumed to be 130 for all
the pipes. The minimum nodal head requirement
Fig. 2 Cubic spline and least-squares fit to the data of
for all demand nodes is 30 m. There are 14
Table 1 commercially available pipe diameters as listed

112 International Journal of Civil Engineerng. Vol. 7, No. 2, June 2009

Table 2. Nodal demand and elevation data for the two-loop
network The problem is solved again, using a cubic
spline approximation of the pipe cost function. To
Node Demand Ground level make the comparison possible, these solutions
(m3 /h) (m) along with the solutions shown in Table 3 are
1 ---- 210.0 converted to a discrete set of commercially
2 100.0 150.0 available diameters using the standard split pipe
3 100.0 160.0 techniques (Fujiwara and Khang 1990) and
4 120.0 155.0 shown in Table 4. In this method, each link with
5 270.0 150.0 a given diameter d is replaced by two pipe with
6 330.0 165.0 diameters du and dl where du and dl are the
7 200.0 160.0 upper and lower commercially available
diameters, respectively. The lengths of these
in Table 1 while Table 2 shows the nodal pipes are then calculated such that the head loss
elevation and demands for the network. in the link remains unchanged. It can be easily
First, the effect of the penalty parameter seen that cubic spline approximation of the cost
treatment is investigated. For this, the problem is function yields a better solution for all different
solved using strategies (1) and (2) described initial solutions. It should, however, be noted that
earlier, with two different initial solutions. Table much of the improvements of the obtained
3 compares the results obtained and the number solutions are due to the use of proposed iterative
of function evaluations required by these penalty method. The results shown in Tables 3
methods. Here dmin and dmax refer to the initial and 4 also indicate that using the most expensive
designs in which diameters of all the pipes are design as an initial guess for the optimization
taken as the minimum and maximum available procedure, represented by dmax in the Table 3,
diameters, respectively. It can be seen that the yields the better solution and hence would be
iterative use of the penalty parameter used in all other tests presented hereafter. The
significantly improves not only the solution for near optimality of the solution obtained with the
all cases but also reduces the number of function proposed method presented in Table 4 is evident
evaluation to get the solution. This is particularly by the fact that the nodal heads at four nodes ,ie;
significant for the cases in which the most nodes 3, 5, 6 and 7, of the network is
expensive solution is taken as the initial guess for approximately equal to the minimum head
the optimization process, where the network cost requirement set by the problem definition. Table
is reduced from 469556 to 431630. This strategy, 5 compares the cheapest solution obtained here
therefore, will be used in all the examples with the results obtained by some other
presented hereafter unless it is stated otherwise. It optimization methods for this problem. It is
should be noted that the results presented in Table clearly seen that the proposed method yield the
3 were obtained with a least-squares fit to second best solution amongst all the methods
approximate the cost function. used in the literature. These solutions are

12
13 12
11
11
5 3 1 33 34 26 27 28 15 14 13 10 9
6 4 2 1 31 32 25 26 27 16 15 14 10 9
16
32 17 8
17 8
30 24 18
7
18
6 4 2 31 19 7

19 6
29 28 23 20 3 4 5 6
30 29 23 20 2 3 4 5
21
21 2
7 5 3 22 1
8 7 22 1

Fig. 3 Two-loop network Fig. 4 Hanoi Network

M. H. Afshar, A. Afshar and M. A. Mariño, Hon. M. Asce 113

 Table 3 Optimal pipe diameters and corresponding nodal heads for the two-loop network

Pipe data Nodal data

Strategy (a) (b) (a) (b)

Initial Guess* dmax dmin dmax dmin dmax dmin dmax dmin

Pipe Diameter (cm) Node Head (m)

1 49.66 50.11 47.77 49.19 1 ------ ------ ------ ------

2 38.30 41.30 26.13 36.24 2 55.48 55.68 54.55 55.27

3 32.46 33.98 39.62 33.78 3 40.91 42.63 31.43 39.29

4 2.54 8.07 2.54 2.54 4 43.83 45.15 44.57 44.80

5 32.30 30.40 37.31 30.86 5 46.80 42.26 30.01 43.79

6 2.54 2.54 25.29 2.54 6 30.00 30.01 30.01 30.01

7 36.37 32.00 23.85 34.26 7 30.01 30.00 30.01 30.01

8 23.70 24.99 2.54 26.72

Cost 469,556 477,266 431,630 462,170

(units)
Evaluations** 1,100 1,532 968 903

* Initial diameters used for the pipe diameters to start the optimization process. Note that NLP

methods require an initial solution to start the process. dmax refers to an initial guess in which

all the pipe diameters are taken as the maximum diameter possible while dmin refers to the

minimum diameters.

** The number of network evaluations required by the NLP method to converge to the final

solution which can be considered as a measure of computational cost of the method.

obtained on a P4-800 MHZ with 5 second of strategies (a) and (b) are shown in Table 7 in its
CPU time. split form along with some of the results obtained
The second test problem considers the Hanoi by other investigators for this problem. It is
network with 34 pipes, 31 demand nodes, and observed that much more can be gained by the
one reservoir as shown in Fig. 4 (Fujiwara and use of the strategy (b) in larger problems. These
Khang 1990; Abebe and Solomatine 1999). The solutions are obtained using the most expensive
minimum nodal head requirement at all demand design as the initial guess. It should be remarked
nodes is 30 m. Table 5 shows diameters of that the solution of Fujiwara and Khang (1990) to
commercially available pipes. The cost of the this problem was reported to be infeasible by
pipes per unit length are calculated based on the Dandy et al. (1996). It is again seen that the
analytical cost function 1.1d1.5 (Fujiwara and proposed method has been able to produce the
Khang 1990). Solutions to this problem using best solution for this problem. These solutions

114 International Journal of Civil Engineerng. Vol. 7, No. 2, June 2009

 Table 4 Optimal split pipe diameters and corresponding nodal heads for the two-loop network.
Pipe data Nodal data

Fit Cubic spline Least squares Cubic spline Least squares

Initial Guess dmax dmin dmax dmin dmax dmin dmax dmin

Pipe Solution [length (m) and diameter (cm)] Node Head (m)

1 474.90 45.72 799.87 45.72 516.59 45.72 250.19 45.72 1 ------ ------ ------ ------

525.02 50.80 200.13 50.80 483.41 50.80 749.81 50.80

2 936.27 25.40 732.71 35.56 777.51 25.40 815.41 35.56 2 54.66 53.78 54.55 55.27

63.73 30.48 267.29 40.64 222.49 30.48 184.59 40.64

3 235.01 35.56 183.32 30.48 143.50 35.56 253.97 30.48 3 30.14 38.06 31.43 39.29

764.99 40.64 816.68 35.56 856.50 40.64 746.03 35.56

4 999.96 2.54 1000.00 2.54 999.77 2.54 1000.00 2.54 4 44.31 43.64 44.57 4480

0.04 5.08 0.00 5.08 0.23 5.08 0.00 2.54

5 470.22 35.56 461.45 30.48 563.35 35.56 889.85 30.48 5 30.00 43.62 30.01 43.79

529.78 40.64 538.55 35.56 436.65 40.64 110.15 35.56

6 9.98 20.32 999.95 2.54 9.84 20.32 999.94 2.54 6 30.01 30.01 30.01 30.01

990.02 25.40 0.05 5.08 990.16 25.40 0.06 5.08

7 105.25 20.32 931.40 35.56 183.70 20.32 177.93 30.48 7 30.01 30.01 30.01 30.01

894.75 25.4 68.60 40.64 816.30 25.4 822.07 35.56

8 999.97 2.54 566.88 25.4 1000 2.54 627.08 25.4

0.03 5.08 433.12 30.48 0.00 5.08 372.92 30.48

Cost 409,954 425,430 410,395 435,025

(units)
1,711 1,558 968 903
Evaluations

are obtained on a P4-800 MHZ with 20 second of

CPU time.
The third test problem concerns the
rehabilitation of the New York City water supply
network with 21 pipes, 20 demand nodes, and
one reservoir as shown in Fig. 5 (Dandy et al.
1996). Commercially available pipe diameters
and their corresponding costs are shown in Table
8 while pipe and nodal data of the network are
shown in Table 9. The inclusion of zero diameter
with zero cost in the table of available pipe
diameter is intended to widen the search space to
include the 'no pipe' option for all of the links
present in the network. Solutions to this problem
using strategies (a) and (b) are shown in Table 10
in its split form along with the cheapest
continuous and discrete pipe solutions reported
by Dandy et al. (1996). Again, the improvement Fig. 5 New York tunnel network

M. H. Afshar, A. Afshar and M. A. Mariño, Hon. M. Asce 115

 Table 5 Optimal pipe diameters obtained with various methods for the two-loop network.
Pipe length (m) and/or diameter (cm)

Pipe Kessler and Shamir Eiger et al. (1994) Savic and Walters Abebe and Solomatine Present work

(1989) (1997) (1999)

GA1 GA2 GA ACCOL*

1 1000.00 45.72 1000.00 45.72 45.72 50.80 45.72 55.88 474.98 45.72

525.02 50.80

2 66.00 30.48 238.02 30.48 25.40 25.40 35.56 45.72 936.27 25.40

934.00 25.40 761.98 25.40 63.73 30.48

3 1000.00 40.64 1000.00 40.64 40.64 40.64 35.56 50.80 235.01 35.56

764.99 40.64

4 713.00 7.62 1000.00 2.54 10.16 2.54 2.54 7.62 999.96 2.54

287.00 5.08 0.04 5.08

5 836.00 40.64 628.86 40.64 40.64 35.56 35.56 40.64 470.22 35.56

164.00 35.56 371.14 35.56 529.78 40.64

6 109.00 30.48 989.05 25.40 25.40 25.40 2.54 10.16 9.98 20.32

891.00 25.40 10.95 20.32 990.02 25.40

7 819.00 25.40 921.86 25.40 25.40 25.40 35.56 45.72 105.25 20.32

181.00 20.32 78.14 20.32 894.75 25.40

8 920.00 7.62 1000.00 25.40 25.40 25.40 30.48 40.64 999.97 25.40

80.00 5.08 0.03 5.08

Cost 417,500 402,352 419,000 420,000 424,000 447,000 409,954

(units)
* Adaptive Cluster Covering with Local Search

Table 6 Available pipe diameter for the Hanoi network

Diameter (cm) 30.48 40.64 50.80 60.96 76.20 100.16

made in the solution by iterative setting of the scale size problems compared to real world pipe
penalty parameter is self-evident. This strategy network examples. Though the applicability of
yields one of the cheapest solution ever achieved the proposed method does not basically depend
for New York tunnels emphasizing on the ability on the size of the problem, its performance might
of the proposed method for the optimal solution be affected by the size of the problem to be
of large scale pipe network optimization solved.
problems. These solutions are obtained on a P4-
800 MHZ with 10 second of CPU time. 6. Concluding remarks
A note should be made here regarding the size
of the problems considered. Although the search A penalty method was presented for
space size of these problems are rather large, but converting the optimal design of pipe networks to
they can only be considered as small to medium an unconstrained problem, which is then solved

116 International Journal of Civil Engineerng. Vol. 7, No. 2, June 2009

 Table 7 Optimal pipe diameters obtained by various methods for the Hanoi network.
Pipe length (m) and/or diameter (cm)

Pipe Present work Fujiwara and Khang Abebe and Solomatine

(1990) (1999)

Strategy (a) Strategy (b) GA ACCOL

1 0.3 76.2 0.01 76.2 100.0 101.6 101.6 101.6

99.7 101.6 100.0 101.6

2 0.62 76.2 0.9 76.2 1350.0 101.6 101.6 101.6

1349.4 101.6 1349.1 101.6

3 13.3 76.2 1.3 76.2 50.0 76.2 101.6 101.6

886.7 101.6 898.7 101.6 850.0 101.6

4 381.3 76.2 1.7 76.2 60.0 76.2 101.6 101.6

768.7 101.6 1148.3 101.6 1090.0 101.6

5 289.7 76.2 2.3 76.2 150.0 76.2 76.2 101.6

1160.3 101.6 1447.8 101.6 1300.0 101.6

6 65.1 76.2 .64 76.2 90.0 76.2 101.6 76.2

385.0 101.6 449.4 101.6 360.0 101.6

7 312.6 76.2 71.9 76.2 350.0 76.2 101.6 101.6

537.4 101.6 778.1 101.6 500.0 101.6

8 392.6 76.2 140.5 76.2 460.0 76.2 76.2 101.6

457.3 101.6 709.5 101.6 390.0 101.6

9 471.4 76.2 217.8 76.2 570.0 76.2 76.2 61.0

328.6 101.6 582.2 101.6 230.0 101.6

10 184.3 60.96 51.5 60.96 690.0 61.0 76.2 101.6

765.7 76.2 898.5 76.2 260.0 76.2

11 902.5 76.2 400.2 60.96 190.0 50.8 76.2 76.2

297.5 101.6 799.8 76.2 1010.0 61.0

M. H. Afshar, A. Afshar and M. A. Mariño, Hon. M. Asce 117

 Table 7 Continued
Pipe length (m) and/or diameter (cm)

Pipe Present work Fujiwara and Khang Abebe and Solomatine

(1990) (1999)

Strategy (a) Strategy (b) GA ACCOL

12 118.4 61.0 189.9 50.8 329.0 50.8 76.2 101.6

3381.6 76.2 3310.1 61.0 210.0 61.0

13 89.1 61.0 388.7 40.6 110.0 40.6 40.6 40.6

713.7 76.2 411.3 50.8 690.0 50.8

14 317.6 61.0 200.0 30.5 100.0 30.5 61.0 40.6

128.4 76.2 300.0 40.6 400.0 40.6

15 427.5 61.0 546.9 30.5 550.0 30.5 76.2 76.2

122.6 76.2 3.1 40.6

16 2439.8 76.2 2712.5 30.5 30.0 40.6 76.2 30.5

290.2 101.6 17.5 40.6 2700.0 50.8

17 1233.2 76.2 716.0 40.6 260.0 50.8 76.2 50.8

516.8 101.6 1034.0 50.8 1490.0 61.0

18 747.4 76.2 233.8 50.8 330.0 61.0 101.6 61.0

52.6 101.6 566.3 61.0 470.0 76.2

19 124.5 61.0 112.0 50.8 150.0 61.0 101.6 76.2

275.5 76.2 288.0 61.0 250.0 76.2

20 143.5 76.2 3.4 76.2 620.0 76.2 101.6 101.6

2056.5 101.6 2196.6 101.6 1580.0 101.6

21 999.6 50.8 587.8 40.64 1260.0 40.6 50.8 76.2

500.4 61.0 912.2 50.8 240.0 50.8

22 348.1 50.8 311.0 30.48 500.0 30.5 50.8 76.2

151.9 61.0 189.0 40.6

118 International Journal of Civil Engineerng. Vol. 7, No. 2, June 2009

 Table 7 Continued
Pipe length (m) and/or diameter (cm)

Pipe Present work Fujiwara and Khang Abebe and Solomatine

(1990) (1999)

Strategy (a) Strategy (b) GA ACCOL

23 456.0 76.2 245.7 76.2 110.0 61.0 76.2 101.6

2194.1 101.6 2404.3 101.6 2540.0 76.2

24 720.1 50.8 685.2 76.2 110.0 40.6 40.6 101.6

509.9 61.0 544.8 101.6 1120.0 50.8

25 1147.2 50.8 1251.5 76.2 1070.0 40.6 50.8 101.6

152.8 61.0 48.5 101.6 230.0 50.8

26 88.9 40.6 750.9 50.8 850 30.5 30.5 61.0

761.1 50.8 99.2 61.0

27 128.8 50.8 222.4 30.5 240.0 50.8 61.0 76.2

171.2 61.0 77.6 40.6 60.0 61.0

28 210.4 76.2 750.0 30.5 210.0 50.8 50.8 30.5

539.6 101.6 0.0 40.6 540.0 61.0

29 149.0 40.6 3.07 30.5 250.0 40.6 61.0 40.6

1351.0 50.8 1496.9 40.6 1250.0 50.8

30 464.7 40.6 1018.8 30.5 1160.0 40.6 76.2 101.6

1535.3 50.8 981.2 40.6 840.0 50.8

31 419.2 30.5 1600.0 30.5 300.0 30.5 76.2 40.6

1180.8 40.6 0.00 30.5 1300.0 40.6

32 55.01 30.5 41.5 30.5 150.0 30.5 76.2 50.8

95.0 40.6 108.5 40.6

33 534.2 30.5 442.6 40.6 860.0 30.5 76.2 76.2

325.9 40.6 417.4 50.8

34 461.9 40.6 191.9 50.8 60.0 40.6 30.5 61.0

488.1 50.8 758.1 61.0 890.0 50.8

Cost ($) 7,086,466 6,142,063 5,562,000 7,000,000 7,836,000

M. H. Afshar, A. Afshar and M. A. Mariño, Hon. M. Asce 119

 Table 8 Pipe cost data for the New York tunnel network

Diameter 0 91.4 121.9 152.4 182.9 213.4 243.8 274.3

(cm)
Cost ($/m) 0 306.8 439.6 577.4 725.1 876.0 1036.7 1197.5

Diameter 304.8 335.3 365.8 396.2 426.7 457.2 487.7 518.2

(cm)
Cost ($/m) 1367.1 1538.7 1712.6 1893.0 2073.5 2260.5 2447.5 2637.8

Table 9 Pipe and nodal data for the New York tunnel network.
Pipe data Nodal data

Pipe Start Node End node Length Existing diameter Node Demand Min. total head

(m) (cm) (l/s) (m)

1 1 2 3535.6 457.2 1 reservoir 91.4

2 2 3 6035.0 457.2 2 2616 77.7

3 3 4 2225.0 457.2 3 2616 77.7

4 4 5 2529.8 457.2 4 2497 77.7

5 5 6 2621.2 457.2 5 2497 77.7

6 6 7 5821.6 457.2 6 2497 77.7

7 7 8 2926.0 335.3 7 2497 77.7

8 8 9 3810.0 335.3 8 2497 77.7

9 9 10 2926.0 457.2 9 4813 77.7

10 11 9 3413.7 518.2 10 28 77.7

11 12 11 4419.6 518.2 11 4813 77.7

12 13 12 3718.5 518.2 12 3315 77.7

13 14 13 7345.6 518.2 13 3315 77.7

14 15 14 6431.2 518.2 14 2616 77.7

15 1 15 4724.4 518.2 15 2616 77.7

16 10 17 8046.7 182.9 16 4813 79.2

17 12 18 9509.7 182.9 17 1628 83.1

18 18 19 7315.2 152.4 18 3315 77.7

19 11 20 4389.1 152.4 19 3315 77.7

20 20 16 11704.3 152.4 20 4813 77.7

21 9 16 8046.7 182.9

120 International Journal of Civil Engineerng. Vol. 7, No. 2, June 2009

 Table 10 Optimal pipe diameters obtained by various methods for the New York tunnel network.

Pipe length (m) and/or diameter (cm)

Pipe Present work Gessler Morgan and Goulter Bhave Dandy et al.

(1982) (1985) (1985) (1996)

Strategy (a) Strategy (b)

1 0 0 0 0 0 0
2 0 0 0 0 0 0
3 0 0 0 0 0 0
4 0 0 0 0 0 0
5 0 0 0 0 0 0
6 0 0 0 0 0 0
7 0 0 254 365.8 0 0
8 0 0 254 0 0 0
9 0 0 0 0 0 0
10 0 0 0 0 0 0
11 0 0 0 0 0 0
12 0 0 0 0 0 0
13 0 0 0 0 0 0
14 4087.3 213.4 0 0 0 0 0
2344.2 243.8
15 0 2832.9 43.8 0 0 136.4 304.8
1891.5 274.3
16 2826.4 365.8 695.8 213.4 243.8 87.4 213.4
5220.3 396.2 7350.9 243.9
17 1618.7 274.3 110.8 213.4 254 243.8 99.2 243.9
7891.1 304.8 9399.0 243.9
18 3179.7 152.4 116.1 182.9 203.2 213.4 78.2 213.4
4135.5 182.9 7199.1 213.4
19 0 529.3 182.9 203.2 152.4 54.4 182.9
3859.9 213.4
20 912.5 243.8 0 0 0 0 0
10791.9 274.3
21 2770.8 365.8 2096.9 152.4 203.2 213.4 81.5 182.9
5275.9 396.2 5949.8 182.9
Cost ($M) 66.90 38.95 41.80 39.20 40.18 38.80

by a general purpose optimisation code. The discrete pipe cost function. Numerical
performance of the method was improved by an experiments showed that the use of cubic spline
iterative use of the penalty parameter, which fit reduces the design cost compared to the design
significantly reduces the design cost compared to obtained via the use of conventional least-squares
the conventional use of the penalty method. The fit. Simplicity of both the basic method and the
method was further improved by using a cubic modification presented herein as well as
spline fit to continuously approximate the comparability of the results with other methods

M. H. Afshar, A. Afshar and M. A. Mariño, Hon. M. Asce 121

makes it a suitable choice for engineering use. and Sediment Control, Univ. of Kentucky,
Research into finding ways to use the same Lexington.
method for discrete optimization of pipe
networks is underway. [10] Goldberg, D. E. (1989). Genetic algorithms in
search, optimization and machine learning.
References Addison- Wesley Publishing Co., Reading,
U.K.
[1] Abebe, A. J., and Solomatine, D. P. (1999).
“Application of global optimization to the [11] Kessler, A., and Shamir, U. (1979). “Optimal
design of pipe networks.” 3rd Int. Conf. on design of water distribution systems.” Water
Hydroinformatics, Copenhagen, Denmark, 989- Resour. Res., 13(6), 885-900.
996.
[12] Mahjoub, Z. (1983). “Contribution a etude de
[2] Afshar, M. H. (2001). “An element-by-element optimization des reaseaux mailles.” Docteur
algorithm for the analysis of pipe networks.” Etat these, Institute Nationale Polytechnique de
Int. J. for Eng. Science, 12(3), 87-100. Toulouse, Toulouse, France.

[3] Alperovits, E., and Shamir, U. (1977). “Design [13] Michalewicz, Z. (1996). Genetic algorithms +
of optimal water distribution systems.” Water data structures = evolution programs. 3rd ed.,
Resour. Res., 13(6), 885-900. Springer-Verlag, New York, N.Y.

[4] Bhave, P. R. (1985). “Optimal expansion of [14] Morgan, D. R., and Goulter, I. C. (1985).
water distribution systems.” J. Environ. Eng., “Optimal urban water distribution design.”
111(2), 177-197. Water Resour. Res., 21(5), 642-652.

[5] Dandy, G. C., Simpson, A. R., and Murphy, L. [15] Quindry, G. E., Brill, E. D., Liebman, J. C., and
J. (1996). “An improved genetic algorithm for Robinson, A. R. (1979). “Comments on
pipe network optimization.” Water Resour. ‘Design of optimal water distrubution systems’
Res., 32(2), 449-458. by E. Alperovits and U. Shamir.” Water Resour.
Res., 15(6), 1651-1654.
[6] Eiger, G., Shamir, U., and Ben-Tal, A. (1994).
“Optimal design of water distribution [16] Quindry, G. E., Brill, E. D., and Liebman, J. C.
networks.” Water Resour. Res., 30(9), 2637- (1981). “Optimization of looped water
2646. distribution systems.” J. of Environ. Eng.,
107(4), 665-679.
[7] Fujiwara, O., Jenchaimahakoon, B., and
Edirisinghe, N. C. P. (1987). “A modified linear [17] Saphir, Y. N. (1983). “Optimization of water
programming gradient method for optimal distribution system.” M.S. thesis (in Hebrew),
design of looped water distribution networks.” Technion-Israel Institute of Technology, Haifa,
Water Resour. Res., 23(6), 997-982. Israel.

[8] Fujiwara, O., and Khang, D. B. (1990). “A two- [18] Savic, D.A., and Waters, G. A. (1997). “Genetic
phase decomposition method for optimal design algorithms for least-cost design of water
of looped water distribution networks.” Water distribution networks.” J. Water Resour.
Resour. Res., 26(4), 539-549. Planning and Management, 123(2), 67-77.

[9] Gessler, J. (1982). “Optimization of pipe [19] Simpson, A. R. and Goldberg, D. E. (1993).
networks.” Paper presented at International “Pipeline optimization using an improved
Symposium on Urban Hydrology, Hydraulics, genetic algorithm.” Research Report No. R109,

122 International Journal of Civil Engineerng. Vol. 7, No. 2, June 2009

 Dept. of Civil and Environ. Eng., Univ. of its application to water distribution systems,” J.
Adelaide, Adelaide, Australia. of Hydraul. Res., vol. 40, pp. 191-203.

[20] Simpson, A. R., Dandy, G. C., and Murphy, L. [22] Wu, Z.Y., Boulos, P.F., Orr, C.H. and Ro, J.J.
J. (1994). “Genetic algorithms compared to (2002) “Using Genetic Algorithms to
other techniques for pipe optimization.” J. Rehabilitate Distribution Systems,” J. Am.
Water Resour. Planning and Management, Water Works Assoc., 93(11), 74-85.
120(4), 423-443.
[23] Yates, D. F., Templeman, A. B., and Boffey, Y.
[21] Solomantine, D. P. (1998). B. (1984). “The computational complexity of
http://www.ihe.nl/hi/sol Vanderplaats, Miura the problem of determining least capital cost
and Associates. (1994). http://www.vrand.com designs for water supply networks.” Eng.
Wu, Z.Y. and Simpson, A.R. (2002) “A self- Optimization, 7 (2), 142-155.
adaptive boundary search genetic algorithm and

M. H. Afshar, A. Afshar and M. A. Mariño, Hon. M. Asce 123