International Journal of Computer Information Systems and Industrial Management Applications. (2013) pp. 308-316 ISSN 2150-7988 Volume 5 (2012) © MIR Labs, www.mirlabs.net/ijcisim/index.html Cross Entropy Optimization for Optimal Design of Water Distribution Networks A. Shibu1 and M. Janga Reddy2 1,2 Department of Civil Engineering, Indian Institute of Technology Bombay, Powai, Mumbai – 400 076, India. Phone: +91 22 2576 7320; Fax : +91-22-2576 7302 E-mail: 1shibuiitb@yahoo.in (Corresponding author) Abstract: This paper presents cross entropy (CE) based methodology for optimal design of water distribution network (WDN). Design of WDN involves selection of suitable diameter for each pipe in the network from the list of commercially available diameters. The CE methodology is applied to two bench mark WDN design problems taken from literature for validation. The first WDN problem deals with determining optimal pipe sizes for planning a new system, while the second WDN deals with rehabilitation of existing WDN by parallel piping. The performance of CE is compared with the results of past studies and it is found that the CE resulted in good optimal solutions. Then, the model is applied to a case study in India. The results suggest that CE method is very effective in optimal design of water distribution networks and has the capability of rapid convergence to optimum solutions. Keywords: water distribution networks, metaheuristics, cross entropy, optimization. I. Introduction Water distribution network (WDN) consists of a set of pipes of different diameters and lengths connected with one another at various junctions called nodes. The diameters and lengths of pipes are designed in such a manner that they deliver the required amount of water with sufficient pressure to the demand nodes without failure. The optimal design of WDN aims to find a combination of the diameters that are feasible and results in minimum cost. Several researchers have formulated different models for optimal design of WDNs. Few studies modeled it as a nonlinear model and solved using Nonlinear Programming (NLP) techniques by treating discrete pipe sizes as continuous variables. The main disadvantage of these NLP methods is the required rounding-off of continuous decision variables to commercially available sizes, sometimes which can lead to network infeasibilities as well as raise questions on optimality of the adjusted solution. Some other studies formulated linear models and solved it using Linear Programming (LP) techniques. However, these methods are capable of maintaining the constraint on discrete pipe sizes (without the need for rounding off solutions), but it requires approximation of non-linear functions, which may not represent the reality as it is. In spite of development of many conventional techniques for optimization, each of these techniques has its own limitations. To overcome those limitations, recently metaheuristic techniques are being used for solving combinatorial optimization problems. By using these techniques, the given problem can be represented more realistically. These also provide ease in handling the non-linear relationships of the formulated model [1]. Genetic algorithms, particle swarm optimization, ant colony optimization algorithm, cross entropy algorithm etc. are some of the techniques fall in this category. These evolutionary algorithms search from a population of points, so there is a greater possibility to cover the whole search space and locating the global optimum. The stochastic search approaches that were used for WDN design include genetic algorithms ([2]; [3]; [4]; [5]; [1]; [6]), Simulated annealing [7], shuffled leaping frog algorithms [8], ant colony optimization algorithms [9], cross entropy algorithms [10] etc. These techniques improve the quality of the solution over the iterations by using heuristics. The cross entropy (CE) method was motivated by an adaptive algorithm for estimating probabilities of rare events in complex stochastic networks, which involves variance minimization. Later it was modified to a randomized optimization technique, where the original variance minimization was changed to cross entropy minimization problem [11]. The CE method was successfully applied to various optimization problems such as traveling salesman, assignment problem etc.[11]. In the present study, CE method is presented for optimization of the water distribution networks. II. Cross Entropy Method The cross entropy method is an iterative technique based on the concept of rare events, which involves two main stages: (i) generation of random sample of initial population (i.e., solution vectors) with a set of parameters, and (ii) updating this set of parameters which control the generation of random data using the sample itself, with the aim of improving the solution in the next iteration. The method derives its name from the cross entropy or Kullback-Leibler distance- a well MIR Labs, USA Shibu and Reddy 309 known measure of ‘information’, which has been successfully employed in various fields of engineering [11]. A. Entropy and Cross Entropy Entropy can be termed as a measure of uncertainty associated with a process (measure of expected information gain from a random variable) [12]. The probability distribution of events if known provides a certain amount of information. Shannon defined a quantitative measure of the distribution in terms of entropy, called Shannon entropy given by (1). n H ( X ) = − K ∑ p r ln p r (1) r =1 where H(X) represents the Shannon entropy corresponding to the random variable X, K is a constant, and pr represents the discrete probability corresponding to the variable at xr. The uncertainty can be quantified with entropy taking into account all different kinds of available information. Thus entropy is a measure of uncertainty represented by the probability distribution and is a measure of the lack of information about a system. If complete information is available, entropy is equal to zero, otherwise it is greater than zero. Cross entropy is a distance measure from one probability distribution to another. One of the well known definitions of Cross entropy is the Kullback–Leibler distance measure [13], serving to assess the similarity between two probability distributions: the assumed distribution q(x) and the actual distribution p(x). Cross entropy [D(P,Q)] is formulated as in (2). n pr (2) qr r =1 The interpretation of (2) is that in order to estimate a probability distribution, the cross entropy should be minimized. The goal is to find a distribution p(x) for which the Kullback – Leibler distance between p(x)* and q(x)* is minimal. D ( P, Q) = ∑ pr ln B. Principle of Minimum Cross Entropy According to Laplace’s principle of insufficient reason, all outcomes of an experiment should be considered equally likely unless there is information to the contrary [13]. Suppose a probability distribution for a random variable X= X = {x1, x2, x3,…,xn} is assumed as Q = {q1, q2, q3,…,qn} based on intuition. This constitutes the prior information in terms of a prior distribution. While estimating the actual distribution P ={p1, p2, p3,…,pn} of random variable X, using all the given information and make the distribution as near as possible to the assumed distribution. Thus, according to the principle of minimum cross entropy (POMCE), the cross entropy (CE), D(P,Q) is minimized as in (3). n p (3) M inim ize D ( P , Q ) = ∑ p r ln r q r =1 r This is referred to as the principle of minimum cross entropy, which minimizes the Bayesian entropy [13]. Here minimizing D(P,Q) is equivalent to maximizing the Shannon entropy. C. Cross Entropy Algorithm The main steps involved in the cross entropy algorithm for solving combinatorial optimization problem is given below. 1. Conversion of the combinatorial optimization problem to a stochastic node network (SNN) problem. 2. Set the trial counter t = 0 and assume equal probabilities for all the options as p0,r , where ‘r’ takes values from 1 to m. The number of stochastic nodes, m=np*nd, where np is the number of variables and nd is the number of available options. 3. Generate Nc sample vectors Xv(x1, x2,…,xm) for v = 1 to Nc using the probability pt,r (i.e., generate a set of Nc possible vectors each of size m, and having zeros and ones, where one corresponds to choosing a specific node, and zero otherwise). The value of Nc is taken as Nc =β*nd, where β is an integer value. The m dimensional vector Xv (x1, x2,…,xm) has the discrete probability of P =(p1, p2,…,pm). 4. Find out the performance function S(Xv) and check for constraints corresponding to each of the random vectors Xv, generated. 5. Now arrange the random vectors Xv, in the ascending order(if the problem is a minimization problem) or descending order (if it is a maximization problem) of their performance function S(Xv) values. Now the top most vectors will be having the best performance value and it is denoted as γt . 6. Choose a set (say ρc) of the top best performing vectors for updating the probability vector pt,r to the probability vector pt+1,r. Here ρc corresponds to percentage of the vectors selected and its value varies between 10% and 20% but may change as a function of the sample size N. The rth component of pt+1,r is obtained as given by (4). pt +1,r = Bt ,r TBt (4) where pt+1,r is the probability of success in the (t+1)th iteration of node r, Bt,r is the total number of times node r was chosen (frequency) out of the best top performance vectors (i.e., TBt the total number of vectors in the elite set) at iteration t. In order to avoid early convergence (stopping criteria of probabilities of potential options approaching ZERO or ONE) to a local optimum solution, a smoothing parameter ( α c ) is used. The probability is modified as given by (5) pt +1,r ← α c pt +1,r + (1 − α c ) pt ,r (5) Using the above probability-updating scheme, the probability of choosing a node at each subsequent iteration increases as the frequency of occurrence of the node in the elite set increases. Updating the entire probability components using (4) in conjunction with the smoothing formula (5) yields the new probability vector pt+1,r. The main reason why such a smoothing updating procedure performs better is that it prevents the incidents of zeroes and ones in the reference vector, as in case such values are obtained they will remain permanently, which is obviously not required. 7. Check stopping conditions: If γt for subsequent iterations remains unchanged and if pt converges to the degenerated case (i.e. all the probabilities pt,r are close to zero and one) then stop. Declare the last γt as the optimal solution γ* and its associated vector X as the design vector X*, otherwise pt,r pt+1,r and return to step 3. Cross Entropy Optimization for Optimal Design of Water Distribution Networks III. Model Formulation The optimization problem is to determine the values of pipe diameters that would minimize the cost of the system without violating any of the constraints. Thus it is required to select one diameter for each pipe from the list of commercially available diameters. The optimization problem can be expressed as, np Minimize Cost = ∑ C ( d i ) * li (6) i =1 subject to, H j ≥ H min , ∀j j q −q in j out j − q j = 0 , ∀j npu L  np L   HLi − h p  = 0 , L = 1,2,3,......., nL   p =1  i =1 L ∑ where, ∑ q inj = 310 a single reservoir which is located at elevation of 100 m. The ground elevation for all nodes is 0. All pipes in the network are of different lengths and the length of pipes is given in Table 1. Data relevant to nodes is given in Table 2. The system constraint on minimum pressure head requirement for all nodes is defined as 30 m. No velocity constraint is taken into account for this network. There are 6 commercially available pipe diameters (nd=6) and unit cost of the pipes used in the case study I are given in Table 3. The study on Hanoi WDN was first carried out by [14]. Thereafter so many researchers [4], [7], [15] and [16] applied various techniques to find optimal solution to Hanoi WDN. The solution search space for the Hanoi WDN is 634. (7) (8) (9) n in ∑Q i (10) ∑Q (11) δ li Qi1.852 (12) i =1 q out j = n out i i =1 HLi = 1.852 4.87 CHW di where C(di) corresponds to the cost per unit length of the pipe having diameter di and li is the length of the ith pipe , Hj and H min are the available and minimum pressure heads at the jth j node; nd =number of demand nodes; q inj = flow entering the jth node; Figure 1. Layout of Hanoi WDN q out =flow leaving from the jth node ; qj= demand at the j jth node; HLi =head loss in ith pipe; npL=number of pipes in a Table 1. Pipe length data for Hanoi WDN loop; h p =head raised by the pump p, npuL=number of pumps in a loop; nL=number of loops in the WDN. nin =number of incoming pipes to the jth node; nout =number of outgoing pipes from the jth node; and Qi = discharge or flow through the ith pipe, δ =constant depending on the units of head loss, length, diameter, and discharge; and CHW=Hazen William’s roughness coefficient. IV. Application of the Model A. Case Study I: Hanoi WDN The Hanoi water distribution network problem [14] as shown in Figure 1, is an extensively studied WDN by many researchers using a variety of optimization methods (such as genetic algorithms, ant colony optimization, simulated annealing etc.) is taken-up as case study I for testing the performance of CE method. This network is a real WDN constructed in Hanoi city at Vietnam, consists of 34 pipes and 32 nodes organized in three loops. The system is gravity fed by 1 Pipe Length (m) 100 18 Pipe Length (m) 800 2 1350 19 400 3 900 20 2200 4 1150 21 1500 5 1450 22 500 6 450 23 2650 7 850 24 1230 8 850 25 1300 9 800 26 850 10 950 27 300 11 1200 28 750 12 3500 29 1500 13 800 30 2000 14 500 31 1600 15 550 32 150 16 2730 33 860 17 1750 34 950 Pipe No. Pipe No. Shibu and Reddy 311 Table 2. Node demand data for Hanoi WDN Node No. Nodal Demand (m3/h) Node No. Nodal Demand (m3/h) 1 - 17 865 2 890 18 1345 3 850 19 60 4 130 20 1275 5 725 21 930 6 1005 22 485 7 1350 23 1045 8 550 24 820 9 525 25 170 10 525 26 900 11 500 27 370 12 560 28 290 13 940 29 360 14 615 30 360 15 280 31 105 16 310 32 805 Table 3. Commercially available pipe diameters and unit cost of pipes for Hanoi WDN Sl. Available Pipe Diameter Unit Cost of Pipe No. inch mm ($/m length) 1 12 304.8 45.73 2 16 406.4 70.40 3 20 508 98.38 4 24 609.6 129.30 5 30 762.0 180.75 6 40 1016.0 278.28 1) Model Run and Output for Case Study I At the start of the algorithm, it is assumed that all the options have equal probability of selection (i.e., P0,r = 1/6). The performance function used for solving the model is np nn i =1 j =1 S ( X v ) = ∑ C ( d i ) × li + ∑ PN × MAX (0, H min − H j) j (13) where S(Xv) is the performance function for the solution vector, and PN is the penalty function rate for violating the nodal pressure constraint. At the end of the final iteration, only 34 (i.e., total number of pipes) options will be having probability equal to one which forms the optimal solution set, and rest of them will be having a probability equal to zero. A program in MATLAB is developed, which is linked to EPANET toolkit for simulation of the WDN and to check the hydraulic feasibility. The stopping criteria is arrived in 31,500 function evaluations with smoothing parameter α = 0.35 and PN =100000000. The output of the model run for Hanoi WDN is given in Tables 4 & 5, and also compared with the past studies. Table 4. Nodal pressure corresponding to the optimal design by Cross Entropy method for Hanoi WDN Node No. Available Nodal Pressure (m) 1(R) 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 100 97.1407 61.6704 57.1713 51.5992 45.7571 44.4013 42.8160 41.5661 40.6585 39.0991 35.6707 31.4625 33.3626 30.5197 30.4795 Remarks Reservoir Avail. pressure is more than the min. pressure required Node No. Available Nodal Pressure (m) Remarks 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 32.9603 49.8247 55.0349 50.0175 40.6683 39.3963 43.4291 37.5807 33.7794 31.7037 30.9604 35.1562 30.7902 30.1112 30.6475 32.0296 Avail. pressure is more than the min. pressure required Table 5. Comparison of Cross Entropy model result of Hanoi WDN with past studies Pipe No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 Total cost ($) Pipe Diameter (inch) as per: Cross Entropy [15] Method 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 30 40 40 30 30 30 30 24 24 24 24 24 16 24 16 16 12 12 12 12 16 16 12 12 20 16 16 24 24 24 24 24 20 40 40 40 20 20 20 12 12 16 40 40 40 30 30 30 30 30 30 20 20 24 12 12 16 12 12 12 16 16 16 16 12 16 12 12 12 12 20 12 16 16 16 20 24 24 6.18×106 6.11×106 6.15×106 [4] Cross Entropy Optimization for Optimal Design of Water Distribution Networks On comparing the results of the Cross Entropy Model for Hanoi WDN problem [14] with results of earlier studies, it is found that the optimum diameters obtained from the present study is coming nearly same for all pipes in the network except for few pipes. Also the optimal cost obtained is closer to optimal costs of previous studies. Thus, the results obtained from present study shows that the CE method is effective and is well suited for optimal design of medium sized WDN like Hanoi WDN. B. Case Study II: Newyork City Tunnel WDN Newyork City Tunnel WDN [15] is taken-up as a case study II, for testing the performance of CE method. The layout of WDN is shown in Figure 2. The network consists of 20 nodes, 21 pipes and 1 loop, and is fed by gravity from a reservoir at a fixed head of 300 ft (91.44 m). The ground elevation for all nodes is 0. This system is in place and requires expansion. The pipe lengths, existing pipe diameters, and nodal demands are given in Table 6, and a Hazen-Williams constant of 100 is assumed for both the old tunnels and new pipes [15]. The system constraint is the minimum pressure head requirement for all nodes which is also given in Table 6. Fifteen commercially available pipe diameters and their unit cost are listed in Table 7. No velocity constraint is taken into account for this network. The objective is to determine whether a new pipe is to be laid parallel to an existing pipe or not, and if needed what will be the diameter of a parallel pipe, while the system is required to provide minimum hydraulic gradients. This network is firstly studied in [17] and thereafter studied by a number of other researchers ([4]; [5]; [6]). Due to pipe aging, the existing gravity flow tunnels are inadequate to meet the pressure requirements at nodes 16, 17, 18, 19, and 20 for the projected demands. Therefore new pipes can be added in parallel to the existing pipes to meet the minimum pressure head requirements. For this problem, 16 possible candidate diameters are available including 15 commercially available diameters and the ‘zero diameter-zero unit cost’ option. Considering all 21 pipes for possible duplication, it results in 1621 possible designs. 312 Table 6. Data for Newyork city tunnel WDN Nodal Demand (m3/h) Minimum Required Nodal Pressure (m) 1 -205665 91.44 2 9419.317 77.72 180 3 9419.317 77.72 2530.49 180 4 8991.166 77.72 2621.95 180 5 8991.166 77.72 6 5823.17 180 6 8991.166 77.72 7 2926.83 132 7 8991.166 77.72 8 3810.98 132 8 8991.166 77.72 9 2926.83 180 9 17329.91 77.72 10 3414.63 204 10 101.941 77.72 11 4420.73 204 11 17329.91 77.72 12 3719.51 204 12 11937.25 77.72 13 7347.56 204 13 11937.25 77.72 14 6432.93 204 14 9419.317 77.72 15 4725.61 204 15 9419.317 77.72 16 8048.78 72 16 17329.91 79.25 Pipe No. Pipe Length (m) Existing Pipe Diameter (inch) Node No. 1 3536.59 180 2 6036.59 180 3 2225.61 4 5 17 9512.2 72 17 5861.588 83.15 18 7317.07 60 18 11937.25 77.72 19 4390.24 60 19 11937.25 77.72 20 11707.32 60 20 17329.91 77.72 21 8048.78 72 Table 7. Commercially available pipe diameters and unit cost of pipe for Newyork city tunnel WDN Sl.No. Pipe Diameter Unit Cost of Pipe 1 2 3 4 5 6 7 8 9 (inch) 36 48 60 72 84 96 108 120 132 (mm) 914.4 1219.2 1524 1828.8 2133.6 2438.4 2743.2 3048 3352.8 ($/foot) 93.5 134 176 221 267 316 365 417 469 ($/metre) 306.7 439.6 577.4 725 875.9 1036.7 1197.5 1368.1 1538.7 10 144 3657.6 522 1712.6 11 12 13 14 15 156 168 180 192 204 3962.4 4267.2 4572 4876.8 5181.6 577 632 689 746 804 1893 2073.4 2260.5 2447.5 2637.7 1) Model Run and Output for Case Study II Figure 2. Layout of Newyork City Tunnel WDN The 21 existing pipes are considered as such and 21 parallel pipes for all the 21 pipes with 16 candidate diameters. At the start of the iteration, it is assumed that all the potential alternatives have equal probability of selection (i.e.,P0,r=1/16), since there are 16 candidate diameters including ‘zero diameter- zero unit cost’ option. While using Shibu and Reddy 313 EPANET, to avoid problems with consideration of zero diameter pipes, negligibly small diameter (i.e., 0.0001 mm) with zero unit cost is considered. The performance function used for solving the model is given by (13). As the iteration begins, some of the candidate diameters becomes superior to the others based on the performance values and their probability increases while for others gets reduced. This step by step iterative procedure for the modification to the probability of candidate diameters will continue until they reach the stopping criteria of approximately ones and zeros in the final iteration. At the end of the final iteration, only 21(i.e., total number of pipes) candidate diameters will be having probability equal to one which forms the optimal solution set, and all the rest will be having a probability equal to zero. The stopping criteria is arrived in 36,000 function evaluations with smoothing parameter α = 0.35 and PN =10000000. The output of the model run for Newyork City Tunnel WDN is given in Table 8 and also compared with the past studies, which is given in Table 9. Table 8. Cross Entropy Model output for Newyork city tunnel WDN Pipe No. Pipe Length (m) Existing Pipe Diamete r (inch) Parallel Pipe Diameter (inch) Node No. Available Nodal Pressure (m) Minimum nodal Pressure required (m) 1 3536.59 180 0 1 91.44 91.44 2 6036.59 180 0 2 89.6743 77.72 Table 9. Comparison of optimal outputs obtained by various approaches for Newyork city tunnel WDN Diameter of parallel pipe (inch) Pipe No. Existing Pipe Diameter (inch) Improved GA [5] Messy GA [6] 1 180 0 0 Cross Entropy Method 0 2 180 0 0 0 3 180 0 0 0 4 180 0 0 0 5 180 0 0 0 6 180 0 0 0 7 132 0 144 144 8 132 0 0 0 9 180 0 0 0 10 204 0 0 0 11 204 0 0 0 12 204 0 0 0 13 204 0 0 0 14 204 0 0 0 15 204 120 0 0 16 72 84 96 96 17 72 96 96 96 18 60 84 84 84 19 60 72 72 72 20 60 0 0 0 21 72 72 72 72 Cost (in million $) 38.8 38.64 38.64 3 2225.61 180 0 3 87.2179 77.72 4 2530.49 180 0 4 86.4983 77.72 5 2621.95 180 0 5 85.861 77.72 6 5823.17 180 0 6 85.3664 77.72 No. of function evaluations 96,750 37,186 36,000 7 2926.83 132 144 7 84.5863 77.72 Feasibility Feasible Feasible Feasible 8 3810.98 132 0 8 84.328 77.72 9 2926.83 180 0 9 83.4469 77.72 10 3414.63 204 0 10 83.4373 77.72 11 4420.73 204 0 11 83.4745 77.72 12 3719.51 204 0 12 83.8627 77.72 13 7347.56 204 0 13 84.7651 77.72 14 6432.93 204 0 14 87.04 77.72 15 4725.61 204 0 15 89.4058 77.72 16 8048.78 72 96 16 79.2747 79.25 17 9512.2 72 96 17 83.1702 83.15 18 7317.07 60 84 18 79.6084 77.72 19 4390.24 60 72 19 77.7403 77.72 20 11707.3 60 0 20 79.4684 77.72 21 8048.78 72 72 The optimal solution obtained in the present study by using the Cross entropy method is satisfying the minimum pressure head requirement at all the nodes, and is resulting in minimum cost. On comparing the results of the CE method for Newyork City Tunnel WDN with the results of past studies on the same WDN, it is found that the number of parallel pipes to be added is nearly same with only slight difference in one of the parallel pipe diameters with approximately same cost for providing parallel pipes. The number of function evaluations taken for producing the optimum solution is less than the other approaches. Thus, the results obtained from present study shows that the cross entropy method is effective and is well suited for the optimal design of large network like Newyork City Tunnel WDN, which involves capacity expansion in terms of adding parallel pipes without disturbing the existing pipes. C. Case Study III: Bengali Camp Zone WDN The Bengali Camp Zone WDN of Chandrapur city in Maharashtra State, India is taken as a case study III. This is a real WDN of Chandrapur water supply system, and whose network details are shown in Figure 3. This WDN is built to serve a new residential area in the city. The network was designed as an extension to the original WDN of Chandrapur city. The projected population for the year 2040 of Bengali Camp zone and the peak factor adopted for the design of WDN are 47126 and 3 respectively. The Cross Entropy Optimization for Optimal Design of Water Distribution Networks Bengali Camp zone WDN consists of 34 nodes, 38 pipes, and is fed by gravity from a tank at a fixed head of 206 m. The existing pipe diameters, and nodal pressure for the Bengali Camp zone WDN are given in Table 10. The pipe details and lengths are given in Table 11; nodal elevations and nodal demands are given in Table 12. A Hazen-Williams constant of 140 is assumed for all the pipes. The system constraint is minimum pressure requirement for all nodes is 11 m. Twelve commercially available pipe diameters and their unit cost are listed in Table 13. No velocity constraint is taken into account for this network. The solution search space of Bengali Camp zone WDN is 1238. 314 Table 11. Pipe details for Bengali camp zone WDN Pipe No. Start Node End Node 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 107 1 3 4 5 6 7 7 4 17 18 19 22 23 24 25 26 27 28 1 3 4 5 6 7 8 9 17 18 19 22 23 24 25 26 27 28 29 Pipe Length (m) 61 413 83 165 715 193 413 220 72 77 165 660 330 715 330 248 468 138 715 Pipe No. Start Node End Node 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 29 24 28 6 11 13 33 13 6 10 22 15 18 20 2 16 17 5 6 30 31 32 11 13 33 14 14 10 15 12 12 20 2 21 5 16 6 7 Pipe Length (m) 385 275 165 138 248 303 193 330 330 165 770 248 220 275 220 83 165 715 193 Table 12. Node details for Bengali camp zone WDN Node No. Figure 3. Layout of Bengali Camp Zone WDN Table 10. Pipe diameters, and nodal pressure as per existing design for Bengali camp zone WDN Pipe No. Existing Pipe Diameter (mm) Pipe No. Existing Pipe Diameter (mm) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 500 500 500 300 100 100 150 150 450 450 450 450 450 450 400 400 400 350 300 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 250 200 150 200 200 100 100 100 150 150 150 100 150 150 100 100 150 300 300 Cost of the WDN as per Existing Design (` ) 25735031 Node No. Available Nodal Pressure (m) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 10.98 11.66 17.33 18 18.26 14.63 16.43 20.12 16.13 15.03 14.8 16.64 14.17 16.25 11.51 18.27 17.97 Node No. Available Nodal Pressure (m) 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 107 18.75 18.89 19.29 19.64 20.69 19.41 22.48 24.22 27.18 25.11 24.88 24.46 21.94 21.48 25.06 15.95 18.75 - 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 107(Resvr) Nodal Elevation (m) 195 194 188.5 187.8 187.5 191 189.2 185.5 189.5 190.5 190.8 188.8 191.4 189.2 194 187.5 187.8 187 186.8 186.4 186 184.8 186 182.8 181 178 180 180.2 180.5 183 183.8 180 189.5 195 (LPS) 0 1.27 0 0.607 2.469 3.517 2.063 1.031 0.55 1.27 0.989 2.61 2.259 1.34 1.059 0.211 0.382 1.186 2.116 1.27 0.564 4.515 8.903 10.055 1.91 2.364 2.939 10.309 11.145 3.901 0.91 1.673 1.27 - Base Demand (m3/h) 0 4.572 0 2.1852 8.8884 12.6612 7.4268 3.7116 1.98 4.572 3.5604 9.396 8.1324 4.824 3.8124 0.7596 1.3752 4.2696 7.6176 4.572 2.0304 16.254 32.0508 36.198 6.876 8.5104 10.5804 37.1124 40.122 14.0436 3.276 6.0228 4.572 - Shibu and Reddy 315 Table 13. Commercially available pipe diameters and unit cost of pipe for Bengali camp zone WDN inch mm 1 4 100 Unit Cost of Pipe (`/m length) 860 2 3 4 5 6 7 8 9 10 11 12 6 8 10 12 14 16 18 20 24 28 32 150 200 250 300 350 400 450 500 600 700 800 1077 1374 1840 2333 2885 3442 4142 4826 6375 8141 10161 Available Pipe Diameter Sl. No. 1) Model Run and Output for Case Study III At the start of the run, it is assumed that all the candidate diameters have equal probability of selection (i.e., P0,r=1/12). The performance function used for solving the model is given by (13). As the iteration progresses, some of the candidate diameters become superior to the others based on the performance values and their probability increases, while for others the probability gets reduced. This step-by-step iterative procedure for updating the probability of selecting a candidate diameter for each pipe will continue until they satisfy the stopping criteria. At the end, the probability of selecting a option for a pipe will be approximately equal to ones and zeros. This means that only 38 decisions (i.e., total number of pipes) will be having probability equal to one which forms the optimal solution set, and the remaining will be having a probability equal to zero. The stopping criteria is arrived in 38,400 objective function evaluations with smoothing parameter α = 0.35 and PN =108. The output of the model run for Bengali Camp Zone WDN is given in Table 14. Table 14. Cross Entropy Model Output For Bengali Camp Zone WDN Pipe No. Optimum Pipe Diameter (mm) Pipe No. 1 600 20 2 600 21 3 600 22 4 300 23 5 100 24 6 150 25 7 100 26 8 100 27 9 500 28 10 500 29 11 500 30 12 450 31 13 450 32 14 400 33 15 400 34 16 350 35 17 400 36 18 350 37 19 300 38 Optimum Cost (`) Optimum Pipe Diameter (mm) 200 100 100 150 200 100 100 100 150 150 100 100 150 150 100 100 100 300 150 25235630 Node No. Available Nodal pressure (m) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 10.99 11.8 17.43 18.19 18.38 14.74 16.52 20.12 16.21 15.11 14.82 16.65 14.19 16.27 11.58 18.38 18.10 Node No. Availabl e Nodal pressure (m) 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 107 18.89 19.05 19.43 19.78 20.86 19.57 22.55 24.29 27.20 25.13 24.90 24.48 21.94 21.49 25.00 15.97 18.89 - On comparing the results of the CE method for Bengali Camp zone WDN with the existing design, it is noticed that the optimal solutions of CE is better than existing design, resulting in 1.94% lesser cost than the existing design. The solution is obtained in 38,400 function evaluations. Also the minimum nodal pressure requirements are well satisfied. The results of present study amply demonstrate that the CE method is an effective optimization method for WDN and has capability to handle larger number of discrete decision variables and various constraints. Thus, CE method is well suited for optimal design of larger water supply networks. V. Conclusions This study presented Cross Entropy (CE) method for solving water distribution network optimization problems. For hydraulic simulation of WDNs, EPANET tool kit is adopted and carried out simulation-optimization modeling for design of WDNs. Initially, the CE method is applied for two benchmark WDN design problems, namely Hanoi WDN and Newyork city tunnel WDN. To evaluate the performance of CE optimization method, the results are compared with the past studies and it is found that the CE method is giving good quality optimal solutions in a few number of objective function evaluations. The results also demonstrated that the CE method can be used effectively for optimal design of new WDN as well as for rehabilitation of existing WDN (i.e., for capacity expansion of WDNs, in terms of adding parallel pipes without disturbing the existing pipes). It is also found that the CE method is capable of handling larger number of discrete decision variables and different types of constraints. After successful validation to standard WDNs, the CE method is applied to a real WDN in India and the results are compared with the existing solutions. It is found that CE method is giving minimum cost solutions (i.e., good quality optimal solutions) in quicker time (i.e., rapid convergence to optimum). Thus, the study concludes that the cross entropy optimization method is an effective optimization method for solving WDN problems, and which can be applied for optimal design of any practical WDN problems. References [1] Gupta, I., Gupta, A., and Khanna, P., 1999, “Genetic algorithm for optimization of water distribution systems,” Environmental Modelling & software, 24(4), pp.437-446. [2] Goldberg, D.E., and Kuo, C.H., 1987, “Genetic algorithms in pipeline optimization,” Journal of Computing in Civil Engineering, 1(2), pp. 129-141. [3] Simpson, A. R., Dandy, G. C., and Murphy, L. J.,1994, “Genetic algorithms compared to other techniques for pipe optimization,” J. Water Resour. Plang. and Mgmt., ASCE, 120(4), pp. 423-443. [4] Savic, D.A. and Walters, G.A., 1997, “Genetic algorithms for least-cost design of water distribution networks,” J. Water Resour. Plang and Mgmt., ASCE, 123(2), pp. 67-77. Cross Entropy Optimization for Optimal Design of Water Distribution Networks [5] Dandy, G.C., Simpson A.R., and Murphy L.J., 1996, “An improved genetic algorithm for pipe network optimization,” Water Resour. Res., 32(2), pp. 449 - 458. [6] Wu, Z. Y., and Simpson, A. R., 2001, “Competent genetic-evolutionary optimization of water distribution systems,” Journal of Computing in Civil Engineering, 15(2), pp. 89-101. [7] Cunha, M.D.C., and Sousa, J., 1999, “Water distribution network design optimization: simulated annealing approach,” J. Water Resour. Plang. and Mgmt., 125(4), pp. 215-221. [8] Eusuff, M. M., and Lansey, K.E., 2003, “Optimization of water distribution network design using the shuffled frog leaping algorithm,” J. Water Resour. Plang. and Mgmt., ASCE, 129(3), pp. 210-225. [9] Maier, H. R., Simpson, A.R., Zecchin, A. C., Foong, W.K.,Phang, K.Y., Seah, H.Y., and Tan, C.L.,2003, “Ant Colony Optimization for Design of Water Distribution Systems,” J. Water Resour. Plang. and Mgmt., ASCE, 129(3), pp. 200-209. [10] Shibu, A. Reddy, M.J., 2011, “Least cost design of water distribution network by Cross entropy optimization,” World Congress on Information and Communication Technologies (WICT), vol.,no., pp..302-306,11-14Dec.2011 doi: 10.1109/WICT.2011.6141262 [11] Rubinstein, R.Y., 1997, “Optimization of computer simulation models with rare events,” European Journal of Operations Research, 99, pp. 89-112. [12] Shannon, C.E., 1948, “A Mathematical theory of communication,” Bell System Tech. Journal, 27, pp. 379 423. [13] Kullback, S, and Leibler, R.A., 1951, “On information and sufficiency,” Ann. Math. Statics, 22, pp. 79-86. [14] Fujiwara, O., and Khang, D.B., 1990, “A two-phase decomposition method for optimal design of looped water distribution networks,” Water Resour. Res., 26(4), pp. 539-549. [15] Dijk, M. V., Vuuren, S. V., and Van, Z., 2006, “Optimizing water distribution systems using a weighted penalty in a genetic algorithm,” ISSN, Water SA, 34(5), pp. 378 - 478. [16] Vairavamoorthy, K., and Ali, M., 2000, “Optimal design of water distribution systems using genetic algorithms,” Computer-Aided Civil and Infrastructure Engineering, Blackwell, 15(4), pp. 374–382. [17] Schaake, J. and Lai, D., 1969, “Linear programming and dynamic programming applications to water distribution network design,”, Rep. No. 116, Dept. of Civil Engineering, Massachusetts Institute of Technology, Cambridge, Mass. Author Biographies 1 Shibu A. is Research Scholar in the Department of Civil Engineering, Indian Institute of Technology Bombay, India. He completed his M.Tech.(Hydraulics Engineering) in 2000 from University of Kerala, India. His field of research is water distribution system modeling under uncertainty by using evolutionary techniques. 2 M. Janga Reddy is Assistant Professor in the Department of Civil Engineering, Indian Institute of Technology, Bombay, India. He 316 completed his Ph.D. in 2006 from Indian Institute of Science, Bangalore, India. His research interests include water resource systems: development of simulation and optimization models; optimal operations of single and multi reservoir systems; irrigation; hydropower; flood control; water distribution systems; multi-criterion decision making.