Multi-objective design of water distribution systems based on the fuzzy reliability index

Many recent investigations have been conducted in many aspects of water resources systems such as reservoir operation (Fallah-Mehdipour et al. 2011, 2012a, 2013a; Aboutalebi et al. 2015; Garousi-Nejad et al. 2016a), hydrology (Orouji et al. 2013; Aboutalebi et al. 2016), project management (Bozorg-Haddad et al. 2010a; Fallah-Mehdipour et al. 2012b), cultivation rules (Bozorg-Haddad et al. 2009; Fallah-Mehdipour et al. 2013b), pumping scheduling (Bozorg-Haddad et al. 2011), hydraulic structures (Bozorg-Haddad et al. 2010b), water distribution networks (Solgi et al. 2015, 2016; Bozorg-Haddad et al. 2016), operation of aquifer systems (Bozorg-Haddad & Mariño 2011), site selection of infrastructures (Karimi-Hosseini et al. 2011), and algorithmic developments (Shokri et al. 2013; Garousi-Nejad et al. 2016b). However, only a few of these works dealt with the fuzzy multi-objective design of water distribution systems (WDSs).

The use of optimization methods to design a WDS began about 30 years ago. Before that time, WDS design was usually performed based on trial and error by experienced engineers and, as a result, because of the complexity of these problems and also their large decision space, those designs were far from optimal. Thus, development of design methods with the aid of optimization techniques was inevitable. Alperovits & Shamir (1977) applied linear programming (LP) for the optimal design of WDS. Subsequently, Quindry et al. (1981), Goulter et al. (1986), and Kessler & Shamir (1989) complemented and further developed the method introduced earlier by Alperovits & Shamir (1977). Fujiwara & Kang (1990) suggested a two-phase decomposition method based on a nonlinear programming (NLP) gradient procedure. LP and NLP consider pipe diameters as continuous parameters while they are discrete commercial diameters. Thus, a modification is needed on the final continuous solutions obtained by LP or NLP methods. Furthermore, optimal solutions found by these methods usually contain one or two pipe segments with different diameters between each pair of nodes and this split-pipe design must be altered into only one diameter (Savic & Walters 1997). These modifications may impact the final optimal solution and may introduce constraint violations (Cunha & Sousa 1999). Therefore, it is advantageous to develop methods that can overcome these problems or at least lessen them.

Savic & Walters (1997) suggested the application of evolutionary algorithms to overcome the aforementioned shortcomings in the least-cost design of WDSs. Evolutionary and meta-heuristic methods are efficient tools in solving nonlinear problems and do not involve linearization or calculation of partial derivatives. Moreover, because of the global sampling of these methods, the probability of reporting local optima as the final answer decreases significantly and also the impact of a starting solution on the final solution is negligible. Other investigators suggested the application of simulated annealing (Cunha & Sousa 1999), taboo search (Lippai et al. 1999), shuffled frog leaping algorithm (Eusuff & Lansey 2003; Seifollahi-Aghmiuni et al. 2011), ant colony optimization (Maier et al. 2003), harmony search (Geem 2006), particle swarm optimization (Suribabu & Neelakantan 2006), and honey-bee mating optimization (HBMO) (Soltanjalili et al. 2010; Ghajarnia et al. 2011), in the optimal design, the rehabilitation, and the calibration of WDSs (Bozorg-Haddad et al. 2008; Sabbaghpour et al. 2012). Those algorithms have been able to report more flexible and realistic designs for WDSs in comparison with traditional mathematical methods. Evolutionary algorithms permit the search for discrete commercial diameters easily in the optimization procedure. In addition, because of their global sampling for solutions these methods are more reliable than traditional methods in finding near-global optima of complex problems.

WDS design is beset by the trade-off between cost and reliability. Mays (1996) considered two types of failures in WDSs: mechanical (such as pipe breakage, pump failure, power outages, control valve failure, etc.) and hydraulic (such as changes in demand or in pressure head, aging of pipes, inadequate pipe sizing, insufficient pumping capacity, insufficient storage capability). Several investigators have tried to combine these two types of failures to define a measure of reliability (Todini 2000). However, there are no universally accepted definitions for risk and reliability for WDSs. Moreover, Walski et al. (1987) pointed out that a reliability measure should reflect the way in which water users are affected. There are various viewpoints on how to apply reliability in WDS design (Rowell & Barnes 1982; Kettler & Goulter 1983; Morgan & Goulter 1985; Goulter & Coals 1986; Goulter & Bouchart 1990; Ostfeld & Shamir 1996; Xu & Goulter 1999; Tabesh et al. 2004, 2009; Geem 2015).

The necessity of building cost-effective WDSs while ensuring consumers' satisfaction with supply during any operation period leads us to consider the design of a WDS as a multi-objective model with the following objectives: (1) minimization of the cost and (2) maximization of the efficiency factors. Walski (2001) lamented that in spite of numerous papers on optimal design of WDS in the past 15 years none has managed to take a leading role in the design of water distribution networks. The former author lamented the common emphasis on minimizing the cost of WDS without enough attention paid to the WDSs' efficiency, this being the main reason that these methods have not gained widespread acceptance. Focusing on the minimization of WDS costs neglects the efficiency factors, such as the reliability in delivering demand with a desired pressure. Hence, the optimal design of WDS must aim at cost effectiveness and achieving maximum system reliability. This explains the pre-eminence taken by the multi-objective design of WDS.

Several researchers implemented multi-objective optimization to the design of WDSs (i.e. Farmani et al. 2005; Fu & Kapelan 2011; Wang et al. 2014; Geem 2015). Todini (2000) stated that if the WDS design is performed based on the complete satisfaction of the hydraulic constraints it may not be capable of delivering water with sufficient pressure and discharge during failures. Therefore, the probability of delivering water with a desirable pressure increases if each node is provided with more power (energy per unit time) than required so as to have a sufficient surplus to be dissipated internally in the case of failures (Todini 2000). This surplus was considered as a reliability measure by Todini (2000) and is named resilience (I_r). The latter author considered the two objective functions of cost minimization and maximization of (I_r) for the multi-objective design of two water networks.

According to Prasad & Park (2004), by maximizing Todini's (2000) resilience index some surplus pressure can be provided in the network and this does not have any influence on the redundancy of the network loops. They indicated that by having a high I_r index in a tree-shaped network is not enough to guarantee a high level of reliability since any failure, and, particularly, a pipe failure, in a tree-shaped or cost-optimized network can have severe consequences in terms of reliability. Thus, resiliency alone is not enough for assessing the reliability of a network and a modification is needed to assess the reliability of the network loops. Prasad & Park (2004) modified the resilience index in a way to be capable of ensuring the network with reliable loops and named the improved index as network resilience (I_n). Neither the I_r index nor the I_n index consider the maximum allowable pressure in their calculations. In fact, as the I_r or I_n reliability indices increase, the nodal pressures of the network increase without any restriction, which may lead to uncontrollable leakage of the pipes in failure conditions.

This paper proposes a new deterministic reliability index based on fuzzy logic. In addition, the multi-objective MOHBMO algorithm is presented to perform multi-objective WDS design. The results are assessed based on the final Pareto front obtained after applying the algorithm on two well-known and benchmark networks. The effectiveness of the introduced reliability measure is also examined by assessing the final solutions from the viewpoint of nodal pressures and having reliable loops. Identification of the payoff between cost and reliability is also another aim of this study, which is obtained by applying the MOHBMO algorithm in complex, nonlinear, and discrete networks.

The importance of high consumption nodes is considered in a reverse order here since the FRI index is defined to be used in an optimization model. In an optimization model intending to maximize an index, solutions with values lower than the index are restrictive. As a result, the model improves solutions with lower values of the desired index. That is, the optimization model focuses on improving the condition of high-consumption nodes over those of other junctions. The sensitivity of high-consumption nodes is increased by decreasing their membership values, which is a desirable feature of WDSs. Thus, by defining the objective function to maximize the FRI index it is possible to improve the value of critical (high-consumption) nodes. This assures that the final optimal answer cannot have low FRI values in high consumption nodes. It is noteworthy that there are other considerations in the design of WDSs. For example, there may be low-demand nodes which are critical and strategically important, e.g. supplying hospitals and other essential facilities. Thus, one must consider site-specific requirements that are implemented in the FRI equations.

The aim of this study is to minimize the cost of WDSs and maximize the FRI. Letting Cost and Reliability denote the cost of network pipes and system's reliability, respectively, the following objective function takes into account the minimization of the network capital costs and the maximization of the FRI index.

MOHBMO is the multi-objective version of HBMO algorithm presented by Bozorg-Haddad et al. (2006). The superiority of the MOHBMO over the multi-objective genetic algorithm and multi-objective particle swarm optimization has been demonstrated in other fields of study (i.e. Niknam et al. 2011), although it has not been applied to design water distribution networks. MOHBMO is inspired by the natural life of honey-bees. Each bee hive can contain one or more queens. The queens exit the hive during the mating season and perform the mating flight by attracting drones. Successful drones which mate with the queen do not necessarily belong to the queen's hive. In other words, the queen might mate with drones belonging to other hives. In addition, two queens might not live in the same hive. Therefore, if a hive contains two queens, one of them leaves the hive and migrates to another bee hive which does not contain a queen. It is possible for the migratory queen to mate also with drones of a new hive, leading to the replacement of honey-bee genes between different bee hives. This mating ritual inspired the development of the multi-objective version of the HBMO algorithm.

In MOHBMO, at first two hives with two (or more) different queens are generated. Each hive performs a single objective optimization with one objective function and attempts to improve its queen after a predefined number of iterations. In a multi-objective problem, the final Pareto solutions found by the algorithm do not extend beyond the area defined by the solutions obtained by solving the single objective problem with each objective function. Thus, to obtain a proper multi-objective Pareto it is necessary for its two extreme points to be as near as possible to the solutions found by the single objective solutions of the problem. To ensure the aforementioned Pareto condition a warm up (WU) period is added to the MOHBMO algorithm (Fallah-Mehdipour et al. 2011). During the WU period, each hive attempts to improve the queen through the global optimum of its own objective function by performing a considerable number of iterations. This produces two different solutions; one having optimal conditions related to the first objective function and the other having optimal conditions related to the second objective function. These solutions are compared with each other. If one of them dominates the other, it is then added to the non-dominated list, otherwise both of them are added to the non-dominated list. The queens are then replaced in the hives. This means that the best solution for first objective function moves to the hive which improves the second objective function, and vice versa. Thereafter, each hive starts evolving once more through its own objective function using a new queen, while at this stage the number of iterations is much less than that in the WU period. After conducting a predefined number of iterations the final solutions are compared and moved to the non-dominated list. Each new member which is added to the non-dominated list is compared with older ones and at the end of the iterations only the non-dominated solutions remain in the list.

MOHBMO is applied to the design of two well-known benchmark problems for the design of WDSs. The following sections present the results of these applications.

Two extreme points in the Pareto of Figure 4 (points A and B) are obtained by a single objective design of the two-loop network, using each objective function separately (WU period). It is seen in Figure 4 that the MOHBMO algorithm was able to cover the area between these two boundary points.

The main aim of the multi-objective design of a WDS is to find a solution which is optimal with respect to the single objective functions. Hence, it is necessary to select one of the solutions on the Pareto set to report it as the final result of the multi-objective design. In this regard, there are different conflict resolution methods to achieve this selection. This study implements the method of Young (1993), in which a desirability or utility function is distributed for each objective function and one of them is chosen by maximizing a mathematical equation based on the gradient of different points of the Pareto (Young 1993). By using this method, point C (Figure 4) is chosen and differentiated by a triangle in Figure 4. The minimum cost solution, maximum FRI solution, and the selected point by Young's (1993) model have been named A, B, and C, respectively, in Figure 4. Their corresponding pipe diameters, pressure head, and FRI values are reported in Tables 1 and 2. Also, the values of both objective functions are shown in these tables. By comparing these values it is seen that the reliability of solution C is 40 times that of solution A, while its cost is only 2.5 times of the cost associated with solution A. A similar comparison on points C and B indicates that solution C achieves 80% of the reliability of solution B while its cost is only a third (0.33) of that of B.

Clearly, the best point on a Pareto set is the point whose objective functions' values are close to the optimal values of each single objective function. In other words, the best design for the two-loop network is the one whose cost is equal to $419,000 and its FRI reliability index is 1.2055. Obviously, such a network does not exist. Showing this point with a rectangle in Figure 4 helps evaluating the Pareto set. It can be seen in Figure 4 that the Pareto set obtained by the MOHBMO algorithm has a proper convexity towards the rectangle. This means that some of the non-dominant solutions have close proximity to this hypothetical optimal point, and even the point selected with the method of Young (1993) fall within that area too.

Point H was selected in this Pareto set with the method of Young (1993), and is shown by a triangle in Figure 8. Capital cost and FRI values for this solution are $7,244,479 and 0.17, respectively. Solution H is able to supply 80% of the reliability of solution G while its cost is only 0.7 of that. By a similar comparison between points H and F it is concluded that although the cost of solution H is 1.2 times higher than solution F, its reliability is 13 times higher. Furthermore, Table 3 shows the diameters of solution H, while the FRI index data for this solution is reported in Table 4. It is seen in Table 4 that Node 2 appears to have a high head. The model chooses pipes' diameters such that the pressure of the nodes immediately downstream the reservoir is relatively high in order to cope with the head losses throughout the gravity network.

This study presented a new deterministic fuzzy reliability index named FRI for the design of WDSs. Another goal of this study was the multi-objective design of WDS. The multi-objective version of the HBMO (MOHBMO) algorithm was developed and applied to the bi-objective design of two benchmark problems with one objective function for cost minimization and the other for reliability maximization. This paper's results demonstrated that the FRI index is able to bring the nodal pressures close to their optimal values (those defined as the average of minimum and maximum permissible values in this study), and is capable of producing redundant loops. The results indicated acceptable performance of the FRI index and MOHBMO algorithm that produced a set of optimal non-dominant solutions for decision makers.

This research was supported by the University of Tehran's Research Institute under Award No. 73148796/1/07. The authors would like to thank the University of Tehran's Research Institute for its support.