Selecting the best location of water quality sensors in water distribution networks by considering the importance of nodes and contaminations using NSGA-III (case study: Zahedan water distribution network, Iran).

Abstract

One of the most effective ways to minimize polluted water consumption is to arrange quality sensors properly in the water distribution networks (WDNs). In this study, the NSGA-III algorithm is developed to improve the optimal locations of sensors by balancing four conflicting objectives: (1) detection likelihood, (2) expected detection time, (3) detection redundancy, and (4) the affected nodes before detection. The research procedure proposed the dynamic variations of chlorine between defined upper and lower bounds, which were determined utilizing the Monte Carlo simulation model. For selecting a contamination matrix with the same characteristics and effects of all possible events, a heuristic method was applied. The coefficients of importance are introduced in this study for the assessment of contamination events and network nodes. The Pareto fronts for each of the two sets of conflicting objectives were computed for benchmark and real water distribution networks using the proposed simulation-optimization approach. Results indicated that sensors should be installed downstream of the network to maximize sensor detection likelihood; however, this increases detection time. For the benchmark network, maximum and minimum detection likelihoods were calculated as 92.8% and 61.1%, respectively, which corresponded to the worst detection time of 11.58min and the best detection time of 5.06min. So, the position of sensors regarding the two objective functions conflicts with each other. Also, the sensitivity analysis related to the number of sensors illustrated that the Pareto fronts became a more efficient tool when the number of sensors increased. The best pollution detection likelihood in the real water network increased by 18.93% and 24.66% by incrementing the number of sensors from 5 to 10 and 5 to 15, respectively. Moreover, adding more than 10 sensors to the benchmark network and more than 15 to the real system will provide little additional detection likelihood.