Water distribution systems (WDSs) are vulnerable to accidental or intentional contamination events due to their large size and complex configurations. In this research, to reduce the impact of pollution on the health of consumers, the optimization-simulation approach was used to determine the optimal locations of quality sensors in a part of Mashhad water distribution network. EPANET software as simulator and multi-objective genetic algorithm as an optimizer are employed to a) maximize the detection likelihood b) minimize expected detection time and c) minimize the mass of contamination consumed. To achieve more realistic answers, the contaminant injection node, injection time, mass rate and duration of injection are considered indefinitely. The results showed that by installing only one or two sensors, 21.3% and 41.6% of contamination events can be detected, respectively. Also, by installing only one sensor in the Mashhad water distribution network, the detection time and the amount of polluted water consumed can be reduced by 17.8 and 37 percent, respectively.
Introduction
Water distribution systems are essential infrastructures that deliver drinking water to consumers. Ensuring their safety is a critical public concern, as these systems are vulnerable to both accidental and intentional contamination.
When an event occurs, it can have a significant impact on society and the economy. For example, in 2016, the water distribution networks of Beijing and Hong Kong were contaminated. Unfortunately, the contamination was not identified until people consumed the contaminated water for a long time. Such incidents highlight the urgent need for rapid contamination detection, accurate source identification, and effective mitigation strategies to minimize adverse consequences.
Previous studies have investigated the optimal placement of contamination detection sensors in water distribution networks. Harif et al. (2023) addressed this problem using NSGA-III, determining optimal sensor locations based on four objective functions: (1) maximizing detection coverage, (2) minimizing detection time, (3) maximizing redundancy, and (4) minimizing the number of contaminated nodes. Shahra and Wu (2023) studied optimal sensor placement in two different distribution networks by randomly selecting contamination scenarios. For these scenarios, they identified optimal sensor locations using an evolutionary optimization algorithm considering two objectives: minimizing the volume of contaminated water consumed and minimizing the detection time of contamination events.
A review of previous studies on sensor placement optimization shows that, significant simplifications have often been applied to reduce computational complexity. These include limiting injection nodes, constraining the number of sensors, and considering only a restricted set of contamination scenarios. Moreover, many uncertainties associated with contamination events, such as the time, duration, and injection rate of pollutants, have been neglected.
To address these limitations, this study proposes a multi-objective optimization approach to determine optimal sensor locations in a part of Mashhad’s water distribution network, aiming to: (a) maximize detection likelihood, (b) minimize detection time, and (c) minimize the volume of contaminated water consumed.

Methodology
In this study, EPANET was employed to simulate the hydraulic behavior and water quality within the distribution network. For the optimization process, both single and multi-objective genetic algorithms were applied. Figures 1 and 2 show, respectively, the general flowchart of the NSGA-II algorithm and the study zone.
Construction of contamination matrix
Contamination scenarios are defined as follows:
a) Intrusion can occur at all nodes except dead ends.
b) Events may start anytime from 00:00 to 24:00.
c) Injection rates range from 1–30 g/min.
d) Injection durations vary between 30–180 minutes.
To define contamination scenarios, the following assumptions are made:
a) Contaminants are introduced every 30 minutes (Start Time).
b) Injection rates of 5, 10, 20, and 30 g/min are considered (Mass).
c) Durations of 40, 80, 120, and 160 minutes are used (Duration Time).
This results in 559,104 possible scenarios (728 × 4 × 4 × 48). Due to this large number, a representative subset of 1,000 scenarios is selected to form the contamination matrix for optimization. Figure 3 presents the flowchart of the steps involved in selecting the top 1000 contaminants for the contamination matrix.

Results and Discussion
Since the number of sensors is not predetermined, it is also considered as an objective function. The allowable range is set between 1 and 20 sensors. Initially, the number of sensors was investigated versus the maximum detection of pollutants. As shown in Figure 4, results showed that the detection likelihood reaches its maximum value of 100% when 18 sensors are deployed. Conversely, the minimum detection likelihood, observed when only a single sensor is used, is limited to 21.3%. Figure 5 illustrates the relationship between the number of sensors and the minimum detection time, showing that the shortest detection time (91 minutes) occurs with the maximum number of sensors (20), whereas with only one sensor, detection time rises to 1184.6 minutes.
The analysis of the number of sensors versus the volume of contaminated water consumed showed that the optimal solution is one that prevents the consumption of contaminated water with the minimum number of deployed sensors. As expected, achieving the lowest volume of contaminated water consumption requires the maximum number of sensors (20), reducing the consumed contaminated water to 24.9 million liters. In contrast, when only a single sensor is installed, the consumed volume reaches its maximum, totaling 237.2 million liters (Figure 6).
The optimal Pareto front was analyzed for two sets of conflicting objectives in sensor placement within water distribution networks. The first set considered maximizing detection likelihood (F1) versus minimizing the expected detection time (F2). As shown in Figure 7, to maximize detection probability, sensors should be located at the downstream nodes of the network. In contrast, minimizing detection time requires placing sensors near input nodes, which conflicts with maximizing detection likelihood. Consequently, sensor locations must be optimized by simultaneously considering both objectives. The second set of objectives involved maximizing detection likelihood (F1) versus minimizing the volume of contaminated water consumed by users (F3). According to figure 8, aggain, while detection probability favors sensor placement at the downstream nodes of the network, reducing contaminated water consumption necessitates locating sensors close to injection points. These trade-offs highlight the need for a multi-objective optimization approach to determine sensor locations that effectively balance detection efficiency and public health protection in water distribution systems.
Conclusion
This study employed a simulation-optimization framework using EPANET and genetic algorithms to determine optimal sensor locations in Mashhad’s drinking water network. Three objectives were considered: maximizing detection likelihood (F1), minimizing detection time (F2), and minimizing the volume of contaminated water consumed (F3).
Results showed that a single sensor could detect over 20% of contamination events, reducing detection time by 17.8% and contaminated water consumption by 37%. To maximize detection likelihood, sensors should be placed at downstream nodes, whereas minimizing detection time and contaminated water volume requires placement near input nodes. These conflicting objectives emphasize the need for multi-objective optimization for effective sensor placement.