Operational response to contamination in water distribution systems: a multi-objective Bayesian optimization approach

Contamination of treated drinking water is a critical public health and safety concern. In this study, a multi-objective Bayesian optimization (MOBO) framework is proposed to optimize operational response to contamination in drinking water distribution systems (WDSs). The optimization framework aims to balance the conflicting objectives of minimizing response time while maximizing water quality metrics after contamination events. This was achieved by simultaneously optimizing two objective functions: the number of field operations (i.e., valve-closings and hydrant-openings), and the total contaminant mass consumed. The framework integrates a WDS simulation model, EPANET, within the proposed framework to simulate the implementation of response actions to various contamination events. Simulation results are then propagated into MOBO to generate Pareto-optimal solutions of the objective functions. A sensitivity analysis was conducted to tune the hyperparameters of the MOBO algorithm, including the covariance kernel of the surrogate model. Two case study WDSs with varying sizes and topological complexities were used to evaluate the performance of the proposed MOBO framework. Additionally, the performance of the MOBO algorithm was compared to the commonly used NSGA-II algorithm. The results showed that the proposed MOBO framework can identify optimal response actions to rapidly and efficiently improve water quality in the wake of contamination events in WDSs.