Bayesian Optimization for Contamination Detection and Response Management in Water Distribution Systems

The contamination of treated drinking water poses a serious public health risk, especially through intrusion or injection into the water distribution systems (WDSs). The unpredictable nature of such contamination complicates detection and response efforts, making immediate action upon contamination detection crucial to minimize its impact on public health. This dissertation develops an emergency management framework for rapid and efficient response to WDS contamination events. The framework comprises three main tasks. The first task involves rapidly detecting and identifying contamination sources using a real-time source identification framework based on Bayesian optimization (BO). This framework quickly identifies contamination source characteristics such as location, injection rates, and patterns. The second task formulates a multi-objective Bayesian optimization (MOBO) framework to determine the best hydraulic actions needed to isolate contaminated zones and flush contaminants from the WDS, aiming to reduce contamination. The third task integrates booster chlorination into emergency responses to identify the optimal chlorine dosage required for disinfecting the WDS. Previous methods, reliant on evolutionary optimization techniques like genetic algorithms, faced high computational costs and scalability issues. This research addresses these limitations by employing BO, which is more efficient and scalable for real-time applications. BO uses a probabilistic surrogate model to estimate function outputs and uncertainties, balancing exploration and exploitation with an explicit acquisition function derived from model predictions. This dissertation conducts a comprehensive sensitivity analysis of BO configurations, comparing Gaussian Process (GP) and Random Forest (RF) surrogate models with acquisition functions like Expected Improvement (EI), Probability of Improvement (PI), and Upper Confidence Bound (UCB). Key findings emphasize the importance of hyperparameter selection for optimal performance. Comparative analysis shows the UCB acquisition function outperforms others in guiding search space convergence. The RF_UCB configuration is preferred in identifying contamination sources and balancing disinfection efficiency while enhancing computational efficiency. MOBO also demonstrates superior performance over multi-objective evolutionary algorithms, requiring fewer evaluations to achieve optimal solutions and enabling real-time response. This dissertation enhances strategies for detecting and responding to contamination in WDSs, advocating the use of BO to ensure safer and more reliable drinking water. It guides future policies for more secure and resilient water systems.