From Leakage Detection to Intervention Planning: Machine Learning and Optimization Approaches for Decision Support in Water Distribution Systems

Water availability and security are critical global challenges: billions of people still lack safe drinking water, while climate change intensifies scarcity and disrupts the stability of freshwater resources. At the same time, high levels of non-revenue water and aging infrastructure affect the operational performance of utilities and compromise their financial sustainability. These combined pressures highlight the need for more efficient, resilient and data-driven strategies in the management of water distribution networks (WDNs). With expanding sensor coverage and the growing availability of operational data, machine learning (ML) and artificial intelligence (AI) provide promising tools to reveal system behavior, identify inefficiencies, and enhance decision-making. This thesis investigates how AI-driven methods can address two important tasks in the WDN domain: leakage detection/isolation and intervention planning. Due to complex dynamics, deep uncertainties, and limited data availability, detecting and isolating leakages is a challenging task. From an ML perspective, leakages can be modeled as concept drift, which refers to the phenomenon of observing the data-generating distribution change over time. Building on this perspective, the methodology is structured around three complementary components: drift detection, to identify when sensor behavior deviates from expected operating conditions; drift localization, to infer where a potential leakage originates in the network; and drift explanation, to clarify the variables and mechanisms driving the detected anomalies. Experiments are conducted using the L-Town benchmark network as a realistic testbed, and beyond confirming results already reported in the literature, the analysis also explores more realistic operational scenarios, including incipient leaks, limited sensor coverage, and multiple simultaneous leaks. Beyond detection, the thesis proposes an optimization-based framework to support utilities in planning maintenance interventions under practical resource constraints. The approach explicitly incorporates budget and time limitations, formulating the decision-making problem as a mixed-integer quadratic program (MIQP) that evaluates the optimal scheduling and prioritization of repairs. This prescriptive model enables utilities to assess the trade-offs between intervention timing, cost-efficiency, and expected improvements in network stability. To demonstrate its applicability, the framework is tested on a synthetic dataset that simulates realistic operational conditions, illustrating how the integration of data-driven diagnostics with constrained planning can enhance the effectiveness of leakage mitigation strategies. Overall, this thesis shows that treating leakage as concept drift, together with complementary machine-learning techniques for detection, localization, and explanation and optimization-based methods for planning interventions, can jointly contribute to more resilient, efficient, and sustainable management of modern water distribution networks.