Aging Water Distribution Networks: A Hybrid Spatial Decision Support and Machine Learning Framework for Leak Detection

Leak detection in aging water distribution networks (WDNs) is a complex engineering challenge influenced by nonlinear hydraulic performance, infrastructure deterioration, spatial heterogeneity, and limited confirmed failure data. Purely sensor-based and data-driven approaches often face scalability constraints and rely on simplified assumptions, limiting robustness under real operating conditions. This study proposes a hybrid spatial predictive framework that integrates pressure-driven hydraulic simulation, GIS-based Fuzzy Analytic Hierarchy Process (FAHP), and supervised machine learning to identify leakage-prone nodes in large-scale WDNs. A pressure-driven EPANET model incorporating emitter coefficients and pipe aging effects simulates realistic leakage under diurnal demand patterns. A four-rule screening process identifies 4,699 high-risk nodes from over 39,000 nodes, followed by sensitivity–correlation analysis to determine influential nodes for efficient sensor placement. Spatial and infrastructural characteristics are quantified using Fuzzy AHP through a weighted evaluation of six criteria, with pipe age identified as the dominant factor (0.382). In addition, GIS-based fuzzy AHP enables the development of a spatial leakage risk map, indicating that more than 30% of the network falls within high (20.98%) and very high (11.06%) risk categories, reflecting significant vulnerability concentrated in aging infrastructure. These features are used to train multiple classifiers, among which Extreme Gradient Boosting (XGBoost) achieves the best performance (accuracy = 0.961; ROC–AUC = 0.989). Application to a full-scale urban WDN in southern Ontario demonstrates that the framework improves leak detection reliability and supports scalable, data-driven infrastructure management.