Hydraulic Modelling and Machine Learning Solutions for Leak Monitoring in Municipal Water Distribution Networks

Urban water distribution networks (WDNs) are essential infrastructures that ensure the continuous supply of clean and safe water to communities. Due to population growth, climate variability, and aging infrastructure, improving operational efficiency and minimizing losses have become critical objectives for water utilities. Among the major challenges faced by modern WDNs, leakages represent a significant source of inefficiency; in Italy, networks lose approximately 42% of the treated water before it reaches consumers. Such losses impose substantial economic costs, reduce system resilience, and hinder the sustainable management of water resources. This thesis work involved the development of a hydraulic digital model and the evaluation of algorithms for leakage monitoring in two municipalities in the Province of Cuneo, Piedmont, Italy: Cavallermaggiore and Marene. The thesis was developed in collaboration with Alpiacque, the water utility manager, and the support of Tesisquare and Fondazione DIG421. The study builds upon real infrastructure documentation, mainly GIS WDNs data, SCADA measurements, operational knowledge and records of historical leakages. The thesis is divided into two complementary parts: In Part I, we developed hydraulic digital models of Cavallermaggiore and Marene using the open source software EPANET. The models integrate reconstructed tank geometries, pump curves, well output data, district-level time patterns derived from SCADA flow measurements, yearly consumer-level demand, and rule-based controls that replicate real operational logic. The models were calibrated to match flow and pressure trends over a representative time period and were used to support the definition of districts in Cavallermaggiore and the development of Part II. In Part II, we evaluated leakage monitoring methods based on machine learning techniques. First, we evaluated the use of a data-driven anomaly detection algorithm using real flow sensor data from the inlets of each district, to implement leakage detection. With this approach we are able to reach over 90% of recall of the historical leakages reported by the utility manager. Second, we used the validated hydraulic model to generate simulated leakage scenarios and pressure measurements, following the approaches described in the BattLeDIM competition. This data was then used to feed the LILA algorithm for leakage detection and localization. By integrating a hydraulic digital model with machine learning techniques, this work provides a practical and scalable framework for monitoring leakage events in medium-sized municipal water distribution networks. The collaboration with Alpiacque and the real-data-informed modelling of Cavallermaggiore and Marene demonstrate how hydraulic simulation, SCADA analytics, and machine learning solutions can support the transition toward smarter, more efficient, and more resilient water management.