Anomaly detection in water distribution networks based on the analysis of differential pressure values

Water distribution networks (WDNs) are complex systems typically featuring thousands of pipes and valves (Marsili et al. 2023), the management of which may be a challenging task for water utilities. In this context, WDN hydraulic models – if properly set up in terms of network features and water-demand patterns – can be useful tools to support water utilities in leakage detection and the identification of anomalies in water consumption (Alvisi 2015, Marzola et al. 2022, Steffelbauer et al. 2022). However, a well calibrated hydraulic model is not always available to water utilities, and sometimes even basic topological or technical aspects (e.g. WDN connectivity, pipe features, etc.) are not perfectly known. Since the last three decades, the advent of WDN digitalization and smart metering has allowed obtaining a large amount of data (e.g. discharge in pipelines, pressure at nodes, water level in storage tanks, etc.) nearly in real-time, opening up new opportunities to water utilities and practitioners. Among these data, pressure measurements may be particularly easy to gather, due to the lower costs of sensors compared to flowmeters, and the possibility of installing them in several sections of the WDN, evenly distributed. Drawing from the work by Mazzoni et al. (2024), this study investigates the opportunity to exploit pressure data to detect anomalies in WDNs without detailed information about WDN characteristics, and in lack of the hydraulic model. In greater detail, a new methodology – exclusively based on the use of pressure data collected in a series of WDN sections – is presented, allowing the detection of both hydraulic anomalies (e.g. unexpected consumption due to unauthorized withdrawal) or mechanical anomalies (e.g. breaks, bursts, or gate valves unintentionally left partially or totally closed after maintenance actions) without requiring detailed information on WDN characteristics or the WDN hydraulic model.