Fast and accurate leakage detection in water distribution systems (WDNs) represent considerable challenges to water utilities that are tasked with minimizing revenue loss, ecological hazards, property damage, supply shortage, and customer dissatisfaction. Widespread methods for leakage detection not relying on smart meter data and achieving high detection accuracy mostly require some form of WDN model calibration. For instance, leakage detection supported by hydraulic models require calibration relying on experimental field data in which ideally no leakages, or other similar anomalies, are present. Moreover, and most importantly, some form of water demand calibration is needed to capture the difference between basic water demand, due to water usage in normal conditions, and additional flows due to leakages. In this work we present a purely data-driven approach to leakage detection and localization based on pressure measurements for which no such calibration is required. Our calibration-free model extends the authors’ earlier work developed in the Battle of the Leakage Detection and Isolation Methods (BattLeDIM), an international competition on leakage detection and localization, focusing on noisy scenarios where temporally highly irregular demands – i.e., demands not following the usual diurnal patterns – are present as is the case of one DMA in the BattLeDIM WDN dataset. The proposed model works by linearizing the temporal regularities of the considered system and isolating each irregular demand as well as each leakage flow into a single dimension in a high-dimensional vector space. Hence, each detected leakage is characterized by a location vector and a flow quantity. The evaluation against ground truth data shows that our method can promptly detect both pipe bursts and growing leaks and determine the location of the leaks with reasonable accuracy.
