A Comprehensive Framework for Leak Characterization in Water Distribution Systems

Municipal water distribution systems (WDSs) face a significant challenge in the form of pipe leaks. Water losses due to leaks may represent up to 35\% of the distributed drinking water volume. As part of management strategies, detecting and locating pipe leaks have been a constant challenge for water utilities and stakeholders due to the underground location of most pipes. Small undetected pipe leaks can turn into major breaks over time, posing an additional burden to a WDS operation. By leveraging machine learning methods, classical optimization methods, and hydraulic modeling, we have developed a comprehensive leak characterization framework that consists of the following three components: (a) Sensor placement, (b) leak detection, and (c) leak localization. For sensor placement (flow and pressure) given a limited budget, we use a classical optimization approach based on coverage and distance metrics. We use machine learning methods for leak detection (time of leak) and leak localization (location and size of a leak). The framework is evaluated on a hydraulic network using simulated data and extended to a medium size system from a community in California. Preliminary results show that, in general, the developed framework detects the presence of leaks successfully. Leak location results are mainly affected by the training data set size. Water utilities and researchers can use the proposed framework to improve the management of water systems with timely leak detection and accurate leak localization