Machine-Learning–Based Risk Assessment Method for Leak Detection and Geolocation in a Water Distribution System

Abstract

Current leak detection practice in a water distribution system consists of monitoring the distributed volume in a district metering area (DMA) and the consumption measured with automated meter reading (AMR) at the building connections. The detection of the occurrence of a potential leak in a DMA is established through a systematic continuous comparison of the real-time distributed volume and the consumption for this DMA and/or, in the absence of AMR, the comparison of the monitored distributed volume and a reference curve based upon past monitoring records of the distributed volume under similar operational conditions. The purpose of this research was to develop, test, validate, and illustrate the application of the machine-learning–based risk assessment method for early detection of high likelihood leaks, their geolocation, and the detection accuracy assessment in the water distribution system of the SUNRISE demonstration site at the University of Lille, France. It illustrates that the proposed algorithm, integrated with a GIS-based spatial flow data analysis, efficiently supports early detection, likelihood severity assessment, and geolocation of leak sources.