Fast Localization of Multiple Leaks in Water Distribution Network Jointly Driven By Simulation and Machine Learning

The leakage control of the water supply network is of highly concerned in the water supply industry. With the aim of achieving an accurate identification and localization of the leakage in the network, a novel Multiple Leaks Detection and Isolation Framework(MLDIF) based on the existing pressure and flow measurements is proposed, combining statistical methods, signal processing and model-based fault diagnosis. First, a gradient iteration algorithm with variable steps is used to calibrate the model parameters of each zone. Second, water consumption of the each region is predicted by analyzing measurement data collected by the AMR instruments. Then, through the joint analysis of the model simulation result and the measured pressure/flow data, the changes of the network state are identified by the STL decomposition and K-means algorithm, and the start time and possible quantity of the leakages are thereby detected. Finally, the pipe with the highest probability of leakage can be locked by ranking the matching degrees between the actual leakage feature and the simulated leakage features. Combined with the framework proposed in this article, a hydraulic model of the system and a dataset of long-term measurements of SCADA, all the leakage events that occurred after 2018 were detected and isolated including several unfixed leaks. The comparative analysis of the recognition results and the provided leak report in 2018 proved the accuracy, efficiency and good applicability of this method in identifying and isolating leaks.