Leak localization in an urban water distribution network using a LSTM deep neural network

Given that water distribution networks are complex systems exposed to factors that induce leaks, it is necessary to implement techniques that allow to locate water leakages as accurately as possible minimizing the required instrumentation. In this paper we propose a leak localization technique based on the use of a long short-term memory (LSTM) deep neural network for classification trained with all possible leak scenarios in the network. As a case study, a real-world district metered area (DMA) is selected. The DMA is first sectorized considering the topological proximity of the nodes. Then, a LSTM is trained with pressure and flow rate data from all the possible leak scenarios in the system obtained from a hydraulic simulator model of the network. To replicate realistic measurements, uncertainty in the demand pattern, nominal water consumption and in the sensor readings is considered. Classification results are presented both for the validation during the training of the LSTM and for measured data of a real induced leak in the system. Copyright (c) 2024 The Authors.