Topology‐Aware Neural Networks for Abnormal Consumption Detection and Location in Water Distribution Networks

Abstract This paper presents a topology‐aware neural network approach for the detection, location, and quantification of abnormal consumptions in water distribution networks. The approach includes two main steps: the optimization of pressure sensor locations to maximize measurement sensitivity and the development of metamodels based on near real‐time data. The metamodel is designed and trained to predict the consumptions at all nodes based on pressure measurements and users' consumption collected by smart meters. These nodal consumptions deduced from the actual measured consumption allow the location of potential abnormal uses in the network. The proposed methodology enables the development of two metamodels, each tailored to specific applications based on the training data. The Static Metamodel relies on pressure head measurements under the assumption of constant nodal consumption, whereas the Dynamic Metamodel accounts for daily consumption variations, enabling the detection and location of abnormal consumption in real‐world scenarios. Both metamodels can detect the location of abnormal consumptions with reasonable accuracy, although this accuracy strongly depends on the number and spatial distribution of sensors, as well as the magnitude and location of the abnormal consumption. As water utilities implement advanced metering systems, the application of the proposed approach becomes more viable, enabling more effective and faster abnormal consumption detection.