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

Le résumé fourni par la source

This repository contains the official implementation of topology-aware neural networks for the detection, location, and quantification of abnormal consumptions in water distribution networks (WDNs). The methodology combines genetic algorithm-based optimization for pressure sensor placement with Graph Neural Network (GNN) metamodels that leverage the network topology through message passing. The repository includes two metamodel variants: (1) a Static Metamodel that relies on pressure measurements under constant nodal consumption assumptions, and (2) a Dynamic Metamodel that accounts for daily consumption variations, enabling real-world abnormal consumption detection. Both metamodels employ an Encoder-Processor-Decoder architecture based on GATv2 (Graph Attention Networks v2) with residual connections. The repository provides pre-trained models for different sensor configurations (5, 20, 45, and 75 sensors), complete datasets for model training and evaluation, Jupyter notebooks to reproduce all results presented in the paper, and the implementation of the GNN architecture. The methodology was validated on the VilaMoura Water Distribution Network (Portugal), demonstrating high accuracy in detecting and locating abnormal consumptions with spatial accuracy ranging from 70-100% within 200 meters for optimal sensor configurations. This work contributes to operational water network management by providing a data-driven, near real-time tool for abnormal consumption monitoring with minimal computational cost.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.