Multi-Leakage Detection Using Graph Attention Networks and Restoration Prioritization in Water Distribution Systems

Ryul Kim, Young Hwan Choi^*
Department of Civil and Infrastructure Engineering, Gyeongsang National University, Jinju, Republic of Korea
* Corresponding Author: Young Hwan Choi. Email: email
(This article belongs to the Special Issue: Explainable AI, Digital Twin, and Hybrid Deep Learning Approaches for Urban–Regional Hydrology, Water Quality, and Risk Modeling under Uncertainty)

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.077480

Received 10 December 2025; Accepted 13 February 2026; Published online 11 March 2026

Download PDF

Abstract

Leakage events occurring at multiple locations simultaneously generate overlapping and topology-dependent pressure signatures, making reliable detection and subsequent restoration planning a persistent challenge in water distribution systems (WDSs). While recent data-driven techniques have improved the ability to identify anomalous hydraulic behavior, most approaches remain limited to the detection stage and offer little guidance on how utilities should prioritize repairs once multiple failures are identified. To bridge this gap, this study proposes an integrated framework that links topology-aware leakage detection with quantitative restoration prioritization. First, a multi-task learning framework based on Graph Attention Networks (GAT) is employed to jointly detect both the location and magnitude of multiple leakages by explicitly incorporating hydraulic responses and network topology into the learning process. The model’s detection robustness is evaluated across networks with contrasting looped, branched, and hybrid topologies to examine how structural characteristics influence detection accuracy under multi-event conditions. Second, the study develops a restoration-planning module that constructs a two-objective decision space combining restoration cost and segment vulnerability, where the latter accounts for disruption potential arising from hydraulic importance and local service connectivity. Non-dominated sorting is used to derive Pareto-optimal restoration sequences, enabling explicit quantification of the trade-offs between operational cost and service disruption. This provides decision-makers with a ranked set of restoration orders that reflect both hydraulic impact and functional risk, rather than relying on heuristics or cost-only criteria. Notably, the proposed framework separates offline training from online inference, requiring only a single forward pass for real-time decision-making without the need for iterative hydraulic simulations. Results demonstrate that topology strongly governs both detection performance and the structure of optimal repair sequences, underscoring the importance of integrating network-aware learning with multi-criteria restoration evaluation.