Topology-guided Graph Neural Networks for Leak Localization in Water Distribution Networks

Leak localization in Water Distribution Networks (WDNs) is a challenging task, particularly in the presence of multiple simultaneous leaks and complex network topologies. Graph Neural Networks (GNNs) have recently emerged as an effective tool by leveraging the network structure to model the spatial propagation of anomalies. However, existing approaches typically rely on pairwise node interactions and do not capture higher-order dependencies induced by the looped structure of real WDNs. In this paper, we propose Top-GGNN, a Topology-guided Graph Neural Networks that incorporates higher-order structural information through a cell-complex representation. Specifically, the method introduces a topological refinement layer based on the first-order Laplacian, which captures both interactions between edges sharing common nodes and dependencies induced by cycles in the network, enabling the model to exploit higher-order structural relationships in addition to standard message passing. Experimental results on simulated WDNs show that the proposed approach improves leak localization performance over standard graph-based models, with more significant gains in multi-leak scenarios and networks with richer structures.