Link Prediction on Water Distribution Networks using Graph Neural Networks

<p dir="ltr">Water Distribution Networks (WDNs) are critical infrastructure for ensuring the reliable and efficient delivery of water to society. Structurally, the network of interconnected pipes of a WDN can be replicated digitally as a graph with the help of a Geographic Information System (GIS), with each pipe representing an edge in the graph. An accurate digital representation is crucial to maintain data integrity and ensure operational reliability, yet they are often incomplete. Historical data gaps and measurement inaccuracies during the digitization of physical infrastructure induce misrepresentations of the network, often manifesting in the form of structural discontinuities, diminishing the reliability of the digital model and impeding the process of deriving meaningful information. In this regard, we propose a Deep Learning (DL) based approach to formulate and tackle structural discontinuities in WDNs as a graph-based link (edge) prediction task using Graph Neural Networks (GNNs). Our approach adapts the SEAL (Subgraphs, Embeddings, and Attributes for Link prediction) methodology, where the GNN learns structural patterns from local enclosing subgraphs extracted around edges (links) of the WDN represented as a graph. We compare the ability of the GNN model to predict the likelihood of link existence while exploring homogeneous and heterogeneous-bipartite graph representations. We validate our framework on a real-world WDN from Denmark and a L-Town network. Experiments showed that the heterogeneous-bipartite representation outperformed the homogeneous graphs, achieving AUC scores of 83–95% vs 78–85%. Overall, the GNN-based approach outperforms traditional heuristics, which performed no better than random chance. This work presents a scalable, data-driven tool for water utilities to identify and remedy structural errors, enhancing data integrity and improving decision-making.</p><p dir="ltr">This paper was presented at the 21st Computing and Control in the Water Industry Conference (CCWI 2025) at the University of Sheffield (1st - 3rd September 2025).</p>