Rapid surrogate modelling of water-hammer in water distribution networks using spatiotemporal graph convolutional networks

Safe drinking water is essential to public health and social stability. Despite advances in modern materials and design, water distribution systems (WDSs) remain vulnerable to failure. Pipe ruptures are particularly consequential because they generate pressure surges (water hammer) that propagate through pipelines, jeopardising upstream and downstream assets. Although theory-driven or physics-based transient models can capture these dynamics, they can be computationally intensive and require specialised expertise, thereby limiting their routine use. Data-driven approaches, such as graph convolutional networks (GCNs), leverage spatial and temporal correlations and have performed well in networked domains such as traffic, finance, and social networks. In this study, we applied a spatiotemporal GCN (STGCN) to simulate water hammer responses under multiple pipe rupture scenarios. We chose STGCN because it explicitly embeds network connectivity and temporal dynamics, enabling topology-aware propagation of rupture signals at significantly lower computational costs than physics-based simulators. Across tests, the STGCN predictions closely matched the observed pressure histories, including peak timing, while substantially reducing the computational cost compared with theory-driven simulations. These results indicate that STGCN are efficient, accurate surrogates for simulating pressure wave dynamics in WDSs under rupture conditions, enabling rapid what-if analyses for utility applications.