Topology-informed survival modelling for predicting future pipe failures in water distribution networks

Pipe failure prediction models are increasingly used to support maintenance and renewal planning in water distribution networks (WDNs), yet most existing models treat pipes as independent assets. Hence, they do not sufficiently capture the influence of network structure on failure risk. Moreover, model evaluation is often based on randomly selected or retrospective data, which may not reflect utility decision-making where historical information is used to prioritize future interventions. Here, we present a topology-informed survival modelling framework for multi-year pipe failure risk prediction. Pipe failures are treated as recurrent events within a gap-time framework and modelled using random survival forests to estimate future failure risk. Network structure is represented through a pipe connectivity line graph, from which graph-based features are derived and combined with conventional pipe input features. The framework was evaluated using pipe asset and failure records from a Swedish water utility, and temporally consistent training and testing data were used to reflect real-world application. The survival outputs separated pipes into distinct risk groups and identified a small high-risk subset in which observed failures were concentrated. In addition, graph-based features provided information beyond conventional pipe attributes and improved the identification of high-risk pipes. The proposed framework provides a practical way to incorporate network topology into pipe failure prediction, supporting risk-based asset prioritization in WDNs.