395 - Predicting the Urban Stormwater Drainage Network State Using Graph Neural Networks

Urban stormwater drainage systems are essential for mitigating flood risks and ensuring the functionality of urban environments. Accurate prediction of drainage system state (flows and depths) is crucial for effective decision-making and resource allocation in flood management. While existing physics-based and data-driven models offer insights, they have limitations in predicting drainage network performance. Graph-WaveNet (GWN), a Graph Neural Network, has shown promise in spatial-temporal data modeling and is a potential candidate for predicting urban drainage flow, capturing the spatial dependencies of complex networks. This study evaluates the accuracy of GWN in predicting urban drainage network state (network inflows, flow rates in pipes, and relative water depths) in a sub-catchment in Kowloon, Hong Kong. Training and testing data were obtained by simulating the responses of the drainage network for 43 rainfall events between 2020 and 2023 using the Storm Water Management Model (SWMM). The GWN model's performance was evaluated using R2, NRMSE, and SMAPE metrics. Key findings highlight that the model achieves accurate predictions of network inflows, flow rates in pipes, and relative water depths up to 20, 20, and 30 minutes in advance, respectively, with outflow conditions impacting prediction accuracy. The network hierarchy plays a significant role in explaining local accuracy, where larger bottom pipes and junctions are better predicted, while end pipes and junctions are more prone to inaccuracies due to the absence of current rainfall information. In conclusion, this study demonstrates the potential of Graph-WaveNet for predicting urban stormwater drainage network state and provides valuable insights for developing more advanced models for urban drainage flow prediction and flood management.