Predicting the urban stormwater drainage system state using the Graph-WaveNet

Graph Neural Networks (GNNs) have been applied to network data such as traffic flow and water distribution systems, yet their use in predicting the state of urban stormwater drainage systems remains rare. This study investigates the application of Graph-WaveNet (GWN), a type of GNN, in forecasting the states of stormwater systems in Kowloon, Hong Kong. Data was sourced from the Storm Water Management Model (SWMM) spanning 43 rainfall events from 2020 to 2023. Based on the preceding 30 to 60 min of network states and rainfall data, GWN predicted junction inflows, pipe flow rates, and relative water depths (fraction of full area filled by flow) for lead times up to 20, 20, and 30 min, with an R2 greater than 0.6, respectively. Prediction accuracy declines with longer forecast horizons. GWN predicts more time steps ahead for pipes’ flow rates and junctions’ inflows, but fewer for relative water depths during peak versus non-peak periods. It is also more effective at predicting states of large pipes and connected junctions downstream, compared to smaller upstream components. GWN's accuracy improves significantly with precise rainfall nowcasting inputs. This study establishes a significant baseline for GWN's performance in predicting urban stormwater systems during rainfall events.