Dynamic Prediction of Rainfall Runoff in Urban Drainage Systems Based on Transformer Modeling

Abstract

Transformer-based modeling has become a transformative approach for capturing spatiotemporal dependencies in urban hydrological systems, addressing limitations in traditional methods. This study introduces a hybrid framework that combines Transformer architectures with the Saint-Venant equations to improve rainfall-runoff prediction in urban drainage systems. By embedding hydrodynamic principles into the model via a physics-constrained loss function, the framework ensures physical consistency while leveraging the self-attention mechanism to identify critical rainfall events and temporal interactions. The model was validated using synthetic dam-break scenarios and real-world case studies in Shanghai, a region characterized by complex drainage networks. Results demonstrate the framework's ability to accurately replicate wave propagation and predict runoff dynamics, even in ungauged or data-scarce areas. The physics-informed design enhances generalization across diverse urban hydrological settings, while attention weight visualizations provide interpretability, offering insights into key rainfall-runoff dynamics. This hybrid approach bridges the gap between physical theories and data-driven techniques, delivering a robust, scalable, and interpretable solution for stormwater management. It enables actionable insights for flood risk mitigation, infrastructure planning, and intelligent water management in dynamic urban environments.