A Data-Driven Surrogate Approach for Real-Time Evaluation of Pumping Control Strategies in Urban Flood Management

Real-time control (RTC) of drainage pumping stations offers substantial potential for urban flood mitigation, yet the computational burden of physics-based models severely limits their utility in rapidly evaluating alternative control strategies during flood events. Although many artificial intelligence methods have achieved rapid urban flood prediction, they have not yet been able to model the flood response to infrastructure control. We develop a data-driven surrogate modeling framework that rapidly predicts flood dynamics (maximum flood extent and peak water depth) under varying pumping operation scenarios. We employ high-resolution finite volume hydrodynamic simulations integrated with pumping control modules to generate an extensive training dataset spanning diverse rainfall events and control configurations across a highly urbanized catchment (approximately 100 km²). Specifically, in the test set, the spatial Root Mean Squared Error (RMSE) is less than 0.05 m, and the structural similarity index (SSIM) exceeds 0.95. The prediction is completed in under one second, representing a three-orders-of-magnitude speed-up compared to the numerical model. This method provides an effective tool for the emergency management of urban flooding