Author(s): Daan Buekenhout; Ricardo Reinoso-Rondinel; Patrick Willems
Linked Author(s): Patrick Willems
Keywords: Flood emulation; Gaussian process; Hybrid model; Pluvial flooding; Urban drainage
Abstract: Urban pluvial flood warning requires near real-time processing of rainfall forecasts due to their short skillful lead times. Detailed physically-based hydrodynamic urban drainage models are too computationally demanding for operational use, especially when working with probabilistic rainfall forecast ensembles. This study explores a hybrid surrogate modelling framework combining simplified hydrodynamic models with machine learning to achieve accurate and timely flood depth mapping. While simplified models significantly accelerate computations by abstracting surface and sewer components, they sacrifice precision compared to the reference hydrodynamic simulations. To bridge this gap, a Gaussian Process (GP) regression model is trained on a comprehensive library of simulation results, enabling efficient bias correction without sacrificing computational speed. Implemented for a case study in Antwerp, Belgium, using radar-derived rainfall data, the framework demonstrated speed improvements with a factor of 8 to 16 depending on the simplified model configuration, while maintaining good accuracy (R² = 0.87).
Year: 2026