Rapid simulation for real-time flood depth prediction using support vector machine

Local Intensive Precipitation (LIP), intensified by climate change, has increasingly caused severe urban flooding. Although traditional hydrodynamic models such as SWMM and FLO-2D offer high accuracy in flood prediction, their computational demands hinder real-time application. This study introduces a rapid flood depth prediction model based on a Support Vector Machine (SVM), trained with data generated from a physically-based 1D-2D coupled simulation. The target area is the Jinheung Apartment intersection in Gangnam, Seoul-an area highly prone to flooding. Cumulative rainfall and manhole overflow data from 1 to 5 h scenarios were used as input variables to predict flood depth. Model validation consisted of two parts: (1) the 1D-2D hydrodynamic model (SWMM-FLO-2D) was validated using observed flood records from September 21, 2010, achieving a 64% match with NDMS inundation points. (2) The trained SVM model was verified by comparing its predictions against FLO-2D results generated using a 3-hour Huff-distributed rainfall scenario. The SVM model showed strong performance with R² = 0.988, NSE = 0.987, % difference = 1.080, and RMSE = 0.098 m. The results confirm that integrating machine learning with physical simulation can provide fast and reliable flood predictions, supporting timely disaster response in urban areas.