Flood Prediction Using Random Forest Trained on Hydrodynamic Simulations Under Hypothetical Design Rainfalls

In correspondence to the increasing need for fast and accurate flood predictions, augmented approach of hydrodynamic (HD) models and machine learning models has been investigated in recent years. Aiming for detailed spatial and temporal predictions of water depth under actual rainfall events, this research used flood simulation results, carried out with hypothetical design rainfalls of various shapes, peak times and intensity using a well-established HD model, as the training data for the construction of random forest model. By including features on runoff coefficients and water levels of river, the augmented method was able to capture the hydraulic differences with respect to land use with coefficient of determination (R2) and accuracy greater than 0.95. When evaluated against historical rainfall events of 2018 and 2021, the model of best performance resulted in an overall RMSE of approximately 0.1 m for both events, even with iterated prediction with lead time over days, indicating a high potential of this model for future real-life applications.