Rapid forecasting of urban flood inundation using multiple machine learning models

Abstract

Urban flood inundation is worsening as the number of short-duration rainstorms increases, and it is difficult to accurately predict urban flood inundation over a long lead time; however, the traditional hydrodynamic-based urban flood models still have difficulty realizing real-time prediction. This study establishes a rapid forecasting model of urban flood inundation based on machine learning (ML) algorithms and a hydrodynamic-based urban flood model. The ML model is obtained by training the simulation results of the hydrodynamic model and rainfall characteristic parameters. Part of Fengxi New Town, China, was used to validate the forecasting model. A comparison of ML predictions and hydrodynamic model simulations shows that when using one ML algorithm (random forest (RF) or K-nearest neighbor (KNN)) for inundation prediction, the accuracy of the inundation water volume and area is insufficient, with a maximum error of 28.56%. Combining the RF and KNN models can effectively improve the prediction accuracy and overall stability, the mean relative errors of the inundation area and depth are less than 5%, and the mean relative errors of the inundation volume can control within 10%. The simulated time of a single rainfall event can be controlled within 20 s, which can provide sufficient lead time for emergency decision-making, thereby helping decision-makers to take more appropriate measures against inundation.