A rapid and efficient method for flash flood simulation based on deep learning

Among the various natural disasters, the death caused by flash flood is the highest. Recently, the combination of deep learning methods and hydrodynamic models has shown superior performance in the simulation of urban and plain areas. However, when dealing with flash flood simulation, the research still faces numerous challenges due to limitations such as data scarcity, small sample sizes, complex terrain, and high levels of uncertainty. Therefore, in this study, we innovatively combined deep learning methods with flash flood simulation and proposed a TCN model to predict the spatiotemporal dynamics of flash floods. First, we extracted the typical rainfall patterns in the study area and used design storm methods to generate a hydrograph dataset, which includes various rainfall patterns and return periods. Then, we developed a Temporal Convolutional Network (TCN) model to predict flash floods. Finally, the benchmark test was carried out by Convolutional Neural Network (CNN), which further proved the performance of TCN. The study found that: (1) The TCN model effectively predicts flash floods, with average MAE, RMSE and NSE reaching 0.04, 0.17 and 0.834 on the validation set. However, the CNN model performed better in small flood scenarios; (2) Error boxplots show that simulation errors for both models increase with the flood volume, and reach the maximum around the flood peak, but the TCN model demonstrated better stability and fewer outliers; (3) For the change of water depth at key points, both TCN and CNN effectively capture the fluctuation of water depth with time in the early stage of flood, but TCN showed higher consistency in the recession period. The results show that the rapid simulation method of flash flood based on TCN can better capture the dynamic characteristics of flash flood, and has been well applied in mountainous areas, which provides a new method for the prediction and early warning of flash flood disasters.