An Explainable Flash Flood Prediction Model in the Qinling Mountains

Abstract

Mountainous river basins, typically located in river source areas, are characterized by steep terrain and dynamic landforms. These regions experience diverse climates due to topographic uplift, making them susceptible to frequent flash floods. The rapid onset and brief response time of flash floods pose significant challenges for achieving accurate and timely forecasting within limited warning periods. Deep learning models have emerged as powerful tools for high-precision streamflow forecasting. This study develops an LSTM-based multi-sliding window flood forecasting model for various lead times and applies it to the Qinling Mountains watershed, with an emphasis on analyzing the model's interpretability. Results from the Maduwang Basin demonstrate the model's excellent performance in flood prediction for 1- and 3-h lead times. While incorporating historical data can enhance model performance for long lead times, excessive historical inputs may be detrimental. Historical runoff significantly influences model performance. However, its contribution neither consistently increases with temporal proximity to the prediction time nor remains uniformly positive. The contribution of input features varies across different flood stages and can be explained by existing hydrological knowledge. This research demonstrates the potential of deep learning for flood forecasting in mountainous basins while providing insights into the interpretation of deep learning models. This provides scientific support for flood warning systems and emergency management.