Study on Machine Learning Method Based on Vector Direction of Flood Process for Flood Forecasting

Accurate flood forecasting is of great significance for flood prevention and mitigation, protection of residents' lives and properties, as well as rational utilization and protection of water resources. To improve the accuracy and reliability of flood forecasting, a deep learning flood process probabilistic forecasting model VD-LSTM-Bootstrap based on the vector direction of the flood process is constructed by coupling the runoff process vectorization method and Bootstrap interval prediction method in the input and output layers of the LSTM model, respectively. Jingle and Lushi watersheds were selected as the study areas, and the model was trained and validated based on 50 and 20 measured flood data according to the 7:3 division ratio, respectively. The results show that, compared with LSTM, the VD-LSTM model has better overall forecasting performance, with NSE above 0.8, RE less than 15%, and RMSE and bias smaller; The discharge simulation results of the VD-LSTM are in better agreement with the measured discharge process lines, and the problems of underestimation of the flood peaks and hysteresis of the model are improved; In terms of probabilistic forecasting, the confidence intervals provided by the VD-LSTM-Bootstrap model exhibit high reliability, with coverage rates in the Jingle and Lushi basins at 90.1%, 85.5%, 80.3%, and 91.7%, 86.2%, 81.6%, respectively, which are above the corresponding confidence level.