Study on long short-term memory based on vector direction of flood process for flood forecasting.

Accurate flood forecasting is crucial for flood prevention and mitigation, safeguarding the lives and properties of residents, as well as the rational use of water resources. The study proposes a model of long and short-term memory (LSTM) combined with the vector direction (VD) of the flood process. The Jingle and Lushi basins were selected as the research objects, and the model was trained and validated using 50 and 49 measured flood rainfall-runoff data in a 7:3 division ratio, respectively. The results indicate that the VD-LSTM model has more advantages than the LSTM model, with increased NSE, and reduced RMSE and bias to varying degrees. The flow simulation results of VD-LSTM better match the observed flow hydrographs, improving the underestimation of peak flows and the lag issue of the model. Under the same task and dataset, with the same hyperparameter settings, VD-LSTM can more quickly reduce the loss function value and achieve a better fit compared to LSTM. The proposed VD-LSTM model couples the vectorization process of flood runoff with the LSTM neural network, which contributes to the model better exploring the change characteristics of rising and receding water in flood runoff processes, reducing the training gradient error of input-output data for the LSTM model, and more effectively simulating flood process.