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

Area of Science:

• Hydrology and Water Resources Engineering
• Artificial Intelligence in Environmental Science

Background:

• Accurate flood forecasting is essential for disaster prevention, public safety, and water resource management.
• Existing hydrological models often struggle with accurately simulating the dynamic characteristics of flood runoff, particularly peak flows and timing.

Purpose of the Study:

• To develop and evaluate a novel hybrid model, Vector Direction-Long Short-Term Memory (VD-LSTM), for enhanced flood process simulation.
• To compare the performance of the VD-LSTM model against the standard Long Short-Term Memory (LSTM) model in flood forecasting.

Main Methods:

• Development of a VD-LSTM model integrating the vector direction of flood processes with LSTM neural networks.
• Training and validation using measured rainfall-runoff data from the Jingle and Lushi basins (50 and 49 samples, respectively, split 7:3).
• Performance evaluation based on Nash-Sutcliffe Efficiency (NSE), Root Mean Square Error (RMSE), and bias metrics.

Main Results:

• The VD-LSTM model demonstrated superior performance over the standard LSTM model, showing improvements in NSE and reductions in RMSE and bias.
• VD-LSTM achieved better simulation of observed flow hydrographs, effectively addressing issues of peak flow underestimation and timing lag.
• The VD-LSTM model exhibited faster convergence and a better fit during training compared to the LSTM model.

Conclusions:

• The proposed VD-LSTM model effectively captures the complex rising and receding water dynamics in flood runoff processes.
• Coupling vectorization with LSTM reduces training gradient errors, leading to more accurate and efficient flood process simulation.
• The VD-LSTM model offers a significant advancement for hydrological forecasting and water resource management applications.