Assessing LSTM algorithm performance for daily runoff prediction at Hoa Duyet hydrological station, Vietnam

Accurate discharge forecasting is crucial for effective water resource management, flood risk mitigation, and hydrological planning, particularly in regions prone to extreme weather events. This study evaluates the performance of a Long Short-Term Memory (LSTM) network in predicting river discharge at the Hoa Duyet hydrology station. The prediction model is developed using rainfall data from the Ngan Sau river basin, collected over a 49-year period from 1975 to 2023. The model's accuracy was assessed across a range of lead times (1-day, 3-day, 5-day, and 7-day) and time lag length (365, 90, 30, 10, and 7 days). It was revealed that short-term forecasts (e.g., 1-day) consistently achieved high accuracy, with the time lag length 90-day yielding the best Nash-Sutcliffe Efficiency (NSE) of 0.864. Seasonal analysis indicated the reliability of the model for the rainy season (NSE = 0.863), but lower accuracy during the dry season (NSE = 0.582), reflecting the challenges of predicting low-flow dynamics. The model also demonstrated reasonable accuracy in predicting annual runoff peaks, with an average error of 91.75 m³/s, although discrepancies were observed in specific years. These findings highlight the LSTM model's capacity to adapt to diverse temporal configurations and hydrological conditions, making it a valuable tool for discharge prediction while emphasizing the need for further optimization in low-flow and extreme event scenarios