Streamflow prediction using Long Short-term Memory networks

Abstract

Abstract Accurate river streamflow prediction is crucial for hydropower operations, agricultural planning, and effective water resources management. However, forecasting reliable streamflow poses challenges due to the intricate nature of weather patterns and non-linear runoff generation mechanisms. The long short-term memory (LSTM) network has gained prominence for effectively simulating non-linear patterns. Despite its popularity, the performance of LSTM in river flow prediction remains insufficiently understood. This study assesses LSTM's effectiveness and explores how different network structures and hyperparameters impact short-term daily streamflow prediction at Kratie stations, a vital hydrological site in the Vietnam Mekong Delta. Training LSTM on historical streamflow data, we find that the size of the training dataset significantly influences network training, recommending a dataset spanning 2013 to 2022 for optimal results. Incorporating a hidden layer with a non-linear activation function enhances learning efficiency, and adding a fully connected layer slightly improves prediction ability. Careful tuning of parameters such as epochs, dropout, and the number of LSTM units enhances predictive accuracy. The stacked LSTM with sigmoid activation stands out, demonstrating excellent performance with a high Nash–Sutcliffe Efficiency (NSE) of 0.95 and a low root relative mean square error (rRMSE) of approximately 0.002%. Moreover, the model excels in forecasting streamflow for 5 to 15 antecedent days, with five days exhibiting particularly high accuracy.