Daily Runoff Forecasting in the Middle Yangtze River Using a Long Short-Term Memory Network Optimized by the Sparrow Search Algorithm

To address the challenge of predicting runoff processes in the middle reaches of the Yangtze River under the influence of complex river–lake relationships and human disturbances, this paper proposes a coupled model based on the Sparrow Search Algorithm-optimized Long Short-Term Memory neural network (SSA-LSTM) for daily runoff forecasting at the Jiujiang Hydrological Station. The input data were preprocessed through feature selection and sequence decomposition. Subsequently, the Sparrow Search Algorithm (SSA) was utilized to perform automated of key hyperparameters of the Long Short-Term Memory (LSTM) model, thereby enhancing the model’s adaptability under complex hydrological conditions. Experimental results based on multi-station hydrological and meteorological data of the middle reaches of the Yangtze River from 2009 to 2016 show that the SSA-LSTM achieves a Nash–Sutcliffe Efficiency (NSE) of 0.98 during the testing period (2016). The Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) are reduced by 49.3% and 51.3%, respectively, compared to the standard LSTM. A comprehensive evaluation across different flow levels, utilizing Taylor diagrams and error distribution analysis, further confirms the model’s robustness. The model demonstrates robust performance across different flow regimes: compared to the standard LSTM model, SSA-LSTM improves the NSE from 0.45 to 0.88 in high-flow scenarios, exhibiting excellent capabilities in peak flow prediction and flood process characterization. In low-flow scenarios, the NSE is improved from −0.77 to 0.72, indicating more reliable prediction of baseflow mechanisms. The study demonstrates that SSA-LSTM can effectively capture hydrological nonlinear characteristics under strong river–lake backwater and human disturbances, providing a high-precision and high-efficiency data-driven method for runoff prediction in complex basins.