Optimizing LSTM networks for improved river hydrological prediction

This study focuses on evaluating the performance of water level prediction utilizing the sparrow search algorithm (SSA) optimization with the long short-term memory (LSTM) model. This hybrid SSA-LSTM model is employed to forecast the Kelantan River water level. Data pre-processing techniques such as mean interpolation and the wavelet denoising technique are utilized to address the missing data issue and ensure data reliability, respectively. The accuracy of the predictions made by the SSA-LSTM model is compared with that of the manually tuned LSTM and PSO-LSTM models. It is demonstrated that the SSA-LSTM model yields superior predictions over both the traditional LSTM and PSO-LSTM models. Utilizing the SSA optimization, this gives rise to improved accuracy, with a reduction in the root mean square error (RMSE) from 0.5829 to 0.5302, which infers a 9.04% reduction as compared with that of the traditional technique. In addition, the coefficient of determination (R2) increases from 0.8160 to 0.8478, thereby indicating improved fitting and generalization abilities. Furthermore, Taylor analysis is performed to assess the water level prediction accuracy, while peak analysis is conducted to evaluate water level prediction during elevated water scenarios.