A Particle Swarm Optimization-Based Deep Learning Framework for Enhanced Prediction of River Water Level

The forecasting of water levels (WLs) in the coastal rivers is critical for effective management of water resources and the execution of flood hazard mitigation strategies. This study aims to develop three DL models that are optimized using particle swarm optimization and to evaluate their effectiveness for projecting river WLs. The models consist of PSO-LSTM, PSO-Bi-LSTM, and PSO-CNN, which have been built and validated at the Khulna hydrologic station of the Rupsha-Pashur River. The models are evaluated using four statistical performance indicators: RMSE, MAE, KGE, and $\mathbf{R}^{\mathbf{2}}$. The objective was to enhance the predictive accuracy of the CNN, LSTM, and Bi-LSTM models integrated with PSO. The findings suggest that the LSTM model has outperformed all the standalone DL models with the highest accuracy in WL prediction. It is also evident from the results that the PSOLSTM model has been identified as the best model in predicting WL among all the standalone and PSO-optimized models. This conclusively proves that the PSO-LSTM-based DL framework is highly effective and could be a sustainable alternative for the enhanced prediction of river WL. The framework detailed in the present study is anticipated to serve as a valuable resource for the enhanced prediction of WL in the rivers of Bangladesh and other parts of the world with limited datasets.