Streamflow prediction using an incremental attention network with LSTM and Chaos optimization techniques

English (OpenAlex)

Accurate streamflow prediction is vital for effective water resource management, flood control, and hydrological modeling. This study presents a novel hybrid framework that integrates an Incremental Attention Network (IAN) with Long Short-Term Memory (LSTM) networks, optimized using Chaos Optimization Techniques (COT), to capture complex temporal dependencies in streamflow data. The IAN employs a dual attention mechanism to model interdependencies between past and target streamflow patterns, while the LSTM learns sequential trends and long-term dependencies. To address the chaotic and nonlinear nature of streamflow, COT is utilized to optimize model hyperparameters, enhancing solution space exploration and convergence. The proposed model was evaluated using real-world datasets from the Altınapa, Demirci, and Ustunler Flow Measurement Stations (FMS) in Turkey’s Konya Closed Basin. Results demonstrate the IAN-LSTM model’s superiority over benchmark models. For Altınapa, it achieved the lowest RMSE (0.03974), MAE (0.01260), and highest R 2 (0.8258) and KGE (0.8977). In Demirci, the model attained an exceptional R 2 of 0.9927 and KGE of 0.9867. Despite the higher variability in Ustunler, it achieved an R 2 of 0.7306 and KGE of 0.8342, outperforming other models. The proposed approach offers a reliable tool for accurate and robust streamflow forecasting.