Hybrid Ensemble AND Deep Learning Models FOR Suspended Sediment Concentration Prediction IN THE Subarnarekha River Basin

Abstract

Accurate prediction of suspended sediment concentration (SSC) is critical for water resources management, sediment control, and reservoir operation. This paper compares the performance of the Random Forest (RF), XGBoost, Long Short-Term Memory (LSTM), and Stacked Ensemble in predicting SSC on a daily basis with different lag-values (1-5 days). Model performance was assessed using the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and Nash–Sutcliffe efficiency (NSE). The results show that the inclusion of antecedent conditions enhances the predictive ability considerably, with a 2-day lag configuration producing the best predictive accuracy (R² = 0.96, NSE = 0.95; N). Tree-based methods (both RF and XGBoost) were more effective in capturing nonlinear responses and extremes compared to baseline models, whereas LSTM was fairly good at capturing sequential dependencies and essentially smooth peaks. The Stacked Ensemble model performed consistently better than all standalone models with respect to accuracy, stability, and variance reproduction between training and testing phases, as shown in scatter plots, Taylor diagrams and time-series simulation. The above results emphasize the need to apply hybrid methods to SSC prediction. The suggested ensemble framework would provide a robust and transferrable approach to the hydrological and water management community to anticipate and manage sediment and ensure that floods are controlled, and basin-scale systems sustainable.