Hybrid deep learning models for daily river discharge prediction: lead-time performance analysis in Mymensingh, Bangladesh

ABSTRACT Accurate river discharge forecasting is critical for flood and drought management in deltaic regions vulnerable to climate extremes. This study evaluates one hyperparameter-optimized machine learning model (XGBoost) against five standard-configuration deep learning (DL) models (CNN, LSTM, GRU, CNN-GRU, LSTM+Attention) for daily discharge prediction in the Old Brahmaputra River. Using 22 years of data (2000–2021) from Mymensingh station, models were assessed across four lag-lead configurations (3-1, 3-3, 7-7, 10-10 days) using nine statistical metrics and PCA-based ranking with normalized composite scores. For 1-day predictions, LSTM+Attention achieved the highest composite score (0.7981), followed by CNN-GRU (0.7749), while all models exceeded R2 values of 0.98. CNN-GRU dominated 3-day forecasts with a composite score of 0.9390 (RMSE: 97.3199 m3/s, R2: 0.9498), and maintained competitive performance for 7-day predictions (R2: 0.8139). For 10-day forecasts, LSTM was superior with a 0.9115 composite score, yet the RMSE's sharp rise from 53.7469 (1-day) to 226.7971 m3/s highlights a significant accuracy loss with longer lead times. Performance degraded substantially with extended lead times, showing systematic underestimation of peak monsoon discharges. Results demonstrate that the DL models match or surpass optimized traditional approaches, with PCA rankings identifying optimal scenario-specific hierarchies for resource-constrained operational environments.