Hybrid Lstm-Xgboost Framework for Multi-Day Water Level Forecasting: A Residual Error Correction Approach

Abstract

Accurate river water level forecasting is crucial for flood risk mitigation, reservoir operation, and sustainable water resource management under increasing hydrological variability. This study proposes a hybrid Long Short-Term Memory–XGBoost (LSTM–XGBoost) framework for multi-day water level forecasting in the Punpun River Basin, India. The framework combines LSTM’s temporal sequence learning with XGBoost’s nonlinear regression and residual correction capabilities. Forecasting performance was evaluated for 1, 3, 5, 7, and 10-day lead times using KGE, R², NSE, RMSE and MAE. The hybrid model consistently outperformed standalone LSTM and XGBoost models. At the 1-day horizon, it achieved KGE, R², and NSE values of 0.976, 0.971, and 0.971, respectively, with RMSE and MAE of 0.315 m and 0.172 m. Although accuracy declined with increasing lead time, the hybrid model exhibited slower performance degradation. Results confirm improved reliability and robustness in capturing nonlinear hydrological dynamics and peak water levels.