A Multi-Model River Water Level Forecasting Framework Under Low, Medium and High-Water Level Regimes

Abstract

This study develops and evaluates a multi-step forecasting framework for the Punpun River using Artificial Neural Network (ANN), Bidirectional Long Short-Term Memory (BiLSTM), XGBoost, and a Dynamic Hybrid model. Water-level forecasts were generated at 1-, 3-, 5-, 7-, and 10-day lead times using historical hydrological time-series data. Model performance was assessed using the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). The forecasts were further examined under low-, medium-, and high-water-level regimes to assess model behaviour across different hydrological conditions. The results showed a progressive reduction in forecasting skill with increasing lead time. Overall R² values ranged from 0.95–0.97 at the 1-day lead time and declined to 0.69–0.70 at the 10-day lead time. XGBoost produced the highest R² values at the 1-, 3-, 5-, and 7-day lead times, whereas ANN produced the highest R² at 10 days. XGBoost also yielded relatively lower RMSE at short and intermediate lead times, while the Dynamic Hybrid model showed comparatively lower MAE from 3 to 10 days. Regime-wise assessment indicated that the Dynamic Hybrid model maintained positive R² values with comparatively lower errors under low-water-level conditions across the evaluated lead times. XGBoost showed relatively better representation of high-water-level conditions at short and intermediate horizons. Overall, incorporating multi-step forecasting with regime-wise evaluation provides a more comprehensive assessment of model behaviour and offers a useful basis for developing river water-level forecasting systems for flood early warning and water-resources management.