Forecasting hourly discharge from Crestuma-Lever dam, Portugal

Accurate forecasts of run-of-river dam releases are essential for estuarine modelling but challenging due to multi-periodic signals driven by hydrology and electricity markets. We forecast the hourly discharge of the Crestuma–Lever dam (Portugal) using five stochastic frameworks: SARIMAX, ARIMAX with Fourier terms (ARIMAX–F), Regime-switching, VAR, and ARDL. Rolling-window validation across the four seasons of the 2019/20 hydrological year demonstrates that ARDL offers the highest fidelity within 48 hours, while all models except VAR remain stable over weekly horizons. In a one-year-ahead forecast, SARIMAX, ARIMAX–F, Regime-switching, and ARDL maintain low RMSEs (223–243 m3 s−1), whereas VAR's error rises significantly. Non-physical negative values are addressed via truncation-to-zero. Findings support a hybrid operational strategy: VAR for daily dispatch (≤24 h) without exogenous data, and ARDL or ARIMAX specifications for extended horizons, enabling robust independent forecasting without requiring proprietary utility data.