Long-term energy forecasting for run-of-river hydropower plants using artificial intelligence

This study presents a deep learning approach for the long-term forecasting of run-of-river (RoR) hydropower plants, addressing the limited focus on long-horizon forecasting in the current state-of-the-art. Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models were developed incorporating catchment-averaged rainfall, hydrological variables (river inflow and hydraulic head) to accurately capture the hydrological variability, and drivers in energy generation. Both models successfully captured the nonlinear time dependent behavior of the hydropower generation and strongly outperformed baseline methods including persistence, SARIMA and XGBoost. The performance of the GRU model was slightly higher compared to the LSTM model with RMSE 110.48 MWh, MAE 74.21 MWh and R2̂ 0.943, RMSE 115.58 MWh, MAE 86.72 MWh and R2̂ 0.938, respectively. The dynamic residual-based prediction intervals for the uncertainty quantification displayed more variability during low flow times and better reliability during high flow scenarios. The long-term recursive forecasts maintained seasonality, but the forecasts diverged gradually as errors propagated, suggesting that forecasts should be viewed as scenario-based forecasts, not as deterministic results. The results showed the potential of deep learning models, especially GRU, for precise and credible forecasting of hydropower. The inclusion of rainfall significantly improved the model’s capacity to accurately capture the hydrological variability. This study supports energy transition by enabling dependable energy planning, decision making and climate resilient operation of RoR hydropower in data-constrained regions.