Multi-Horizon Forecasting of a Hybrid Wind–Solar Plant Using Machine Learning and SHAP

In response to the rising share of renewable sources in modern power systems, accurate generation forecasting has become essential for grid stability and the reliable operation of hybrid plants. Wind and solar power remain difficult to forecast because they strongly depend on changing weather conditions, while most machine learning models behave as black boxes, which reduces their usefulness for real operation. This study aims to forecast the production of a co-located wind–solar plant over three horizons: 1h, 6h, and 24h, using quality-controlled hourly data from a single site, while keeping the models interpretable. We employed four machine learning techniques: XGBoost, random forest, multilayer perceptron, and support vector regression, applied separately to wind and solar power and compared with a persistence baseline, a standard benchmark for short-term forecasting, using the mean absolute error (MAE) and the coefficient of determination (R²). SHAP (SHapley Additive exPlanations) was then applied to explain the predictions. The results indicate that the best machine learning models outperformed the persistence baseline across all considered horizons. At the 1h horizon, the best models reached an R² of 0.962 for wind using random forest and 0.961 for solar using XGBoost. For solar power, the reduction in error compared with persistence was about 44% at 1h and reached 57% at 6h, where persistence is weakest. Wind power was mainly driven by wind speed, with a negligible effect of air density, while solar power was governed by ground irradiance. The SHAP analysis showed that recent production dominates at 1h, whereas wind speed and ground irradiance become more important as the horizon increases. Future improvements in forecasting accuracy could include developing horizon-specific architectures, incorporating probabilistic forecasts to quantify uncertainty, and exploring jointly trained wind–solar models to capture cross-source interactions.