A lightweight interpretable ensemble for short-term wind speed forecasting

Abstract Using five years of hourly meteorological data from Basel, Switzerland, a lightweight, interpretable ensemble method that combines complementary base learners-Decision Tree (DT), Orthogonal Matching Pursuit (OMP), and Huber Regressor-is introduced through a validation-guided linear blending strategy. Unlike traditional fixed-weight or complex stacking ensembles, optimal combination weights are learned by the proposed approach on a held-out validation window, respecting temporal dependencies via a rolling-origin expanding-window evaluation. The inputs are further enriched by domain-specific feature engineering (cyclical encodings for wind direction and season, rolling statistics). An R $$^{2}$$ of 0.879, MSE of 16.37, and RMSE of 4.04 are achieved by the proposed ensemble on a held-out test set, surpassing all individual models. Although Gradient Boosting yields marginally higher accuracy (R $$^{2}$$ = 0.889), the proposed ensemble is recommended for operational use due to its 8 times faster training, 4 times faster inference, 95% fewer parameters, superior interpretability, and greater robustness to input noise, offering a practical balance between accuracy and deployability. It is confirmed by an ablation study that each base model contributes uniquely to the ensemble’s robustness, and resilience to input perturbations is demonstrated by sensitivity analysis. By balancing accuracy, computational efficiency, and interpretability, a practical solution for short-term wind forecasting, supporting better grid scheduling and renewable energy management, is offered by the proposed ensemble.