Short-term wind power forecasting integrating wake effect modeling with variational mode decomposition enhanced deep learning architectures

This work presents a robust hybrid framework for short-term wind power forecasting, validated on the GECAMA wind farm in Spain, which comprises 69 turbines and has a nominal capacity exceeding 300 MW. The proposed approach combines three established physics-based wake models (Jensen, Bastankhah, Larsen) with five deep learning methods, further enhanced by input preprocessing via variational mode decomposition. Hourly energy production is forecasted using wind data from ECMWF, AROME, and ICON EU meteorological databases. Including wake models as input features helps reduce bias from meteorological signals by accounting for available wind turbines and physical effects, such as wake interactions and terrain. Adding previous forecast errors as features further boosts short-term accuracy. The hybrid models achieve error reductions of 40%–50% for one-hour-ahead forecasts, tapering to 1%–10% by 24 h. With variational mode decomposition (VMD), improvements reach 74%–77% for 3–6 h horizons and about 8% at 36 h. Across all horizons, VMD-enhanced models consistently outperform both standard hybrids and pure physical models. These results show that integrating wake modeling, deep learning, and advanced preprocessing is a practical way to improve wind power forecasts and support reliable electricity market decisions. • Hybrid wake–deep learning model enhances short-term wind power forecasts. • Integrates physics-based wake effects with DL models and VMD preprocessing. • Achieves up to 77% accuracy gain over traditional physical-only approaches. • Validated on Spain’s largest wind farm, improving robustness under terrain effects.