Machine learning-based prediction model of wind turbine power generation

Renewable energy sources have become central to the transition toward cleaner energy systems, with wind energy demonstrating the most rapid global growth since 1990. However, its production is inherently dependent on variable and uncontrollable factors such as weather conditions and wind dynamics. In this work, we analyze a dataset spanning two and a half years, collected from wind turbines, and apply extensive exploratory data analysis and preprocessing to enable accurate forecasting of wind power generation. Initially, the dataset was evaluated using multiple regression models for baseline predictions, while the Prophet model was employed to extract long-term trends and seasonality. The processed data were then integrated and used as input for CatBoost and Random Forest models, incorporating a windowing mechanism informed by autocorrelation and partial autocorrelation analysis to optimize temporal dependencies. Forecasting was conducted across three horizons: 15 min, 1 day, and 1 week ahead. The proposed hybrid approach achieved a root mean square error of 30.6 for 15 min forecasting, 50 for one-day forecasting, and 41 for one-week forecasting, representing at least a 50% improvement over the best standalone regression. Results further confirm the expected trend that longer forecasting horizons increase RMSE and reduce R 2 , due to resampling constraints and the need for more extensive input data. Nonetheless, the hybrid methodology consistently outperformed standalone models, demonstrating stability and robustness across different horizons. By leveraging the complementary strengths of multiple regressors within a unified framework, this study highlights the potential of hybrid machine learning approaches to significantly enhance the predictive accuracy of wind energy forecasting.