Power Prediction in a Wind Farm Using Machine Learning Models

Resumen

This paper presents the development and evaluation of an advanced power prediction system for a wind farm using machine learning and deep learning models, based on historical meteorological data. The analyzed dataset comprises 25,596 hourly records corresponding to the period 2022–2024, which were subjected to a rigorous process of cleaning, physical validation, and generation of derived variables, such as air density and power according to the Betz limit. Based on a Pearson correlation analysis, wind speed, solar radiation, and air temperature were identified as the most influential variables in power generation. Multiple predictive models were trained and compared, including linear regression, Random Forest, XGBoost, MLP neural networks, LSTM, and Transformers, evaluated using RMSE and coefficient of determination (R²). Additionally, lagged variables were incorporated to improve prediction at 1-, 3-, and 6-hour horizons. The results show that a hybrid LSTM + XGBoost architecture, based on residual error correction, consistently outperforms individual models, achieving an RMSE of 61.98 W and an R² of 0.779 in one-hour prediction. Finally, the proposed model was implemented in an operational environment, demonstrating its viability for real-world energy management and planning applications.