Comparative Analysis of Machine Learning Techniques in Wind Power Generation Forecasting: A Data Case Study from Brazil

Wind energy is strategic in the energy transition, but its intermittency makes generation forecasting challenging. This study proposes machine learning models for wind power forecasting over short-term horizons (1 to 6 hours), using real data from turbines located in Espírito Santo (Jan/2023 to Feb/2025) and Rio Grande do Sul (Feb/2020 to Jan/2021). Ensemble algorithms (Random Forest, Gradient Boosting, XGBoost, LightGBM) and deep neural networks (MLP, LSTM, GRU, TCN, Transformer) were compared, with input features based on meteorological variables and temporal lags. Temporal cross-validation was applied to preserve the chronological order of the data, and models were evaluated using statistical metrics (RMSE, MAE, and R²). In the RS case, the TCN model reduced error by up to 3.6% compared to the second-best model. LightGBM combined competitive accuracy with the highest inference speed, running up to 99 times faster than TCN. SHAP analysis identified wind speed as the most important predictive variable.