Uma nova variante de stacking ensemble baseada em aprendizagem de máquina para previsão da velocidade do vento

Abstract

Wind power generation has experienced a significant increase in recent years, making wind speed forecasting tools an essential task to enhance the integration of renewable energy sources. This thesis proposes a novel Stacking ensemble variant, named Data Stacking, based on Machine Learning algorithms for wind speed forecasting. The model simultaneously combines original input data with the predictions of base learners to produce the final estimate. The methodology was validated using two datasets: the Korea Composite Stock Price Index (KOSPI) benchmark and a wind speed database from a Brazilian city, including several meteorological variables. The process involved data preprocessing, feature selection, and evaluation under different forecasting horizons with multiple error metrics. Results showed that Data Stacking consistently outperformed individual algorithms and traditional ensembles such as Stacking and Multi-level Stacking. For the Brazilian wind speed dataset, the model achieved a mean absolute error (MAE) of 0.4855 and a normalized root mean square error (nRMSE) of 0.2389, with error reductions ranging from 0.7% to 4.4% compared to Stacking, depending on the selected base learners. Furthermore, the study demonstrated that Data Stacking can yield satisfactory results even when including base learners with poor predictive performance, highlighting its robustness.