Ensemble Learning Models for Wind Power Forecasting

Wind power, a clean and sustainable energy source, has experienced substantial growth in Brazil’s energy capacity over recent decades. Accurate wind power forecasting is crucial for effectively harnessing wind energy and ensuring the reliable operation of power systems. However, due to the unique characteristics of wind power generation time series, developing statistical models for forecasting can be a challenging endeavor. This paper introduces a comprehensive approach by proposing forty-two ensemble learning models designed for forecasting wind power time series. These ensembles are created by combining various machine learning models and utilizing different aggregation methods, which incorporate various statistical measures such as the arithmetic average, harmonic average, median, and weighted average, with weights determined through metrics like mean absolute percentage error (MAPE), mean absolute error, and root mean squared error, was employed to evaluate its performance in forecasting wind power time series. This evaluation was conducted at time intervals of 10, 30, 60, and 120 min for two wind farms situated in Bahia, Brazil. The findings suggest that the ensemble method, which combines forecasts from individual models using weighted averages based on MAPE-derived weights, proved to be effective achieving the lowest percentage error in 87.5% of the evaluated cases. Conversely, the ensemble utilizing the harmonic average exhibited a higher error rate compared to the alternatives in 75% of the cases.