Previsão da geração de energia eólica: uma abordagem utilizando técnicas de combinação de modelos de previsão

Abstract

Wind power has a large role in society, as it represents 11% of the energy matrix in Brazil. Furthermore, has an important environmental and economic character. In this sense, obtaining wind power forecast it is important for making decisions regarding to exploration of this kind of energy and for operations planning, aiming the use of energy resources. However, due to the characteristics of the series related to wind power generation, such as intermittence and non-standard external data, such as the climate and the geographical characteristics of wind farms, the development of mathematical models in the context of a series of temporal factors becomes a difficult task. Thus, in this work forty-two learning model ensembles were proposed for wind power time series forecast. The ensembles were generated from different base models combined by means of arithmetic mean, harmonic mean, median and weighted mean (with weights defined by Root Mean Squared Error, Mean Absolute Error and Mean Absolute Percentage Error). The ensembles were applied in order to generate a wind power time series forecast for the forecast horizons of 10, 30, 60 and 120 minutes for two wind farms, located in Bahia, Brazil. The results show that the combination of forecasts of individual models performed through weighted average with weights defined from the Mean Absolute Percentage Error was considered the one with the best performance in terms of the lowest percentage error in 87.5% among the combinations performed. While the combination performed using the harmonic mean presented a larger error in relation to the others in 75% of the situations. The model that weights (with weights defined using the Mean Absolute Percentage Error) the predictions of the support vector machine, extreme learning machines, Gaussian processes and regression with regularization approaches was considered to have the best performance in terms of the smallest error at 62.5 % of ratings.