Metodologia híbrida para previsão da geração de energia eólica para curto e médio prazo utilizando inteligência computacional em região tropical e subtropical

Abstract

It is known that one of the main constituents of modern society is energy, which is necessary to create consumer goods based on natural resources and to supply many of the services that human beings have favored. It is a fact that since the first industrial revolution there has been an exponential increase in emissions of greenhouse gases into the atmosphere, potentiating global warming and, consequently, climate change, air pollution, and health problems. Therefore, scientific studies applied to sustainable technologies are justified to guarantee quality and increase the generation of energy from alternative sources to supply this demand. Inserted in this context, wind energy is a sustainable alternative in full development in Brazil and Uruguay, sites that contemplated this research on the short-term and medium-term wind power forecasting using a hybrid model based on computational intelligence and wavelet decomposition. The general objective of this study was to evaluate and implement improvements in the forecast of wind power generation in the short and medium-term, 1 h to 168 h ahead, in microscale spatial resolution using computational simulation methodology applying supervised machine learning by artificial neural network and the decomposition of temporal signals using Wavelets transform. This effort aimed to meet the shortage in this matter and the demand of the electric energy production and distribution sector in Brazil and Uruguay, to enable an improvement in the use of wind power in current projects and future exploration, production, and commercialization of this energy source. It is noteworthy that this hybrid forecasting model originated a low computational cost tool designed to provide such forecasts to electric energy concessionaires, generators and distributors, and even to the electric system operators. The results achieved in this research proved that the discrete Meyer wavelet function among 48 studied functions has less associated error for the application of filtering and decomposition of wind speed signals, becoming the most efficient for such application, and the use of these filtered data in the feed of recurrent neural networks it was effective for medium-term wind speed forecasting and medium-term wind power forecasting, and short-term wind power ramp forecasting in tropical and subtropical sites.