Estrutura Híbrida Combinando Wavelets e Ridge Regression Ensemble de Modelos de Aprendizado de Máquina Aplicados à Previsão Solar e Eólica no Brasil e na Espanha

• Ano de defesa: 2023
• Autor(a) principal: Carneiro, Tatiane Carolyne
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Aprendizado de Máquina; Energia renovável; aprendizado profundo; regressão penalizada; teoria do portfólio; recurso eólico; Inteligência artifícial; Regularização; Machine learning; deep learning; penalized regression; portfolio theory; wind resource; solar resource
• Link de acesso: http://repositorio.ufc.br/handle/riufc/75040

In recent years, with the rapid development of electricity generation through wind and solar sources, some problems have gradually emerged (uncertainty and unpredictability in the final generation) and are often inherent to intermittency. Currently, one of the essential methods to solve these problems is the application of forecasting methodologies. These methods learn the behavior of the analyzed series and implement the acquired knowledge to predict future values, being able to perform this task with individual methodologies, by combining different methodologies (hybrid models) or by integrating different results from individual models (ensemble models). This thesis proposes the application of two ensemble/integration methodologies: a) Ridge Regression Ensemble (RRE); and b) Hybrid model that combines data decomposition through the Wavelet Transform (WT), Machine Learning (AM) models and the RRE (WD-RNA-RRE). To compare and validate the performances of the proposed models, the model of integration through Portfolio Theory (PrevTP) is also applied. The objective of the applications is to integrate consolidated wind and solar forecasting methodologies applied to two locations with different latitudes and climate profiles. Based on the simulations developed (PrevTP,WD-RNA-RRE and RRE), the methodologies proved to be efficient in improving the forecasting performance of isolated methods and applicable to different locations around the world. In terms of MAPE and RMSE, in the application to solar data and in both locations, the ensemble models (WD-RNA-RRE and RRE) achieved better accuracy than the best Cascade Forward Back Propagation (CFBP), model that had the best performance among the individual applications, and than the PrevTP model. In the applications to wind data, in terms of MAPE and RMSE, the ensemble models (WD-RNA-RRE, RRE and PrevTP) managed to improve the performance in relation to the models applied individually. Thus, PrevTP can be highlighted as the best application to Spanish data and RRE with the best results in the application to data from Brazil. PrevTP improved the accuracy of individual methodologies in two of the four applications developed. The WD-RNA-RRE and RRE modeling reduced the forecast errors in all applications and can be useful in optimizing the planning of the use of intermittent solar and wind resources in the electric matrices.