Evolving spatio-temporal forecasting models for renewable energy systems

• Ano de defesa: 2020
• Autor(a) principal: Carlos Alberto Severiano Junior
• Orientador(a): Frederico Gadelha Guimarães
• Banca de defesa: Rosângela Ballini, Eduardo Pestana de Aguiar, Rodrigo César Pedrosa Silva, Laura Corina Carpi
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Minas Gerais
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia Elétrica
• Departamento: ENG - DEPARTAMENTO DE ENGENHARIA ELÉTRICA
• País: Brasil
• Palavras-chave em Português: Renewable energy; Solar energy; Wind energy; Fuzzy time series; Spatio-temporal forecasting; Evolving models
• Link de acesso: http://hdl.handle.net/1843/35669

Renewable energy systems such as solar photovoltaics and wind are sources of energy very sensitive to climate variations, which can affect their generation patterns. It is very important to use mechanisms that can help to anticipate such variations and enable more informed decision-making. Forecasting methods can contribute to this task and therefore their application in this area has been widely studied. Forecasting methods usually take as input historical data from the time series generated by the point of interest. For a further improvement in forecasting accuracy, the information available in space has been also added to forecasting methods. These approaches, called spatio-temporal methods, make use of all the available data collected from different locations. In renewables, variations observed at neighbor locations may occur in the near future at some point of interest, since many of these events are result of climatic phenomena. This reinforces the possibility that spatio-temporal data analysis can improve forecasting performance in renewable energy systems. In addition, climatic events tend to influence the patterns observed in the time series related to energy production in the system, thus presenting non-stationarity. Such scenario demands the development of mechanisms that allow the forecasting model to adapt to changes in time series patterns. In this thesis, proposals for the treatment of such problems related to renewable energy forecasting are presented. From the extension of Fuzzy Time Series (FTS) models, proposals are applied to first deal with the nonstationarity problem using a model adaptation mechanism. Then, another model that also proposes an adaptation mechanism, aligned with the processing of multivariate data, is presented and evaluated regarding the forecast of solar and wind energy. The analyzed adaptation mechanisms were able to provide a performance gain for the proposed models, as well as the use of multivariate data arranged in the context of a spatio-temporal problem. The e-MVFTS model, which integrates an evolving clustering technique with an FTS model to perform spatio-temporal forecasting, presented results comparable to models of greater complexity and scope. The model also presents the advantages of its robustness of parameters and the ability to adapt to changes in data without the need for new training steps. Its adaptation mechanism provides greater flexibility to FTS models, since it does not require a previous configuration of its partitioning scheme and is able to adapt this structure dynamically during runtime. In addition, the clustering algorithm used in the model was originally developed for data stream problems, and is therefore able to handle large volumes of information. These characteristics position it as an extension of the FTS models applicable to the renewable energy forecasting problem. In addition, its FTS-based rationale makes it a model whose representation of rules is easier to understand, which is an additional motivation for being adopted in support of decision making in renewable energy systems. The e-MVFTS also showed good results on artificially generated non-stationary data sets. Therefore, an assessment of the model applied to other forecasting problems can be a direction for future work.