Abordagem para seleção de atributos e determinação de atrasos de tempo em modelos autorregressivos de previsão baseados em redes neurais artificiais

• Ano de defesa: 2020
• Autor(a) principal: Bezerra, Francisco Elânio
• Orientador(a): Pereira, Fabio Henrique
• Banca de defesa: Pereira, Fabio Henrique, Teixeira, Julio Carlos, Di Santo, Silvio Giuseppe, Dias, Cleber Gustavo, Alves, Wonder Alexandre Luz
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Nove de Julho
• Programa de Pós-Graduação: Programa de Pós-Graduação de Mestrado e Doutorado em Engenharia de Produção
• Departamento: Engenharia
• País: Brasil
• Palavras-chave em Português: modelo autorregressivo; rede neural multicamadas; seleção de atributos; PCA; transformada wavelet
• Palavras-chave em Inglês: autoregressive model; time series; attribute selection; artificial neural network; principal component analysis; wavelet transform
• Área do conhecimento CNPq: ENGENHARIAS::ENGENHARIA DE PRODUCAO
• Link de acesso: http://bibliotecatede.uninove.br/handle/tede/2457

In autoregressive models, the current value of time series is regressed to past values of the same series, and past and present values of one or more exogenous variables, with several applications in the prediction of future values, such as: electricity prediction, condition monitoring equipment, financial market, weather prediction, among others. Although, there are different approaches in the literature proposing improve the accuracy of prediction, there is a research gap regarding the selection of the number of past values of the time series of interest and exogenous variables used in the model. Moreover, the usual approaches in the literature adopt the same delay value for all the variables involved, which may not adequately represent the underlying relationship of the process under study. Thus, the objective of this thesis is to propose an approach for selecting attributes and determining time delays in autoregressive prediction models, based on artificial neural networks, principal component analysis and wavelet. To achieve the objective proposed in this thesis, different approaches were tested, from the effect of time delays on the response variable, the effect between time delays and sampling rate, selection of the most relevant variables and approximations created by type of wavelet transform for build approximations to create time delays for an autoregressive forecasting model. The results have showed the selection of attributes and the use of wavelet transform to create time delays, with a sparse and simplified version of the data, improves the understanding between the input and output variables and, consequently, the prediction accuracy in auto-regressive models. These approaches have involved applications in the electric energy sector, such as, oil dissolved gas in power transformer, electric energy from an electricity distributor company, electric energy from an electricity power generator. The results of these approaches have resulted in publication in a national congress and international scientific journals.