Emprego de redes neurais artificiais para predição de temperatura de topo de óleo e perda de vida útil em transformadores de potência

• Ano de defesa: 2020
• Autor(a) principal: Kaminski Júnior, Antônio Mário
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; Engenharia Elétrica; UFSM; Programa de Pós-Graduação em Engenharia Elétrica; Centro de Tecnologia
• Palavras-chave em Português: Redes neurais artificiais; NARX; Predição de temperatura; Temperatura de topo de óleo; Transformadores de potência; Temperatura de ponto mais quente; Perda de vida útil; Artificial neural networks; Temperature prediction; Top-Oil temperature; Power transformers; Hot-spot temperature; Loss of life; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: http://repositorio.ufsm.br/handle/1/22470

The precise temperature prediction in power transformers allows a better use of its nominal capacity, extending the equipment's useful life and strategic planning based on the expected future operating conditions. The proposal of new models that present a good predictive capacity is, therefore, of great interest to those responsible for power transformers. The present work presents a method of developing Artificial Neural Networks (ANNs), justifying the parameters chosen based on the thermal behavior of power transformers, for prediction of top-oil temperature using the NARX neural network model, not yet used for temperature prediction of transformers. All data sets used for training and testing the predictive ability of ANNs are real monitoring data from five elevating transformers in a hydroelectric plant. Tests of prediction ability were performed for all transformers, combining trained networks from one of the transformers and applied to the inputs of others, addressing in which situations the best and worst performances occurred. Afterwards, the methods for calculating the loss of life of transformers proposed by standards are presented and a comparison is made between the one calculated from the monitoring data and from the temperature values provided by the neural network. In order to validate the prediction capacity for expected future scenarios, six fictitious scenarios of long duration are proposed and then their useful life is estimated. All the results obtained are satisfactory, with errors below 4%, or 2 °C on absolute values, most of the periods in which the tests were carried out, capable of proving the predictive capacity of the ANNs developed using the method presented not only in its application for temperature monitoring, but also from the perspective of loss of life.