Rede neural convolucional autoconfigurada para identificação de cargas elétricas similares em Smart Grid

• Ano de defesa: 2021
• Autor(a) principal: Vianna, Vinicius Wittig
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Mestrado em Energia; Centro Universitário Norte do Espírito Santo; UFES; Programa de Pós-Graduação em Energia
• Palavras-chave em Português: NILM; aprendizado de máquina automatizado; otimização heurística; custo computacional; subject.br-rjbn; Engenharia/Tecnologia/Gestão
• Link de acesso: http://repositorio.ufes.br/handle/10/14398

Convolutional Neural Networks (CNN) have been shown to be very efficient tools in the task of identifying similar electrical charges in a smart grid. However, these networks currently depend on highly skilled labor to be properly designed, in view of the large number of hyperparameters and variety of adjustments, which makes this undertaking highly laborious and costly. In addition, traditionally manual and completely empirical adjustments do not guarantee the achievement of an optimal architecture, due to the impossibility, in general, of testing all adjustment combinations for a set of hyperparameters, within a previously defined value space. Therefore, this work has as main objective to add an automated adjustment mechanism of the hyperparameters of a CNN dedicated to the autonomous identification of highly similar electrical charges in a smart grid. For this purpose, a classification system based on a CNN architecture manually obtained from previous work is initially used, in order to find the minimum number of necessary and sufficient cases to strategically allow a classification accuracy of at least 95%. Then, the optimum number of cases is used to optimize the number of CNN convolutional and dense layers, in addition to the number of neurons in such layers, without compromising the performance of the reference architecture (the manually adjusted one). The system was tested using two sets of data, one based on arrays of up to four technically identical fluorescent lamps and the other based on arrays of up to four microcomputers also technically identical. With the first set, the reduction in the number of cases required for training the reference CNN was 90%, while in the second case, it was 33.4%. Then, the respective minimum number of cases were used to adjust hyperparameters from the reference CNN, resulting in a reduction of 56.02% and 90.41% over the number of trainable parameters of such networks, from the respective databases.