Classificação de faltas em linhas de transmissão utilizando métodos de aprendizado de máquina

• Ano de defesa: 2021
• Autor(a) principal: Fonseca, Gabriel Aparecido
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Lavras; Programa de Pós-Graduação em Engenharia de Sistemas e Automação; UFLA; brasil; Departamento de Engenharia
• Palavras-chave em Português: Energia Elétrica - Linhas de transmissão; Classificação de faltas; Filtro notch; MiniRocket; Aprendizado de máquina; Electric Power - Transmission lines; Fault classification; Notch filter; Machine learning; Automação Eletrônica de Processos Elétricos e Industriais
• Link de acesso: http://repositorio.ufla.br/jspui/handle/1/49241

Power transmission lines are components highly susceptible to faults. Several factors such as animals, human failure and lightning can lead to the occurrence of a fault. In addition, the increasing demand for electricity generation, distribution and transmission has contributed to this becoming a recurrent problem. Several works have already explored the use of computational intelligence, signal processing and other techniques in the construction of protective methods for quick verification and action in the occurrence of transmission line fault. Many of these works focus on approaches using signal processing such as Fourier or wavelet transforms. With the advance of machine learning, some classic techniques started to be used in this area with success. This work focuses on the offline classification of ten (AG, BG, CG, AB, AC, BC, ABG, ACG, BCG and ABC) types of faults that arise when a short circuit occurs in the transmission line, investigating the use of classical techniques such as notch filter and random forests. For comparative purposes, recently created techniques, called Rocket and MiniRocket, were used to extract features in time series and good results were obtained in the identification of faults that occurred in the transmission line. As a result of this dissertation, accuracies greater than 93% were obtained considering up to 1/16 cycle post-fault. For signals with 1 and 1/2 cycle post-fault, accuracies higher than 97% were obtained.