Detecção de faltas de alta impedância: uma abordagem utilizando diferentes técnicas de aprendizado de máquina

• Ano de defesa: 2024
• Autor(a) principal: Kilian, Eduardo Davila
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Pampa; UNIPAMPA; Mestrado Acadêmico em Engenharia Elétrica; Brasil; Campus Alegrete
• Palavras-chave em Português: Engenharia elétrica; Impedância (Eletricidade); Aprendizado do computador; Adaptação de impedância; Impedance (Electricity); Impedance matching; Machine learning; Electrical engineering; CNPQ::ENGENHARIAS
• Link de acesso: https://repositorio.unipampa.edu.br/jspui/handle/riu/10017

This study presents an approach for applying machine learning (ML) techniques in the context of high-impedance fault (HIF) detection in power distribution networks. The proposed methodology aims to detect the occurrence of HIFs by extracting features from the substation current signal while avoiding false alarms triggered by other events, such as capacitor bank switching, transformer energization, and load connection. To achieve this, the selected features are extracted from the statistical distribution of the data within a moving window that scans the current signal. Additionally, a feature selection technique is applied to assess the impact of each feature on the performance of the ML algorithms and to determine whether the removal of any feature negatively affects the model’s classification capability. A total of 1,722 simulations were conducted, comprising 864 cases of HIF and 856 cases of other events occurring in the distribution system. Furthermore, a validation set was created, consisting of 574 cases, with 288 HIF cases and 286 other events. Finally, a test set was used, containing the same number of cases as the validation set. In the test set, to evaluate the robustness of the methodology against different noise levels, SNR values ranging from 20 dB to 100 dB were introduced. The results demonstrated that the methodology, when combined with different ML models, maintained performance metrics above 90% for noise levels above 40 dB, proving to be a robust tool for accurate HIF detection. Keywords: High Impedance Fault, Machine Learning, Electric Power Systems Protection, Distribution System, Electric Arc.