Localização de faltas em linhas de transmissão utilizando processamento estatístico de sinais e aprendizado de máquina

• Ano de defesa: 2025
• Autor(a) principal: Pereira, Thalita Kely
• Orientador(a): Ferreira, Danton Diego
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Lavras; Escola de Engenharia – EENG
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia de Sistemas e Automação
• País: brasil
• Palavras-chave em Português: Localização de faltas em linhas de transmissão; Identificação cega de segunda ordem; Análise de componentes independentes; K-Nearest Neighbors; Extreme Gradient Boosting; Multilayer perceptron; Árvore de Decisão; Fault location in transmission lines; Second-order blind identification (SOBI); Independent component analysis; Decision tree; Random forest
• Área do conhecimento CNPq: ENGENHARIAS
• Link de acesso: https://repositorio.ufla.br/handle/1/60033

This work proposes an approach for fault location in transmission lines using signals simulated in ATP, the SOBI algorithm for extracting independent components, and three denoising techniques: Principal Component Analysis (PCA), Wavelet Transform, and Moving Average Filter. The third component extracted by SOBI was used as the basis for preprocessing. After noise removal, supervised regression algorithms were applied: K-Nearest Neighbors (KNN), Multilayer Perceptron (MLP), MLP optimized by Genetic Algorithm (GA), Decision Tree (DT), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). The dataset consisted of 7,920 windows with 700 samples, reduced to 350 through subsampling. The models were evaluated using 10-fold cross-validation, and the performance metrics used were Relative Error percentage (ER%) and the Coefficient of Determination (R²). Experiments were conducted via computational simulation, with training and inference times recorded. KNN achieved the best overall performance, especially for SNR levels above 50 dB. The Wavelet technique performed better under low noise, while the Moving Average Filter was more effective under high noise conditions. The results demonstrate the feasibility of the proposed method, with a focus on low computational cost and high accuracy.