Estudo de algoritmos para classificação de séries temporais: uma aplicação em qualidade de energia elétrica

• Ano de defesa: 2007
• Autor(a) principal: MORAIS, Jefferson Magalhães de
• Orientador(a): KLAUTAU JÚNIOR, Aldebaro Barreto da Rocha
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Pará
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação
• Departamento: Instituto de Ciências Exatas e Naturais
• País: Brasil
• Palavras-chave em Português: Algoritmos de computador; Linha de transmissão (Energia); Energia elétrica; Controle de qualidade
• Área do conhecimento CNPq: CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO::ANALISE DE ALGORITMOS E COMPLEXIDADE DE COMPUTACAO
• Link de acesso: http://www.repositorio.ufpa.br:8080/jspui/handle/2011/1700

It concerns automatic classification of short circuits in transmission lines. Most trans-mission systems use three phases: A, B and C. Hence, a short-circuit between phases A and B will be identified as AB". Considering the possibility of a short-circuit to ground" (T), the task is to classify a time series into one among eleven possibilities: AT, BT, CT, AB, AC, BC, ABC, ABT, ACT, BCT, ABCT. These faults are responsible for the majority of the distur-bances in electric power systems. Each short circuit is represented by a sequence (time-series) and both online (for each short segment) and offline (taking in account the whole sequence) classification are investigated. To circumvent the current lack of labeled data, the Alternative Transient Program (ATP) simulator is used to create a public comprehensive labeled dataset. Some works in the literature fail to distinguish between ABC and ABCT faults. Then, results differentiated these two faults types adopting preprocessing techniques, different front ends (e.g., wavelets) and learning algorithms (e.g., decision trees and neural networks) are apresented. The computational cost of the some classifiers during the test stage is investigated and the choosing parameters of classiffers is done by automatic model selection. The results indicate that decision trees and neural networks outperform the other methods.