Identificação de perdas não técnicas em sistemas de distribuição agregando dados exógenos e inteligência artificial

• Ano de defesa: 2022
• Autor(a) principal: Capeletti, Marcelo Bruno
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; Engenharia Elétrica; UFSM; Programa de Pós-Graduação em Engenharia Elétrica; Centro de Tecnologia
• Palavras-chave em Português: Perda não técnica; Distribuição de energia elétrica; Mineração de dados; Aprendizado de máquina; Dados exógenos; Redes neurais artificiais; Exogenous data; Non-technical loss; Electricity distribution; Data mining; Machine learning; Artificial neural networks; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: http://repositorio.ufsm.br/handle/1/27059

Identifying of Non-Technical Losses (NTL) is one of the most challenging tasks in power distribution systems due to its technical difficulty, adversities encountered in locus and cost of moving units for verification in the field. It is found in large quantity in developing countries such as India and Brazil and generates financial losses in order of billions for consumers and actors of electrical power systems. Make use of exogenous data in the identification of NTL in energy distribution systems is recognize a field with great added potential, since nowadays when data are increasingly available in large quantities, such data must be analyzed and applied in a concise way, so that intelligent machine learning systems are able to identify and generate possible NTLs targets, moving inspection teams to consumer units with a higher probability of NTL being identified. This work proposes a complete work of data oriented NTL identification, employing a database of electricity distribution utility and an exogenous (climatic) base together with a model of artificial neural networks, performing a supervised classification model. The dataset is characterized by the unbalanced database of power distribution companies labeled according to previous inspections. The classification by a multilayer perceptron neural network optimized through Bayesian optimization, obtained an overall accuracy of 72,20%, receiver operating characteristic area under the curve with a score of 0.684 and an irregular UC success rate of 43,66% in the test bench labeled by previously inspected units. It is emphasized that the methodology was compared without the use of exogenous data, obtaining an improvement of 6,26% in recall rate of irregular units. Therefore, it is concluded that the model is significantly improved when temperature information is incorporated as an input variable.