Detecção de fraude em unidades consumidoras não telemedidas com uso de técnicas de aprendizado de máquina

• Ano de defesa: 2020
• Autor(a) principal: Paulo, Fernanda Rodrigues
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Engenharia Elétrica; Programa de Pós-Graduação em Engenharia Elétrica; UFPB
• Palavras-chave em Português: Perdas comerciais; Perdas não técnicas; Fraude de energia; Classificação de padrões; Aprendizado de máquina; Inteligência Artificial; Detecção de fraude; Recuperação de consumo; Commercial losses; Non-technical losses; Energy fraud; Pattern classification; Machine learning; Artificial intelligence; Fraud detection; Consumption recovery; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/18759

In 2018, estimates that about 310 TWh were destined to supply irregular connections and measurements in Brazil, approximately R$ 9 billion losses for distributors. The concessionaire of this study faces challenges to detect fraud, mainly due to the volume of data and the limitation on finding patterns without a structured tool. Considering this scenario, the development of an automated methodology is proposed to detect fraud in low voltage customers, without telemetry, using artificial intelligence tools. Information was extracted from the company's database, attributes were implemented, the main variables were selected and then the models were evaluated. The main variable proposed compares the average consumption of the unit with the closest geographic neighbors with similar size characteristics. Variables are also proposed aiming to detect the moment of a reduction in the energy consumption, as well as its value. The most common Machine Learning techniques were tested and four models were proposed: Support Vector Machine was used for consumers with an indication of possible fraud; for residential units without this indication, Gradient Boosting was used; for rural units, Random Forest was used; for the other classes, a Multilayer Perceptron Neural Network was used. The models were qualified based on a new metric, proposed as an alternative to the usual evaluation metrics, which computes the percentage of the energy benefit theoretically recovered by the model in relation to all the energy that could have been recovered. In theoretical tests, it was possible to obtain an accuracy of 39.4%, surpassing 19.5% the current methodology of the company, with 69.8% greater recall. The energy benefit metric also shows that the proposed methodology was able to recover 59.5% of the total amount of energy available, 153.2% higher than the company's current model. New research involves the application of the proposed methodology to the company's base for the classification of the consumers and inspections will be sent to verify the results.