Classificação de eventos em monitoramento NILM de cargas elétricas residenciais utilizando rede neural convolucional

• Autor(a) principal: Medeiros, Aérton Pedra
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Manancial - Repositório Digital da UFSM
• Texto Completo: http://repositorio.ufsm.br/handle/1/19725
• topic: CNN; Monitoramento não intrusivo; NILM; Rede neural convolucional; CNN; Non-intrusive load monitoring; NILM; Convolutional neural network; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA

In the context of the expected advances with the implementation of Smart Grids, enabling the modernization of the electricity billing system and making the management of residential electric power systems attractive, the detailed knowledge of the residential consumption profile gains importance as an element to subsidize the decision making by Home Energy Management Systems. One of the techniques for detailing the residential electrical consumption profile is Non-Intrusive Load Monitoring (NILM), a low cost installation technique that presents its complexity in the development of the disaggregation algorithm. Given this challenge, this dissertation presents a methodology based on low frequency sampling electrical monitoring to perform the extraction of characteristics of the activated and deactivated electric loads during the operation of the electric network. Through these characteristics, an evaluation of the event classification performance is performed using the Convolutional Neural Network - CNN, a type of artificial intelligence specially used for visual pattern recognition. To evaluate the developed method, a case study is performed using monitoring data from the United Kingdom Domestic Appliance Level Electricity database - UK-DALE, a data source widely used in NILM surveys. The performance of the developed method is evaluated using the Precision, Recall, F1-Score and Accuracy metrics as well as the confusion matrix to present the classification errors. To compare the classification performance obtained by the developed method is also modeled classification method called Decision Tree. Through the performance analysis of the developed method it is observed that it presents some restriction to deal with behaviors not foreseen in the training phase, but presents the ability to learn new behaviors through new training phases. Also, it presented good performance in the classification of low power events, when compared to the Decision Tree classification method. In conclusion, the use of convolutional neural network presents positive performance to be used in the event classification in NILM applications.