Classificador não invasivo de cargas elétricas residenciais com acionamento simultâneo

• Autor(a) principal: Lima, Rodrigo Botelho de
• Data de Publicação: 2018
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFLA
• Texto Completo: http://repositorio.ufla.br/jspui/handle/1/31740
• topic: Monitoramento não invasivo; Identificação de cargas elétricas; Estatística de ordem superior; Florestas aleatórias; Redes neurais artificiais; Non-invasive monitoring; Electrical loads identification; Upper order statistics; Random forests; Artificial neural networks; Engenharia de Software

Thinking about the future of the world's, especially about energy resources, scientists are seeking alternatives that allow management of these resources. This concern increases with the high consumption of electric energy by the people. Beside of this, in January 2018, electrical energy consumers were able to choose between staying in the normal tariff or switching to a white tariff, but to do this choice the needs to know their profile consume. This study presents a proposal to classify individually driven loads or even when another one is already in operation. For this a dataset was set up with five different types of loads that make up five different classes. Five classes are also assembled which use the triggering of one load while another is in operation. As a classifier, machine learning algorithms were implemented using: Artificial Neural Networks, Vector Support Machine and Random Forests. After developed the model, the results of the collected metrics, showed that the average accuracy for the Classifier based on Vector Support Machine was 99.8%, the average accuracy was 99.31%, and the sensitivity was 99.8%. The Classifier based on Artificial Neural Networks reached the average accuracy of 98.85%, the average precision of 98.82% and the average sensitivity of 98.5%. As for the classifier based on Random Forests, the classifier reached the average values for accuracy of 98.95%, 98.81% for precision and sensitivity of 98.8%. For all classifiers the mean value reached for the F1 score metric was 0.99, this result analysis showed that the models performed very well. Thus, as observed, the SVM proved to be better in all metrics. The SVM still presented better performance in training time, suggesting a lower computational effort and hence a lower cost to produce. Another evaluation to be carried out is that the Higher Order Statistics also showed to be efficient in the extraction of parameters in this work.