Classificadores não intrusivos de cargas elétricas industriais utilizando técnicas de inteligência computacional

• Autor(a) principal: Damasceno, Douglas Roberto Fernandes
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFLA
• Texto Completo: http://repositorio.ufla.br/jspui/handle/1/56777
• topic: Monitoramento de carga não intrusivo; Perceptron de multicamadas; Máquinas de vetores de suporte; Algoritmos de otimização; Métodos de clusterização Fuzzy; Non-intrusive load monitoring; Multilayer perceptron; Support vector machines; Optimization algorithms; Fuzzy clustering methods; Sistemas Elétricos de Potência

Improving energy management has required performing fundamental tasks such as monitoring electrical loads, due to the current economic situation and growing ecological trends. This work presents a method of identification and classification of five industrial loads of a automotive shock absorber production line, namely: a valve press, an oil doser, a traction test, a dynamometer and a roller. In order to collect the training data of the proposed classifiers, the loads were triggered individually and the electrical current signal data were obtained through the Non-Intrusive Load Monitoring technique. As classification methods, the following machine learning algorithms were implemented: Artificial Neural Networks (ANN) of the Multilayer Perceptron (MLP), Support Vector Machines (SVM) and also the fuzzy clustering methods K-Means (KM), Fuzzy C-Means (FCM) and Gustafson-Kessel (GK). In order to obtain the main parameters of the MLP and SVMs, three optimization techniques were applied, namely Particle Swarm Optimization (PSO), Differential Evolution (DE) and the Gray Wolf Optimizer (GWO). As for the clustering methods, to determine the efficient number of clusters, the validation indices Xie-Beni Criterion (XB), Classification Entropy (CE), Partition Index (SC) and Dunn Index (DI) for each proposed method. The best classifier obtained, comparing the MLP classifiers and the SVMs, was the MLPPSO, which presented among the main performance metrics a precision of 0.9556, F1-score of 0.9478, accuracy of 0.9474 and the Kappa coefficient of 0.9345 demonstrating the effectiveness of the classifier. Regarding the clustering methods, the GK stood out, which presented precision of 0.8472, accuracy 0.8378, F1-score 0.8398 and Kappa coefficient 0.7991, these values being lower than expected, and therefore not being applicable for classification of loads.