Classificação para o monitoramento não-intrusivo de cargas em sistema embarcado com rede neural convolucional

Abstract

Electric power is very important for the economic development of all countries, and its consumption has been growing at an impressive rate, doing so faster than other types of power. Along with the increase in consumption, there are also concerns about environmental sustainability; after all, ensuring access to electric power in a reliable, sustainable, modern, and affordable way for all is one of the objectives of the Sustainable Development Goals, an agenda proposed by the United Nations (UN). In addition to encouraging the usage of renewable and less environmentally impactful energy, there are also concerns to create increasingly more power efficient devices, and to reduce the waste of electric power by seeking alternatives for a more efficient use of it. The active involvement of consumers results, for the most part, in a more efficient use of electric power, which increases interest in the development of technologies that make them aware of their habits. Studies show that, the greater the detail of information about electrical power consumption, the greater the amount of electric power saved by consumers. One of the most used techniques to analyze such details is Non-Intrusive Load Monitoring (NILM), who, by disaggregating the loads, distinguishes between each of the appliances and explores the electric power consumption of each one individually. Therefore, in order to contribute to this technique, and considering the ever-growing progress in the electronic and machine-learning areas, this study proposes a set of training strategies using a deep-learning method for load classification in an embedded system, and therefore, contribute to a more efficient use of electric power. Based on literature and experiments, we adopted the binary image of the voltage-current as the distinguishing feature, as it obtained the best results. In order to classify the devices, we used said images as input to the Convolutional Neural Network (CNN), which was chosen after obtaining the best results in the tests that were performed. After we used the leave-one-out cross-validation method, our CNN model was evaluated using the PLAID dataset and obtained an F-Score macro-average of 74.76% for PLAID1, 56.48% for PLAID2, and 73.97% for PLAID1+2, and those results were very close to literature. The novelty of this study is the quantization of the CNN model using TensorFlow Lite, and its application in a resource-constrained embedded system (ESP32). The accuracy rate achieved in testes performed with all data from the PLAID1+2 dataset was 98.55%, which shows that the embedded device can be used to perform the load classification with a high accuracy rate.