Identificação de eletrodomésticos com base no consumo de corrente elétrica utilizando aprendizagem profunda

Abstract

Electric energy consumption in the world has been changing progressively every year. Energy concessionaires have a recurring concern with meeting the increasingly growing demand, trying to avoid rationing, the use of thermoelectric plants, or even the popularly called “blackouts”. Such events are also reflected in concern on the part of the end consumer, since energy costs are high in these scenarios, and, in a worst case, services are interrupted. This work aims to study a practical implementation to bring information about which household appliances consume the most energy within a residential environment, through artificial intelligence. Another proposal is to also be able to offer information about the value of the electricity bill, even before being informed by the concessionaire. Together, this data can objectively provide control over more conscientious consumption by the population. For this implementation to be carried out, Machine Learning techniques were used, such as Artificial Neural Networks, to analyze the electrical current of household appliances. The feasibility of this implementation was validated according to the hypothesis of a characteristic in the current consumption of the devices, in which, due to their specific construction and internal circuitry, they have a pattern, peculiar to each one, in the waveform of the current consumed, which generates a kind of spectral “signature”, when analyzed in the frequency domain. Experiments and data collection were carried out using specific laboratory equipment connected to common household appliances, with the aim of training a neural network, so that it can, based on the current consumed, provide information about which appliance is connected to the energy at a given time. Through the experiments, the best configuration for processing the data was selected, as well as the architecture of the neural network used. The result obtained was satisfactory for most of the household appliances tested, in which it was possible to predict with good accuracy which equipment was analyzed.