Identificação não-intrusiva de cargas similares em smart grid usando rede neural convolucional

Abstract

This work deals with the problem of identifying equipment monitored from a common coupling point, which equipment is technically identical from an electrical point of view, listed here as highly similar. So, experimentally, four fluorescent lamps and four computers are used, none or even all of which can be in simultaneous operation, resulting in two sample banks called Bank A, with about 8 million voltage and current samples required for each of the 16 possible configurations of the lamps, and Bank B, with 999600 voltage and current samples required by each of the 16 possible configurations of the computers. Such samples are acquired at 99960 samples per second, quantized in 16 bits. The objective is to use part of these randomly selected samples and, through manually and empirically configured convolutional neural networks, to train such networks to obtain accuracy compatible with those observed in the literature. An index is proposed to assess network performance. This index considers the number of network parameters and the training time so that the neural network can achieve a reference accuracy. In addition, the robustness of the methodology was verified in the face of variations in the nature of the behavior of the electrical equipment under identification, the decrease in the number of samples and the limitations in the acquisition hardware in order to sample the data in lower frequencies and resolutions. The results show that the networks maintained their performance, even with loads of a variable nature over time, reaching accuracy between 92.45% and 100%. They also show that the reduction in the number of samples has a negative impact on accuracy. However, it becomes significant from a 40 % reduction in the total number used in the network configuration process. Regarding the reduction of the sampling rate, it is possible to verify the non-commitment of the system with rates up to 8 times lower. Finally, decreasing the resolution of the samples causes significant degradation when the resolution is less than 10 bits. Therefore, this work proves that the non-intrusive method is also efficient to identify highly similar loads and shows that the presented methodology is a viable alternative when it is possible to deal with the high cost of identification involved, that is, the ability to obtain, store and process large masses of data in a non-prohibitive time