Reconhecimento não-intrusivo de equipamentos elétricos empregando projeção vetorial

Abstract

Electricity consumption in homes and workplaces has been growing steadily over the decades and attitudes to reduce these costs should be taken. An interesting solution is to provide to electricity users, and also to the energy company, detailed data of individual consumption of each electrical appliance. To accomplish this, researchers in the field have focused their efforts on non-intrusive methods of load identification, where a single energy meter is able to desagreggate the appliances by monitoring the total consumption of electricity of that location. Non-intrusive methods are easy to install and demand little maintenance, but require a robust method for identifying these loads. Therefore, the aim of this work is to investigate nonintrusive methods of recognition of electrical appliances to find the desaggregated consumption of these loads. Among these methods, there are the already widely used image recognition pattern methods, that now are been used also to detect electrical devices. In this paper, two of these techniques are discussed, the Principal Component Analisys, a classical method in the literature, and the Vector Projection Length, a completely new method and never used in the loads recognition field before. Current and voltage data were collected from 16 residential appliances, involving all types of loads (resistive, inductive, electronic and hybrid/other types). These data were used as training samples and test samples (unknown samples). A study is carried out using the current and also the power, independently, as load signatures. Also, a comparative analysis of the results of signatures in the time domain and time-frequency (Stowkwell transform) is conducted. As the main contributions to this work, we verified that the Vector Projection Length for load identification is quite feasible, with results up to 96% of tested appliances being identified. However, the results with Principal Component Analisys did not presented the same performance, reaching only 81% of accuracy rate. Comparing the signatures, it became clear that one should use the current in the time-frequency domain for better performance. Neither the use of power, or the time domain obtained satisfactory results of load identification when applying image pattern recognition techniques to load recognition.