Identificação não intrusiva de cargas por floresta aleatória com otimização fogos de artifício

Abstract

The control of expenses related to electricity has been showing a great growth, especially in residential environments. Monitoring of electrical loads that are turning on and off from a home are often performed using smart-plugs, providing to the consumers information about operation intervals and power consumed by each device. Despite a practical solution to control and reduce electricity costs, it has a high cost due to the amount of meters required. The high cost problem can be worked around by using a non-intrusive load monitoring proposal (NILM), where voltage and current measurements are taken at the home entrance, in counterpart demand a extra processing step. In this extra step, it is necessary to calculate the powers, identification of the occurrence of events and finally, the identification of which equipment was turned on or off. The proposals of this work were to use a new power calculation standard proposed by the IEEE (1459-2010), the elaboration of a heuristic event detector using floating analysis windows to locate stability zones in the power signals after indicating a power variation above a predetermined value, testing the best way to dispose of event identifier data to identify which load has been added or removed from the monitored circuit, and optimization of the parameters of the Random Forest classifier using the fireworks optimization algorithm (FA). The proposed event identifier and classifier tests were performed on the dataset BLUED, which contains data collected at a north american residence over a period of one week. For the classifier tests, four different forms of data entry were used, and subsequently the two forms that obtained the best performances were used in the classifier optimization process. The event identifier results were compared with other publications that used different approaches and obtained satisfactory results. And the results of the classifications were compared to each other, for using different data entry forms, and also as an ideal classifier, where an improvement in the results was also observed when compared with the results of a classifier with commonly used parameters, presenting a larger number of trees used in each RF, but with a limited depth of each tree. And the importance of the variables involved in the classification process were calculated.