Feature Extraction and Power Quality Disturbances Classification Using Smart Meters Signals

This paper presents a methodology aimed at extracting features to obtain information that will highlight disturbances related to the field of power quality. Due to the concept of smart grids, it is clear that the classification of the disturbances should be undertaken using smart meters, so that a large amount of data corresponding to the voltage and current waveforms are not exchanged using the communication channels, i.e., between smart meter and Utility's database server. Thus, it is necessary to ensure a balance between computational effort (arising from the implementation of algorithms on hardware) and the satisfactory performance of the algorithm for the classification of disturbances. Based on the assumption that the classification task is onerous, this paper proposes a step of feature extraction that may be calculated and analyzed offline using synthetic waveforms/signals, which are subsequently validated using field measurements. It should be noted that this offline analysis is required to determine the most relevant features for the process of classifying each disturbance. However, in order to establish the effectiveness of the feature extraction step, the response of decision trees of the C4.5 type and of artificial neural networks of the multilayer perceptron type were verified during the phase of disturbance classification. In short, good results were obtained that corroborate the hypothesis that the feature extraction step is necessary to classify disturbances effectively and with low computational effort.