Segmentação, classificação e detecção de novas classes de eventos em oscilografias de redes de distribuição de energia elétrica

Abstract

This work presents new approaches for two of the fundamental steps in automatic waveform analysis in electrical distribution systems: transient time detection and its classification. Two datasets were used to compare and validate the proposed methods. The first is composed by simulated waveforms, by using the Alternative Transient Program, while the second is formed by real data from a monitoring system developed for overhead distribution power lines. The real data present a set of relevant events for the analysis proposed here, mainly due to the variety of events, including lightning-related transients. Regarding transient detection (waveform segmentation), the experiments involve usual segmentation methods, such as Kalman filtering, standard Discrete Wavelet Transform, and autoregressive models, besides two new techniques based on the Teager Energy Operator and Support Vector Data Description. The results obtained on both simulated and real world data demonstrate that the method based on Support Vector Data Description outperforms other methods in the transient identification task. Regarding the automatic waveform classification, a new approach including the detection of classes not defined in the training stage (called novelties) is presented. Also, the classifier is able to discriminate among multiple known classes, normally defined as multi-class classification. Two different approaches are compared, by using multi-class classification and novelty detection in two subsequent stages and in a simultaneous way. The following classifiers were assessed: X-Means, K-Nearest-Neighbors, and Support Vector Data Description with different formulations, besides the Support Vector Machine for multi-class classification. Furthermore, a technique for the post-processing of the novelties is presented, in order to provide some useful information for the experts, regarding possible similarities in the novelty set. To accomplish this task, automatic clustering methods were used. The final results, especially for the dataset with real examples, show that it is possible to obtain a relevant classification performance (above 80%) in each one of the three stages of the classification process: multi-class classification, novelty detection, and the post-processing applied to the novelties (automatic clustering).