New Trends in Power Quality Event Analysis: Novelty Detection and Unsupervised Classification
- PDF / 831,220 Bytes
- 10 Pages / 595.276 x 790.866 pts Page_size
- 52 Downloads / 147 Views
New Trends in Power Quality Event Analysis: Novelty Detection and Unsupervised Classification André Eugenio Lazzaretti1 · Vitor Hugo Ferreira2 · Hugo Vieira Neto1
Received: 29 January 2016 / Revised: 24 May 2016 / Accepted: 2 August 2016 © Brazilian Society for Automatics–SBA 2016
Abstract A new method for automatic event–cause classification in power distribution networks for the detection and clustering of previously unknown classes of transient voltage waveforms is presented. The approach performs the detection of novelties—events that are not present during modeling of the classifier—in addition to the classification of known events, using a formulation based on support vector data description. Additionally, an unsupervised clustering method for novelties is proposed, in order to collect relevant information about their features and allow identification of new classes of events, which constitutes the main contribution of this work. Two different automatic clustering methods are compared: X-Means clustering and Rival Penalized Expectation Maximization. Experiments using both simulated and real data for the entire classification process, which includes multi-class classification with novelty detection and identification of new classes, are presented. The results obtained demonstrate that the proposed method fully agrees with current trends in smart distribution networks, in which automatic identification, characterization, and mitigation of events are critical for network operation and maintenance. Keywords Automatic clustering · New class identification · Novelty detection · Waveform classification
B
André Eugenio Lazzaretti [email protected]
1
Federal University of Technology – Paraná (UTFPR), Avenida Sete de Setembro 3165, Curitiba, PR 80230-901, Brazil
2
Federal University Fluminense (UFF), Rua Passo da Pátria 156, Niterói, RJ 24210-240, Brazil
1 Introduction A common feature in most classifiers already developed for automatic waveform analysis in electric power systems is the use of supervised learning for multi-class classification models, which is used to discriminate among different power quality (PQ) events, such as flickers, sags, swells, harmonics, (Mahela et al. 2015; Ferreira et al. 2015; De Yong et al. 2015; Biswal et al. 2014). In power distribution systems, these PQ disturbances are normally related to a specific cause, or a set of subsequent causes (Koziy et al. 2013), e.g., short circuit, lightning and switching events (Mohanty et al. 2008; Lazzaretti et al. 2013, 2015). In this sense, one can say that as important as automatic characterization of the effects of PQ phenomena in electrical networks is the automatic identification of causes that are responsible for waveform variations (Dahal et al. 2015), especially for power utilities (Ray et al. 2012; Lidula and Rajapakse 2012). Automatic event–cause classification is the general objective of this research. However, the possibility of occurrence of previously unseen events, which cannot be included a priori in the training phase o
Data Loading...
