Novelty detection and multi-class classification in power distribution voltage waveforms

• Accurate classification of events in waveforms from electrical distribution networks.
• Novelty detection: dynamic identification of new classes of events.
• SVDD using negative examples and maximal margin separation: better generalization.
• Experiments using real data: significant improvements in classification accuracy.
• Direct application as part of tools to assist mitigation processes in power utilities.

The automatic analysis of electrical waveforms is a recurring subject in the power system sector worldwide. In this sense, the idea of this paper is to present an original approach for automatic classification of voltage waveforms in electrical distribution networks. It includes both the classification of the waveforms in multiple known classes, and the detection of new waveforms (novelties) that are not available during the training stage. The classification method, based on the Support Vector Data Description (SVDD), has a suitable formulation for this task, because it is capable of fitting a model on a relatively small set of examples, which may also include negative examples (patterns from other known classes or even novelties), with maximal margin separation. The results obtained on both simulated and real world data demonstrate the ability of the method to identify novelties and to classify known examples correctly. The method finds application in the mitigation process of emergencies normally performed by power utilities’ maintenance and protection engineers, which requires fast and accurate event cause identification.