A Dissimilarity Learning Approach by Evolutionary Computation for Faults Recognition in Smart Grids

Abstract

In a modern power grid known also as a Smart Grid (SG) its of paramount importance detecting a fault status both from the electricity operator and consumer feedback. The modern SG systems are equipped with Smart Sensors scattered within the real-world power distribution lines that are able to take a fine-grain picture of the actual power grid status gathering a huge amount of heterogeneous data. The Computational Intelligence paradigm has proven to be a useful approach in pattern recognition and analysis in facing problems related to SG. The present work deals with the challenging task of synthesizing a recognition model that learns from heterogeneous information that relates to environmental and physical grid variables collected by the Smart Sensors on MV feeders in the real-world SG that supplies the entire city of Rome, Italy. The recognition of faults is addressed by a combined approach of a multiple weighted Dissimilarity Measure, designed to cope with mixed data types like numerical data, Time Series and categorical data, and a One-Class Classification technique. For Categorical data the Semantic Distance (SD) is proposed, capable of grasping semantical information from clustered data. The faults model is obtained by a clustering algorithm (k-means) with a suitable initialization procedure capable to estimate the number of clusters k. A suited evolutionary algorithm has been designed to learn from the optimal weights of the Dissimilarity Measure defining a suitable performance measure computed by means of a cross-validation approach. In the present work a crisp classification rule on unseen test patterns is studied together with a soft decision mechanism based on a fuzzy membership function. Moreover a favorable discrimination performance between faults and standard working condition of the (One-Class) classifier will be presented comparing the SD with the well-known Simple Matching (SM) Distance for categorical data.