Unsupervised Classification for Non-Technical Loss Detection

Abstract

Electricity theft has been a major issue for DSOs for years. However, despite efforts to detect it and the application of legal deterrents, the phenomenon insists. In this paper, unsupervised data mining NTL detection techniques are tested on a smart meter data set provided by the Greek DSO, HEDNO and a publically available smart meter data set. Frauds are simulated and the Twitter breakout detection library is used for extracting features which will be combined with typical features already found in literature. Then, unsupervised classifiers such as rule systems, multi-variate Gaussian distribution (MGD), local outlier factor (LOF), k-means, fuzzy c-means, DBSCAN and SOM are demonstrated. Finally, a number of metrics are calculated for each data set and classifier such as accuracy, detection rate, precision, false positive rate and F <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub> score. In addition, the amount of stolen energy detected and detection delay are proposed as metrics for NTL detection system studies.