Improving Electricity Theft Detection with a Stacked Ensemble Model

Abstract

Energy theft is a major problem for utility providers as it results in huge  revenue loss and interruption in power supply. It is shown that traditional detection  models cannot adequately cope with different methods of theft and unequal data distributions. This work provided a novel machine learning approach to use a stack ensemble of Random forest, Gradient boost with XGBoost, and Logistic regression as meta classifiers. To solve the problem related to data imbalance, the ADASYN oversampling method is used to increase the number of samples of the minority class. The results are then measured with Recall, FNR, ROC AUC in order to see that the model works effectively to detect theft with good levels of TPR without crossing FDR and FNDR levels. This work provides a useful application to various utility companies designed to identify fluctuations in energy consumption, protect power distribution networks, and minimize energy losses.