ANN-Based Electricity Theft Classification Technique for Limited Data Distribution Systems

Electricity theft has been a challenge for distribution systems over the years. Theft presents a massive cost to the system operators and other issues such as transformer overloading, line loading, etc. It has become crucial for measures to be implemented to combat illegal electricity consumption. This work sought to develop an artificial neural network-based electricity theft classifier for distribution systems with limited data, i.e., systems that can only provide consumption data alone and no auxiliary data. First, a novel data pre-processing method was proposed for the systems with consumption data only. Again, synthetic minority oversampling is employed to deal with the unbalance problem in the theft detection dataset. Afterwards, an artificial neural network (ANN)-based classifier was proposed to classify customers as normal or fraudulent. The proposed method was tested on actual electricity theft data from the Electricity Company of Ghana (ECG) and its performance compared to random forest (RF) and logistic regression (LR) classifiers. The proposed ANN-based classifier performed exceptionally by producing the best results over RF and LR regarding precision, recall, F1-score, and accuracy of 99.49%, 100%, 99.75%, and 99.74%, respectively.