Implementation and Performance Evaluation of Machine Learning-Based Apriori Algorithm to Detect Non-Technical Losses in Distribution Systems

The emergence and augmentation of nontechnical losses (NTLs) in a power distribution system has always been considered a critical issue in the global electricity market. Being considered as uncharged, unlawfully or unfairly charged consumed electricity, NTLs due to fraudulent activities can impose a serious burden on the grid and revenue loss in the state’s budget. The detection of NTLs based on the analysis of huge consumer dataset using machine learning is an acknowledged approach due to its expedient performance compared to manual inspection. This paper addresses a comprehensive investigation of monthly electricity consumption data of ~15000 consumers over three years using machine learning for NTL detection. The data is acquired by meters having automatic meter reading (AMR) capability and is processed using support vector machine (SVM) classifier, deep neural network (DNN), gradient boosted reinforcement learning (GBRL) and apriori algorithm. The results of ML algorithms are assessed by various performance metrics based on the confusion matrix and compared among themselves as well as with the findings from the published works. The least recall score of 0.8926 is exhibited by SVM classifier and poorest scores of accuracy (0.9376), specificity (0.9451), precision (0.7521), F1 (0.8183) and false positive rate (0.0549) are given by DNN program. With apriori algorithm, the scores of accuracy, recall, specificity, precision, F1 and false positive rate are observed as 0.9964, 0.9927, 0.9971, 0.9843, 0.9885, 0.0029 respectively. In all the performance based six domains, a significant improvement of 6%, 10%, 5%, 23%, 17% and 5% respectively compared to the least scores is demonstrated by the apriori algorithm. Therefore, in all these domains, the apriori algorithm being reported for the first time in this research work outperforms all other methods of this paper as well as the similar published works.