Performance evaluation of selected machine learning techniques in the detection of non-technical losses in the distribution system

Electricity is an essential source in acquiring industrial and economic development in South Africa. Power distribution systems face daily challenges in tracing and estimating technical and non-technical losses. Non-technical losses (NTLs) like energy theft, poor meter readings and inadequate payments lead to anomalous spending and patterns. This work uses the buying data to assess the efficacy of Decision Trees (DT), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Logistic Regression, Artificial Neural Networks, and K-Nearest Neighbors in predicting NTLs using a South African dataset. Comparatively, all the tree-based models (DT, RF, and XGBoost) achieved perfect scores across all evaluation metrics in classifying honest and dishonest customers.