Anomaly Detection of Non-Technical Losses in Smart Meter Data Using K-Means Clustering and Isolation Forest: A Case Study From Indonesia

Non-technical losses (NTL) in electricity distribution, often caused by theft or faulty metering, represent a significant challenge for utility companies worldwide. With the growing deployment of smart meter data systems, large-scale voltage and current data provide new opportunities for intelligent anomaly detection. This study investigates using two unsupervised machine learning techniques such as K-Means Clustering and Isolation Forest to detect NTL anomalies in smart meter data from Indonesia's National Electricity Utility customers. The research focuses on customers with current transformer (CT)-based meters, analyzing datasets spanning from 2021 to 2023. The detection framework was implemented using the KNIME Analytics Platform, enabling efficient data preprocessing, model tuning, and evaluation. Simulation results show that K-Means achieved recall up to 98.53%, particularly effective in detecting voltage anomalies. Isolation Forest, while initially underperforming, improved significantly after parameter tuning, reaching up to 91.54 % recall. From a power system perspective, the models successfully identified abnormal patterns such drop voltage, over voltage and unusual current flows indicative of potential meter tampering or load bypass. These findings highlight the potential of machine learning-based anomaly detection to support utility operations in reducing energy losses and improving billing accuracy.