Anomaly Detection in AMR Load Profiles Using Artificial Intelligence

Non-technical energy losses due to electricity theft and meter errors significantly impact utility efficiency. Existing anomaly detection methods in Automatic Meter Reading (AMR) lack speed and accuracy. This paper evaluates three AI methods-Isolation Forest, One-Class SVM, and Autoencoder Neural Networks-to detect anomalies in AMR load profiles from major customers of PLN Banjarmasin. Performance metrics indicate Autoencoder Neural Networks achieved the highest accuracy (92.90 %) and F1-Score (94.26 %), making it the most suitable method for practical deployment. Isolation Forest, with perfect precision (100 %), is recommended when minimizing missed anomalies is critical. This study highlights AI's potential in effectively reducing non-technical losses.