Towards Detecting Non-Technical Losses in Electrical Power Systems with Machine Learning

In this paper, Machine Learning (ML) techniques are applied to address the critical challenges of non-technical losses (NTLs) in electrical power systems. The NTL study focuses on a comprehensive dataset obtained from the Hellenic Electricity Distribution Network Operator (HEDNO) in Greece, mainly collected from traditional metering systems. A careful preprocessing and feature engineering framework is implemented to prepare the (imbalanced) data for ML analysis. Several ML models are examined including Multinomial Naive Bayes, Logistic Regression, Decision Trees, Random Forest and advanced ensemble techniques like XGBoost and CatBoost. The obtained results demonstrate that ensemble models, particularly CatBoost, perform best in detecting NTLs, showing high effectiveness in differentiating between normal and fraudulent consumption patterns, leading to improved detection and prevention of NTLs in power distribution networks.