<![CDATA[Electricity theft is a primary source of non-technical losses (NTL) in power distribution systems, causing substantial economic damage to utilities. This study develops a machine learning approach to detect electricity theft using real smart meter data from the State Grid Corporation of China (SGCC), comprising 42,372 consumers over 1,034 days with an 8.53% theft rate. Nineteen statistical features were engineered from the daily consumption profiles after handling 25.64% missing values through linear interpolation and outliers through the three-sigma rule. Four classifiers, namely Logistic Regression, Decision Tree, Random Forest, and histogram-based Gradient Boosting, were trained with the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance and evaluated on a stratified hold-out set. The Gradient Boosting model achieved the best performance with an ROC-AUC of 0.750, an F1-score of 0.323, and a recall of 0.416, outperforming the other classifiers, while cross-validation confirmed the stability of the results. Feature importance analysis revealed that the missing-data ratio, the low-consumption ratio, and the longest zero-consumption run are the most discriminative indicators of theft. The results show that statistical feature engineering combined with ensemble learning provides an effective and computationally efficient method for electricity theft detection in smart grids.]]>
Copyrights © 2026
