Electricity Theft Detection Based on Load Profile Using XGBoost and SMOTE

Electricity theft is one of the main causes of non- technical losses, leading to significant financial losses for electricity companies. Therefore, an effective method is needed to detect and reduce the potential for theft. Various existing detection approaches, such as Decision Trees, Logistic Regression, and AdaBoost, have been used to detect electricity theft. However, these methods still suffer from several limitations in handling imbalanced and complex data, often resulting in number of errors. Data imbalance presents a major challenge in the classification process, where the number of normal customers greatly exceeds the number of customers committing theft. To address this issue, this study proposes a new approach by integrating XGBoost with the data resampling method SMOTE. SMOTE is utilized to balance the distribution of data between normal customers and those exhibiting anomalous consumption patterns, while XGBoost is selected for its ability to handle complex data and effectively identify suspicious consumption behaviors. Experimental results demonstrate that the combination of XGBoost and SMOTE enhances detection accuracy to 93.63 percent, improves model sensitivity, and reduces detection errors, including false positives and false negatives. This approach significantly strengthens the stability and reliability of the model in detecting electricity consumption anomalies, thereby assisting utility companies in mitigating non-technical losses, improving operational efficiency, and supporting efforts to detect electricity theft more effectively and accurately.