Improved Extremely Random Tree Based Low Voltage User Electricity Theft and Breach of Contract Identification Technology

Electricity theft, electricity default (including over-capacity electricity consumption and unauthorized power transfer), and metering device anomalies cause significant line losses and economic losses to power companies. This study proposes an intelligent method for detecting electricity theft, electricity default, and metering anomalies in low-voltage distribution networks. It fully utilizes the advantages of the fast training and generalization capabilities of the Extra-Trees algorithm, and integrates the RandomForest, AdaBoost, GradientBoosting, and SVC algorithms to make up for its shortcomings. This method adopts a two-layer ensemble model: the first layer uses the above five algorithms to model, fit, and predict user behavior and metering data; the second layer uses the prediction results of the first layer model as features, combined with the test set labels, and uses XGBClassifier as the base classifier to further optimize the detection performance. Experimental results based on a dataset with complex user behavior and metering challenges in Meizhou, China show that the detection accuracy of the system for electricity default exceeds 87%, significantly improving line loss management and grid reliability.