A CNN-BiLSTM-XGBoost Framework for Real-Time Electricity Theft Detection and Abnormal Energy Consumption Analysis

Abstract

Electricity theft and abnormal energy consumption are significant problems that lead to high losses and a decrease in the reliability of the power system. However, existing detection methods are insufficient for the complexities of the consumption pattern, time series characteristics, and nonlinear information of the smart-meter data. To address these issues, this paper introduces the hybrid CNN-BiLSTM-XGBoost system that utilizes convolutional neural networks for feature extraction, bidirectional long short-term memory networks for learning the temporal aspects of electricity consumption, and extreme gradient boosting for classification. The proposed model differs from traditional machine learning and single deep learning techniques. According to the experimental results, the hybrid approach reached 98.20% accuracy, 97.60% precision, 97.10% recall, 97.35% F1-score, and 99.10% ROC-AUC. However, actual experiments should be conducted to justify the correctness of the obtained results.