Electricity Theft Detection via Self-Attention Feature Encoding and GSA-Optimized XGBoost

To address the issues of difficult implementation and low accuracy in current electricity theft detection methods, this paper proposes an electricity theft detection method that integrates self-attention encoding with XGBoost-GSA. Firstly, it conducts periodic fluctuation analysis on the data and screens the daily electricity consumption data using the coefficient of variation. Then, it employs a Transformer-based self-attention autoencoder for feature extraction from the screened daily electricity consumption data. Subsequently, the gravitational search algorithm (GSA) is utilized to optimize some hyperparameters of XGBoost. The extracted features are fed as input into XGBoost to achieve the detection of electricity theft behavior. In the experiments, the proposed method achieved the best performance across all indicators.