Anomaly Detection Model for Electricity Consumption Based on Adaptive Threshold Estimation and Ensemble LSTM Autoencoders

The promotion and application of smart meters provide massive basic data support for abnormal electricity consumption detection for power users, but they also face challenges such as a shortage of electricity theft samples in traditional abnormal user detection models and difficulties in real-time abnormal electricity consumption monitoring. Therefore, this article proposes an ensemble long short-term memory autoencoder power user abnormal electricity consumption detection model based on adaptive threshold estimation. Firstly, a long short-term memory autoencoder model is used to learn the normal electricity consumption patterns of regular power users. The Bootstrap sampling method is employed to diversify individual learners, thereby increasing the learning diversity of the model for normal behavior. Then, the electricity consumption data of the power user is input into the ensemble long short-term memory autoencoder for reconstruction, and during the testing phase, the reconstruction error is used to determine whether the user being detected is abnormal. Secondly, to determine the optimal error threshold without prior knowledge, the adaptive selection of the theft threshold is achieved through kernel density estimation. Users whose reconstruction error exceeds the upper bound of kernel density estimation are identified as abnormal users. Finally, simulation examples based on actual data were constructed for verification. The experimental results showed that the performance indicators of the proposed abnormal electricity consumption detection model were superior to those of traditional autoencoders, with an accuracy increase from 82.98% to 95.74%, proving the superiority of the proposed model.