Anomaly detection in electric power metering based on data mining technology

The application of intelligent algorithms in power system data analysis continues to deepen, providing new technical pathways for metering anomaly identification. Addressing the limited recognition accuracy of traditional detection models under complex time-series features and noise interference, this study develops a data-mining-based anomaly detection model for power metering. The model employs an autoencoder network as its core structure and incorporates a probability-density calibration mechanism to identify anomalous samples through joint discrimination of reconstruction error and density distribution. At the feature level, it integrates time-domain statistical indicators with frequency-domain energy features to enhance the representation of high-dimensional data. Building upon the Autoencoder architecture, the model refines network depth and activation functions and introduces Gaussian Kernel Density Estimation (KDE) for probability calibration. Experiments on a State Grid metering dataset comprising 5.2×106 samples compare the proposed approach with Isolation Forest, One-Class Support Vector Machine (OC-SVM), and a baseline Autoencoder model. Results show that it achieves an accuracy of 96.4% and an F1-score of 0.94, improving by approximately 1.7% and 1.6% over the baseline Autoencoder. The model also demonstrates real-time inference efficiency, reducing average processing time by about 26% compared with Support Vector Machine (SVM), confirming its effectiveness and engineering practicality in power-metering anomaly detection.