Robust Clustering and Anomaly Detection of User Electricity Consumption Behavior Based on Correntropy

Abstract

Anomaly detection in power systems is crucial for ensuring the safety and stability of electrical grids. Traditional methods struggle to extract meaningful features from electricity consumption data due to significant differences in usage patterns across various user types, such as residential and industrial users. Applying a single model for all user categories increases feature complexity and computational demands. Additionally, non-Gaussian outliers caused by equipment and measurement noise can significantly deviate from normal data patterns, making them difficult to filter using standard methods. To address these challenges, this paper proposes a robust, user-type-specific anomaly detection method. After data preprocessing, a correntropy-based K-means clustering method is used to separate users with noisy data. A two-stage detection framework combining fuzzy logic and a convolutional neural network (CNN)-long short-term memory (LSTM) model enhances both detection efficiency and accuracy. The experiments were conducted using open-source datasets, and the results demonstrated that our method achieved an accuracy of 95%, which is approximately 4% higher than the traditional Isolation Forest method. This indicates that our approach effectively balances efficiency and accuracy in anomaly detection, with its generalizability further validated on an additional dataset.