Abnormal Electricity Detection of Users Based on Improved Canopy-Kmeans and Isolation Forest Algorithms

Aiming at the existing user abnormal electricity consumption detection methods that have the problem of difficult classification of user similar electricity consumption patterns, this paper proposes an unsupervised isolation forest abnormal electricity consumption detection model based on the Canopy-Kmeans algorithm with weighted density improvement. To start, we propose a composite parameter analysis method for user electricity consumption patterns, volatility, trends, and correlations using Irish smart meter data. This method involves joint data cleaning, interpolation, and feature construction. Additionally, principal component analysis is introduced to fuse features across layers and reduce dimensionality in user electricity consumption. Subsequently, we introduce the weighted density improvement Canopy-Kmeans clustering algorithm. This algorithm determines the K value and clustering centers using the maximum weight product method, based on definitions of sample density, average intra-class sample distance, and inter-class distance in the multilayer fusion feature data. Finally, we propose a fusion mechanism of weighted density improvement Canopy-Kmeans and isolation forest algorithms to jointly construct a model for detecting abnormal power usage based on multilayer fusion feature data analysis. The results demonstrate that multilayer fusion feature parameters vary in size and discretization among different user types, enabling classification of users with diverse electricity consumption patterns. Moreover, the anomaly detection model based on multilayer fusion feature data analysis improves accuracy rates, recall rates, and F1 scores compared to other algorithms.