•  
  •  
 

Keywords

distribution network; trimmed window; principal component analysis; K-means; INNE; anomaly detection

Abstract

A large amount of anomaly data of voltage can be generated during the transformer district operation of distribution network. These anomaly data cannot correctly reflect the operating conditions of the distribution network and can seriously affect the analysis of the voltage characteristics of the transformer district. Therefore, anomaly detection of the voltage operation data in the transformer district is of great significance. Due to the problem of low accuracy of traditional anomaly detection methods, an algorithm, combining trimmed window feature extraction clustering (TFEC) and isolation-based anomaly detection using nearest-neighbor ensembles (INNE), is proposed. The daily features of the initial data are first extracted by using the trimmed window of TFEC. Then, based on principal component analysis (PCA) and K-means, the daily feature data are downgraded and clustered to obtain operation data clusters of initial voltage based on multiple daily fluctuation types. Moreover, the INNE algorithm is used to construct an integrated INNE detector in the data space of each cluster and compute a composite anomaly score for each sample. Finally, the anomaly samples are determined based on the anomaly scores. The main advantages of the model lie in integrating the trimmed window and clustering, which further enhances the advantages of the INNE algorithm in terms of local and global anomaly detection capabilities. By using the actual voltage operation data of the transformer district in a city's distribution network for validation and comprehensively comparing with other algorithms in terms of several evaluation indices, the results show that the TFEC-INNE model improves the detection effects in various anomaly scenarios.

DOI

10.19781/j.issn.1673-9140.2025.06.012

First Page

122

Last Page

135

Download

Share

COinS