Anomaly Detection of Distribution Network Based on Adversarial Dual Autoencoder

Abstract

Anomaly detection on attributed graphs aims to identify individuals or groups whose patterns deviate significantly from most samples, which is very important for internet security, power system, and other applications. Especially in the power system, the energy flow pattern of the distribution network is very similar to the graph topological constraints. Taking each equipment and bus as nodes and the relevant electrical measurement data as attributes, the distribution network fault location problem is mapped to the anomaly detection task on the attributed graph. The traditional GNN-based anomaly detection algorithm for distribution networks has some problems, such as over-smoothing and inefficient embedding learning. To solve the problem of anomaly detection in the distribution network, an anomaly detection algorithm based on adversarial dual autoencoder (DGAE_GAN) is proposed, which is composed of a structural encoder and attribute encoder based on residual connection GAT. Through joint learning, the attributed graph is mapped to the low-dimensional space, and the decoders reconstruct the attributed graph. Then, a discriminator is invited to learn the sample similarity, and the game strategy can effectively alleviate the over-smoothing problem. Finally, the anomaly node detection task is realized by calculating the reconstruction errors from both the structure and attribute levels. We conducted multiple experiments on three real-world datasets, and the IEEE standard example shows that our proposed model is superior to the existing methods.