Study of power grid subnet partition based on graph neural network

With the increasing scale of power systems, their reliability analysis and calculation become more complex and difficult. Community structure, as an important topological characteristic of complex networks, plays a prominent role in power grid research and application. The current methods for community division of power networks are mainly based on the topological characteristics of the network, with less consideration of the power balance of the subnetwork, which requires larger-scale machine-cutting or load-cutting operations when the subnetwork operates independently after the grid is unbundled. To solve this problem, this paper proposes a community segmentation method for power networks based on graph neural networks that integrally considers the topology of the network and the power balance of the network. Node attributes such as node degree, betweenness, and power value are selected as node features to help the model capture more correlations between nodes. The traditional K-means algorithm is also optimized and improved, and the method of selecting generator nodes as the clustering centers is proposed to ensure that there are generator nodes supplying energy in each community. Experiments are conducted on the IEEE standard test systems, and the effectiveness of the method proposed in this paper is verified by comparing it with other community segmentation methods.