EGDS-CNN: a Deep Learning Framework for Critical Node Identification and Vulnerability Analysis of Power Systems

With the continuous expansion of power grids, power grids have become more and more complicated, and their vulnerability has increased significantly. It is known that node failures may trigger large-scale blackouts; hence, the accurate identification of critical nodes is crucial for safeguarding power grid security. Traditional identification methods either over-rely on topological structures or focus on single electrical features, which limits their ability to capture the comprehensive impact of critical nodes on power grids and results in incomplete characterization of node importance. To overcome these limitations, this paper proposes an electrically guided dual-stream convolutional neural network (EGDS-CNN) framework to identify the critical nodes of power grids. Specifically, we develop an electrically guided neighborhood sampling strategy that prioritizes nodes with high power flow centrality (PFC) to be selected as the generalized neighbors and extracts their features; then, we use position encoding to weight the node feature, and a structured feature matrix for each node can be constructed. On this basis, EGDS-CNN reformulates critical node identification as a nonlinear regression task via convolutional neural networks. By integrating an adaptive attention mechanism into dual-stream CNN architecture, EGDS-CNN can achieve deep fusion of electrical and topological information, and the precise identification of critical nodes can be obtained. Experimental results show that EGDS-CNN significantly outperforms the existing methods in vulnerability analysis, which confirms that EGDS-CNN is a more effective approach for critical node identification.