A Graph Neural Network-Based Approach for Critical Node Detection in Dynamic Satellite Networks

To address the challenges in critical node identification within complex satellite networks, such as difficulties in dynamic modeling and limited computational resources, this study proposes an innovative spatiotemporal feature learning-based identification scheme. First, by coupling a dynamic geographic constraint mechanism with complex network modeling, a generative satellite network model (BS-TN) incorporating satellite orbit characteristics was developed, enabling generation of controllable simulation datasets. Second, a spatiotemporal fusion node importance evaluation model (ST-FNIM) was introduced. A lightweight L-GAT network was designed by incorporating nonlinear attenuation into the attention mechanism and pruning the multi-head attention to reduce the computational overhead, thereby extracting the multi-dimensional spatial features of satellite nodes across time steps. An event-driven ET-LSTM architecture was also proposed, enabling responsive sequential analysis of node behaviors via an optimized gating mechanism. By fusing spatiotemporal features, the model enabled the accurate dynamic identification of critical nodes. The experimental results demonstrated that compared with conventional centrality algorithms and existing deep learning methods, the proposed approach significantly improved computational efficiency and dynamic adaptability, while increasing identification accuracy by approximately 10%. It enabled real-time assessment and response to sudden events such as link failures and traffic surges, offering a novel technical framework for satellite network management, resource optimization, and security assurance.