Spatio-Temporal Correlated Network State Prediction and Dynamic Routing for Satellite Networks

Most existing routing algorithms for Low-Earth-Orbit (LEO) satellite networks neglect the spatio-temporal features inherent in the network states, leading to suboptimal performance in scenarios with dynamic network topologies. In this paper, we propose a Spatio-Temporal Graph Attention Network (STGAN) architecture for the extraction of spatio-temporal features from satellite network states. Building upon this, we propose a State Prediction based Dynamic Routing Algorithm (SP-DRA), which combines STGAN with the enhanced Shortest Path First algorithm (eSPF). SP-DRA utilizes STGAN to predict future network states by exploiting the spatio-temporal correlations of historic observations. Based on state predictions, each link is assigned with a delay-based prediction weight, which is used as the input for eSPF. The proposed eSPF dynamically selects the path with the smallest prediction weight to facilitate congestion avoidance and reduce transmission delay. Simulation results show that our STGAN architecture provides up to 18.6% prediction accuracy improvement than the existing deep learning-based methods, namely STGCN and ST-MGAT. Meanwhile, the SP-DRA algorithm outperforms existing routing strategies including SPF, Explicit load balancing algorithm (ELB), and Satellite networks Link State Routing algorithm (SLSR) in terms of packet loss rate and average end-to-end delay. Specifically, the performance gains increase with network traffic.