STG-SR: Spatio-Temporal Graph based Load Balancing Routing for Satellite Networks

Abstract

Low Earth Orbit (LEO) satellite networks face challenging routing conditions due to dynamic topology evolution and uneven spatio-temporal traffic distributions. To address the difficulty of jointly supporting network-wide path adaptation and localized congestion mitigation, this paper proposes STG-SR, a cohesive load-balancing routing framework for softwaredefined satellite networks. Specifically, STG-SR combines a global spatio-temporal graph with deep reinforcement learning for basic path provisioning, while employing a local spatio-temporal graph for forwarding-table reconstruction and congestion-aware multipath traffic splitting. Simulation results under low-density and high-density spatio-temporal traffic patterns show that STG-SR achieves lower delay, reduces the congestion node ratio, and improves traffic distribution compared with Dijkstra, DBPR, and CGR, demonstrating the effectiveness of coordinated global-local routing in dynamic LEO satellite networks.