GNN-Enhanced Routing in SD-WAN Over Realistic Space–Ground Networks With Constellation Dynamics and Weather-Driven Link Variability
TL;DR
A learning-based routing framework for Space-Ground Integrated Networks deployed at the SD-WAN customer-premises equipment that generalizes well to satellite configurations and traffic demands, achieving up to a 16.5% reduction in end-to-end delay compared to a KNN baseline.
Abstract
The integration of LEO satellite networks into Software-Defined Wide Area Networks (SD-WANs) enables near-global coverage and low latency, but also introduces challenges due to highly dynamic topologies, capacity-limited satellite access, and time-varying delays. Conventional reactive SD-WAN traffic steering mechanisms struggle to cope with these dynamics, particularly for latency-sensitive applications. This paper proposes a learning-based routing framework for Space-Ground Integrated Networks deployed at the SD-WAN customer-premises equipment. The framework performs per-request end-to-end path selection based on network state, including residual link capacities and propagation delays. Routing decisions are learned using a Double Deep Q-Network (DDQN) augmented with a Graph Neural Network (GNN) to capture the time-varying satellite–terrestrial topology. The proposed approach is evaluated on Iridium-NEXT constellation using real Two-Line Element (TLE) data, together with real-world delay and capacity measurements from the Wetlinks dataset. Results show that the GNN–DDQN model generalizes well to satellite configurations and traffic demands, achieving up to a 16.5% reduction in end-to-end delay compared to a KNN baseline. Additional experiments under clear-sky, rainy, and extreme rainfall conditions demonstrate the robustness of the proposed framework in capacity-constrained regimes.