Distributed Dynamic Routing for LEO Satellite Networks with Temporal Graph Convolutions and Imitation Acceleration

Large-scale Low Earth Orbit (LEO) constellations offer promising solutions for global network access with low latency and high-speed transmission. However, their dynamic topology poses challenges for routing. Traditional centralized routing schemes enable global optimization but entail high overhead and computational complexity, making them unsuitable for large-scale LEO networks. We propose an algorithm named MATGCIR that leverages Multi-Agent deep reinforcement learning and Temporal Graph Convolution for distributed Intelligent Routing, extracting dynamic topology features and incorporating spatiotemporal information. A pretraining mechanism using shortest path rules enhances training efficiency and decision quality. Simulation results demonstrate that MATGCIR effectively addresses LEO network complexities and accelerates the training process for large-scale constellations.