Graph Reinforcement Learning With Inertia-Aware Temporal Aggregation for Domain Partitioning in LEO Satellite Networks

Abstract

Low Earth orbit (LEO) satellite networks constitute a critical space-segment infrastructure for integrated space-aerial-terrestrial networks, where autonomous and reliable control is essential for mission-critical service orchestration. To support scalable space-layer control, LEO satellite networks require dynamic domain partitioning. However, existing snapshot-based partitioning schemes recompute domains based solely on the instantaneous topology, often triggering frequent controller handovers. With time-varying inter-satellite links, such handovers incur signaling and migration overhead and degrade control-plane stability. To address this, we propose a graph reinforcement learning with inertia-aware temporal aggregation (GRL-ITA) framework to achieve topology-adaptive domain partitioning with reduced controller handovers. The objective is to reduce a period-level composite cost including propagation delay, controller load, and migration overhead. GRL-ITA leverages a graph neural network to capture spatial topology and a reinforcement learning agent that observes topology snap-shots and produces slot-level controller candidate preferences. Furthermore, an inertia-aware temporal aggregation mechanism converts the agent’s observation-driven preferences into a stable period-level DC set, suppressing frequent handovers. Simulation results show that the proposed approach outperforms several traditional baselines in reducing controller handover frequency while maintaining competitive propagation latency and load balancing of controllers.