QoS-Aware Resource Allocation in LEO Satellite Networks: A Graph Neural Network Approach

Abstract

Efficient radio resource allocation is pivotal for maximizing the service capability of Low Earth Orbit (LEO) satellite networks. However, the high mobility of satellites and the rapidly time-varying channel conditions pose significant challenges to traditional resource management schemes. Conventional optimization methods suffer from prohibitive computational complexity, while heuristic approaches typically prioritize users with superior channel gains to maximize the sum-rate. Consequently, this strategy severely degrades the service reliability and fairness for disadvantaged users. To address these challenges, this paper proposes a QoS-Aware Primal-Dual graph neural network (QPD-GNN) framework for downlink resource allocation. By capturing the dynamic topological mapping between fluctuating user demands and channel states, the proposed method learns an intelligent scheduling policy to optimize resource distribution. Extensive simulations demonstrate that QPD-GNN outperforms relevant baseline schemes in terms of service reliability. Crucially, while conventional sum-rate maximization benchmarks achieve an upper bound in average spectral efficiency by concentrating resources on strong users, our method guarantees a substantially higher user service satisfaction rate and edge-user spectral efficiency. These results confirm that QPD-GNN achieves a superior trade-off between system throughput and fairness, establishing it as a robust solution for ensuring universal coverage in highly dynamic LEO constellations.