In the emerging landscape of next-generation satellite communications, Low Earth Orbit (LEO) satellite constellations have garnered significant attention due to their low latency, high revisit frequency, and global coverage. However, the dynamic topology, limited energy budget, and fluctuating traffic patterns pose critical challenges for real-time routing, load balancing, and power control. This paper presents a novel hybrid framework that integrates Graph Neural Networks (GNN) and Deep QNetworks (DQN) to address distributed power-aware routing and load balancing in LEO satellite systems. First, a virtual node model is constructed to abstract the evolving satellite topology, enabling a static graph representation over time. GNNs are employed to learn edge-level representations from inter-satellite link features such as bandwidth, delay, and betweenness centrality. These features are then embedded into a DQN-based decision module, which dynamically selects optimal routing paths based on a utility function that balances throughput and latency. Furthermore, a two-hop load-sensing distributed routing algorithm (RMLBR) is proposed to overcome local optima caused by limited neighborhood awareness, enabling each satellite to perceive indirect link load and reroute traffic more effectively. Simulation results under realistic orbital and traffic conditions demonstrate that the proposed DQN+GNN scheme reduces link handover frequency by 42. 2 \% compared to elevation-prioritized strategies while achieving near-maximum throughput (63. 08 Mbps). The method also maintains low packet loss (<4 \%), minimal average latency (100 ~ms), and high SINR stability under closed-loop power control.
Fan et al. (Sun,) studied this question.