Graph Neural Network Assisted Deep Reinforcement Learning Based Hierarchical Routing in Mega LEO Satellite Constellation
Shenzhan Xu, Rongke Liu, Hangyu Zhang, Aryan Kaushik
Beihang University China Mobile Research Institute
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Mega low-Earth-orbit (mLEO) constellations with inter-satellite links (ISLs) enable wide-area data delivery, but routing in such systems is challenged by large network scale, time-varying link availability, and spatially concentrated hotspot traffic. This paper studies task route setup over the satellite ISL segment and develops a learning-based hierarchical framework that decomposes source-to-egress ISL route setup into inter region route setup and intra-region path construction. At the upper layer, regions perform synchronized neighbor message passing on the slot snapshot to generate local embeddings, and a parameter-sharing graph neural network (GNN)-assisted Proximal Policy Optimization (PPO) policy determines the inter region route through sequential local decisions of the currently active region agent. This design avoids satellite-level real-time global search while preserving sensitivity to regional load and boundary-link conditions. To improve training under sparse success feedback, a heuristic teacher-guided warm start provides soft guidance in the early stage of policy learning. At the lower layer, a load-aware shortest-path method constructs executable intra-region segments, and each segment is revalidated at its actual start slot so that local repair can be triggered by link disruptions or capacity blocking. Using a simulation environment built from real two-line element (TLE) data and the Simplified General Perturbations 4 (SGP4) propagator, we evaluate the proposed method under different task sizes and runtime ISL failures. The results show that the proposed method achieves a favorable trade-off among task success rate, penalized average delay, and load balance, while maintaining robust performance under runtime ISL failures.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
工程Satellite Communication Systems
Space Satellite Systems and Control · Opportunistic and Delay-Tolerant Networks