计算机科学
机制(生物学)
卫星
布线(电子设计自动化)
计算机网络
路由算法
自适应路由
分布式计算
实时计算
多路径路由
算法
静态路由
方案(数学)
算法设计
基于策略的路由
地球观测卫星
动态源路由
通信卫星
卫星系统
路由协议
作者
WENTAO HE,Huayi Li,SHI QIU
标识
DOI:10.1016/j.ast.2026.112361
摘要
Quality-of-Service-aware routing in Low Earth Orbit (LEO) satellite networks is challenging because the rapidly varying topology and unexpected link failures make real-time multi-objective optimization computationally expensive. Existing methods incur high online computation overhead and adapt slowly to dynamic network conditions. This paper proposes a two-stage dynamic routing framework that decouples topology evolution modeling from online routing decisions. In the first stage, a Transformer-based spatio-temporal model predicts each satellite’s establishable inter-satellite link set and link distances, serving as a learning-based surrogate for the O ( N 2 ) pairwise geometric calculations required at each time step. In the second stage, a hierarchical graph attention mechanism extracts multi-dimensional network state features and learns congestion-aware edge weights that jointly encode delay, packet loss, bandwidth, and hop count objectives, while a recursive partitioning strategy reduces the priority-queue overhead in shortest-path computation from O ( m log n ) to O ( m log 2/3 n ). Simulations on the Starlink Phase-1 constellation (1,584 satellites) show that the topology prediction module achieves an F1-score of 0.958 and a distance prediction root mean square error of 6.40 km with a 177 × inference speedup over analytical computation. Compared with representative baselines including Dijkstra-based shortest path first routing, the proposed algorithm reduces average end-to-end delay by 50.5%, packet loss by 39.4%, and improves throughput by 57.1%. Under random link failure rates up to 20%, the framework maintains a 100% routing success rate, whereas all four baselines degrade to below 31%.
科研通智能强力驱动
Strongly Powered by AbleSci AI