【作者】 刘云龙；

【导师】 谢人超；

【作者基本信息】 北京邮电大学 ， 信息与通信工程， 2025， 硕士

【摘要】 近年来,低轨卫星星座与边缘计算的深度耦合驱动星载边缘云向广域实时服务范式演进,成为支撑空天地一体化服务的关键基础设施。然而,卫星节点资源的高度受限性、星间拓扑的分钟级动态重构特性以及负载的时空异质性特征,导致传统云计算框架在服务编排效率、资源利用率及系统弹性等方面面临严峻挑战。无服务器计算虽具有事件驱动的函数级资源调度与毫秒级弹性扩缩能力,但其在动态服务部署、资源自适应管理及不确定性决策等方面的固有局限性,仍难以适配卫星网络极端动态性与资源约束的耦合环境。因此,针对上述挑战,本文围绕星载边缘云中无服务器计算的服务编排与弹性资源管理问题,分别从如下两方面展开研究:(1)针对星载环境中时延敏感性与资源成本的强耦合问题,提出多目标联合优化模型。通过构建多维资源约束及星间链路动态特性的混合整数非线性规划问题,创新性地设计基于水滴表面优化(Drops on Surface Optimization,DSO)与强化学习的混合智能算法,通过液滴群协同搜索机制增强全局探索能力,结合Q-leaming的在线奖励反馈实现动态决策优化。实验结果表明,在复杂网络规模的卫星网络中,相较于灰狼优化(Grey Wolf Optimizer,GWO)和粒子群优化(Particle Swarm Optimization,PSO)算法,该方法将端到端时延降低30%到70%,部署成本减少6%到12%,收敛速度提升40%,有效平衡时延与成本间的权衡,并适应高动态拓扑变化。(2)针对星载网络流量突发性与时空相关性引发的资源震荡问题,提出LSTM-GP-RL混合架构。基于多尺度小波分解与图卷积网络(Graph Convolutional Networks,GCN)构建时空流量预测模型,有效捕捉流量序列的长短期依赖特征与空间相关性。通过高斯过程回归(Gaussian Processes Regression,GPR)量化预测不确定性,并结合注意力机制强化关键特征提取。进一步设计不确定性感知的深度强化学习机制,构建包含资源状态转移概率矩阵的动态决策模型,实现资源扩缩容的主动式调节。仿真实验表明,该方法在资源震荡次数较静态阈值法减少68%,响应时间违规率下降82%,能够在保障服务稳定性的同时显著抑制资源震荡,提升系统弹性与资源利用率。更多还原

【Abstract】 In recent years,the deep integration of Low Earth Orbit(LEO)satellite constellations and edge computing has driven the evolution of satellite edge clouds toward a wide-area,real-time service paradigm,establishing itself as a critical infrastructure for integrated space-air-ground networks.However,the highly constrained resources of satellite nodes,the minute-level dynamic reconstruction of inter-satellite topologies,and the spatiotemporal heterogeneity of service loads present significant challenges to traditional cloud computing frameworks in terms of service orchestration efficiency,resource utilization,and system resilience.Although serverless computing features event-driven function-level resource scheduling and millisecond-level elastic scaling,its inherent limitations in dynamic service deployment,adaptive resource management,and uncertainty-aware decision-making make it difficult to adapt to the coupled environment of extreme dynamics and resource constraints in satellite networks.To address these challenges,this dissertation focuses on service orchestration and elastic resource management for serverless computing in satellite edge clouds,conducting research in the following two aspects:(1)To tackle the strong coupling between latency sensitivity and resource cost in satellite environments,a multi-objective joint optimization model is proposed.By formulating a mixed-integer nonlinear programming(MINLP)problem incorporating multidimensional resource constraints and the dynamic characteristics of inter-satellite links,a hybrid intelligent algorithm is developed,combining Drops on Surface Optimization(DSO)and reinforcement learning.This approach enhances global search capabilities through a collaborative droplet swarm mechanism and achieves dynamic decision optimization via Q-learning-based online reward feedback.Experimental results demonstrate that,in complex satellite network scenarios,this method reduces end-to-end latency by 30%to 70%,decreases deployment cost by 6%to 12%,and accelerates convergence by 40%,outperforming Grey Wolf Optimizer(GWO)and Particle Swarm Optimization(PSO)in balancing latency-cost trade-offs and adapting to highly dynamic topologies.(2)To mitigate resource oscillations caused by bursty and spatiotemporally correlated traffic in satellite networks,an LSTM-GP-RL hybrid architecture is proposed.A spatiotemporal traffic prediction model is constructed using multi-scale wavelet decomposition and Graph Convolutional Networks(GCNs)to capture both long-and short-term dependencies and spatial correlations in traffic sequences.Gaussian Processes Regression(GPR)is employed to quantify prediction uncertainty,while an attention mechanism is incorporated to enhance key feature extraction.Furthermore,an uncertainty-aware deep reinforcement learning mechanism is designed,establishing a dynamic decision model with a resource state transition probability matrix to enable proactive resource scaling.Simulation results show that this method reduces the number of resource oscillations by 68%compared to static threshold methods and decreases the service-level agreement(SLA)violation rate by 82%,significantly enhancing system elasticity and resource efficiency while ensuring service stability.更多还原