卫星跳波束通信具有灵活性高、抗干扰能力强、资源利用率高的特点，是新一代卫星通信系统的发展方向。随着具有突发性与波动性的卫星互联网接入业务爆发式增长，以及非静止轨道星座的蓬勃发展，卫星跳波束按需资源管理面临巨大挑战。论文聚焦于星座通信系统跳波束资源管理，研究跳波束实时智能决策、星座集中式调度决策、星座分布式智能决策等技术方法，解决卫星时、空、频、功、轨等资源与多样化业务按需匹配问题。主要创新工作如下：首先，针对时变业务下多波束资源管理实时性要求高的问题，研究了基于决策分解的快速跳波束资源管理方法。率先建立了多智能体合作决策模型，将指数决策空间降维至线性决策空间，提出了基于共享奖励和全局状态的多智能体强化学习算法，为实现星上低复杂度决策计算提供设计参考。仿真结果表明，相比于传统方法，所提方法可有效提高卫星业务吞吐率性能，实现毫秒量级的卫星时、空、频资源管理。其次，针对星间负载不均与星间干扰的问题，研究了负载均衡、干扰规避和动态切换等多类约束下的集中式星座通信按需跳波束资源管理方法。建立了干扰夹角、空间隔离图案和图案干扰惩罚模型，并将多星集中式跳波束资源管理问题分解为多个调度子问题，有效降低了求解复杂度。进一步地，建立了用户候选接星集合、小区动态权重图模型，避免了星座拓扑动态变化导致的用户频繁切换。仿真结果表明，相比于传统方法，所提方法可有效缓解星座系统间干扰，提高系统内负载均衡性能和用户平均业务满足度性能，用户平均切换次数降低约30%。最后，针对多星分布式资源管理中的决策冲突问题，研究了缺乏全局信息条件下的多星协同分布式按需资源管理方法。建立了半中心式的星地两级资源管理架构模型，提出了在信关站集中调度业务流向基础上，采用集中式训练、分布式部署学习架构的星上自主资源管理方案。仿真结果表明，所提算法可有效缓解多星决策冲突导致的星间干扰，提升系统业务吞吐率，并支持毫秒量级的多星分布式实时资源管理。

Satellite beam hopping communication systems, characterized by high flexibility, strong interference resistance, and efficient resource utilization, are regarded as the direction for the development of next-generation satellite communication systems. However, with the explosive growth of satellite internet access services characterized by bursts and fluctuations, and the vigorous development of non-geostationary satellite constellations, on-demand resource management for satellite beam hopping systems faces significant challenges. This doctoral thesis focuses on beam hopping resource management in constellation communication systems, studying technological methods such as real-time intelligent decision-making for beam hopping, centralized scheduling decisions for constellations, and distributed intelligent decision-making for constellations to address the issue of on-demand matching between satellite resources in time, space, frequency, power and orbit dimensions, and diverse service requirements. The main innovative work is as follows: Firstly, to address the problem of high real-time requirement of multi-beam resource management in time-varying traffic, a rapid beam hopping scheduling method based on decision decomposition is studied. We first establish a multi-agent cooperative decision-making model, reducing the exponential decision space to a linear decision space, and propose a multi-agent reinforcement learning algorithm based on a shared reward and state mechanism to provide design references for achieving low-complexity on-board decision-making calculations. Simulation results demonstrate that the proposed method effectively improves the performance of system throughput, achieving millisecond-level management of time, space and frequency resources compared to traditional methods. Secondly, to address the problem of uneven traffic distribution and inter-satellite interference, an on-demand beam hopping scheduling method for constellation communication is studied under multiple constraints such as load balancing, interference avoidance, and dynamic handover. We establish models for interference angle, spatial isolation patterns, and pattern interference penalties, and decompose the multi-satellite centralized beam hopping resource management problem into multiple scheduling sub-problems, effectively reducing computational complexity. Furthermore, we establish models for user candidate satellite sets and dynamic weight graphs of cells to prevent frequent user handover caused by dynamic changes in constellation topology. Simulation results demonstrate that the proposed method effectively mitigates inter-constellation interference, and improves the performance of inter-satellite load balancing and user average service satisfaction, with an approximately 30% reduction in average handover frequency for users compared to traditional methods. Finally, to address the problem of multi-satellite distributed decision-making conflict, a multi-satellite distributed on-demand scheduling method under the condition of lacking global information is studied. We establish a semi-centralized two-tier resource management architecture and propose an on-board autonomous scheduling scheme based on centralized training and distributed deployment learning architecture, with ground station centralized scheduling of service flow. Simulation results demonstrate that the proposed algorithm can effectively alleviate the inter-satellite interference caused by multi-satellite decision conflict and improve the system throughput, supporting millisecond-level distributed real-time scheduling.