低轨卫星（low-Earth-orbit, LEO）的迅猛发展为未来通信服务注入了全新活力。然而，由于网络流量固有的波动特性，在高度动态的网络环境中保障差异化服务质量仍面临严峻挑战。本文提出一种基于在线学习的星地融合网络资源调度方案，旨在以最小化资源消耗实现按需服务。具体而言，本文聚焦三大核心问题： 1) 如何准确表征动态星地链路资源；2) 如何精准预测业务需求并保障不确定性；3) 如何实现动态网络资源与随机流量的按需匹配。在星地链路建模方面，采用第三代合作伙伴计划（the 3rd Generation Partnership Project, 3GPP）制定的非地面网络（non-terrestrial network, NTN）信道与天线模型。在业务需求预测方面，提出了结合一维卷积（one-dimensional convolution, 1D-Conv）、长短期记忆（long short-term memory, LSTM）和注意力机制的预测模型以实现平均业务需求预测，并引入保形预测（conformal prediction, CP）理论以应对突发流量带来的不确定性。在资源按需匹配方面，构建了一种双时间尺度资源调度框架，包括大时间尺度的资源预留和小时间尺度的资源调整，并设计了一种基于在线凸优化（online convex optimization, OCO）的资源调度算法，能够在有限网络信息的条件下提供长期的性能保障。基于真实数据集的实验结果验证了所提预测算法的准确性和高效性。基于高保真卫星互联网仿真平台，以星地信道模型的网络模拟器3（Network Simulator 3, NS3）实现为基础，实验结果表明所提出的资源调度方法在性能上接近理想条件下的全局最优解，在保障用户服务质量（quality of service, QoS）需求的同时显著降低了系统资源占用率。

The rapid growth of low-Earth-orbit satellites has injected new vitality into future service provisioning. However, given the inherent volatility of network traffic, ensuring differentiated quality of service in highly dynamic networks remains a significant challenge. In this paper, we propose an online learning-based resource scheduling scheme for satellite–terrestrial integrated networks (STINs) aimed at providing on-demand services with minimal resource utilization. Specifically, we focus on: ① accurately characterizing the STIN channel, ② predicting resource demand with uncertainty guarantees, and ③ implementing mixed timescale resource scheduling. For the STIN channel, we adopt the 3rd Generation Partnership Project channel and antenna models for non-terrestrial networks. We employ a one-dimensional convolution and attention-assisted long short-term memory architecture for average demand prediction, while introducing conformal prediction to mitigate uncertainties arising from burst traffic. Additionally, we develop a dual-timescale optimization framework that includes resource reservation on a larger timescale and resource adjustment on a smaller timescale. We also designed an online resource scheduling algorithm based on online convex optimization to guarantee long-term performance with limited knowledge of time-varying network information. Based on the Network Simulator 3 implementation of the STIN channel under our high-fidelity satellite Internet simulation platform, numerical results using a real-world dataset demonstrate the accuracy and efficiency of the prediction algorithms and online resource scheduling scheme.