Leveraging Lyapunov-Guided Contextual Bandits for Scalable Distributed Vehicular Access in Satellite–Terrestrial Integrated Networks

Abstract

The integration of Low Earth Orbit (LEO) satellite with terrestrial Road Side Units (RSUs) offers a promising architecture for 6G-enabled Intelligent Transportation Systems by combining wide-area coverage with low-latency access for Connected Autonomous Vehicles. However, realizing efficient multi-vehicle cooperative access in such Heterogeneous Satellite–Terrestrial Integrated Networks (HSTINs) is hindered by three fundamental challenges: the spatiotemporal non-stationarity of satellite–terrestrial channels, the scalability bottleneck in distributed coordination among large vehicle populations, and the difficulty of enforcing long-term Quality-of-Service (QoS) constraints across periods. To address these coupled challenges, this paper proposes LyMCTS, a Lyapunov-guided Mean-Field Contextual Thompson Sampling framework that hierarchically decouples the original intractable problem. Lyapunov optimization transforms long-term QoS constraints into per-step virtual queue stability objectives, Mean-Field Game approximation reduces multi-vehicle coordination overhead to lightweight population-level statistics, and adaptive Contextual Thompson Sampling enables principled online learning under non-stationary satellite–terrestrial channel dynamics. Simulation results based on real LEO satellite ephemeris data show that LyMCTS achieves higher system utility, better delay performance, and more stable QoS satisfaction than benchmark and ablation methods across different mobility and network scales.