Prediction-Driven Task Scheduling in Satellite IoT: A Constrained Reinforcement Learning Approach

Abstract

The incorporation of Mobile Edge Computing into satellite systems is a highly promising approach to enabling large-scale intelligent Internet of Things services in remote areas. However, the high-speed mobility of satellites and the extreme scarcity of on-board resources pose significant challenges, as traditional reactive scheduling often fails to handle the rapidly shifting traffic demands, leading to severe resource fragmentation and latency spikes. Since the deterministic mobility of satellites also creates exploitable spatiotemporal traffic patterns, we propose a prediction-driven Constrained Reinforcement Learning approach to minimize average response time while balancing satellite loads. Specifically, we first develop a spatiotemporal prediction model that characterizes the traffic load across the satellite constellation. By integrating the prediction model, we formulate the scheduling problem as a Constrained Markov Decision Process and transform it into a standard Markov Decision Process by leveraging Lyapunov optimization to meet the resource constraint, whereby an Actor-Critic scheduling method with Proximal Policy Optimization is employed to learn the optimal policy. Empirical results show up to $4 \times$ higher scheduling efficiency than baseline methods.