Prediction-Assisted Multi-UAV Online Service Migration and Trajectory Control for MEC-Empowered Vehicular Networks
Abstract
The rapid proliferation of intelligent vehicular networks and mobile edge computing (MEC) demands substantial improvements in computation efficiency, latency reduction, and service continuity, especially under high mobility and stochastic task generation.Conventional single-uncrewed aerial vehicle (UAV) solutions with static trajectory design lack the adaptability required in such dynamic environments. To address these challenges, we develop a collaborative multi-UAV optimization framework that jointly coordinates continuous UAV movement and service migration decisions. The framework incorporates a stacked long short-term memory (LSTM) model to predict short-term vehicle trajectories from historical motion data, which enables mobility-aware planning and reduces unnecessary cross-UAV migrations. To balance long-term migration cost and instantaneous delay, a Lyapunov-based technique is integrated into the per-slot optimization. Furthermore, a multiagent deep deterministic policy gradient (MADDPG) algorithm is adopted to facilitate cooperative policy learning across UAV agents in stochastic and partially observable environments. By unifying prediction, trajectory design, and adaptive migration control, the proposed method effectively adapts to time-varying network conditions. Extensive simulations demonstrate that the proposed method significantly outperforms baseline schemes in terms of delay, reward stability, and scalability under dynamic network loads.