Digital twin-driven collaborative task offloading method for vehicular edge computing
Abstract
Vehicular edge computing (VEC) has emerged as a critical paradigm for meeting the demands of computation-intensive and delay-sensitive vehicular applications by pushing computational capabilities to the network edge. However, in highly dynamic traffic environments, vehicle mobility, fluctuating wireless channels, limited resources, and stochastic task arrivals make offloading decisions vulnerable to insufficient state awareness, unbalanced resource allocation, and degraded service continuity. To address these issues, this paper proposes a digital twin (DT) driven collaborative dynamic task offloading framework for VEC. Leveraging DT technology to construct a real-time virtual mapping of the physical VEC network, the proposed framework facilitates the collaborative allocation of computing resources between edge servers (ESs) and opportunistic edge servers (OESs) through global state monitoring and prediction. Specifically, it fully exploits the collaborative computational capabilities of roadside units (RSUs) and opportunistic vehicles, leveraging idle resources of opportunistic vehicles to construct a vehicular ad-hoc cloud to alleviate the load pressure on ESs. Based on this architecture, a task execution model including local computing, V2R offloading, and V2V offloading is formulated. Transmission delay, computing delay, energy consumption, resource capacity, and task completion constraints are jointly considered to minimize system delay and energy consumption while improving the task completion rate. To solve the multi-vehicle continuous offloading and resource allocation problem, a federated learning-assisted multi-agent deep deterministic policy gradient (FL-MADDPG) algorithm is developed. MADDPG enables cooperative decision-making among vehicles, while FL aggregates vehicle-side model parameters to reduce local training bias, preserve data privacy, and enhance policy generalization. Experimental results demonstrate that the proposed algorithm can balance system load more effectively and outperform mainstream baseline algorithms in terms of convergence performance, average system latency, and task completion rate.