Cache-assisted task offloading in Vehicular Edge Computing: A spatio-temporal deep reinforcement learning approach

Vehicular Edge Computing (VEC) faces significant challenges in jointly managing caching and task offloading due to dynamic network conditions and resource constraints. This paper proposes a novel framework that addresses these challenges through a synergistic three-stage process. The innovation lies in the tight integration of our modules: first, a Spatio-Temporal Fast Graph Convolutional Network (ST-FGCN) accurately forecasts task demands by capturing complex spatio-temporal correlations. Second, these predictions guide a Prediction-Informed Edge Collaborative Caching (PIECC) algorithm to proactively optimize resource placement across edge servers. Finally, a Genetic Asynchronous Advantage Actor–Critic (GA3C) strategy performs robust task offloading within this optimized environment. Unlike traditional reinforcement learning methods that often struggle with the large state–action spaces in VEC and converge to local optima, our framework simplifies the decision process via predictive caching and enhances exploration with the GA-infused GA3C algorithm. Simulation results demonstrate that our proposed framework significantly reduces long-term system cost, outperforms baseline methods in both latency and energy efficiency, and offers a more adaptive solution for dynamic VEC systems. © 2025 Elsevier B.V.