PHVC: A Novel Prior-Guided Hybrid Machine Learning-Based VANET Caching Strategy

Vehicular Ad Hoc Networks (VANETs) have become an important communication technology for intelligent transportation systems, where efficient caching mechanisms are required to reduce content delivery latency and communication load. However, caching strategies inherited from conventional Mobile Ad Hoc Networks (MANETs) are often insufficiently adaptive to rapid topology changes, heterogeneous vehicular content interests, and the different caching characteristics of vehicles and Roadside Units (RSUs). This paper proposes PHVC, a Prior-Guided Hybrid Machine Learning-based VANET Caching strategy that combines lightweight Q-learning with supervised learning and conventional caching knowledge. PHVC separates cache replica placement and cache replacement into coordinated learning processes. Support Vector Machine (SVM)-based historical pattern learning provides auxiliary guidance for cache replica placement, whereas Least Recently Used (LRU) information is introduced as temporary prior guidance for cache replacement during the initial reinforcement-learning stage. Each vehicle and RSU independently maintains a local Q-table whose state representation is based on cache positions rather than global content identities, thereby limiting the state-space and storage requirements. A request-triggered delayed reward mechanism is further employed to associate caching decisions with subsequent cache-use outcomes. The strategy is evaluated using SUMO, OMNeT++, Veins, and INET under representative urban and highway scenarios. At the final simulation points, compared with the best-performing conventional benchmark for each metric, PHVC improves the cache hit ratio by approximately 13.0–26.1%, reduces average content delivery latency by 23.9–48.7%, and reduces normalised link load by 7.6–23.4%. The results demonstrate that the proposed prior-guided hybrid learning framework provides an effective performance–complexity trade-off for adaptive cache management in dynamic vehicular networks.