Revenue-Aware Vehicular Edge-Cloud Offloading via Two-Stage DRL and Dynamic Server Clustering

Modern vehicles generate large volumes of sensor data that can exceed onboard processing capabilities. While task offloading to edge servers reduces latency and alleviates resource scarcity at vehicles, vehicular mobility and dense traffic often create localized overloads where multiple neighboring edge servers become congested simultaneously. Such regional overloads exhaust edge resources, degrade success rate, and diminish service-provider revenue. Edge-server cooperation is necessary under these conditions; however, dynamically forming cooperation groups and jointly deciding how much to offload and where to process enlarges the state–action space and complicates learning when these coupled actions are selected in a single step. To address this revenue-aware offloading problem under latency constraints and multi-tier subscription service, we formulate a joint offloading and resource-allocation problem that maximizes service-provider revenue and propose a hierarchical edge–cloud framework with three connected components: (i) a cloud-assisted dynamic edge-server clustering algorithm that groups edge servers by resource availability to form resource-balanced cooperation structures that remain effective under regional overloads; (ii) a two-stage deep reinforcement learning (DRL) policy that sequentially determines the offloading fraction and then selects the processing server to reflect their dependency and stabilize training; and (iii) a credit-based computing resource allocator that tracks cumulative service outcomes and subscription tier to enforce long-term fairness. The simulation results show that the proposed framework converges 1.75× faster than the single-stage DRL baseline, achieves an approximately 10% higher success rate and increases revenue by 30% under heavy-load conditions.