
The Internet of Vehicles (IoV) enables cooperative and intelligent transportation through Vehicle-to-Everything (V2X) communications, but its practical deployment remains constrained by high mobility, rapidly varying topologies, and stringent energy limitations in dense urban environments.This paper presents SERENE, a federated reinforcement learning framework for New Radio Vehicle-to-Everything (NR-V2X) vehicular networks, and introduces SERENE-Prox, a proximal extension designed to improve convergence stability under non-independent and non-identically distributed (non-IID) vehicular data distributions. While SERENE defines the overall learning and networking architecture, SERENE-Prox constitutes the main algorithmic contribution by explicitly mitigating client drift and data heterogeneity across roadside units (RSUs).The proposed framework integrates Software-Defined Networking (SDN), edge intelligence, and multi-hop clustering to jointly optimize Packet Delivery Ratio (PDR), energy efficiency, and cluster stability under realistic NR-V2X Mode 2 with sensing-based semi-persistent scheduling (SPS) constraints. At its core, a multi-objective Control Policy Optimization Algorithm (CPOA) balances these competing objectives while maintaining low vehicle-side computational overhead.The framework is evaluated through co-simulation with the ns-3 network simulator and Simulation of Urban MObility (SUMO), including urban shadowing, NR-V2X Medium Access Control (MAC) behavior, 30 independent runs, and 95% confidence intervals (CIs). The results show that SERENE maintains high reliability (above 93% PDR) while reducing energy consumption relative to state-of-the-art baselines, whereas SERENE-Prox further improves convergence stability and robustness under heterogeneous RSU observations with only marginal additional energy cost. Overall, the results confirm the practicality of the proposed framework for energy-efficient and resilient IoV deployments.
Vehicle-to-infrastructure (V2I) communications involve interactions between mobile vehicles and static roadside unit (RSU) nodes, resulting in limited connectivity and short interaction windows, particularly at high speeds. While there is consensus on using IEEE 802.11p and C-V2X radio technologies at the physical layer, the upper layers are technology-agnostic. These upper-layer communication technologies exhibit varied capabilities, each with distinct strengths and limitations in dynamic environments, leading to one-size-fits-all inefficiencies in V2I ecosystems. Existing methodologies primarily focus on lower-layer strategies, such as scheduling, radio access, dynamic resource allocation, and message prioritisation, to enhance network performance. Additionally, upperlayer approaches often assume ideal network conditions and fail to adapt to real-time variability. To overcome these limitations, this study introduces a novel protocol fusion module (FM) that dynamically leverages the strengths of twenty combinations of Internet of Things (IoT) application-layer protocols (ALPs) and transportlayer protocols. The FM is guided by fuzzy-based multi-criteria decision analysis (MCDA) and contextual multi-armed bandit machine learning models as decision engines (DEs). Extensive performance evaluations across multiple case studies and node densities indicate that the framework can optimise critical performance metrics under both single and shared-spectrum conditions. All DEs achieved an accumulated PDR averaging of at least 75% under high vehicle penetration, with higher values under low-and-mild-traffic conditions, aggregated over the set of protocol combinations selected for each communication session. These improvements were most visible under shared-spectrum, where fuzzy MCDA-based DE recommendations maintained desirable latency (in the 3 ms range) and PDR averages within 90% and higher across both the SCH and CCH channels is most simulated scenarios and node densities. Overall, the framework demonstrated that it could improve critical metrics across various V2I scenarios.