Scalable and Reliable GNN-based Network Traffic Management in Software-Defined Vehicular Networks
Year: 2025
Type: dissertation
Abstract: Software-Defined Vehicular Networks (SDVNs) face bursty traffic loads, frequent handovers, and unstable cluster membership. This thesis introduces a predictive, cluster-aware SDVN architecture that combines hierarchical SDN control (edge/local and global tiers) with graph neural networks(GNN) for proactive traffic steering and RSU load balancing. Control objectives target stability, cluster longevity, and load fairness through a multi-objective training regime. A spatiotemporal GNN predicts short-horizon network states on vehicle–RSU graphs. The edge tier issues adaptive data re-routing and handover actions to mitigate emerging congestion, while the global tier coordinates inter-region consistency. A complete simulation pipeline is implemented in OMNeT++, Veins, INET, and SUMO with OpenFlow4Core, modeling mobile vehicles, RSUs, and SDN switches. A lightweight clustering layer groups vehicles by mobility and communication affinity and exposes cluster-level features to the predictor, enabling anticipatory control rather than reactive throttling. Experiments across multiple traffic densities and sampling windows compare the proposed STGNN\_Cluster controller with an SVM-driven baseline. Results show consistently lower end-to-end latency, higher throughput and delivery ratios, improved RSU load fairness, and fewer disruptive cluster changes, indicating enhanced stability under mobility. Ablation studies attribute gains to hierarchical control, predictive graph learning, and the multi-objective training design. The resulting framework provides a modular and reproducible foundation for real-time, learning-augmented SDVNs.
Author: Nadia Karimzadeh
Institution: Brock University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
Funder: Brock University
OpenAlex ID: W7223077328
