IRNet: INVANETs Performance Prediction via Spatio-Temporal Graph Attention Networks

Intelligent Vehicular Ad-hoc Networks (INVANETs) optimize road travel in Intelligent Transportation Systems (ITS), exploiting smart technologies to enhance safety, efficiency, and reduce congestion. Network modeling, as a fundamental tool in network analysis and optimization, is critical to the ongoing enhancement of INVANETs. However, affordable models that can accurately capture topological dependence among network performances in INVANETs and adapt to their dynamics are currently lacking. This paper proposes the INVANETs Routing Modeling Network (IRNet). The IRNet is based on an elaborate Spatio-Temporal Graph Attention Networks (STGAT) to capture the intricate relationships between topology, network states, and Key Performance Indicators (KPIs), e.g., delay, jitter, and drops. IRNet first constructs models for per-packet delay and drops mathematical distributions, then calculates these KPIs for source-destination pairs, and derives global network indicators. It adapts to INVANETs' dynamic topologies using snapshots. Furthermore, IRNet generalizes well across different topologies, routing schemes, and traffic conditions. Evaluations reveal IRNet's superior prediction accuracy, outperforming baseline models with a 17.64% improvement in average-case MRE, highlighting its effectiveness. Use cases are also involved to illustrate IRNet's optimization impact on INVANETs.