Dij-ST-GAT: Dijkstra-based spatio-temporal graph attention network for predictive congestion-aware routing in vehicular IoT networks

Vehicular Internet of Things (IoV) networks require efficient routing algorithms capable of rapidly adapting to dynamic traffic conditions. Conventional shortest-path algorithms rely on current traffic information. These algorithms cannot predict future road congestion and traffic jams. This paper proposes Dijkstra-Spatio Temporal-Graph Attention Network (Dij-ST-GAT) routing frame work, designed for predictive congestion-aware routing in IoV networks. The proposed model utilizes node features including latitude, longitude, and node degree, together with edge features comprising vehicle count, speed, and travel time to predict future traffic status. The proposed model uses Historical traffic information from the previous five snapshots. This historical data is incorporated to predict future congestion. It will dynamically update routing costs before path computation. An offline training stage generates the initial prediction model. Online training stage updates the model whenever traffic conditions changes by more than 10 percentage or traffic incidents occur. The proposed framework is evaluated on OpenStreetMap(OSM) road networks with 35, 149, and 222 Road Side Units (RSUs) using a simulator written using Python. Experimental results demonstrate that Dij-ST-GAT achieves higher prediction accuracy and F1-score while reducing travel time, end-to-end delay, and congestion exposure compared with conventional GAT-Dijkstra (Dij-GAT) approach, demonstrating its effectiveness and scalability for intelligent vehicular networks.