A Novel Graph Convolution–Based Immediate Neighborhood Extraction in VANET Stability Enhancement

The promise of vehicular ad hoc networks (VANETs) to improve traffic safety and pave the way for autonomous driving systems has attracted a lot of interest. Cluster heads (CHs) and stable cluster formation are essential components of VANET. However, due to the dynamic, unreliable nature of VANET (unstable drivers), it becomes a challenge for efficient clustering and CH selection. To address this, we redefine the clustering problem using immediate neighborhood extraction using graph convolution, a graphical deep learning technique, to generate low-dimensional embeddings. The model is trained on real-world traffic pattern data from the Connaught area of Delhi, collected using Simulation of Urban Mobility (SUMO), with attributes such as location, ID, and speed. We introduce first-order approximation method for extracting the immediate vehicle neighborhood. Here, the adjacency matrix of the vehicle graph is enhanced with self-connections using convolution layers. These embeddings are then used for stable CH selection and clustering. We introduce novel driver behavior–driven CH selection parameter based on cluster impulsiveness value (CIV) along with relative speed, cluster neighbor consistency strength (NCS), and CH residency duration (CHRD), to maintain reliable vehicles as CH promoting high cluster lifetime and achieve high network performance. The paper employs gray relational analysis (GRA) to refine CH selection, ensuring robust performance by mapping the above CH parameters to ideal cluster characteristics and generating a CH. Experimental result shows that the proposed model achieves 92% of cluster lifetime and achieves 18.46% improvement in throughput and low packet delay compared to conventional methods, validating the effectiveness of advanced deep learning approaches in optimizing VANET operations in dynamic urban environments.