AI-Assisted Identification and Prediction of Dynamic Vehicular Communication Network States Using Taxi GPS Trajectories

Abstract

Empirical monitoring of potential Vehicle-to-Vehicle (V2V) connectivity remains less studied than traffic-flow prediction and simulation-based vehicular networking. This study develops a taxi-Global Positioning System (GPS)-driven framework for identifying and predicting dynamic vehicular communication-network states in Hohhot, China. Taxi positions were aggregated into 672 15-min snapshots and converted into potential V2V proximity networks under communication radii of 50, 100, and 150 m. Standardized topology features were clustered into sparse-fragmented, transitional, and relatively-connected states, and the next-state prediction was evaluated using Persistence, Random Forest, and eXtreme Gradient Boosting (XGBoost). Increasing the radius raised the mean degree from 3.2930 to 9.3370, whereas the largest connected component ratio remained below 0.05, indicating local connectivity but global fragmentation. State self-transition probabilities were 0.8680, 0.8716, and 0.7713 for 50, 100, and 150 m, respectively. XGBoost performed best at 50 m, Random Forest at 100 m, and Persistence at 150 m. Under a common supplementary horizon-aligned protocol, average best Macro F1 decreased as the prediction horizon increased from 15 to 30, 60, and 90 min. The framework therefore provides a lightweight basis for short-term communication-state monitoring, while more realistic wireless constraints and fully inductive validation remain future work.