Forecast-Driven UAV Deployment and Deep Q-Network Routing for Scalable Vehicular Networks

Abstract

Traditional VANETs struggle with coverage gaps and routing instability due to high mobility and static infrastructure. This paper proposes a congestion-aware, UAV-assisted framework integrating Predictive K-Means clustering for deployment and Deep Q-Network (DQN) routing with Conservative Q-Learning (CQL). The deployment strategy uses vehicle velocity vectors to proactively position UAVs at forecasted network hotspots. Simultaneously, the routing agent utilizes a multi-objective reward function and a novel Packet Survivability Score to prioritize endangered packets, while CQL ensures stable convergence by suppressing Q-value overestimation. SUMO simulations demon-strate that the framework outperforms PSO and Cell-based baselines, achieving over 90% Packet Delivery Ratio (PDR) in dense scenarios. Results show a reduction in end-to-end delay and average hop counts by over 60% and 50%, respectively, providing a scalable solution for resilient Intelligent Transportation Systems.