A Hybrid CNN–LSTM–DQN Framework for Intelligent Routing and Traffic Prediction in Internet of Vehicles under Dynamic Urban Environments

Abstract

The fast advancement of Internet of Vehicles (IoV) technologies has presented new possibilities of managing traffic in dynamic cities in an intelligent way. But the nature of vehicular networks, which is characterized by high mobility, non-stationary traffic flows and various communication constraints, makes traditional prediction and routing techniques insufficient. This paper presents a new hybrid architecture that is synergistic in combining Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks and Deep Q-Networks (DQN) in order to tackle both issues of precise traffic prediction and optimal route in the IoV system. LSTM module learns long-range temporal dynamics and periodic traffic patterns necessary to predict traffic using multiple steps and CNN module learns hierarchical spatial features of grid-based representations of traffic. DQN module represents the vehicle routing as a Markov Decision Process (MDP), which is learned by reinforcement learning interactions with a realistic simulator based on the SUMO-based urban traffic. The extensive experiments, performed on three benchmark datasets, i.e., UCI ML Repository Traffic Dataset, synthetically-generated VANET mobility trace, and SUMO urban simulation environment, prove the better performance of the proposed framework than the state-of-the-art baselines. In particular, the hybrid model reduces traffic prediction RMSE by 22-25, traffic average vehicle traveling time by 23-30, and network congestion index by 33-36 with the situation of different traffic density such as low-density (200-500 vehicles), middle-density (500-1,000 vehicles), and high-density (1,000-1400), the suggested framework provides a realistic and scalable framework to manage real-time IoV-based traffic including its ability to operate in varying urban traffic scenarios.