Connected Vehicles Traffic Prediction

The development of information technology is transforming the automotive industry, leading to the emergence of connected vehicles that integrate advanced sensing, communication, and control systems to enhance connectivity between vehicles, people, and infrastructure. Connected vehicles enable real-time data exchange, thus improving safety, efficiency, and urban traffic management. They collect and analyze traffic data to optimize routes, reduce congestion, and support smart city planning through interactions with infrastructure like intelligent signals. Traffic flow prediction, a critical area in connected vehicle research, combines vehicle networks, big data, and AI to address spatial-temporal variability and external factors such as location, weather, and events, which impact traffic patterns. This book presents five chapters, starting with foundational knowledge on AI and graph neural networks in connected vehicles, followed by specific traffic prediction methods, and concluding with a discussion on future challenges.

This chapter explores the application of spatial-temporal graph convolutional networks (STGCNs) for connected traffic flow prediction. By integrating spatiotemporal dependencies with graph structures, STGCNs are effective in capturing complex static and dynamic spatial correlations within traffic flow data while accounting for historical and multi-factor dynamic changes. This chapter covers various methods, including the static and dynamic spatial correlation neural network (SDSCNN), spatial-temporal complex graph convolution network, prior knowledge enhanced time-varying graph convolution network, and residual attention-enhanced multi-factor graph convolutional network. Each approach aims to improve prediction accuracy by leveraging distinct mechanisms for capturing spatiotemporal features and addressing the heterogeneity introduced by multi-dimensional factors.

This chapter focuses on the role of dynamic and multi-graph models in connected traffic prediction, emphasizing how multi-graph convolution networks aggregate various graph structures to model spatiotemporal correlations. Dynamic multi-graph approaches capture dependencies across different graph structures and temporal scales, reflecting the evolving nature of traffic flow and its complex spatial structures. Covered methods include the multi-sequential temporal convolution gated graph neural network, dynamic spatiotemporal correlation graph convolutional network, and dynamic multi-graph synchronous aggregation framework. By incorporating multi-graph convolution structures, these approaches efficiently handle multi-scale spatiotemporal data, enhancing the models’ adaptability and prediction stability in complex traffic environments.

This chapter introduces hybrid models that integrate both local and global spatial correlations, modeling the dynamic changes in traffic flow across different spatial scales. The approaches presented focus on combining local feature extraction with global spatial characteristics through hybrid graph convolution structures to achieve more accurate predictions of connected traffic speed and flow. The key methods discussed include the hybrid model integrating local and global spatial correlation, urban road network connected traffic speed prediction model based on global spatiotemporal characteristics, and the perturbation learning enhanced U-shaped multi-graph convolutional network (PLU-MCN). By effectively combining multi-level spatial information, these methods improve the models’ ability to capture both macro and micro features, resulting in more flexible and precise predictions.

This chapter discusses the challenges and future directions in graph neural networks (GNN) for traffic prediction. Key challenges include data heterogeneity, multimodality, and dynamism, highlighting the need for integrating multi-source data, handling diverse transportation modes, and modeling evolving traffic patterns. GNN models, while effective, face challenges in interpretability, requiring frameworks that combine traditional traffic theories with data-driven insights. Further, small sample learning and uncertainty quantification remain critical issues. The book offers future research recommendations, such as optimizing model efficiency, incorporating emerging technologies like federated learning, and improving interpretability, ultimately aiming to enhance traffic prediction for intelligent transportation and connected vehicle systems.