Traffic Flow Prediction Using Deep Learning: Advances, Challenges, and Future Directions

Abstract

As urbanization accelerates globally and intelligent transportation technologies rapidly progress, the accurate prediction of road traffic has emerged as a critical issue in the field of smart transportation. This paper introduces a comprehensive review of current research on traffic flow prediction utilizing deep learning and related technologies. It analyzes the limitations of current methods, such as poor generalization to spatiotemporal heterogeneity, reliance on external influencing factors, data quality and quantity issues, and insufficient model explainability and computational scalability. Furthermore, the paper outlines the development trends of the field from the perspective of multimodal data fusion, ultimodal data fusion and federated learning. The study also discusses fuproposing a three-layer fusion framework of data, models, and systems. Emphasis is placed on ensuring data security and privacy through mture directions, including dynamic feature modeling, system deployment at the edge, and real-time prediction. By analyzing the architecture and challenges of current predictive models, this article offers theoretical direction and technological insights for the advancement of intelligent transportation technology.