A Review of Time Series Prediction Research in the Transportation Field

Abstract

This review looks at the new progress of time series prediction in the field of transportation, especially the multimodal fusion driven by the Large Language Model (LLM) and the new challenges encountered. It reviews the development of forecasting methods, from traditional statistical models to deep learning methods, and now to the framework based on Transformer. The research especially points out the great potential of large models in dealing with ultra-long-term prediction, cross-modal reasoning and uncertainty quantification, and also talks about how they can adapt to spatiotemporal traffic data through discrete coding and cue engineering. We also analyzed the multimodal strategy of integrating GPS trajectory, video surveillance and social media texts, which can help us better understand traffic scenes and find abnormal situations. The main applications, such as traffic flow forecasting, adaptive signal optimization, autonomous driving decision support and safety early warning system, all show the practical functions of these technologies. We also carefully analyzed the challenges, such as high computing cost, limited real-time reasoning, difficult model understanding, and privacy issues. The future direction will focus on the development of edge cloud collaboration, federated learning and digital twin integration to promote the progress of intelligent transportation systems. This review summarizes the current progress in an all-round way, and points out the way to realize an extensible, easy-to-understand and privacy-preserving traffic forecasting scheme.