Research Progress and Future Prospects of Large Language Models in the Field of Traffic Flow Prediction
DOI:
https://doi.org/10.61173/wjqdgc28Keywords:
Large language models, Traffic Forecasting, Data Transformation, Model ScaleAbstract
With the enhancement of computer performance, the predictive capabilities of large language models have also improved, and these models have been applied in many fields. Currently, the application of large language models in the field of traffic flow prediction has entered a rapid development stage. Thanks to its powerful feature learning and generalization capabilities, it has provided new ideas and methods for traffic flow prediction, and a series of achievements have been achieved. This article reviews the research progress of mainstream large language models in the field of traffic flow prediction. This introduces the background and significance of applying large language models to traffic flow prediction, as well as the current mainstream large language models, and elaborates on the relevant model architectures and methods in detail. Analyzed its advantages in enhancing predictive accuracy and interpretability, as well as the challenges it faces. At the same time, an outlook on the future development directions of large language models in this field has been provided, aiming to offer a reference for further research and application of traffic flow prediction in intelligent transportation systems.
References
gradual ineffectiveness of traditional traffic planning mod- 15.3 2024: 1-45. els. Leveraging natural language processing capabilities, [3] Miglani A and Neeraj K. Deep learning models for traffic large language models transform complex traffic flow data flow prediction in autonomous vehicles: A review, solutions, and
into structured language, capturing spatial and temporal challenges. Vehicular Communications 20 2019: 100184. features to enable predictions, thus providing new insights [4] Wu Y Y et al. A survey on origin-destination flow prediction. for intelligent traffic planning. 2024 11th International Conference on Soft Computing & At present, research in this field has made certain prog- Machine Intelligence (ISCMI). IEEE, 2024. ress. The TP-LLM model demonstrates advantages com- [5] Ren Y L et al. TP-LLM: A traffic prediction framework pared to traditional models, while the advanced ST-LLM based on pretrained large language models. arXiv preprint achieves explicit spatial modeling, parameter efficiency, arXiv:2403.02221. 2024 . and robust cross-domain upgrades through graph-en- [6] Gao N et al. Generative adversarial networks for spatiohanced spatiotemporal LLM and LoRA fine-tuning tech- temporal data: A survey. ACM Transactions on Intelligent
niques. It outperforms most existing models in real data- Systems and Technology (TIST) 13.2 2022: 1-25. sets as well as in few-shot and zero-shot scenarios. [7] Yu Y et al. A review of recurrent neural networks: LSTM However, research still faces key issues: noise affects the cells and network architectures. Neural computation 31.7 2019: accuracy of predictions and the reliability of decisions 1235-1270. at the levels of data input, training, and output; improve- [8] Liu C X et al. ST-LLM+: Graph Enhanced Spatioments in prediction accuracy come with increased model Temporal Large Language Models for Traffic Prediction. IEEE complexity, leading to a decrease in interpretability; infor- Transactions on Knowledge and Data Engineering. 2025 . mation loss can easily occur during the data transforma- [9] Zhao Y & Zhou Y. B. Research on the application of tion process, hindering practical application implementa- reinforcement learning in traffic flow prediction. Information tion. and Computers (Theoretical Edition), 36(03) 136-138. 2024. In response to these issues, the study proposes a multi-lev- [10] Ferianc M et al. Navigating noise: A study of how noise el data integration approach, semi-supervised learning, influences generalisation and calibration of neural networks. and ensemble prediction to address noise problems; an arXiv preprint arXiv:2306.17630. 2023. improved attention mechanism is employed to balance [11] Gloeckle F et al. Better & faster large language models via accuracy and interpretability; and data transformation multi-token prediction. arXiv preprint arXiv:2404.19737. 2024. is optimized through semantic rule mapping and human [12] Yuan X L, Xin L and Jian Z. The development of new feedback. In the future, efforts will focus on optimizing energy vehicles for a sustainable future: A review. Renewable
model structures, upgrading noise handling, enhancing in- and Sustainable Energy Reviews 42 2015: 298-305.
Downloads
Published
Issue
Section
License
Copyright (c) 2025 by the authors.
This work is licensed under a Creative Commons Attribution 4.0 International License.
