A knowledge-informed dynamic correlation modeling framework for lane-level traffic flow prediction

Lane-level traffic prediction forecasts near-future conditions at specific lane segments, enabling real-time traffic management and particularly aiding autonomous vehicles (AVs) in precise tasks such as car-following and lane changes. Despite substantial advancements in this field, some key challenges remain. First, the traffic state of a lane segment exhibits dynamic, nonlinear spatial correlation with other segments, making accurate modeling complex in real-world environments. Second, existing deep learning models depend heavily on specific datasets, leading to poor generalization. Third, while recent studies have shown that Large Language Models (LLMs) exhibit superior performance in generating reliable traffic prediction results, their direct application is hindered by inefficiency, high computational costs, and difficulties in capturing dynamic traffic features. To address these challenges, we propose the Knowledge-informed Dynamic Correlation Modeling (KIDCM) framework, which integrates pre-trained LLMs with traditional predictive methodologies to achieve a balance between generalization and prediction accuracy. Specifically, we introduce a General Spatial Dynamics Modeling (GSDM) method, which leverages the unbiased traffic data generated by LLM to analyze the general law dynamic spatial correlations. By integrating traditional time-series models with attention mechanisms, GSDM effectively models both linear temporal dependencies and nonlinear spatial interactions, ensuring robust generalization across varying conditions. Additionally, we develop a surrogate model that distills the traffic prediction function of LLMs. This surrogate model can be fine-tuned with small sample sizes, preserving the generalization advantages of LLMs while mitigating their typically high resource demands. Extensive evaluations demonstrate that our framework outperforms state-of-the-art models in terms of generalization, small-sample training, and computational cost.