A novel hybrid macroscopic fundamental diagram-informed deep learning method for lane-level traffic prediction

Lane-level traffic prediction forecasts traffic conditions at the lane level, offering detailed insights into traffic flow, congestion, and vehicle behavior on individual lanes. This enhanced granularity enables autonomous vehicles (AVs) to perceive the surrounding traffic environment with greater precision, supporting autonomous driving functions like car-following and lane-change decisions. However, current research faces challenges that require further investigation. First, most studies utilize data-driven methods for lane-level traffic prediction, which are prone to common issues such as overfitting and limited generalization. Second, the existing prediction methods often fail to model the spatiotemporal correlations between lanes deeply, limiting their ability to fully capture the dynamic nature of traffic. To address these issues, we propose a novel hybrid macroscopic fundamental diagram (MFD)-informed deep learning method (MFD-IDL) for lane-level traffic prediction. By integrating the MFD to represent the physical relationship between traffic flow and density, we embed this correlation within our deep-learning model training framework to mitigate the influence of historical data quality on prediction accuracy and bolster model generalization ability. Moreover, we introduce a multi-scale graph fusion technique to fully model the spatiotemporal features of traffic data across lanes by using multi-source traffic states and lane-level traffic network topology. Extensive experiments based on real-world datasets demonstrate that MFD-IDL achieves a 14.3% reduction in RMSE and a 15.3% improvement in MAE compared to state-of-the-art baselines. Furthermore, empirical results indicate that the proposed method not only enhances computational efficiency but also exhibits low overfitting and superior generalization.