A hybrid stochastic car-following model based on data-driven and theory–driven method

As a fundamental framework in micro traffic flow analysis, car-following (CF) behavior modeling is essential for understanding the mechanisms underlying traffic congestion formation and evolution. Data-driven CF models, based on machine learning, offer high prediction accuracy but lack interpretability and safety. In contrast, theory-driven CF models, rooted in micro traffic flow theory, provide clear physical interpretability and inherent safety but exhibit limited predictive accuracy. Furthermore, deterministic CF models struggle to capture stochastic factors influencing human driving behavior. To overcome these limitations, this study proposes a 2I-PBL hybrid stochastic CF model, integrating the data-driven PSO-Bi-LSTM model with the stochastic theory-driven 2D-IDM model. Firstly, the Bi-LSTM model is enhanced using the particle swarm optimization (PSO) algorithm, resulting in the PSO-Bi-LSTM machine learning CF model with superior predictive performance. Secondly, a combined forecasting method is employed to fuse the PSO-Bi-LSTM and 2D-IDM models, incorporating stochastic factors. This integration produces the 2I-PBL hybrid model, which accurately predicts driving behavior while ensuring safety. Finally, the proposed 2I-PBL model is validated by demonstrating its effectiveness in capturing heterogeneous driving behaviors.