A Novel Hybrid Car-Following Model Combining Kinetic Dynamics and Deep Learning Networks

Car-following (CF) model has been extensively studied from various aspects and by different methods in the past few decades. Moreover, with the help of big data and advanced deep learning technologies, recent work seeks data-driven car-following modeling. Nonetheless, the interpretability and reliability of the data-driven model need to be further addressed. In this paper, we propose a novel hybrid car-following model that combines traditional mathematical kinetic dynamics and a deep learning network, in which the mathematical model is used to statistically describe general human driving behavior and the deep learning network is used to capture the complexity of random driving behavior caused by specific human factors. In particular, we define a general network to train the residual data calculated by the estimated kinetic model and real driving trajectories. Specifically, the typical Intelligent Driving Model (IDM) is adopted as the mathematical model, and a Long Short-Term Memory (LSTM) neural network is used as the basic network to build our hybrid model - RES-ILSTM. We compare the calibration and prediction results with several models using the NGSIM dataset to evaluate the proposed model. The experiments show that our calibration algorithm achieves better accuracy of results and RES-ILSTM is able to maintain high modeling accuracy and performs well in terms of interpretability, data efficiency, and generalization.