Research On Stability And Generalization Ability Of Car-Following Model Based On Deep Learning

The car-following model is one of the theoretical foundations for the algorithm of the car-following function in the future advanced driving assistance system of automobiles.The rapid development of machine learning and the reduction of difficulty in obtaining high-fidelity vehicle trajectory data have opened up new ideas for the research of car-following models.Existing research mainly focuses on modeling car-following behavior,but there is still insufficient research on the stability and generalization ability of the model.A car-following model with higher stability and stronger generalization ability can not only reduce the collision risk during the car-following,but also improve vehicular riding comfort.Therefore,this paper conducts a study on the stability and generalization ability of car-following models based on deep learning.The main research content and results are as follows:1.Research on the extraction and processing methods of car-following data.Research was conducted on the extraction and processing methods of car-following data,and the sources,accuracy,extraction standards,and processes of car-following trajectory data used in modeling were analyzed,this provides a basis for the developing and parameter calibration of data-driven car-following models in the future.Studying car-following behavior and characteristics through natural driving datasets is more conducive to analyzing the characteristics of drivers and traffic flow regularities.2.Research on the developing and parameter calibration of car-following models.The car-following models,LSTM,CNN-LSTM,and PSO-LSTM,were developed respectively,and their parameters were calibrated.The LSTM car-following model was established based on the LSTM network algorithm,its input and output variables are determined after analyzing the correlation coefficients of various parameters in the car-following trajectory data and the characteristics of car-following behavior,and the model is configured based on the nonlinear characteristics during the car-following.Cross validation method was used to calibrate the model parameters.After analyzing the impact of various parameters on model performance,a CNN-LSTM car-following model was developed by combining a CNN and a LSTM network algorithm.Firstly,a CNN is used to extract car-following behavior features,and then the LSTM network for driving behavior prediction to improve the LSTM model’s ability to extract car-following behavior features.In response to the difficulty of neuron learning tasks caused by the difference in data volume in each car-following period and the difficulty of model convergence caused by the discrete Batch size values distribution during the training process,and the forgetting rate is one of the key parameters in the LSTM unit forgetting gate,the PSO algorithm is selected to optimize the parameters of the LSTM network(number of neurons,Batch size,forgetting rate),and the optimized parameters are used to establish the PSO-LSTM car-following model.3.Research on the stability of CNN-LSTM car-following model and PSO-LSTM car-following model.Conducting the stability study of the car-following pair and car-following queue,respectively.The simulation experiment of the car-following pair can analyze the characteristics of driver’s car-following behavior from a micro level,which is beneficial for exploring the internal mechanism of car-following behavior.Multi vehicles queuing is a bridge between micro driving behavior and macro traffic flow.Conducting simulation experiments on multiple groups car-following pair and car-following queue,the stability of CNN-LSTM car-following model and PSO-LSTM car-following model was studied by analyzing the distribution law of test errors for each car-following period and calculating the ratio of the second norm of acceleration of each following vehicle to the second norm of acceleration of the head vehicle during the car-following.The results show that the CNN-LSTM car-following model and the PSO-LSTM car-following model have better stability than the LSTM car-following model.Both the CNN-LSTM car-following model and the PSO-LSTM car-following model can improve the stability of the platoon in congested traffic.4.Research on the generalization ability of CNN-LSTM car-following model and PSO-LSTM car-following model.The generalization ability of the car-following model in this paper includes the ability to reproduce heterogeneity,the ability to reproduce traffic hysteresis phenomena,and the ability to adapt to mixed traffic flow.This paper studied the generalization ability of CNN-LSTM car-following model and PSO-LSTM car-following by conducting simulation experiments on heterogeneous driving behavior,traffic hysteresis phenomenon,and mixed traffic flow.The results show that the generalization ability of the PSO-LSTM car-following model is stronger than that of the CNN-LSTM car-following model;the heterogeneity of driving behavior is one of the key reasons for the phenomenon of traffic hysteresis;the PSO-LSTM car-following model can not only accurately reproduce the hysteresis phenomenon in congested traffic flow,but adapt to car-following scenarios in mixed traffic flow.