Research On Categorized Car-following State Prediction Model Based On Deep Convolutional Fusion

Car-following model aims to provide an in-depth understanding of traffic flow and driving behavior from a microscopic perspective.Traditional theory-driven car-following models are mainly described by physical parameters such as acceleration and headway.The sufficient physical parameters can improve the accuracy of car-following model but inevitably increase the complexity of car-following model.Deep learning based data-driven car-following models can improve the generalization and accuracy of car-following models with the powerful learning capabilities.However,the differentiation of vehicle driving style usually makes it difficult for a single model to accurately describe the following behavior under the influence of different driving style.Therefore,in order to propose a car-following state prediction model with high prediction accuracy and strong generalization ability,by combining the advantages of both theory-driven and data-driven car-following models,the influence of different driving style on car-following behaviors was analyzed,and a categorized car-following state prediction model based on deep convolutional fusion was proposed.Meanwhile,a simulation system based on data-driven car-following state prediction model was designed and implemented.The main contents are as follows:(1)Considering the shortcomings of theory-driven and data-driven car-following modeling approaches,a car-following state prediction model based on convolutional fusion of GRU and IDM(GRU-IDM-CNN)was proposed.The convolutional neural networks were utilized to nonlinearly fuse the acceleration outputs of GRU and IDM model,which improved the model prediction accuracy while solving the problem that data-driven methods cannot accurately fit the extreme cases of vehicle acceleration.The comparisons among the proposed model,the single model and the linear fusion model were executed in the vehicle trajectory dataset of NGSIM and based on the evaluation metrics of Root Mean Square Error(RMSE)and Mean Absolute Error(MAE).The test results showed that GRU-IDM-CNN could reduce the RMSE of position by 23.1% and the RMSE of velocity by 18.5%,compared to GRU-based model.And the RMSE of position was reduced by 9.5% on I-80 test set and by12.0% on US-101 test set,compared with AKF-LSTM-IDM,respectively.(2)Considering the influence of different driving styles on the following behaviors,a categorized GRU-IDM-CNN car-following model was proposed based on driving style characteristics.Firstly,the driving styles were classified into aggressive,standard and conservative by utilizing cluster analysis.Secondly,a driving style labeling module was introduced in the front of GRU-IDM-CNN.Finally,comparison experiments on I-80 dataset was conducted with GRU-IDM-CNN without considering driving styles.The test results showed that the RMSE of position in the categorized fusion car-following model was reduced by 7.3% and the RMSE of velocity was declined by 16.3%.Therefore,the categorized GRU-IDM-CNN car-following model based on driving style features was helpful to further improve the prediction accuracy of car-following states.(3)A simulation system for data-driven car-following state prediction was constructed based on the categorized car-following state prediction model.Considering the weakness of data management and human-computer interaction in the construction of data-driven car-following model,Flask framework,Vue front-end framework and My SQL database were utilized to design the simulation system and realize the three functional modules: data management,model training and result prediction.The proposed deep convolutional fusion-based categorized car-following state prediction model can well improve the accuracy of car-following state prediction.The developed simulation system for data-driven car-following state prediction can further enhance the simulation efficiency of the data-driven car-following model.The relevant research results in this paper will provide a theoretical basis for the subsequent research and its application of the car-following behaviors modeling for connected vehicles and the longitudinal collision avoidance for autonomous vehicles.