Research On Asymmetric Driving Behavior Based On A Hybrid Physical Model-driven And Data-driven Approach

In congested driving environments,the decline in traffic capacity,traffic oscillation,and even traffic accidents can be attributed to drivers’ asymmetric driving behavior.Microscopic traffic flow theory,the essential characteristics of following behavior,the influencing factors of traffic congestion generation,and the prevention and control measures of traffic congestion are widely studied by scholars.However,it is difficult to promote the existing physical carfollowing models because of the large error between them and the real driving environment.And the physical phenomenon of acceleration and deceleration asymmetry cannot be explained by the data-driven car-following models,although these models are accurate in prediction.Therefore,a hybrid physical model-driven and data-driven approach combining the advantages of the above two types of models is developed to study the asymmetric driving behavior during car-following.First of all,based on the concept of asymmetric driving behavior theory,the microscopic expressions of acceleration and deceleration curves are proposed,and the driving states of the car-following vehicle are classified into five macroscopic traffic phases: free flow,acceleration,deceleration,stationary and coasting.On this basis,the main causes of the asymmetry phenomenon are explored,including differences in driving style and reaction time,performance differences between vehicles,and discrete driving characteristics during acceleration and deceleration.Secondly,the development process of the optimal velocity model is presented,and the asymmetric full velocity difference model(AFVDM)is established considering the effect of asymmetric driving behavior.The stability of the model AFVDM is verified by loop simulation experiments.Furthermore,the model parameters were calibrated by filtering the suitable carfollowing data from the NGSIM US-101 dataset.The results show that the model with a better fitting,smaller calibration error and validation error is the model AFVDM that considers the asymmetric properties.After that,a hybrid physical model-driven and data-driven car-following model is constructed,based on the data-driven model Bi LSTM network framework and the computational graph of the physical car-following model.The training and testing of the hybrid physical model-driven and data-driven car-following model are accomplished by selecting the appropriate network parameters,loss function,and training algorithm.The results show that the advantages of the robustness of the physical model and high prediction accuracy of the datadriven model are inherited by the hybrid physical model-driven and data-driven car-following model.In addition,the model with the most accurate prediction of acceleration and trajectory is the hybrid physical model-driven and data-driven car-following model.Finally,the micro-characteristics and macro-characteristics of asymmetric driving behavior are selected as indicators to further evaluate the prediction ability of the physical model,data-driven model,and hybrid physical model-driven and data-driven model.The results demonstrate that the best model for calculating intensity differences,plotting hysteresis curves,capturing discrete driving characteristics,restoring stop-and-go traffic phenomena,and simulating the propagation of congestion is the hybrid physical model-driven and data-driven car-following model.