Research On Traffic Flow Operation Risk Evaluation Based On Deep Learning

With the rapid development of China's economy,travel demands have increased significantly,the contradiction of road supply and demand has become increasingly prominent,and traffic accidents have frequently occurred.The severe road traffic safety situation urgently requires the researches on active prevention and control of road transportation network operational risks.In the actual road environment,different stages of the traffic accident will have different degrees of impact on the operation of the road network traffic flow.Therefore,it is necessary to explore the internal relation between traffic accidents and dynamic traffic flow characteristics,study and explore risk evaluation and prediction methods for the potential risks,occurrence process and post-effect stages of accidents,all these will contribute to proposing and implementing active safety prevention and control strategies for traffic accidents,improving the level of emergency management and in reducing the hazards of traffic accidents.This study is based on traffic flow operating data and accident data,using deep learning methods as research tools to make a deeply exploration the inherent relations between traffic flow operating characteristics and different stages of accidents,then establish real-time and effective traffic flow operating risk evaluation methods to realize accurate and effective estimations and predictions of different traffic accident stages.The results will help drivers to take more reasonable travel decisions and choose proper driving behaviors,also will reduce the probability of accidents and improve travel reliability;meanwhile,it can help traffic managers to take more comprehensive and effective traffic accident risk prevention and control measures.Focusing on the research purpose of traffic flow operation risk evaluation,the main research contents and results of this paper are as follows:(1)A traffic flow missing data filling network model—TGAINAiming at the problem the multi-modal distribution solution of missing traffic flow data and the uncertainty in the missing filling process,a new time-series generative adversarial network model—TGAIN is proposed to realize the filling of missing traffic flow data.The model builds a TGAIN network structure and uses adversarial generative training to learn the multi-modal distribution of missing data in traffic flow;a multiple filling strategy based on TGAIN is designed to deal with the uncertainty in the missing value filling process and improve the performance of the algorithm filling.Finally,actual traffic flow datasets are used to verify the missing data filling performance of the proposed method.(2)A traffic risk state deep clustering network model—TRDCNIn order to have a more effectively finding of the potential risk characteristics of traffic flow and improve the traffic risk status clustering effect,a new unsupervised deep clustering network model—TRDCN is proposed to realize the classification of traffic operation risk status.In order to solve the problem of mismatch in traditional state division methods due to the separation of feature extraction and clustering processes,TRDCN fuses feature extraction and clustering tasks into the End-to-End network model,and constructs a multi-channel autoencoding network to extract traffic risk feature information,and a deep feature clustering layer is designed to iteratively optimize the clustering effect and feature extraction process.Taking the actual accident data set as an example,the traffic risk state clustering results are statistically analyzed,and the accident risk level under each traffic state is quantified.The quantitative results of the traffic risk state are compared with the actual road accident situation to verify the model effectiveness of the division of risk status.(3)A traffic risk state prediction network model based on M-B-LSTMTo solve the uncertainty and over-fitting problem in the deep learning process caused by the random volatility and uneven distribution of traffic flow parameters,a new short-term traffic flow parameter prediction network model-M-B-LSTM is proposed;and verifies the traffic parameter prediction performance of the model through actual cases.Based on the M-B-LSTM traffic parameter prediction model,combined with TRDCN traffic flow risk state division network,M-B-LSTM+TRDCN traffic risk state prediction hybrid network structure is proposed;the actual traffic flow operation datasets and accident datasets are adopted,the comparative experiments verify that the proposed method can effectively predict the traffic risk state.(4)A spatial-temporal graph traffic abnormal detection network model—STGADCombining the propagation process and characteristics of traffic incidents in the road network in time and space,a new spatial-temporal graph traffic abnormal event network model—STGAD is proposed to realize the automatic detection of traffic abnormal events in the road network environment.The model abstracts the road network as a graph structure,and introduces the spatial-temporal convolution operation and the attention mechanism to capture the spatial and temporal dynamic characteristics of road traffic flow.At the same time,a multicomponent fusion prediction module and an event anomaly evaluation module are designed to realize the flow mode learning and prediction process and abnormal evaluation process based on prediction results respectively.Finally,the actual traffic flow data is used to verify that the proposed method can effectively detect abnormal traffic events in the road network environment.(5)A traffic incidents impact prediction network model—X-G-PGCFBased on the analysis of the evaluation indicators and factors of the impact of traffic incidents on traffic flow;borrowing and integrating the algorithm advantages of Xgboost,GAN and GCF,a new network model for predicting the impact of traffic incidents is designed—XG-PGCF.The model first uses Xgboost to screen the main accident characteristics according to the task contribution;in view of the incomplete and expandable characteristics of the accident characteristics in the information collection,the GAN-based accident feature enhancement module is designed to complement and reinforce the incomplete accident characteristics under actual conditions.Then,enhance and improve the deep forest model,adding a multi-granularity feature arrangement and a weighting mechanism for decision results to realize the prediction and estimation of the impact of traffic incidents.Finally,the actual accident data is used for empirical analysis,which verifies that the proposed method can effectively predict the impact of traffic incidents.To sum up,this study focuses on the research of traffic flow operation risk evaluation methods,and develops deep learning methods from five aspects: filling of missing traffic flow data,traffic risk state division,traffic risk state prediction,automatic traffic incident detection,and traffic incident impact estimation.The research results will enrich the theoretical system of traffic flow operation risk evaluation and provide more effective theoretical tools for the prevention and control of actual traffic incidents.