Study On Freeway Real-time Crash Risk Modeling And Reliability Of Model Transferability For Multi-scale Data Conditions

Traffic safety has always been an important but difficult point in the work of traffic management departments with the rapid development of freeways.As dynamic traffic data become easier to acquire,"active" traffic safety management has been a focus of current research,which improves the defect of traditional "passive" traffic safety research that can’t timely and effectively identify the risk of traffic crashes.However,in the current active traffic safety research,the type of dynamic traffic data is relatively single.Researches on real-time crash risk modeling based on multi-scale data are not taken into consideration;The influence of limited traffic data on the realtime crash risk prediction model is ignored;And the reliability of model transplantation of risk prediction models under limited data and different spatial conditions is paid little attention to.It keeps the application of active traffic safety management away from elaborative guidance.Based on the above problems,this paper has done the following research:(1)Freeway traffic data preprocessing and multi-scale data set constructionAiming at the analysis and prediction of real-time crash risk,a series of relevant data processing steps were established,including crash data extraction,traffic flow data extraction,loop data extraction,traffic flow variable selection,loop location selection and data matching.And a multi-scale data set was constructed for real-time crash risk analysis and prediction.(2)Real-time crash risk analysis and prediction with small sample dataUnder the condition of different sample sizes of traffic crashes,significant crash symptom factors of small sample data sets were screened based on Logistic stepwise regression,and comparative analysis was conducted.Then,Bayesian Logistic real-time crash risk prediction models were constructed and the differences of prediction models under the condition of different sample sizes were analyzed.The results showed that:When the sample sizes were different,the significant crash symptom factors were different,but the upstream loop speed(up＿s)was the same.In general,the performance of the real-time crash risk prediction model based on Bayesian Logistic regression improved with the increase of the sample size.(3)Real-time crash risk analysis and prediction under different precision data conditionsBased on statistical method and machine learning method,the high precision and low precision data sets were analyzed.Based on Logistic stepwise regression method and random forest algorithm,the two data sets were screened for crash symptom factors.The results showed that: in the same screening method,the risk variables of different precision data sets were different.In the same data set,the stepwise regression method can get significant crash symptom factors,while the random forest algorithm can only get the importance of variables,which requires subjective factor screening.Based on Bayesian Logistic regression and support vector machine,real-time crash risk prediction models were constructed for the two data sets respectively.The results show that the models constructed with high precision data set had better performance.The AUC value of high precision data set of the Bayesian Logistic regression was 0.737,0.081 higher than that of the low precision data set.In the support vector machine model,the AUC value of high precision data set was 0.85,0.05 higher than that of the low precision data set.In addition,the prediction result of machine learning algorithm was better than statistical method,but Bayesian Logistic had the advantage of quantitative interpretation.(4)Reliability analysis of freeway real-time crash risk prediction model transferabilityBased on Bayesian Logistic regression and Bayesian method,the real-time crash risk prediction models constructed with different precision data sets and different spatial regions were validated and analyzed.The results show that: Based on the Bayesian updating method,the model parameters of the low precision data set were updated to the real-time crash risk prediction model of the high precision data set,and the AUC value of the model was increased from 0.656 to 0.682.At the same time,the model parameters based on I-405 updated the real-time crash risk prediction of I-5,which improved the AUC value of the model from 0.737 to 0.751.The analysis result showed that based on the Logistic regression model of Bayesian method,the prediction performance can be improved to a certain extent when new data is obtained.