Study On Crash Frequency Prediction Considering The Time-varying Characteristics Of Factors

In recent years,road traffic safety is still facing huge challenges.It is necessary to build the crash frequency model developed by traffic,weather conditions and roadway geometric,to predict crash number for a given long period.The outcomes of the crash frequency model are used to identify hotspot,evaluate road safety and propose effective countermeasures.However,in crash frequency prediction studies,since the dependent variables are the total number of crashes in a long period,time-varying factors,such as traffic and weather conditions,are also aggregated over such a long period.As a result,traditional crash frequency model may lead to lost the influence of the variation of timevarying factors on crash frequency.If the time-varying factors with smaller time scales is used to predict the long-period accident frequency,which can consider the influence of the characteristics of the time-varying factors,the accuracy of the crash frequency may be improved.Therefore,with refined-scale data including daily-based,hourlybased and 5min-interval-based traffic and weather data,this thesis proposed hourlybased method,daily-based method and the crash-risk-based method,respectively.The evaluation of the proposed methods for a long period of crash frequency prediction is also performed.First,crash,time-varying factors and geometric information were aggregated into four types: 5min intervals,hourly,daily and monthly.After collecting all the data,four datasets including real-time-based dataset,hourly-based dataset daily-based dataset and monthly-based dataset were established.Secondly,according to traffic and weather data at 5min intervals,the real-time crash model and the steps of real-time crash prediction,the crash-risk-based method was proposed by this thesis.Meanwhile,the true-crashlikelihood-based method,which also uses real-time crash model and is proposed by other study,was introduced.Then,according to hourly data,daily data and traditional crash frequency model,hourly-based method and daily-based method were proposed.And monthly-based method which is traditional crash frequency prediction method,was also inroduced.Finally,after the methods were estimated,the comparison of forecasting accuracy among the crash-risk-based method,the true-crash-likelihoodbased method,hourly-based method,daily-based method and monthly-based method were performed.The results showed that among the five methods,the true-crashlikelihood-based method has the best accuracy.It indicated that microscopic traffic and weather data at 5 min intervals can reflect the influence of time-varying factors on crash frequency,which was beneficial to increase the accuracy of monthly crash frequency prediction.Comparing hourly-based method,daily-based method and monthly-based method,it was found that the accuracy of monthly-based method is the best.