Abstract: This research aims at exploring the influence of multi-scale data sets on real-time traffic crash risk modeling towards freeways, and realizing the real-time risk model transplantation for freeways with spatial differences. First, multi-scale data sets were constructed via extracting freeway sections with different detector characteristics: highresolution data sets, small-sample data sets, low-resolution data sets, and data sets with spatial differences under the same scale conditions (both are high precision and large sample size); Furthermore, the influence of various sample sizes on the prediction performance of the traffic crash risk model was quantified by Bayesian Logistic regression, and statistical methods and machine learning methods were introduced to model the high and low resolution data sets respectively. Finally, the real- time traffic crash risk migration model based on the Bayesian updating method was established, and the freeway real-time crash risk prediction model was spatially transplanted, simultaneously its reliability was verified. The results show that: the performance of the model based on Bayesian Logistic regression improves with the increasing sample size; under the condition of high resolution data, the Area Under Curve (AUC) values of the Bayesian Logistic regression model and Random Forest- Support Vector Machine (RF- SVM) model are 0.092 and 0.037 higher than those under the condition of low resolution data, respectively; in the spatial migration with various data resolution, the AUC value of the low-resolution road segment model can be improved from 0.645 to 0.714 by the Bayesian updating method, and in the spatial migration with the same data scale, the AUC value of the updated road segment model can be improved from 0.737 to 0.751 by applying the Bayesian updating method. The conclusions indicate that: the model from the freeway section with a big sample size can boost the model classification accuracy but cannot significantly improve the performance of the prediction model, the results have some fluctuations, while the model from the freeway section with high data resolution can have higher accuracy of classification and prediction of the model; statistical methods have more advantages in model interpretation, and machine learning has better prediction performance under the condition of low resolution data; The Bayesian updating model can improve the accuracy of model spatial transplantation to a certain extent.

Key words: traffic engineering, dynamic traffic safety, model transplantation, freeway, Bayesian Logistic regression; machine learning

摘要： 本文旨在探究不同尺度数据集对高速公路实时事故风险建模的影响，并实现不同数据特征路段间事故风险模型的空间移植。首先，提取不同特征路段检测器信息，以动态交通流匹配事故数据构建多尺度数据集：高精度数据集、小样本数据集、低精度数据集以及同尺度条件下（同为高精度大样本量)的空间差异数据集；进而，通过贝叶斯Logistic回归量化不同样本量对事故风险模型预测性能的影响；分别采用统计学和机器学习手段建模分析高精度和低精度数据集；最后，基于贝叶斯更新方法建立实时事故风险移植模型，对高速公路实时事故风险预测模型进行空间移植，并验证其可靠性。结果显示：贝叶斯Logistic回归的性能随着样本量增大而有所提升；高精度数据条件下，贝叶斯Logistic回归和RF-SVM(Random Forest-Support Vector Machine)模型的AUC(Area Under Curve)值比低精度条件下分别高出0.092和0.037；在不同数据精度的空间移植中，贝叶斯更新方法可令低精度路段模型的AUC值从0.645提至0.714，在相同数据尺度的空间移植中，该方法可将被更新路段模型的AUC值从0.737提至0.751。结论表明：高样本量路段的模型虽然具备较高的分类准确度，但无法稳定地提升模型的预测性能，而高精度数据路段的模型可获得更高的分类准确度和预测精度；统计学方法在模型解释层面更具优势，机器学习手段在低精度数据路段下的预测性能更好；贝叶斯更新模型能够在一定程度提升空间移植效果。

关键词: 交通工程, 动态交通安全, 模型移植, 高速公路, 贝叶斯Logistic回归, 机器学习