基于机器学习技术的逐时雾事故判别气象模型

为进一步提高雾天交通安全气象保障精细化能力，以江苏、安徽高速公路雾事故多发路段为例，利用2012—2018年事故信息与气象资料，建立一种基于变量选择和特征提取的逐时雾事故判别支持向量机模型。模型参照递归特征消除思路选择事故发生时间、地理位置、气象环境等重要变量，使用主成分分析提取重要变量的主要特征，并以径向基为核函数、以网络搜索确定最优参数。结果表明：结合重要变量选择和主成分分析的支持向量机混合模型能够成功识别出训练集81.4%和测试集83.0%的事故样本，AUC分数均为0.946;判别效果优于支持向量机单独算法，以及仅基于重要变量选择或主成分分析的支持向量机算法；3个典型实例分析也说明该模型对于阶段性或持续性大雾天气下的交通事故发生有一定判识与警示意义。

An Hourly Meteorological Model for Fog Accident Discriminant Based on Machine Learning Technology

In order to further improve the ability of refined meteorological services for traffic safety in foggy weather, this study takes Jiangsu and Anhui expressway sections where frequent fog caused accidents happen as examples, with the application of the disaster information and weather data from 2012 to 2018 to establish a support vector machine hybrid model for hourly fog accident detection based on variable selection and feature extraction. The model uses the recursive feature elimination method to select the important variables from accident time, geographic location, and meteorological environment, and then extracts the main features of the important variables by principal component analysis. The radial basis is used as the kernel function, and the optimal parameters are determined by network search. The results show that this support vector machine hybrid model can successfully identify 81.4% of the accident samples in the training set and 83.0% of the test set, and the AUC scores are both 0.946. The ability to identify fog accidents is superior to the support vector machine algorithm and the support vector machine algorithm based only on main variable selection or principal component analysis. The analysis of three typical examples also shows that the support vector machine hybrid model has certain identification and warning significance for the occurrence of traffic accidents under periodic or persistent foggy weather.