【作者】 柴林；

【导师】 陆振波； 王树盛；

【作者基本信息】 东南大学 ， 交通运输（专业学位）， 2023， 硕士

【摘要】 准确了解和掌握道路交通安全风险及变化趋势是交通安全管理的重要内容。当前,道路交通事故风险相关研究多针对高速公路和城市道路,普通国省干道交通事故风险预测研究相对较少。论文以普通国省干道为研究对象,融合交通事故和天气等数据,分析了国省干道交通事故影响因素并提取相应特征,构建了交通事故严重程度预测模型,提出了一种考虑事故频数、严重程度以及路段特征的交通事故综合风险指标,实现了不同时空颗粒度下国省干道交通事故风险的预测。论文的主要研究工作如下。首先,基于可获取数据特点,论文提出了一种融合事故数据和天气数据的普通国省干道交通事故特征提取方法。考虑数据的时空相关特性,论文对现实可获取数据进行了预处理。接着,论文结合国省干道货车比例高、车速快等特点,分析了国省干道交通事故影响因素,从“人、车、路、环”四个方面提取了交通事故特征,为后续交通事故严重程度预测和交通事故风险预测建模提供了数据基础。其次,针对既有模型无法有效预测国省干道交通事故严重程度问题,论文构建了基于贝叶斯网络的国省干道交通事故严重程度预测模型,为后续交通事故风险预测提供了模型基础。具体的研究工作包括:提出一种考虑直接财产损失、人员伤亡情况、造成道路拥堵等因素的交通事故经济损失指标,对交通事故严重程度进行量化并分类;基于评分搜索方法和相关性分析方法,构建了事故严重程度与影响因素特征关系的贝叶斯网络,并采用最大似然估计算法进行参数学习;最后,对构建模型进行性能评估,并对影响交通事故严重程度的影响因素进行贝叶斯网络后验概率推理和敏感性分析。研究结果表明,论文提出方法可以有效预测国省干道交通事故的严重程度,准确性达92.4%。最后,针对既有方法无法有效预测国省干道交通事故风险及其变化趋势的问题,论文提出了考虑严重程度的国省干道交通事故风险预测方法,实现了在不同时空颗粒度下对国省干道交通事故风险的动态预测。具体的研究工作包括:构建了一种考虑事故频数、严重程度以及路段特征的交通事故风险指标;在划分预测区域并对输入数据进行时空匹配基础上,构建了基于事故严重程度预测模型的交通事故风险动态预测模型;最后,对构建模型进行性能评估,探讨在不同时空颗粒度下的交通事故风险变化趋势,研究对比不同时空颗粒度下模型的预测准确度。实例结果表明,论文提出方法可以有效实现对国省干道交通事故风险的动态预测。更多还原

【Abstract】 Accurate mastery of road traffic safety risks and changing trends are essential aspects of traffic safety management.Current research on road traffic accident risks primarily focuses on highways and urban roads,with relatively fewer studies on traffic accident risk prediction for ordinary national and provincial trunk roads.This thesis takes ordinary national and provincial trunk roads as the research object,integrates traffic accident and weather data,analyzes the influencing factors of national and provincial trunk road traffic accidents and extracts corresponding features,constructs a traffic accident severity prediction model,and proposes a comprehensive traffic accident risk index that considers accident frequency,severity,and road segment characteristics,achieving prediction of national and provincial trunk road traffic accident risk at different spatial-temporal granularities.The main research work of the thesis is as follows.Firstly,based on the available data characteristics,this thesis proposes a method for extracting features of ordinary national and provincial trunk road traffic accidents by integrating accident and weather data.Considering the spatiotemporal correlation of the data,the thesis preprocesses the available data.Subsequently,by combining characteristics such as high truck ratios and high speeds on national and provincial trunk roads,the thesis analyzes the influencing factors of traffic accidents and extracts features from the "people,vehicle,road,and environment" perspectives,providing a data foundation for subsequent traffic accident severity prediction and traffic accident risk prediction modeling.Secondly,to address the issue that existing models cannot effectively predict the severity of national and provincial trunk road traffic accidents,the thesis constructs a severity prediction model based on Bayesian networks,providing a model foundation for subsequent traffic accident risk prediction.Specific research work includes: proposing an economic loss index for traffic accidents that considers direct property damage,casualties,and resulting road congestion,quantifying and classifying accident severity;constructing a Bayesian network of accident severity and influencing factor features using scoring search methods and correlation analysis,and parameter learning using the maximum likelihood estimation algorithm;finally,performance evaluation of the constructed model and Bayesian network posterior probability inference and sensitivity analysis of factors influencing traffic accident severity.The research results show that the proposed method can effectively predict the severity of national and provincial trunk road traffic accidents,with an accuracy of 92.4%.Lastly,to address the issue that existing methods cannot effectively predict the risk and changing trends of national and provincial trunk road traffic accidents,this thesis proposes a risk prediction method that considers accident severity,achieving dynamic prediction of traffic accident risk at different spatial-temporal granularities.Specific research work includes:constructing a traffic accident risk index that considers accident frequency,severity,and road segment characteristics;constructing a dynamic traffic accident risk prediction model based on the accident severity prediction model,after dividing prediction regions and spatial-temporal matching of input data;finally,evaluating the performance of the constructed model,exploring traffic accident risk trends at different spatial-temporal granularities,and comparing prediction accuracy at different granularities.The example results show that the proposed method can effectively achieve dynamic prediction of national and provincial trunk road traffic accident risks.更多还原