摘要 高速公路具有运行速度快、运输成本低、通行能力大、运输效率高等特点，以快捷、安全在国民经济建设和日常生活中起着重要的作用。发达国家高速公路事故数量约为一般公路的30~51%，而我国高速公路每百公里交通事故数量比一般公路高2倍以上，安全形势十分严峻。近十几年来，通过加强安全教育、严格法律法规等诸多措施，高速公路的交通安全状况有了明显的改善；但近几年高速公路事故率下降趋势相对平缓甚至有所提高，因此需要对传统的研究方法进行改进，以更加有效地提高高速公路安全管理的针对性和有效性，降低高速公路交通事故数量。 本研究通过分析高速公路不同路段的道路属性、交通参数等对相应路段事故发生的影响程度，从而建立高速公路不同类型路段的事故预测模型，以供公路管理部门、交通管理机构制定相应的管理对策、提高应急救援措施的针对性和合理性；根据高速公路事故预测结果结合经验贝叶斯（EB）方法对行车危险路段进行判定，并为改造优先性决策提供数据支撑。 论文依托广东省公路管理局科技项目“道路交通事故黑点评估及综合治理技术研究”、广东省交通运输厅科技项目“高速公路事故预测及综合治理决策系统研究”，研究工作中得到合作单位的大力支持，获得了大量的高速公路设计资料、事故资料、交通流量等数据，其中包括27条高速公路2200多km，近5年发生的13.5万余起交通事故事故资料。研究内容主要包括事故资料完整性恢复、特征路段分类、事故预测模型及事故修正系数（CMFs）的确定、标定流程、预测结果的修正及应用等工作。 不同管理部门的事故资料，因其侧重点不同，数量和质量都会存在一定的差异；论文利用数据挖掘中异常检测及关联分析技术对路政及交警部门的记录数据进行重复性检测、数据复原。针对海量的事故、设计资料、交通流量等筛选、分类工作，论文研究过程中编制了数据挖掘、特征路段判定及相应事故资料处理的应用程序，有效地减轻了数据整理的工作强度；提高了对不同特征路段、不同道路属性等海量数据处理的有效性和数据的准确性。 本文利用行车危险性描述模型，考虑临近路段单元之间的相互影响及行车危险性有限扩展等原理，对特征路段进行分类；通过分析长直线、小半径平曲线、隧道、连续纵坡、互通立交及服务区等不同特征路段事故类型、影响因素、事故空间分布规律、对临近线形单元的影响区范围等，利用间接、直接指标分析不同特征路段的行车危险性分布规律，建立了不同特征路段对应的行车危险性描述模型，并确定了各自的事故预测模型的基本解释变量；最后根据95%风险涵盖量，确定了特征路段在高速公路事故预测模型中的分析区范围。 高速公路事故预测模型是本研究的核心内容，论文首先明确了高速公路事故预测的流程及基本路段划分条件，然后针对解释变量弹性为常数（恒弹性）的不足，引入解释变量柔性的分析方法。利用线形对数函数、超对数函数建立基于变量恒弹性及柔性情况下负二项、广义负二项等6种模型的基本形式，通过模型拟合效果对比分析，最终选择了基于解释变量柔性的广义负二项分布建立高速公路事故数量预测基本模型（SPF）。随后利用广义负二项分布建立了长直线路段、小半径平曲线路段、隧道路段、连续长下坡路段、连续长上坡路段等5种特征路段的事故预测模型；同时建立了AADT、车道数量、货车比例、互通立交及服务区、夜间照明、强制限速等6项事故修正系数（CMFs）的数值或函数模型。利用确定的高速公路事故预测基本模型（SPF）、5种特征路段事故预测模型及6种事故修正系数（CMFs）可以对路段的事故数量进行预测。 由于不同地区气候条件、车型比例、交通流量、驾驶人特点等与样本高速公路所处地区的差异，论文提出了高速公路事故预测模型计算结果的标定流程，并明确标定周期、资料要求等；提出了标定系数的计算方法，即某地区高速公路标定周期内的年平均事故数量与预测模型预测值的比值。 为充分发挥历史事故资料对高速公路事故预测模型的补充、修正作用，论文通过具体地点经验贝叶斯（SSEB）方法和项目级经验贝叶斯（PLEB）方法与路段、项目整体预测结果相结合，有效提高了预测事故数量期望值的精度；对广东省山区和平原区2条高速公路2012年的事故数量进行了预测和EB修正，经与实际事故数量对比，证明本文的研究成果是可行的，预测结果是可靠的。 论文最后利用高速公路事故预测模型及具体地点经验贝叶斯（SSEB）方法，结合行车危险性描述模型对高速公路行车危险路段的判定方法进行了研究，利用广东省某平原区高速公路进行了计算分析，对其采用不同方法、不同工况下的行车危险路段划分结果进行了对比，结果证明经具体地点SSEB方法修正后的高速公路事故预测模型预测结果可以将实际事故资料与理论分析相结合，能有效地反映不同路段、不同行驶方向的行车危险水平、提高判定的准确度，并能为安全保障措施改造优先性决策提供支持。 关键词：高速公路；事故预测模型；数据挖掘；行车危险性描述模型；特征路段；解释变量柔性；广义负二项分布；事故修正系数；经验贝叶斯方法；行车危险路段判定 ABSTRACT Expressway is characterized by high speed, low cost, large traffic capacity, advanced transport efficiency. It plays an important part in national economic construction and people's lives with its shortcut and safety. Generally, in developed country, traffic crashes of freeway are 30~51% of that of ordinary highway. In contrast with developed country, traffic crashes of freeway are over 2 times than that of ordinary highway, and safety situation of freeway is extremely critical. In the last decade safety environment has been obviously improved by strengthen safety education, rough laws and regulation, and so on. However, in recent years the downtrend of the crash rate has been relatively gentle, even increased again. Thus traditional research method should be improved to decrease freeway crashed effectively. By analyzing highway attributes and traffic parameters of different sections impaction on crashes, Crash Prediction Models (CPMs) of different segments types of expressway are established to for highway management department, traffic management agencies to improve appropriate management strategies and emergency rescue measures. Combining results of CPMs with empirical Bayes (EB) method can judge dangerous sections of expressways, and support decision-making priorities for security improvements of those dangerous sections. Supported by the project ‘ Assess and Comprehensive Treatment Technology of road accident prone’ of Guangdong Highway Administration and “Research on freeway accident prediction and comprehensive treatment decision system” of Guangdong Communication Department, large number of Design-information, more than 135 thousand accidents data and traffic volume of 27 expressways(about 2200km) are collected from cooperators, which includes Transport Authority of Guangdong Public security department and Guangdong Communication Group Co., Ltd. Restore integrity of accidents data, Division method of specific segments by driving-danger-description (DDD) model, CPMs and crash modified factors (CMFs)， Calibration Process and calibration factor and determination of drive-dangerous(DD) segments , improve-priority decisions are studied. For different focus of traffic police and expressway administration, some differences are found in quantity and quality of crash records. Anomaly detection and correlation analysis of Data Mining are adopted to detect reproduce and recover origin data. Application program is crafted by C# to screen, classify accidents, design information and traffic volume, which can release manpower on data processing. DDD model, which can reflect the interactions of adjacent sections, is adopted to character spatial variations of crash possibility on different special segments and fixed influence scopes. By DDD model, special segments are classified. Analyzing crash types, influence factors, spatial distribution of crashes, affected zone on adjacent sections of different special segments that includes long straight lines, sharp curves, tunnels, consecutive grades, interchanges or service areas, DDD models are established and Parameters of DD distribution functions are determined, and then special segments analysis zones in CPMs were determined with 95% risk. Based on Predictive Method Processing and qualification of segments division, road characteristics and crash data are classified to different base segments. Discrete data statistical methods are contrasted to choose the most appropriate method for expressway crashes. To improve the deficiencies of assume constant elasticity for estimation parameters in log-linear function model, flexibility which is widely used in econometrics and manufacturing, is introduced to assess the possibility that explanatory variables had non-constant elasticity. Through analysis and contradistinction on 6 trans-log function forms with constant elasticity and flexible explanatory variables, the generalized negative binomial (GNB) model with trans-log functional form is adopted to assess flexible explanatory variables. The base segments of SPF are from four-lane expressway without lighting. By the data consist of accident data and road characteristics, the values of the coefficients for SPF and over-dispersion parameter are determined. The CPMs and over-dispersion parameters of different special segments that includes long straight lines, sharp curves, tunnels, consecutive grades, are established. To common segments which are different with fundamental segments, CMFs for AADT, lanes, percentage of trucks, lighting, interchanges or service area, speed enforcements are assigned values or functions. Total number of expressway segments crashes per year is calculated by SPF with CMFs and CPMs of special segments. To adjust the CPM which was developed with data from Guangdong Province for application in another jurisdiction, calibration provided a method to account for differences between jurisdictions in factors such as climate, driver characters, and vehicle type and traffic volumes. Calibration factor, procession; cycle and data requirements are analyzed. With the help of Site-Specific (SSEB) and Project-Level (PLEB) Empirical Bayes, the historical crash data are combined with the predictive crash numbers of expressway, then the estimate of expected crashes were supplemented and amended to improve accuracy. The SPF with CMFs and CPMs of special segments are applied to mountainous and plain area expressway in Guangdong Province, the predictions are combined with crashes of 2011 by SSEB or PLEB method. The revised predictions (RP) are contrasted to actual crashes of 2012, which indicate that research results are feasible and predictions are reliable. Based on DDD model, the predictive crash numbers with SSEB that combined historical and predictive crash data are adopted to determine DD segments in expressway. Appling different methods and working conditions to plain area expressway in Guangdong Province, the predictions of DD segments are contrsted. The results proved that DD level of different segments and determination accuracy is improved effectively, and priority of safety and security improvements can be decided. Key Words: Expressway; Crash Predictive Models(CPMs); Data mining, Driving-Danger-Description (DDD) model; Special segments; Flexible explanation variables; Generalized Negative Binomial(GNB) distribution; crash modified factors (CMFs); Site-Specific (SSEB) and Project-Level(PLEB) Empirical Bayes; Determination of DD segments. |
硕士论文