基于负二项回归分析的高速公路神经网络事故预测模型

发布时间：2018-07-24 13:58

【摘要】：我国高速公路近年来的交通安全形势有所改善,但群死群伤事件仍时有发生,交通安全问题依然突出。为加强高速公路交通安全管理,可通过事故预测进行事故发生原因分析,从而预判事故发生规律并制定针对性措施,降低事故率及事故发生严重程度。因此本文基于事故数据统计分析及神经网络技术开展高速公路事故预测研究。基于事故数据的统计分布特性及神经网络技术,本文对高速公路事故预测展开研究,给出了各条高速公路的统计分布特性并基于此提出了基于负二项回归分析的变量选择方法,分别构建了山岭重丘区、平原微丘区高速公路神经网络事故预测模型,提出了基于敏感性分析的网络模型验证方法,最终通过模型的应用验证了模型的可靠性与可移植性。由于高速公路地形条件主要分为平原微丘区和山岭重丘区两类,且这两类地形条件下的高速公路设计标准存在差异,其事故影响因素也有所不同,因此本文考虑了高速公路地形条件,分别对山岭重丘区及平原微丘区事故预测模型的构建进行研究。首先,选择了辽宁省及广东省的9条高速公路作为本文的研究对象,并对其事故数据及其关联数据进行处理与组织,根据公路几何线形按同质法将高速公路划分成满足研究需求的事故预测单元,据此构建包含事故数据及其关联数据的山岭重丘区、平原微丘区高速公路交通事故基础数据系统。其次,分别对各条高速公路进行事故数据统计分布特性研究,研究结果表明基于几何线形划分的高速公路事故预测单元事故数主要服从负二项分布。据此,选择根据负二项分布进行事故影响因素的分析及变量选择工作。分别分析山岭重丘区及平原微丘区高速公路各影响因素与预测单元事故数之间的统计关系,确定理想线形标准,根据理想线形标准提出预测单元指标空值项的合理赋值方法,从而使其满足变量选择算法要求,最后基于负二项回归分别完成了山岭重丘区、平原微丘区预测模型自变量的选择。然后,根据所研究问题特性,采用Elman神经网络标定山岭重丘区及平原微丘区高速公路神经网络事故预测模型,分别进行模型的训练与测试,由测试结果可知,这两类地形条件下预测模型网络的泛化能力均较强。通过已标定完成的事故预测模型分析各输入变量的灵敏度,从而确定各变量与预测单元事故数的内在规律,从交通安全机理角度验证了这两类模型均具有正确性与有效性。最终,分别对山岭重丘区、平原微丘区事故预测模型进行应用分析,应用平均相对误差分别为8.705%和6.651%。应用结果表明这两类地形条件下的事故预测模均型具有一定的有效性和较大的可移植性。
[Abstract]:The traffic safety situation of expressway in China has improved in recent years, but the mass death and injury incidents still occur from time to time, and the traffic safety problems are still prominent. In order to strengthen the management of expressway traffic safety, the cause analysis of accidents can be carried out through accident prediction, so as to predict the occurrence rules of accidents and formulate targeted measures to reduce the accident rate and the severity of accidents. Therefore, this paper based on statistical analysis of accident data and neural network technology to carry out highway accident prediction research. Based on the statistical distribution characteristics of accident data and neural network technology, this paper studies the prediction of expressway accidents, presents the statistical distribution characteristics of each expressway, and puts forward a variable selection method based on negative binomial regression analysis. The neural network accident prediction model of expressway in mountain and hill area and plain micro-hill area is constructed, and the method of network model verification based on sensitivity analysis is put forward. Finally, the reliability and portability of the model are verified by the application of the model. Because the topographic conditions of expressway are mainly divided into two types: plain micro-hill area and mountain heavy hill area, and the design standards of expressway under these two kinds of terrain conditions are different, the influencing factors of the accidents are also different. Therefore, considering the topographic condition of expressway, this paper studies the construction of accident prediction model in mountainous area and plain area. First of all, 9 highways in Liaoning Province and Guangdong Province are selected as the research object of this paper, and the accident data and their associated data are processed and organized. According to the homogeneity method of highway geometry, the expressway is divided into accident prediction units to meet the needs of research. Based on this, the basic data system of expressway traffic accidents in mountain heavy hill area and plain micro-hill area is constructed, which includes accident data and its correlation data. Secondly, the statistical distribution characteristics of each expressway accident data are studied respectively. The results show that the number of accidents in the expressway accident prediction unit based on geometric line partition mainly depends on the negative binomial distribution. According to this, the analysis of the influencing factors and the selection of variables are carried out according to the negative binomial distribution. The statistical relationship between the influence factors of expressway and the number of accidents of prediction unit in mountain heavy hill area and plain micro-hill area is analyzed, and the ideal linear standard is determined. According to the ideal line shape standard, the reasonable assignment method of empty value term of prediction unit index is put forward. Finally, based on the negative binomial regression, the selection of independent variables in the prediction model of mountain heavy hill region and plain micro-hill area is completed. Then, according to the characteristics of the studied problem, the neural network accident prediction model of expressway in mountain heavy hill area and plain micro-hill area is calibrated by using Elman neural network, and the model is trained and tested respectively. The generalization ability of the predictive model network is strong under these two kinds of terrain conditions. The sensitivity of each input variable is analyzed by the calibrated accident prediction model, and the inherent law of each variable and the number of accidents in the prediction unit is determined. The correctness and validity of the two models are verified from the view of traffic safety mechanism. Finally, the accident prediction models of mountain heavy hill area and plain micro-hill area are analyzed, and the average relative error is 8.705% and 6.651respectively. The application results show that the average pattern of accident prediction under these two kinds of terrain conditions is effective and transplantable to some extent.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U491.3

【参考文献】