高速公路交通事故模糊逻辑预测模型改进研究

以AADT、路段长度、车道数、大型车比例和地形条件作为模型输入变量,以每公里事故数作为模型输出变量,结合辽宁省高速公路数据构建基本的交通事故模糊逻辑预测模型.考虑到模糊集合结构和模糊控制规则对预测结果可能产生的影响,提出调整模糊集合和融入先验知识构建规则库的模型改进方法.以粤赣高速和开阳高速为案例,分析了基本模型与改进模型的可移植性.最后,应用同样的数据构建了负二项分布事故预测模型,并与模糊逻辑预测模型进行了对比分析.研究结果表明,纵向比较,模糊集合细化一定程度提高模型预测精度,细分模型相较于基本模型,总体平均相对误差减少8.3%,模型优度提高0.357;横向比较,融入先验知识构建模糊规则库能一定程度提高模型预测精度,基本先验模型相较于基本模型,总体平均相对误差减少1.9%,模型优度提高0.164.融入先验知识后模型的可移植性增强,平均预测精度高于基本模型,相对误差大于0.5的样本数减少3.8%,总体误差减少3.4%,总体平均相对误差减少4.1%,模型优度提高0.385;但细化集合的模型可移植性较低,与粗分和基本模型相比各个指标值均不同程度变差;而模糊逻辑事故预测模型与负二项分布事故预测模型在预测精度和可移植性方面均无显著差异. 

An Improved Model of Accident Prediction on Freeways Based on Fuzzy Logic

AADT, length of segments, number of lanes, proportion of oversize vehicles, and road alignment of each road section are collected from a freeway network in Liaoning Province, which are set as the inputs of a basic fuzzy logic model for accident prediction, and the frequency of accidents per kilometer is set as the outputs.In consideration of that structures of Fuzzy Sets and rules of Fuzzy Control may have possible negative impacts on prediction results, a method to improve the fuzzy logic model is established by resizing structures of Fuzzy Sets and setting up rules of Fuzzy Control with prior knowledge, respectively.The portability of the basic fuzzy logic model and the improved model is analyzed.This improved model and a negative binomial prediction model are both used in a case study of Yuegan and Kaiyang freeway as a comparison.The results show that, subdividing Fuzzy Sets can increase the accuracy of prediction.Compared with the basic model, the average relative error in total (Zt) of the subdivided model decreases by 8.3%, and model goodness (MC) increases by 0.357.With prior knowledge, the accuracy of prediction can be increased.Compared with the basic model, the Zt of the basic prior model decreases by 1.9%, and MC increases by 0.164, respectively.Prior knowledge also increases the portability of this model, and improves the average accuracy of prediction, Ra decreases by 3.8%, overall error decreases by 3.4%, Zt decreases by 4.1%, and MC increases by 0.385, respectively.However, compare with the model of rough sets and basic model, the portability of this model decreases when apply subdividing Fuzzy Sets.The basic fuzzy logic model and negative binomial prediction model for crash prediction are almost at the same level in accuracy of prediction and portability