交通事件持续时间预测贝叶斯网方法研究

摘要

The prediction of traffic incident duration is an important part of Incidents Management System. The main factors of incident duration were identified using stepwise regression according to 1-880 field data. Naive Bayes model (NB), Tree Augmented Naive Bayes model (TAN), and Bayesian Networks Model (BN) were developed for incident duration prediction. This paper also determined the algorisms of inference, structure learning and parameter learning based on data character. The models were evaluated respectively under different sample sizes and missing data extent. The results showed that with 30% missing of data, the 30 minutes accuracy of Bayesian network predictions model was a-bove 80%%交通事件持续时间的预测是事件管理系统的重要组成部分,根据I-880实测数据集,利用逐步回归分析的方法确定事件持续时间的主要影响因素,分别建立了应用于事件持续时间预测的朴素贝叶斯(NB)模型、加树朴素贝叶斯(TAN)模型以及一般贝叶斯网(BN)模型,在分析数据特点的基础上确定了贝叶斯网的推理算法、参数学习以及结构学习方法.在不同数据缺失的程度和不同训练样本规模下,分别对三种模型的预测准确率进行了评价,结果表明贝叶斯网预测模型在数据缺失30％的情况下30min准确率高于80％.