【作者】 刘玉洁；

【导师】 何丕廉；

【作者基本信息】 天津大学 ， 计算机应用技术， 2009， 博士

【摘要】 航班延误一直困扰是国际国内民航业的一个热点问题。近年间我国航空延误日益加重,已经影响到民航业的发展,改善延误状况迫在眉睫。航班延误多发生在繁忙的枢纽机场,枢纽机场又是多数航班的转乘点,是航班链中的关键环节。当航班延误发生在繁忙的枢纽机场时,延误在航班链中的波及将不可避免。减轻繁忙枢纽机场的延误,可以使整条航班链,继而整个民航系统的运行状态得到改善。本文将某繁忙的枢纽机场AIRPORTP作为主要研究对象,对其中的进、离港延误,延误波及进行建模与预测。本文基于贝叶斯网络的理论,面向枢纽机场和航班链内的航班延误与波及,提出了多种算法与学习方法。首先,面向繁忙的枢纽机场,本文将贝叶斯网络参数学习算法应用到航班延误的预测领域。针对基于专家经验的网络结构维度过高、运算量过大的问题,且会使每种状态下的样本量不足的问题,提出了两种解决方案:一是独立属性抽取,通过分析各属性间的相关性,分割网络结构。此方案运行速度快,方便处理大数据集;二是精简优化模型,通过回归分析,剪除关联性低的节点,以降低网络的维度,此方案正确率高,速度比方案一慢,适合处理较小的数据集。本文通过分析和建模发现进、离港拥有同样的属性和网络结构,基于方案二建立了延误波及的对称性模型PMofA。其次,面向航班链中的延误波及问题,本文在利用贝叶斯结构学习算法对进、离港延误进行建模的基础上,使用混合学习方法建立MSP模型。同时还提出了一种改进型的K2结构算法TFK2,TFK2比传统的K2更适合为航班延误的建模,速度更快,准确率更高。继而基于TFK2算法,提出一种能够产生冗余节点的协商结构算法,由该算法生成的网络结构中,经过协商,不同条件下的节点间会存在竞争或冗余两种状态,使网络的预测准确率和速度都得以提高。再次,基于集成学习理论和TFK2贝叶斯结构算法,本文还提出了一种带有自反馈的航班预测集成学习系统SEFS。该系统中包含有三个子学习器,通过对三个子学习器的训练,可以对航班延误情况进行预测。依据预测的结果,用颜色和概率表示对航班延误进行预警,使预警更具人性化。最后,应用SEFS系统,以AIRPORTP为例,对其中一定时间段内的延误航班数量进行预测,面向繁忙的枢纽机场发布预警;应用SEFS系统对特定航班的延误时间进行预测,面向航空公司与乘客发布预警。更多还原

【Abstract】 In domestic and overseas, flight delay is one of the problems that the staffs inindustry devote themselves to research. In recent years, the delays of flights indomestic become more and more serious, and have effected on the developing of thecivil aviation industry. Flight delay status needs to be researched and improvedimmediately. The flight delays are often happened in those busy hub-airports whichare important in flight chains. Many flights turn around in hub-airports. Therefore,when the delay happened in a busy hub-airport, the delay propagation will beinevitable. Relieving the delay status of busy hub-airports will lighten the pressure ofwhole flight chain, even the whole civil aviation system. Therefore, focusing on oneof those busy hub-airports named AIRPORTP, arrival/departure delays and delaypropagation have been discussed in this paper separately.Orienting to the hub-airports and flight trains, many algorithms and methods areadvanced based on Bayesian Network:Firstly, orienting to the busy hub-airports, Bayesian Network parameter learningalgorithm is used in the field of flight delay estimation. In order to solve the problemof high-dimension, large-computing-scale in the network constructed by experts’experiments, two methods are raised. One is named Independent Attribute TakingOut,which separates the network based on analyzing of the correlation between allattributes. This method is suit for dealing with huge data set since its speed incalculating is high. The other is named Optimized Model. The nodes with lowcorrelation are eliminated after regressive analyzing. The correct rate of estimatingwith this method is high. This method is suit for dealing with small data set since itsspeed in calculating is slower than the first method stated above. Then the dimensionof model’s structure is reduced. After analyzing and modeling, the same nodes andstructure are discovered in the arrival and departure delays models, therefore asymmetrical model named PMofA is established for the delay propagation based onOptimized Model.Secondly, orienting to the delay propagation in flight chains, after modelingarrival delay and departure delay separately based on the Bayesian structure learningalgorithm, MSPmodel is established bya Mixed Learning Method. Then an improvedK2 structure learning algorithm named TFK2 is raised. It is proved to be more suitable to model and estimate the flight delays than the traditional K2 algorithm bothin learning speed and correct rate. Based on TFK2, a Negotiating Structure Learningalgorithm is raised. The network’s structure built by this Negotiate Structure Learningincludes redundancy nodes. The nodes change their working states as competition orredundancy in different conditions to further enhance the correct rate and the speed incalculating ulteriorly.Thirdly, a Self-feedback Ensemble-learning Flight-delay Estimating System(SEFS) is raised based on Ensemble Learning Method and TFK2 algorithm. Thesystem includes three sub-learners. After trained,the system can be used to estimateflight delays. And according to the estimated result of SEFS, the delay of specificflight can be pre-warned with color and probability, to make pre-warning morehumanity.Finally, pre-warning is issued for a busy hub-airport, based on the estimatednumber of delayed flights during a period time in AIRPORTP by the SEFS System.And pre-warning is issued for air companies and guests, based on the estimated delaytime of specific flights by the SEFS.更多还原