【作者】 李星毅；

【导师】 侯忠生；

【作者基本信息】 北京交通大学 ， 交通信息工程及控制， 2010， 博士

【摘要】 近年来,交通系统在全球范围内面临着前所未有的挑战,解决交通拥堵,提高交通灵活性成为全球关注的课题。交通流量分析是智能交通系统的基础,对交通组织和控制具有重要意义。本文主要讨论时间序列相似性问题及其在交通流量分析中的应用。基本思想是：以历史数据为基础,通过相似性特征分析发现交通流潜在的模式,根据检测数据与模式之间的关系实现对交通流量状态的评价、未来发展趋势的预测和分析。文章对四个问题进行了研究和讨论：相似时间序列的模式发现；交通流离群数据的发现和清洗方法；交通流预测方法；交通状态判别与评价方法,并通过对实际交通流采样数据的分析,验证了本文提出的各种分析方法在实际应用中的可行性和可靠性。本文的主要工作和创新点如下：1.提出了交通流分解方法,在分解的基础上定义了相似类的基准和偏离的概念。基准反映了相似类成员的共性,偏离反映了成员的差异性,充分刻画了相似的本质。通过实际数据验证和讨论了交通流相似性,讨论了交通流模式,给出了交通流高峰期的定量判别方法和相关参数标定算法。2.提出了基于相似的交通流预测方法,分别就短期预测和中长期预测设计两类预测算法。其中,中长期预测方法MFBS实现了交通流中长期的定量预测。文章通过实际数据验证了两类算法的可行性。并使用实际数据与传统的三种短期预测方法进行了比较,结果表明MFBS算法的精度可以到达传统短期预测的精度。3.提出了基于相似的交通流状态判别和评价方法。状态判别方法SFT可以实现拥挤状态发现和发生原因判断,比传统的判别方法更为直接有效。基于路段负荷、基于拥挤状态变化和基于高峰期特征属性变化的三种评价方法分别从微观和宏观层面实现了对交通流趋势的评价。通过实际数据验证了相应方法的可行性。4.提出了基于相似的离群数据判别方法,并针对静态系统和实时系统设计了基于相似的离群数据修复算法。文章通过实际数据验证了修复算法的可行性。更多还原

【Abstract】 To solve traffic congestion, improve traffic flexibility, has been becoming a subject of global concerns because of worldwide transportation systems facing unprecedented challenges in recent years. Traffic stream analysis is the basis of intelligent transportation systems and has great significance for traffic organization and control. This dissertation focuses on time series similarity problems and its application in traffic flow analysis.The basic ideas of the papers are as follows:finding the potential pattern of traffic flow through analyzing the characteristics of similarity, achieving assessment of traffic flow, forecast and analysis of the future trends according to the relationship between the detection data and models, are based on historical data. The dissertation studies and discusses the following four questions:the discovery of pattern of time series similarity, the detection and cleaning method of qutlier data of traffic flow, the methods of traffic flow forecasting, traffic status discrimination and evaluation methods.Then, this dissertation verifies the feasibility and reliability of all analytical methods in practical applications through analyzing sampled data of actual traffic flow.The main works and contributions in the dissertation are as follows.1. Idea about time series decomposition is present. The flow series are decomposed into benchmark series and deviation series. Benchmark reflects common property of class members and deviations reflects differences of members, and they reveal the essence of similarity. The dissertation also gives quantitative checking methods of peak traffic flow and relevant parameter calibration algorithm and discussion about traffic flow similarity and traffic flow model through validation of actual data.2. The dissertation proposes traffic flow forecasting methods based on similarity. Specifically, it designs two types of prediction algorithm working for short-term and medium-term prediction separately. The long-term prediction MFBS completes long-term quantitative prediction of traffic flow. At the same time, the dissertation verifies the feasibility of two types of algorithms through actual data, and compares actual data with three traditional short-term prediction methods, the results show that MFBS algorithm can reach the precision as traditional short-term prediction do.3. The dissertation proposes identification and evaluation methods of the state of trafflic flow based on similarity. Status identification method (SFT) can complete finding crowded state and judging causes, and it is more effective than traditional identification methods. Three kinds of evaluation methods based on road load, the changes of crowded state, and the peak characteristics changes achieve evaluation of traffic flow trends from the micro-and macro-level separately. The practical feasibility of corresponding methods is validated by actual data.4. The dissertation proposes checking methods of outlier data based on simliarity and designs recovery algorithms to recover outlier data for static systems and real-time systems. The feasibility of recovery algorithm is validated by actual data.更多还原