基于海量出租车轨迹数据的旅行时间预测

发布时间：2018-05-29 23:17

本文选题：车辆轨迹数据 + 车辆行程　；参考：《华东师范大学》2017年博士论文

【摘要】：近年来,随着中国城市化水平的不断提高,机动车数量有加速增长的趋势,而道路基础设施的建设却相对缓慢,城市交通的供需矛盾日益加剧,在很多大中型城市中,交通拥堵正在逐渐成为常态。在这种情况下,出行者的旅行时间复杂多变,出行成本越来越高。如何准确预测未来的旅行时间,对于出行者和交通管理者,都是一个具有重大现实意义的问题。此时,智能交通系统(intelligent transportation system,ITS)的日益成熟和各种海量车辆轨迹数据的出现给旅行时间预测的研究带来了巨大的机遇。在这种背景下,本文希望以海量车辆轨迹数据为数据支撑,对高度城市化地区的旅行时间(travle time)预测问题进行探索。虽然车辆轨迹数据能提供完整时空覆盖的交通信息,但其海量性也给数据的维护和检索带来困难。另外,旅行时间预测的准确度不仅依赖于预测模型的性能,也受限于数据本身的复杂性。因此,在进行旅行时间预测研究之前,还需要从数据质量的角度来研究历史旅行时间序列的复杂度,对历史旅行时间序列的可预测性进行评价。基于此,针对海量车辆轨迹数据的旅行时间预测研究应涵盖轨迹数据建模与索引、旅行时间可预测性分析、以及旅行时间预测的实现等多方面内容。基于行程的轨迹数据模型能给车辆轨迹数据提供便于管理的组织形式,而针对车辆行程的索引结构能明显改善检索效率,为行程信息的获取提供可行的解决方案;在高效获取行程数据的基础上进行历史旅行时间序列的可预测性分析,是对预测数据的检验和评价,为旅行时间的预测提供保障;而考虑影响交通的各种条件去设计预测模型实现旅行时间预测则是本文最终的研究目的。基于此,本文针对车辆轨迹数据建模、车辆轨迹数据索引、旅行时间可预测性和旅行时间预测模型等四方面展开研究。在车辆轨迹数据建模方面,本文使用基于车辆行程的轨迹数据模型来组织车辆轨迹数据,并根据基于车流方向的道路网络模型提出了基于道路拓扑的轨迹数据地图匹配算法。为了获取用于旅行时间预测的行程数据,本文定义"车辆行程"(vehicle-based trip)来表达出行者的一次出行经历,并以其为逻辑单位组织车辆轨迹数据,通过轨迹提取、地图匹配、行程划分等步骤来实现轨迹数据的建模。其中,针对一般道路与快速路系统并存的复杂城市道路网络,本文提出了一种基于道路拓扑的轨迹数据地图匹配算法,该方法通过最短距离法筛选轨迹的备选路段集;然后进行轨迹分段,消除轨迹中的环形结构;接着对每个轨迹段利用有向路段的拓扑关系选取匹配路径,实现轨迹数据在复杂道路网络中的地图匹配。实证研究描述了数据建模的整个过程,不仅证实了本文使用的路网模型和轨迹数据模型的可用性,还对轨迹数据建模的性能进行了分析。针对海量车辆轨迹数据的高效存取问题,本文提出了一种面向行程的车辆轨迹数据索引方案——TripCube,它使用三维的时空索引立方体维护车辆行程数据,并根据车辆行程的起止点和出发时间来快速检索车辆行程信息。与通用索引结构的多组性能对比实验表明,TripCube结构简单、易于维护,且对车辆轨迹数据的存取效率远远优于通用的索引结构。接着,本文讨论了历史旅行时序列的复杂度对旅行时间预测的影响。在分析历史旅行时间序列复杂性的基础上,把"旅行时间可预测性"(travel time predictability)定义为使用历史旅行时间序列正确预测旅行时间的概率,并提出一个基于熵的方法去测量旅行时间可预测性的最大值。首先,使用多尺度熵(Multiscale Entropy,MSE)的改进算法——RCMSE(the refined composite multiscale entropy algorithm)计算不同时间尺度下旅行时间序列的复杂度;然后,关联旅行时间序列的熵和序列的旅行时间可预测性最大值,求解旅行时间可预测性的最大值。实证研究分析了时间尺度、容差和序列长度等因素对旅行时间序列的熵和旅行时间可预测性的影响,讨论了旅行时间预测的精度,还进行了旅行时间可预测性与实际预测结果的对比实验。实验结果证实了本文提出的旅行时间可预测性的有用性及其计算方法的可靠性。在上述研究的基础上,本文提出了面向行程的旅行时间预测模型。在出行者更关注特定起止点行程的旅行时间的背景下,本文使用基于反向传播神经网络模型的预测方法,充分考虑多种影响旅行时间的因素(出行时间、天气条件、空气质量等),实现城市道路网络中任意起止点间的面向行程的旅行时间预测。实证研究使用13个月的海量出租车数据进行,其中12个月的数据用于训练模型,1个月的数据用于验证预测结果。实验结果证实了本文提出的预测模型的有效性和准确性,同时也表达了气象条件对旅行时间的影响是不可忽略的。最后,对上述研究成果进行总结,明确了本文研究的主要贡献和局限性,并对未来进一步的研究工作进行了展望。
[Abstract]:In recent years, with the continuous improvement of the level of urbanization in China, the number of motor vehicles has increased rapidly, but the construction of road infrastructure is relatively slow, the contradiction between supply and demand of urban traffic is increasing. In many large and medium-sized cities, traffic congestion is becoming normal. In this case, the traveler's travel time is complex and changeable. The travel cost is getting higher and higher. How to accurately predict the future travel time is a significant problem for both the traveler and the traffic manager. At this time, the growing maturity of intelligent transportation system (ITS) and the study of the occurrence of a variety of mass vehicle trajectory data to the travel time prediction In this context, this paper hopes to explore the problem of travle time prediction in highly urbanized areas with massive vehicle trajectory data as data support. Although vehicle trajectory data can provide traffic information with a complete and space-time coverage, the mass character also brings difficulties to the maintenance and retrieval of data. In addition, the accuracy of travel time prediction depends not only on the performance of the prediction model, but also on the complexity of the data itself. Therefore, the complexity of the history travel time series needs to be studied from the point of view of the data quality before the travel time prediction research, and the predictability of the history travel time series is evaluated. In this case, the travel time prediction for mass vehicle trajectory data should cover the modeling and index of the trajectory data, the predictability of travel time, and the realization of the travel time prediction. The trajectory data model based on the travel can provide a convenient management organization for the vehicle trajectory data, and for the vehicle travel. The index structure can obviously improve the retrieval efficiency and provide a feasible solution for the acquisition of travel information. On the basis of efficient acquisition of travel data, the predictability analysis of the history travel time series is the test and evaluation of the forecast data and the guarantee for the travel time prediction; and the various conditions that affect the traffic are considered. The purpose of this paper is to study the four aspects of vehicle trajectory data modeling, vehicle trajectory data index, travel time predictability and travel time prediction model. In the aspect of vehicle trajectory data modeling, this paper uses the number of trajectories based on vehicle travel. According to the model, vehicle trajectory data are organized, and a path map matching algorithm based on road topology is proposed based on road network model based on the direction of traffic flow. In order to obtain travel time prediction, this paper defines "vehicle travel" (vehicle-based trip) to express a traveler's trip experience and take it as logic. In this paper, a track data map matching algorithm based on road topology is proposed in this paper. In this paper, a path map matching algorithm based on road topology is proposed. The selected section of the path is selected, then the trajectory is segmented to eliminate the ring structure in the trajectory. Then, the matching path is selected for each track segment using the topology of the directed section to realize the map matching in the complex road network. The empirical study describes the whole process of modeling the data, which not only confirms the use of this paper. The availability of road network model and trajectory data model is also analyzed, and the performance of trajectory data modeling is also analyzed. In view of the efficient access problem of mass vehicle trajectory data, this paper proposes a travel oriented vehicle track data index scheme, TripCube, which uses a three-dimensional spatio-temporal index cube to maintain vehicle travel data, and The multi group performance comparison experiment with the general index structure shows that the TripCube structure is simple and easy to maintain, and the access efficiency of the vehicle trajectory data is far superior to the general index structure. Then, this paper discusses the complexity of the history travel time sequence. The influence of travel time prediction. On the basis of analyzing the complexity of historical travel time series, "travel time predictability" is defined as the probability of using historical travel time sequence to predict the travel time correctly, and an entropy based square method is proposed to measure the maximum predictability of travel time. First, An improved algorithm of Multiscale Entropy (MSE) - RCMSE (the refined composite multiscale entropy algorithm) is used to calculate the complexity of the travel time series at different time scales; then, the maximum predictability of the entropy and the travel time of the sequence of associated travel time series is calculated, and the most predictability of travel time is solved. The empirical study analyses the influence of time scale, tolerance and sequence length on the predictability of travel time sequence entropy and travel time. The accuracy of travel time prediction is discussed, and the comparison experiment between travel time predictability and actual prediction results is carried out. The test results confirm the travel time proposed in this paper. On the basis of the above study, the travel time prediction model of travel oriented is proposed on the basis of the above research. In the background of the traveler's more attention to the travel time of the specific stop point travel, this paper uses the prediction method based on the back propagation neural network model to take full consideration of many kinds of travel time. The factors (travel time, weather condition, air quality, etc.) are used to predict the travel time of travel time between any starting points in the urban road network. The empirical study uses 13 months of mass taxi data, of which 12 months of data are used for training model, and the data of 1 months are used to verify the prediction results. Experimental results confirm the results. The validity and accuracy of the prediction model presented in this paper also show that the influence of weather conditions on travel time can not be ignored. Finally, the above research results are summarized, the main contributions and limitations of the study are clarified, and the further research work in the future is prospected.
【学位授予单位】：华东师范大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：U491.14

【参考文献】