数据稀疏和数据缺失情况下的旅行时间预测研究

Recently, many new problems are emerging and challenging researchers in urban travel time prediction field. These problems include:. 1) Sparse data problem. Most recent studies assume we can collect the travel time data of most vehicles that had passed through a certain road segment in a certain time period. However, all the intelligent transportation systems in Chinese cities are still under constructing, and we can only detect and record a few vehicles. The collected travel time data are thus sparse.. 2) Missing data problem. Because of the physical limits of sensors and data transmission networks, intelligent transportation systems around the world are suffering from missing data problem. In many Chinese cities, the missing ratios are often higher than usual. Many previous travel time prediction models cannot give satisfactory results by using the data with such high missing ratios.. 3) Small sample size problem. Since the intelligent transportation systems in many Chinese cities are under developing, we do not always have a large amount of historical travel time data available. This usually leads to inappropriate identification and training of forecasting models.. Considering these three problems, we aim to integrate the research on individual vehicle 's travel time and the research on overall traffic flow dynamics on the same road segment in this project. Particularly, we will emphasize the Bayesian Networks based travel time prediction model. The sparse data, missing data, and small sample size will be viewed as certain posterior knowledge with special forms. As a result, we could quantitatively examine the influence of data sampling problem on the Bayesian inference results and then solve these three problems in a united way. Based on this new idea, we can finally discuss how to improve the prediction performance under the constraints of sparse data, missing data and small sample size problems. The obtained results are useful to many other traffic engineering problems (e.g. ramping metering), and have important academic values and promising practical perspectives.

近年来，国内城市道路旅行时间预测研究遇到了许多新的挑战，其中包括：. 1）数据稀疏的问题。国外已有研究往往假定能够检测到一段时间内通过某一路段的大多数车辆的旅行时间。但我国各城市的智能交通系统均在完善中，能检测的车辆比例经常较低。. 2）数据缺失的问题。由于传感器和数据传输网络的物理条件所限，目前国内外的智能交通系统均存在数据缺失问题。而国内很多城市的数据缺失情况更为严重。. 3）历史数据不足，影响预测模型辨识校正的问题。. 有鉴于此，本项目将城市道路个体车辆旅行时间和整体交通流运动规律研究有机结合在一起，着力发展基于贝叶斯网络的旅行时间预测模型。我们将数据稀疏、数据缺失和小样本转化为特定结构的后验知识，统一考察后验知识结构和数量的变化对于贝叶斯推理结果的影响，最终改进目前的旅行时间预测算法。相关结果可用于匝道控制等多方面，具有重要的学术意义和广阔的应用前景。

本项目的资助下，我们收集分析了大量公交车、出租车和网约车的旅行时间数据。.首先，我们研究了不同交通状态下的旅行时间变化规律，研究旅行时间偏峰和多峰分布的成因，建立了交通流时空分布特征和旅行时间分布模型之间的关系。.其次，我们研究了交通流的波动特性，并揭示了交通数据压缩、交通异常数据检测、交通缺失数据补偿和交通预测等多个以往独立研究问题存在着隐含的共性。其中，交通数据压缩强调长期趋势，寻求数据的共性。交通异常数据检测反其道而行之，强调数据的差异。交通缺失数据补偿和交通预测则更多的是研究去掉趋势之后的波动特征。实践表明：基于去趋势方法，可以显著提高基于数据建模的旅行时间预测精度。.再者，我们对城市道路的旅行时间不确定性和可预测性进行了深入研究，提出了多种预测模型，并在微巴公司出租车旅行时间预测与调度的实际应用中取得了良好效果。.最后，我们还分析了如何通过设置交通信号等方式来优化城市道路的旅行时间，以及如何根据具有不确定性的旅行时间时间进行路径选择。.部分成果获得教育部2016年自然科学二等奖（项目负责人为第一完成人）。相关成果帮助项目负责人获得2017年IEEE Fellow。