CASIA OpenIR  > 毕业生  > 博士学位论文
基于深度学习的交通预测方法研究
段艳杰
学位类型工学博士
导师王飞跃
2017-05-21
学位授予单位中国科学院研究生院
学位授予地点北京
关键词智能交通 深度学习 交通预测
其他摘要

        交通预测是智能交通系统的关键问题之一,对交通设施规划、交通管理与控制、个体出行选择都具有重要的意义。随着经济的发展,世界范围内城市化、现代化、机动化进程的加快,城市交通承载着巨大的运输压力,各类交通问题日益凸显。人们对交通问题的关注度持续提升。在此形势下,发展智能交通系统刻不容缓。另一方面,随着技术的进步,交通数据规模日益扩大,采用更加智能化的方法进行交通预测符合智能交通系统发展的需求。

    本文在平行交通系统的框架下,面向交通大数据的发展形势,研究基于深度学习的交通预测方法。针对历史、当前以及未来的交通预测需求,本文分别对交通数据弥补、短时交通流量预测以及路段行程时间预测这三个典型问题进行了研究。本文的主要工作包含以下几个方面:

    1. 提出了基于降噪堆叠自动编码机(DSAE)模型的交通数据弥补方法,弥补缺失的交通数据。

    该方法视交通数据为一个整体,把缺失数据看成噪声影响的结果,首次将交通数据弥补转变成从含有噪声的不完整数据预测除去噪声的完整数据的过程。利用深度学习方法中的降噪堆叠自动编码机,将不完整交通数据作为模型输入,完整交通数据作为模型输出目标,实现交通数据的自动恢复。该方法充分发挥深度学习模型对于高维数据的处理优势。计算实验结果验证了该方法在大规模交通数据弥补问题上的可行性及较传统方法的优越性。

    2. 提出了降噪堆叠自动编码机(DSAE)模型用于交通数据弥补问题的优化实现机制。

    该实现机制采用层次训练算法,首先利用包括路网区域内所有观测站点的大规模交通数据对模型进行第一个阶段的训练,然后利用特定观测站点的数据对模型进行第二阶段的训练。第一阶段训练完的模型能够用于所有观测站点的数据弥补,第二阶段训练完的模型则能够提升对特定观测站点的数据弥补效果。该实现机制充分发挥了大规模交通数据在训练深度模型上的优势,同时又兼顾特定对象数据的重要性。计算实验结果表明该实现机制的弥补效果优于其他对比实现机制。

    3. 提出了基于堆叠自动编码机(SAE)模型的交通流量预测方法,预测路网中的站点流量和道路流量。

    该方法首次结合深度学习中堆叠自动编码机的预训练技术和交通流量预测的时空信息融合技术,实现多层神经网络对大规模路网中站点流量和道路流量的一次性预测。该方法无需人工分析各站点或道路之间的相关性,能够通过深度网络自动学习各站点或道路之间的空间和时间关联。该方法开拓了交通大数据背景下减少人工干预的交通流量预测新方法,计算实验结果表明该方法能够取得较小的预测误差,达到了目前最先进的水平。

    4. 提出了基于长短时记忆递归神经网络(LSTM-RNN)模型的行程时间预测方法,预测路网中各路段的行程时间。

    该方法首次将深度学习中能够考虑样本之间时序相关性的模型用于复杂的行程时间预测问题。该方法能够实现自动学习记忆长时的行程时间数据模式和短时的行程时间数据模式,并通过模型中隐层的连接给当前预测传递历史信息。该方法免去人工分析影响行程时间的社会因素和自然因素的繁琐过程,能够进行多步行程时间预测。计算实验结果表明,该方法在单步预测中能够取得较小的预测误差,在多数路段的多步预测中也能取得较好的预测效果。实验对比结果表明该方法能够提高适用路段的行程时间预测的准确率。

 
;

    Traffic prediction is one of the key problems of the Intelligent Transportation System (ITS), which is of great significance to traffic planning, traffic management and control, and individual travel choice. With the development of economy and the accelerating process of urbanization, modernization and motorization around the world, transportation bears a huge pressure, and various traffic problems have become increasingly prominent. People's attention to the traffic problems continues to improve. In this situation, developing ITS is of great urgency. On the other hand, with the progress of technology, the scale of traffic data continuously increases, the use of more intelligent method for traffic prediction meets the needs of ITS development.

    In the framework of the parallel transportation system based on the ACP theory and for the development situation of big data, traffic prediction based on deep learning is studied in this dissertation. In view of the historical, current and future traffic prediction needs, this dissertation studies the three typical problems of traffic prediction: traffic data imputation, short-term traffic flow prediction and link travel time prediction respectively. The main work and contributions of this dissertation are as follows:

    1. The DSAE based approach for traffic data imputation is proposed. This approach treats traffic data containing observed normal data points, corrupted or missing data points as a whole corrupted vector, and for the first time transforms traffic data imputation into clean data recovering or corrupted data denoising. Based on the denoising stacked autoencoders (DSAE) in deep learning, this approach uses the corrupted traffic vector as the model input and the clean traffic vector as the model output target in order to realize the automatic recovery of the traffic data. This approach gives full play to the advantages of deep learning for high dimensional data processing. The computational experiments verify the feasibility of the proposed approach for large scale traffic data imputation and its superiority over the traditional methods.

    2. An efficient realization of deep learning for traffic data imputation is proposed. This realization adopts the hierarchical training algorithm. The first step is training the model using large-scale data from all vehicle detector stations (VDS) in the region. The second step is training the model obtained in the first step using data from the current single VDS. The model obtained in the first step can be employed to impute traffic data from all VDS in the region. The model obtained in the second step can improve the imputation performance on the current VDS. This realization not only gives full play to the advantages of large scale traffic data in training deep learning models, but also pays attention to the specific object data. The computational experiments show that this realization mechanism is better than other contrastive realizations.

    3. The SAE based approach for traffic flow prediction is proposed. This approach for the first time combines the pre-training technique of stacked autoencoders (SAE) in deep learning and the space-time information fusion idea in traffic flow prediction, to realize the multi-point simultaneous prediction of station flow and road flow in a multi-layer neural network. This approach gets rid of manually analysis of the relevance of each station or road, and can automatically learn the spatial and temporal association between stations or roads through the deep learning model. This approach opens up a brand new view to reduce manual engineering for traffic flow prediction under the background of big data. The computational experiments show that this approach can achieve a state of the art performance.

    4. The LSTM-RNN based approach for travel time prediction is proposed. This approach for the first time employs the long short-term memory recurrent neural network (LSTM-RNN) which can consider the serial correlation between data samples, to solve the complex travel time prediction problem. This approach can automatically learn the long term and short term patterns in the travel time data, and transfer the historical information through the connections between hidden layers of adjacent neural nets. This approach eliminates the cumbersome process of manually analyzing influence of social factors and natural factors on travel time, and can make multi-step travel time prediction. The computational experiments show that this approach can obtain small prediction error in single step prediction and good performance in multi-step prediction for most links. Comparisons demonstrate its capacity and potential in improving the accuracy of travel time prediction, especially for links with long distance.

文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/14817
专题毕业生_博士学位论文
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
段艳杰. 基于深度学习的交通预测方法研究[D]. 北京. 中国科学院研究生院,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
DYJ-phd.pdf(3128KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[段艳杰]的文章
百度学术
百度学术中相似的文章
[段艳杰]的文章
必应学术
必应学术中相似的文章
[段艳杰]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。