CASIA OpenIR  > 毕业生  > 硕士学位论文
基于深度学习的交通流预测方法研究
王春翔
2022-05-17
Pages67
Subtype硕士
Abstract

交通预测作为智能交通系统的关键技术之一,能够为交通信号控制、出行路线选择、车辆调度等提供技术支撑,一直受到领域内学者的广泛关注。近年来,深度学习方法被广泛应用于交通预测领域,并取得了一系列成果。在实践中,采集到的交通数据会存在缺失或损坏等问题。如何从缺失、不完整的交通数据出发,构建基于深度学习的交通预测方法,是当前的研究热点,涉及交通数据弥补、以及弥补预测同步进行的端到端深度学习网络设计等问题。本文针对这些问题进行研究,主要工作如下: 

(1) 基于图采样与聚合框架的交通数据弥补方法。通过对交通路网中不同站点的交通数据进行相关性分析生成空间邻接矩阵。在此基础上,设计了一种时空弥补模型,通过堆叠的空间信息聚合层和时间卷积层提取交通数据时空特征,并以高斯聚合函数作为权重聚合缺失数据节点的邻居信息得到其特征表示。进而生成缺失数据,完成数据弥补任务。实验结果表明,该模型可有效利用邻居节点信息提升弥补效果。

(2) 基于循环注意力单元的交通流预测方法。受到注意力机制调节信息传递作用的启发,将其嵌入到循环神经网络中代替长短期记忆网络或门控循环单元的“门”结构,设计了一种能够在保证预测性能的同时大幅度减少模型参数的交通流量预测模型-循环注意力单元。实验结果表明,该模型在处理交通时间序列数据时,注意力机制能够使模型关注重要信息,以较少的参数实现较高的预测性能。

(3) 端到端交通流同步弥补-预测方法。考虑到交通数据弥补和预测任务的相关性,设计了一种多任务同步网络结构,将两者整合在一个网络中。该网络以非完整数据为输入,同步完成交通数据弥补和预测任务,输出最终的预测结果,实现了端到端的交通流预测。为了更好地融合两个任务,设计了一种自适应加权融合损失函数来动态调整弥补和预测任务的权重。实验结果表明,该模型可以有效融合数据弥补和预测任务,提升预测性能。

Other Abstract

As one of the key technologies of intelligent transportation systems, traffic prediction can provide technical support for traffic signal control, travel route choice, vehicle scheduling, etc., and has attracted a lot of interest. In recent years, deep learning methods have been widely used in the field of traffic prediction and have achieved good results. In practice, the collected traffic data may be missing or corrupted. How to construct a traffic prediction method based on deep learning from the missing and incomplete traffic data is a hot research issue at present, which involves traffic data imputation and end-to-end deep learning network design for simultaneous imputation and prediction. The main work of this thesis is as follows.

(1) A traffic data imputation method based on graph sampling and aggregation. Through the correlation analysis of traffic data of different stations in a traffic network, the spatial adjacency matrix is generated. On this basis, a spatio-temporal imputation model is designed to extract spatio-temporal features of traffic data through the stacked spatial information aggregation layer and time convolution layer for generating. The Gaussian aggregation function is used as the weight to aggregate the neighbor information of the missing data node to obtain its feature representation. Experimental results show that the model can effectively use the neighbor node information to improve the imputation effect.

(2) Traffic Flow prediction based on recurrent attention unit. Inspired by the role of attention mechanism in regulating information transmission, this thesis embedded the attention module into the recurrent neural network to replace the "gate" structure of long short-term memory network or gated recurrent unit, and designed a traffic flow prediction model - Recurrent Attention Unit, which can greatly reduce the model parameters while ensuring prediction performance. Experimental results show that the attentional mechanism can make the model focus on important information and achieve high prediction performance with less parameters when processing traffic time series data.

(3) End-to-end traffic flow synchronous imputation-prediction method. Considering the correlation between traffic data imputation and prediction tasks, a multi-task synchronous network structure is designed to integrate the two tasks within one deep network. The network takes missing data as input, completes traffic data imputation and prediction tasks synchronously, outputs the final prediction result, and realizes end-to-end traffic flow prediction. In order to better fuse the two tasks, an adaptive weighted fusion loss function is designed to dynamically adjust the weights of imputing and forecasting tasks. Experimental results show that the model can effectively integrate data imputation and prediction tasks to improve prediction performance.

Keyword交通流预测,数据弥补,时空特征提取,图神经网络,深度学习
Subject Area控制理论其他学科
MOST Discipline Catalogue工学
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/48851
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
王春翔. 基于深度学习的交通流预测方法研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2022.
Files in This Item:
File Name/Size DocType Version Access License
基于深度学习的交通流预测方法研究.pdf(4552KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[王春翔]'s Articles
Baidu academic
Similar articles in Baidu academic
[王春翔]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[王春翔]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.