基于图卷积神经网络的城市交通流量预测方法研究
方深
2021-11-30
页数148
学位类型博士
中文摘要

城市交通流量预测任务在现代化城市生活中扮演重要角色,受到学术界与工业界普遍关注。

城市交通流量预测作为智能交通系统的核心功能之一,为多项智能交通服务提供了关键技术支撑。准确的预测结果可以帮助城市居民设计出行方案,协助管理部门优化交通控制策略,提升交通服务质量。

然而,由于城市交通网络存在复杂的时空关联特征,城市交通流量预测任务仍存在以下挑战。

首先,城市交通流量通常呈现多类型时序依赖关系。一般地,交通流量既存在短时(约一小时)近邻时序特征,也存在长时(一天、一周、甚至更长)周期时序特征。如何针对多类型时序依赖关系进行建模以提升城市交通流量预测性能,是需要进一步研究的问题。其次,城市交通流量存在复杂的空间依赖关系。由于交通节点流量特征很大程度上受其周围土地利用类型、城市功能影响,因此其可能同时存在局部与非局部空间关联。此外,天气条件、节假日通知、重大临时事件等因素也会影响城市交通流量。这些事件所关联的数据多源异构,如何融合这些数据是一个难点研究问题。


针对以上问题,本文开展城市交通流量预测方法研究。主要工作与创新如下:

  1. 提出了一种基于全局时空网络的城市交通流量预测方法。该方法通过构造多分辨率时间模块与全局关联空间模块,提取交通网络中的多类型时序特征,以及局部与非局部空间特征。具体地,时间模块由多层堆叠的张量因果卷积构成。其中,底层卷积感受野较短,提取短时近邻时序特征;高层卷积感受野较长,捕捉长时周期特征。另外,空间模块由局部谱域图卷积和非局部节点关联机制构成。其中,局部谱域图卷积用于学习近邻交通节点空间关联,而非局部节点关联机制挖掘远距离节点空间关系。全局时空网络融合时间与空间模块,在地铁、公交、出租车三个真实数据集上验证了有效性。
  2. 提出了一种基于多源时空网络的城市交通流量预测方法。其核心思想是,提取并融合交通流量时空特征与多源异构外部数据特征。具体地,该方法首先设计了一种膨胀注意力图卷积,可提取非局部空间特征。其次,针对交通流量特征提取问题,采用多层堆叠的膨胀注意力图卷积同时捕获交通网络局部与非局部空间特征,并融合了短时近邻与长时周期时序特征。然后,针对外部数据嵌入问题,依据数据属性嵌入了天气、节假日、和城市主要兴趣点空间分布等特征。最后,针对多源数据融合问题,先在空间维度融合交通流量时空特征与城市兴趣点分布特征,再在时间维度融合天气、节假日特征。在三个真实数据集上验证了所提方法有效性。
  3. 提出了一种基于元学习多源时空网络的城市交通流量预测方法。其核心思想是,采用元学习融合交通数据与多源异构时空外部数据。针对多源数据存在语义间隔问题,分别在时间与空间维度提出了元学习特征融合模块。在元学习框架下,外部数据与交通数据融合强度并非预先固定,而由元学习器训练习得。所提方法不仅消除多源异构数据间语义间隔,还可减少融合操作对数据源先验知识依赖。在此基础上,天气信息、节假日通知、和城市主要兴趣点分布等外部因素嵌入特征均可与历史交通流量数据融合。针对多类型时序特征提取问题,以及多步预测任务中的误差累积问题,提出的预测方法被设计为一种编码器——解码器结构。其中,编码器提取交通流量短时近邻趋势,解码器以历史周期数据作为近似估计,结合短时近邻趋势调整估计结果。在三个包含多源异构外部数据的真实交通数据集上,验证了所提方法有效性。
  4. 提出了一种基于结构自动搜索多源时空网络的城市交通流量预测方法。其核心思想是,采用网络结构自动搜索技术,提取交通网络中局部与非局部空间特征。具体地,针对不同交通场景下非局部空间关联强度不一致问题,网络结构搜索技术可自动选择并组合具有不同感受野的图卷积算子,构造最优网络结构。此外,该方法继承了上一工作中的两项重要特点:基于元学习的多源特征融合策略,与模型整体的编码器——解码器结构。因此,该方法考虑了局部与非局部空间关联、多类型时序特征,实现了多源异构数据融合,解决了多步预测任务中的误差积累等题。在北京市地铁、公交、出租车等不同类型的真实交通场景下的广泛实验,验证了所提方法有效性。
英文摘要

As one of the core component of Intelligent Transportation Systems, urban traffic flow prediction provides critical technical support for multiple intelligent transportation services.

For instance, accurate prediction results can help urban residents design travel plans, assist traffic management departments in optimizing traffic control strategies, and improve the quality of traffic services. In general, the task of urban traffic flow prediction plays a significant role in modern urban life, and has attracted widespread attention from the fields of academic and industry.

However, due to the complicated spatio-temporal correlations of urban traffic networks, this task is still faced with the following challenges. First, urban traffic flow usually presents multiple types of temporal dependencies. To be specific, traffic flow renders both short-term (about one hour) neighboring and long-term (one day, one week, or longer) periodic temporal features. How to model multiple types of temporal dependencies to improve the performance of urban traffic flow prediction is a key issue that require further research. Second, there exists complicated spatial dependences on urban traffic flow. Since the flow features on traffic nodes are largely affected by the nearby urban functions, both local and non-local spatial correlations could be implied. Third, external factors such as weather conditions, holiday notices, and major events affects urban traffic flow. The data associated with these events are multi-source and heterogeneous. How to fuse these data is a difficult research problem.

 

To address the above problems, this dissertation proposes a variety of solutions, and the main contributions and innovations are summarized as follows:

  1. A global spatio-temporal network (GSTNet) for urban traffic flow prediction is proposed. By constructing the multi-resolution temporal module and the global correlated spatial module, the GSTNet can extract multi-type temporal features, as well as the local and non-local spatial features on traffic networks. Specifically, the temporal module is composed by multi-layer stacked tensor causal convolution.   The low-layer convolution has a shorter receptive field and extracts short-term neighboring temporal features, while the high-layer convolution has a longer receptive field and captures long-term periodic features. On the other hand, the spatial module consists of localized spectral graph convolution and non-local correlated mechanism. Among them, the localized spectral graph convolution can model the spatial correlations of neighboring traffic nodes, while the non-local correlated mechanism mines the relationships of long-distance traffic nodes. Experiments on three real-world datasets verify the effectiveness of the proposed method.
  2. A multi-source spatio-temporal network (MSNet) for urban traffic flow prediction is proposed. The core idea of MSNet is to extract and integrate the traffic spatio-temporal features and the characteristics of multi-source heterogeneous external data. Specifically, an dilated attentional graph convolution~(DAGC) is first designed to extract the non-local spatial features. Second, to consider traffic flow feature extraction, the multi-layer stacked DAGC is adopted to capture the local and non-local spatial features, which also combines the short-term neighboring and the long-term periodic temporal features. Then, to consider external data, weather information, holiday notices, and the spatial distribution of points of interest are embedded based on the attributes. Finally, to consider multi-source data fusion, the traffic features and the distribution features of points of interest are first merged on spatial dimension, and then the weather and holiday features are fused on temporal dimension. The effectiveness of the proposed method is verified on three real-world datasets.
  3. A meta-learning based multi-source spatio-temporal network (MetaMSNet) for urban traffic flow prediction is proposed. The core idea of MetaMSNet is to adopt meta-learning to fuse traffic data with multi-source heterogeneous spatio-temporal external data. To consider the semantic gap between multi-source heterogeneous data, two meta-learning based feature fusion modules are proposed on temporal and spatial dimensions, respectively. Under the framework of meta-learning, the fusion intensity of between external and traffic data is not fixed in advance, but is learned by meta-learner. It not only eliminates the semantic gap between multi-source heterogeneous data, but also reduces the dependence of the fusion operation on the prior knowledge of the data source. To consider the different functions provided by the neighboring and periodic temporal data, as well as the error accumulation in multi-step prediction tasks, the MetaMSNet is designed as an encoder-decoder structure. The encoder extracts the short-term neighboring trend of traffic flow, and the decoder adoptes the periodic historical data as an approximate estimate, and adjusts the estimation through combining the short-term neighboring features. The effectiveness of the proposed method is verified on multiple real-world traffic datasets containing multi-source heterogeneous data.
  4. A novel multi-source spatio-temporal network via automatic neural architecture search (AutoMSNet) for urban traffic flow prediction is proposed. The core idea is to extract local and non-local spatial features based on the technique of automatic network structure search. Specifically, to consider the different strengthes of non-local spatial correlations in different traffic scenarios, the technique of neural architecture search is adopted to automatically select and combine graph convolution operators with different receptive fields to construct the optimal model structure. In addition, the proposal AutoMSNet inherits two significant features from the previous MetaMSNet: the meta-learning based multi-source feature fusion strategy, and the overall encoder-decoder structure of the model. Therefore, it can simultaneously solve the issue of local and non-local spatial correlations, multi-types of temporal features, multi-source heterogeneous data fusion, and error accumulation in multi-step prediction tasks. Extensive experiments under different real traffic scenarios such as Beijing subway, bus, and taxi have verified the effectiveness of the proposed method.
关键词深度学习 图卷积神经网络 城市交通流量预测 时空数据挖掘
语种中文
七大方向——子方向分类机器学习
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/46591
专题模式识别国家重点实验室_先进时空数据分析与学习
通讯作者方深
推荐引用方式
GB/T 7714
方深. 基于图卷积神经网络的城市交通流量预测方法研究[D]. 自动化大厦十三层第一会议室. 中国科学院自动化研究所,2021.
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
方深-基于图卷积神经网络的城市交通流量预(24749KB)学位论文 开放获取CC BY-NC-SA
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
[方深]的文章
百度学术
百度学术中相似的文章
[方深]的文章
必应学术
必应学术中相似的文章
[方深]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。