【作者】 王鑫；

【导师】 李炜；

【作者基本信息】 安徽大学 ， 计算机技术（专业学位）， 2021， 硕士

【摘要】 近年来,民用汽车保有量的持续增长,导致交通供需不平衡问题日益严重。交通流数据能够直观反映出实时的交通状况,若能提前了解可靠的交通信息,将有助于交通管理者制定和实施交通规划策略,有效地降低公共安全风险。同时,还可以帮助旅行者更好地规划出行路线,减少时间成本和经济损失。因此对道路的管理者和使用者都具有十分重要的意义。然而,由于道路之间的时空相关性复杂多变,交通流数据的非线性和随机性特征,交通流预测变得非常具有挑战性。现有的大部分预测方法没有根据实时路况信息有效地进行时空相关性分析,并且没有考虑人类活动导致不同时间段的数据周期存在差异性。为了解决上述问题,本文基于时空图神经网络提出了自适应卷积和多元数据关联的交通流预测方法,主要贡献如下:(1)本文提出一种自适应卷积的图神经网络模型(Adaptive Convolution Graph Neural Network model,ACGNN)。该模型由自适应卷积块和门控循环单元组成,首先通过自适应卷积块中的广度函数以道路为中心,从广度的角度对邻域道路进行探索,再通过深度函数进行选择性遗忘,以此来学习到自己的卷积路径并根据该路径捕捉空间依赖关系。最后,融合门控循环单元进行时空相关性分析。(2)本文提出了一种新的多元数据关联的图卷积网络模型(Multisensor Data Correlation Graph Convolution Network model,MDCGCN)。该模型由近周期、日周期和周周期三个组件构成,每一组件又由两部分组成:1)基准自适应机制块;2)多元数据关联卷积块。第一部分可以有效地消除周期性数据之间的差异,使得模型根据实时路况信息自适应地选择合适的周期数据作为输入,有效地确保了输入数据的质量。第二部分通过分析道路间实时的交通模式变化关系,构建交通模式关系图并结合原始交通拓扑图,有效地分析时空相关性。(3)为验证本文方法的可靠性,在真实城市交通流数据集上进行了大量的实验,并与其他方法进行了对比。实验结果表明,本文所提出的方法在各类评价指标上都优于对比方法。更多还原

【Abstract】 Recently,the number of car keepers are increasing,which causes a serious problem of imbalance on traffic supply and demand.The traffic flow data can directly reflects real-time traffic conditions.If reliable traffic information can be known in advance,it will help traffic managers to formulate and implement traffic planning strategies,effectively reducing public safety risks.At the same time,it can also help travelers to better plan their travel routes,reducing time cost and economic loss.Therefore,it is of great significance to the road managers and users.However,the spatio-temporal correlation between roads is complex and variable,and the traffic flow data is highly nonlinear and random,which makes the traffic flow prediction very challenging.At present,most of the existing methods do not effectively analyze the spatiotemporal correlation based on real-time traffic information and do not take into account the differences in data cycles at different time periods caused by human activities.To address these issues,this thesis proposes an adaptive convolution and multivariate data association traffic flow prediction method based on spatio-temporal graph neural network.The main contributions are as follows:(1)This thesis proposes Adaptive Convolution Graph Neural Network model(ACGNN).The model constructed by adaptive convolution block and gating recycling units.First,from the Angle of breadth,explore neighboring roads by adaptive breadth of convolution block function centering on the road,and then forget selectively through the depth function,in order to learn self convolution path and capture spatial dependencies according to the study of the path.Finally,the spatio-temporal correlation analysis was performed by combining the gated cyclic unit.(2)This thesis proposes Multisensor Data Correlation Graph Convolution Network(MDCGCN)model.The MDCGCN model consists of three parts: recent,daily period and weekly period components,and each of which consists of two parts: 1)benchmark adaptive mechanism block;2)multisensor data correlation convolution block.The first part can eliminate the differences among the periodic data and effectively improve the quality of data input.By analyzing the real-time changing relationship of traffic patterns among roads,the second part constructs a traffic pattern relationship graph and combines the original traffic topology graph to effectively analysis the spatio-temporal correlation.(3)To verify the reliability of the proposed method,a large number of experiments were carried out on traffic flow data sets,and compared with other methods.The experimental results show that the proposed method is superior to the comparison methods in all kinds of evaluation indexes.更多还原