面向动态交通流多步预测的时空图模型
Spatio-temporal graph model for multi-step forecasting of dynamic traffic flow
杨平 1李成鑫 1刘宜成 1吕淳朴2
作者信息
- 1. 四川大学 电气工程学院,四川 成都 610065
- 2. 清华大学 自动化系,北京 100084
- 折叠
摘要
为更好表征交通路网中节点之间的动态隐式关系,提出一种基于时空数据嵌入的动态图卷积交通流预测模型.基于路网中节点之间的共现关系,利用深度游走算法将时空数据映射到嵌入空间中学习节点的向量表示;引入时隙嵌入特征与二维空间嵌入特征共同构建三维嵌入邻接张量,用于捕获时空依赖关系;在图卷积网络中添加自适应更新机制,利用循环组件演化图卷积网络的参数,以捕获图序列的动态性.将所提模型应用于基于真实交通数据集的交通流预测,结果验证了其有效性和提取路网隐式关系的准确性.
Abstract
To characterize dynamic implicit relationship between nodes in traffic network,a traffic flow forecasting model was proposed based on the dynamic graph convolution network considering the spatio-temporal data embedding.In view of the co-occurrence relationship between nodes in the road network,a method of the vector representation assisted by the deepwalk algorithm was designed that could map spatio-temporal data into embedding space.The time slot embedded feature and the two-dimensional spatial feature were introduced to jointly construct the three-dimensional embedding adjacency tensor for capturing the spatio-temporal dependencies.An adaptive updating mechanism was added to the graph convolution network,which used RNN components to evolve the parameters of the graph convolution network to catch the dynamic characteristic of the graph sequence.The model was applied to traffic flow forecasting based on real traffic data set,and the results verified the effectiveness of the proposed model and the accuracy of it in extracting the implicit relationship of the road network.
关键词
交通流预测/时空数据嵌入/深度游走算法/节点向量表示/时空依赖/动态图卷积/自适应更新机制Key words
traffic flow forecasting/spatio-temporal data embedding/deepwalk algorithm/node vector representation/spatio-temporal dependence/dynamic graph convolution/self-adapting updating mechanism引用本文复制引用
基金项目
四川大学-泸州战略合作基金项目(2020CDLZ-4)
出版年
2024
NETL
NSTL
万方数据