MCAGCN: Multi-component attention graph convolutional neural network for road travel time prediction

文献类型： 外文期刊

第一作者： Zhao, Zhihua

作者： Zhao, Zhihua;Li, Chao;Zeng, Qingtian;Zhang, Xue;Zeng, Qingtian;Li, Chao;Xie, Nengfu

作者机构：

关键词： data analysis; data mining; graph theory

期刊名称：IET INTELLIGENT TRANSPORT SYSTEMS （ 影响因子：2.7； 五年影响因子：3.2 ）

ISSN： 1751-956X

年卷期： 2023 年

页码：

收录情况： SCI

摘要： With the development of intelligent transportation technology, road travel time prediction has become an important research direction. Owing to the complex periodic dependence and non-linear features of road travel time series, achieving accurate and effective predictions remains a challenging task. Most existing traffic sequence prediction methods lack modelling of the dynamic correlation between multiple period information, resulting in unsatisfactory prediction results. To address this, a multi-component attention graph convolutional network (MCAGCN) is proposed for road travel time prediction. First, the spatial-temporal features of three historical components (hourly, daily and weekly) are modelled individually. A skip attention layer is then used to fuse multi-scale spatial-temporal features to enhance the model's feature extraction capabilities. Secondly, a component attention layer is proposed to calculate the correlation between different components using the temporal features of the prediction moment, to achieve dynamic modelling between different period information. The experimental results on the Tianchi, METR-LA, and PeMS-BAY datasets, which are real-world traffic forecasting datasets, demonstrate the superiority of MCAGCN. Multi-component attention graph convolutional network consists of three parallel components that extract spatiotemporal features from Xh, Xd, and Xw, respectively. Each component contains a skip attention layer to fuse spatiotemporal features of different scales. Subsequently, the three components along with the time feature of the prediction moment are fed into a component attention layer for dynamic period modelling. Finally, the prediction result is obtained through the output layer.image

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序