Spatial-temporal upsampling graph convolutional network for daily long-term traffic speed prediction

Abstract

The daily long-term traffic prediction is an important urban computing issue, and can give users a global insight into traffic. Accurate traffic prediction is conducive to rational route planning and efficient traffic resource allocation. However, it is challenging to capture the global spatial-temporal correlations for daily long-term traffic prediction. In this paper, we propose a spatial-temporal upsampling graph convolutional network (STUGCN) for daily long-term traffic speed prediction. STUGCN uses an innovative upsampling method to capture the global spatial-temporal correlations. Specifically, in spatial dimension, we construct an upsampled road network by adding virtual nodes to the original road network to capture local and global spatial correlations. In temporal dimension, we build a time graph to capture the temporal correlations among adjacent time steps. Besides, we construct a knowledge base, and the global temporal correlations can be captured by upsampling the current day from the knowledge base. Therefore, STUGCN not only preserves the local spatial-temporal correlations, but also has the ability to learn global spatial-temporal correlations. The experimental results on two real-world datasets demonstrate that our approach is approximately 16.4%–17.1%, 14.1%–17.0% and 17.4%–22.4% better than the state-of-the-art in terms of MAE, RMSE and MAPE metrics, respectively.

Keywords

• Long-term traffic speed prediction
• Spatial-temporal upsampling
• Graph convolutional network
• Intelligent transportation system