MmgFra: A multiscale multigraph learning framework for traffic prediction in smart cities

Traffic prediction is an important part of smart city projects. Due to the complex topology of urban road network and the dynamic change of traffic data, establishing a spatio-temporal model to accurately predict traffic volume remains a challenging task at present. Recently, Graph Convolution Networks (GCN) have been widely used to extract features from non-grid data, and time sequence models have been used to learn temporal features of traffic distributions. However, current GCN based methods only make use of the natural structure of road network, while ignoring the information of administrative units, neighborhood units and other hierarchical structures of spatial interaction. Therefore, traditional models are typically developed under one single scale, which are far from the prediction of multi-scale systems. To address this issue, we propose a novel framework, named MmgFra, to merge multi-scale information for high-precision urban traffic flow prediction. MmgFra consists of three components: a spatial feature extraction module, a feature clustering fusion module, and a temporal feature extraction module. Specifically, we use stacking GCNs to extract spatial structure information of the city road network, administrative unit road network, and neighborhood units, and employ a DIFFPOOL module to cluster and fuse the above information. Finally, we introduce GRU to capture temporal features. We evaluated its performance on a real city dataset using different time scales. The experimental results indicate that compared with state-of-the-art neural network models and GCN-based variant models, our model exhibits higher predictive accuracy. MmgFra improves by approximately 8.4%-29.5% and 7.5%-30.6% in terms of RMSE and MAPE metrics, respectively.

Automated summary

Traffic prediction is a challenging task due to the complex urban road network and dynamic traffic data. In this study, we propose a novel framework, named MmgFra, to merge multi-scale information for high-precision urban traffic flow prediction. The experimental results show that our model outperforms state-of-the-art neural network models and GCN-based variant models in terms of predictive accuracy.