Abstract: In the field of traffic prediction,there are complex and long-range spatial-temporal relationships between data.The existing graph structures fail to fully explore the implicit spatial-temporal relationships between data.To address the above problems,this study makes a series of improvements on the multivariate time series forecasting with graph neural networks(MTGNN),and proposes a spatial-temporal propagation graph neural network(STPGNN) for traffic prediction.Firstly,it captures information at different time scales through the multi-scale convolution module,and uses the feature fusion module to fuse to capture complex temporal information.After that,on the basis of the unidirectional adaptive graph structure of MTGNN,a bidirectional graph learning layer is designed and added to deeply explore and utilize the implicit bidirectional spatial relationship between data.Next,for the information transfer between network layers,a new information transfer method is proposed,which transmits the multi-scale time information in each layer to the next layer in turn,so as to better explore the complex and long-range spatial-temporal relationship.Finally,according to the temporal and spatial information of each level of the network,the prediction results are obtained by output convolution.Experiments are carried out on datasets such as METR-LA,PEMS-BAY and NE-BJ.The results show that STPGNN can effectively improve prediction accuracy and is better than some existing methods in three commonly indicators.This is especially true when it comes to longer-term forecasts.

Key words: Traffic prediction,Information transfer,Graph neural network,Multi-scale convolution