Prediction of Multivariate Spatial-Temporal Series Data Based on Adaptive Spatial-Temporal Information

This paper proposes an Prediction of Multivariate Spatial-Temporal Series Data Based on Adaptive Spatial-Temporal Information(ASTCN), which learns complex spatio-temporal information from multivariate spatio-temporal series through trainable temporal embeddings and graph adjacency matrices. A gated fusion mechanism is employed to control the proportion of different temporal embeddings to improve prediction accuracy. Visualization of the temporal embeddings reveals that weekly temporal embeddings have the greatest impact on prediction accuracy, followed by daily temporal embeddings, while monthly temporal embeddings have the least impact. Additionally, a novel method for constructing graph adjacency matrices is introduced. Ablation experiments demonstrate that the two types of graph adjacency matrices proposed in this method have varying degrees of influence on improving the prediction accuracy of the dataset. Consequently, this paper integrates the two graph adjacency matrices, enabling ASTCN to achieve superior prediction accuracy on traffic speed, traffic flow, and air quality datasets compared to when either matrix is used alone. In comparative experiments, ASTCN ultimately achieves excellent prediction performance with relatively low training costs.