MSTSAGCNet: Multi-scale spatio-temporal feature fusion model for multi-variable time series prediction

In the context of the big data era, accurate prediction of multivariate time series is crucial for enhancing the efficiency of management decision-making across various industries. Currently, graph neural networks have emerged as popular models for multivariate temporal prediction due to their capability to model both spatial (inter-variable) and spatio-temporal relationships. Recent research on multi-scale time series prediction using graph neural networks has considered features at multiple scales; however, studies on the types of multi-scale features and the interactions between different scales remain insufficient. To address these issues, this paper proposes a Spatio-Temporal Adaptive Graph Convolutional Network model for multiple scales (MSTSAGCNet) and conducts related experiments. First, a multi-scale learning and generation module for spatio-temporal feature maps is constructed to model three types of spatio-temporal relationships at different scales and generate corresponding graph structures. Next, a cross-scale relation module is built to further learn the relationships between different scales. Finally, comparative experiments and ablation studies are performed on six public datasets to validate the effectiveness and superiority of the proposed model. Experimental results demonstrate that the proposed model achieves better prediction performance than baseline models, and the two novel modules positively impact the model’s prediction accuracy.