Dynamic graph structure correction with nonadjacent correlations for multivariate time series forecasting

Effectively modeling the relations between variables in multivariate time series is of utmost importance for accomplishing accurate predictions. In real-world scenarios, in addition to sequential correlations, the evolution of relations between variables also exhibits nonadjacent correlations at different scales. However, existing methods primarily focus on constructing dynamic graph structures at each time step using temporal features extracted by continuous temporal models, which cannot capture above latent dependencies. In this study, we introduce the Dynamic Graph Structure Correction (DGC) model, leveraging a multi-scale framework with dilated convolution. To take full advantage of nonadjacent correlations in the evolution of relations between variables, we adaptively select history-related graph structures to correct initial graph structure constructed by Gate Recurrent Units. In addition, we design a time-decay-based attention mechanism to address the influence of time intervals between history-related and current time steps. Finally, the evolved graph structures are fed into graph neural networks to handle the multi-scale and complex structural relations. Our proposed model achieves superior performance compared to state-of-the-art methods in multivariate time series forecasting, as evidenced by the evaluation results on four widely used benchmark datasets.