Time-lagged relation graph neural network for multivariate time series forecasting

Abstract

Recently, Graph Neural Network-based approaches (GNNs) have been widely studied in Multivariate Time Series (MTS) prediction, which could extract information from the closely related variables for prediction. The variables contained in MTS data are lagged correlated, and the future trends of the lagging variables are guided by the leading variables. However, as the existing approaches only focus on delay-free relations, they cannot utilize the guidance information in leading variables to achieve accurate prediction. To address this issue, we propose a novel frame called the Time-Lagged Relation Graph Neural Network (TLGNN) including two key components: the time-lagged relation graph and the time-lagged relation graph learning. The time-lagged relation graph could explicitly model the time-delay relations among MTS variables by connecting variable nodes at lag intervals. The graph learning module could adaptively extract the time-delay relations among MTS variables. Based on the novel designed graph structure, the TLGNN could extract the guidance information from previous values of leading variables to generate more efficient feature representations for prediction. In experiments, the prediction accuracy is significantly improved due to the full exploration of the time-delay relations. Compared with existing methods, the TLGNN achieves the best results in both the single-step prediction and the multi-step prediction tasks.