Multivariate long sequence time-series forecasting using dynamic graph learning

Xiaohu Wang, Yong Wang, Jianjian Peng, Zhicheng Zhang

Abstract: Time series prediction is a subset of temporal data mining, which seeks to forecast its values in the future by using the accessible historical observations within the specified time periods. Deep neural networks have shown their superiority in predicting time series according to recent studies. Unfortunately, most models overlook differences and interdependencies between variables when tackling the multivariate long sequence time-series forecasting problem. In order to explicitly learn them and take them into account at a fine-grained level to resolve the dynamics of variable dependencies, we introduce Graph Convolutional Networks into Transformer in this article. To counteract the local insensitivity of the Transformer, we further integrate the Temporal Convolutional Network as a component of the self-attention layer. The experimental evaluation results on four open datasets show that our model performs noticeably better than a diverse range of state-of-the-art benchmarks.