Abstract:Time series forecasting provides insights into the far future by utilizing the available history observations. Recent studies have demonstrated the superiority of transformer-based models in dealing with multivariate long-sequence time series forecasting (MLTSF). However, the data complexity hinders the forecasting accuracy of current deep neural network models. In this article, a hybrid framework - Waveformer - is proposed, which decomposes fluctuated and complex data sequence into multiple stable and more predictable subsequences (components) through the entire forecasting process. Waveformer interactively learns temporal dependencies on each pair of decomposed components, which enhances its ability of learning their temporal dependencies. Moreover, Waveformer treats the implicit and dynamic dependencies among variables as a set of dynamic direct graphs. Based on which, an attention adaptive graph convolution net (AAGCN) is designed, which combines self-attention and adaptive direct graph convolution to capture multivariate dynamic dependencies in a flexible manner. The experimental results on six public datasets show that Waveformer considerably outperforms a varied range of state-of-the-art benchmarks, with at the most 54.3% relative improvement.

A hybrid framework for multivariate long-sequence time series forecasting