Dsformer: A Double Sampling Transformer for Multivariate Time Series Long-term Prediction

By Chengqing Yu et al.

Read the original document by opening this link in a new tab.

Table of Contents

1 INTRODUCTION
2 RELATED WORK
2.1 Deep learning based methods
2.2 Transformer based methods
3 METHODOLOGY
3.1 Preliminaries
3.2 Overall framework of the proposed model
3.3 Double sampling block

Summary

Multivariate time series long-term prediction aims to predict data changes over a long period. Transformer-based models have been effective but often do not fully utilize global information, local information, and variable correlation. To address this, the DSformer model is proposed, consisting of a double sampling block and a temporal variable attention block. The DS block uses down sampling and piecewise sampling to focus on global and local information, while the TVA block combines temporal and variable attention to mine and integrate key information. Experimental results show the DSformer outperforms existing models on real-world datasets.