SageFormer: Series-Aware Framework for Long-Term Multivariate Time-Series Forecasting

This repository contains the code for the paper "SageFormer: Series-Aware Framework for Long-Term Multivariate Time-Series Forecasting" by Zhenwei Zhang, Linghang Meng, and Yuantao Gu, published in the IEEE Internet of Things Journal.

Introduction

SageFormer is a novel series-aware graph-enhanced Transformer model designed for long-term forecasting of multivariate time-series (MTS) data. With the proliferation of IoT devices, MTS data has become ubiquitous, necessitating advanced models to forecast future behaviors. SageFormer addresses the challenge of capturing both intra- and inter-series dependencies, enhancing the predictive performance of Transformer-based models.

Usage

To train and evaluate the SageFormer model:

• Clone this repository
• Download datasets from Google Drive or Baidu Drive and place them in the ./dataset folder
• Create a virtual environment and activate it
• Install requirements pip install -r requirements.txt
• Run scripts in the ./scripts folder to train and evaluate the model, for example:

  sh scripts/long_term_forecast/ECL_script/SageFormer.sh

• Model checkpoints and logs will be saved to outputs folder

Contacts

For any questions, please contact the authors at zzw20 [at] mails.tsinghua.edu.cn or write a discussion on github.

Citation

If you find this code or paper useful for your research, please cite:

@ARTICLE{zhang2024sageformer,
  author={Zhang, Zhenwei and Meng, Linghang and Gu, Yuantao},
  journal={IEEE Internet of Things Journal}, 
  title={SageFormer: Series-Aware Framework for Long-Term Multivariate Time Series Forecasting}, 
  year={2024},
  doi={10.1109/JIOT.2024.3363451}}

Acknowledgement

This library is constructed based on the following repos:

• https://github.com/thuml/Time-Series-Library
• https://github.com/PatchTST/PatchTST