Non-stationary Transformers

This is the codebase for the paper: Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting, NeurIPS 2022. [Slides], [Poster].

🚩 News (2023.02) Non-stationary Transformer has been included in [Time-Series-Library], which covers long- and short-term forecasting, imputation, anomaly detection, and classification.

Discussions

There are already several discussions about our paper, we appreciate a lot for their valuable comments and efforts: [Official], [OpenReview], [Zhihu].

Architecture

Series Stationarization

Series Stationarization unifies the statistics of each input and converts the output with restored statistics for better predictability.

De-stationary Attention

De-stationary Attention is devised to recover the intrinsic non-stationary information into temporal dependencies by approximating distinguishable attentions learned from unstationarized series.

Showcases

Preparation

1. Install Python 3.7 and neccessary dependencies.

pip install -r requirements.txt

2. All the six benchmark datasets can be obtained from Google Drive or Tsinghua Cloud.

Training scripts

Non-stationary Transformer

We provide the Non-stationary Transformer experiment scripts and hyperparameters of all benchmark dataset under the folder ./scripts.

# Transformer with our framework
bash ./scripts/ECL_script/ns_Transformer.sh
bash ./scripts/Traffic_script/ns_Transformer.sh
bash ./scripts/Weather_script/ns_Transformer.sh
bash ./scripts/ILI_script/ns_Transformer.sh
bash ./scripts/Exchange_script/ns_Transformer.sh
bash ./scripts/ETT_script/ns_Transformer.sh

# Transformer baseline
bash ./scripts/ECL_script/Transformer.sh
bash ./scripts/Traffic_script/Transformer.sh
bash ./scripts/Weather_script/Transformer.sh
bash ./scripts/ILI_script/Transformer.sh
bash ./scripts/Exchange_script/Transformer.sh
bash ./scripts/ETT_script/Transformer.sh

Non-stationary framework to promote other Attention-based models

We also provide the scripts for other Attention-based models (Informer, Autoformer), for example:

# Informer promoted by our Non-stationary framework
bash ./scripts/Exchange_script/Informer.sh
bash ./scripts/Exchange_script/ns_Informer.sh

# Autoformer promoted by our Non-stationary framework
bash ./scripts/Weather_script/Autoformer.sh
bash ./scripts/Weather_script/ns_Autoformer.sh

Experiment Results

Main Results

For multivariate forecasting results, the vanilla Transformer equipped with our framework consistently achieves state-of-the-art performance in all six benchmarks and prediction lengths.

Model Promotion

By applying our framework to six mainstream Attention-based models. Our method consistently improves the forecasting ability. Overall, it achieves averaged 49.43% promotion on Transformer, 47.34% on Informer, 46.89% on Reformer, 10.57% on Autoformer, 5.17% on ETSformer and 4.51% on FEDformer, making each of them surpass previous state-of-the-art.

Future Work

We will keep equip the following models with our proposed Non-stationary Transformers framework:

• Transformer
• iTransformer
• Informer
• Autoformer
• FEDformer
• Crossformer
• Reformer
• ......

Note: Series Stationarization as an architecture-free module has been widely applied for addressing non-stationarity in time series. Please refer to time-series-library for the implementationdetails.

Citation

If you find this repo useful, please cite our paper.

@article{liu2022non,
  title={Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting},
  author={Liu, Yong and Wu, Haixu and Wang, Jianmin and Long, Mingsheng},
  booktitle={Advances in Neural Information Processing Systems},
  year={2022}
}

Contact

If you have any questions or want to use the code, please contact liuyong21@mails.tsinghua.edu.cn.

Acknowledgement

This repo is built on the Autoformer repo, we appreciate the authors a lot for their valuable code and efforts.