Non-stationary Transformers

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

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
• Autoformer
• Informer
• LogTrans
• Reformer
• FEDformer
• Pyraformer
• ETSformer

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 [email protected].

Acknowledgement

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