TimeBridge: Non-Stationarity Matters for Long-term Time Series Forecasting

[中文解读1] [中文解读2] [BiliBIli Video]

Updates

🚩 2025-05-01: TimeBridge has been accepted as ICML 2025 Poster.

🚩 2025-04-18: Release the detailed training logs (see _logs).

🚩 2025-02-11: Release the code.

🚩 2024-10-08: Initial upload to arXiv [PDF].

Usage

1. Install the dependencies

   pip install -r requirements.txt

2. Obtain the dataset from Google Drive and extract it to the root directory of the project. Make sure the extracted folder is named dataset and has the following structure:

   dataset
   ├── electricity
   │   └── electricity.csv
   ├── ETT-small
   │   ├── ETTh1.csv
   │   ├── ETTh2.csv
   │   ├── ETTm1.csv
   │   └── ETTm2.csv
   ├── PEMS
   │   ├── PEMS03.npz
   │   ├── PEMS04.npz
   │   ├── PEMS07.npz
   │   └── PEMS08.csv
   ├── Solar
   │   └── solar_AL.txt
   ├── traffic
   │   └── traffic.csv
   └── weather
       └── weather.csv
   

3. Train and evaluate the model. All the training scripts are located in the scripts directory. For example, to train the model on the Solar-Energy dataset, run the following command:

   sh ./scripts/TimeBridge.sh

Bibtex

If you find this work useful, please consider citing it:

@article{liu2025timebridge,
      title={TimeBridge: Non-Stationarity Matters for Long-term Time Series Forecasting}, 
      author={Liu, Peiyuan and Wu, Beiliang and Hu, Yifan and Li, Naiqi and Dai, Tao and Bao, Jigang and Xia, Shu-Tao},
      journal={International Conference on Machine Learning},
      year={2025},
}

Contact

If you have any questions, please get in touch with [email protected] or submit an issue.