Awesome

Spatio-Temporal Differential Equation Network

This is a Pytroch implementation of Spatio-temporal Differential Equation Network (STDEN) for physics-guided traffic flow prediction, as described in our paper: Jiahao Ji, Jingyuan Wang, Zhe Jiang, Jiawei Jiang, and Hu Zhang, STDEN: Towards Physics-Guided Neural Networks for Traffic Flow Prediction, AAAI 2022.

The training framework of this project comes from chnsh. Thanks a lot :)

Requirement

• scipy>=1.5.2
• numpy>=1.19.1
• pandas>=1.1.5
• pyyaml>=5.3.1
• pytorch>=1.7.1
• future>=0.18.2
• torchdiffeq>=0.2.0

Dependency can be installed using the following command:

pip install -r requirements.txt

Model Traning and Evaluation

You can run the code by

# traning for dataset GT-221
python stden_train.py --config_filename=configs/stden_gt.yaml

# testing for dataset GT-221
python stden_eval.py --config_filename=configs/stden_gt.yaml

The configuration file of all datasets are as follows:

Note the data is not public, and I am not allowed to distribute it.

Cite

If you find the paper useful, please cite as following:

@inproceedings{ji2022stden, 
  title={{STDEN}: Towards physics-guided neural networks for traffic flow prediction}, 
  author={Ji, Jiahao and Wang, Jingyuan and Jiang, Zhe and Jiang, Jiawei and Zhang, Hu}, 
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence}, 
  year={2022}, 
  volume={36}, 
  number={4},   
  pages={4048-4056}
}