mahp-scut / STICM

Code and datasets for paper "Spatiotemporal convolutional network for time-series prediction with causal-factor inference"

STICM: Spatiotemporal information conversion machine for time-series forecasting

This respository includes codes and datasets for paper "Spatiotemporal information conversion machine for time-series forecasting".

The paper is online on Fundamental Research journal.

Spatiotemporal information conversion machine for time-series forecasting

• A spatiotemporal information conversion machine (STICM), was developed to efficiently and accurately render a multistep-ahead prediction of a time series by employing a spatial-temporal information (STI) transformation.

Data avalability

• We put data file including the locations of traffic loops in folder datasets/traffic.
• The locations of all 155 meteorological stations used in wind speed dataset are provided in folder datasets/ws.
• Other dataset are uploaded to Google Drive, and can be downloaded here.

Environment requirements

• python = 3.6
• tenforflow = 2.1
• cuda-version = 10.1
• cudnn-version = 7.6.5

Training and making predicitons

• We release the sample training codes and predicting codes corresponding to the Lorenz dataset, which is located at folder experiment/multi_sample_symmetric/. The script train.py is used for training and the script eval.py is used for evaluation after training the model.

• We can make predictions on other datasets by modify the given sample codes for Lorenz dataset.

Examples

• The example movie in terms of traffic speed prediction mentioned in our paper is given in folder example_movies/traffic_movie.mp4.

Citation

If you find this repository useful in your research, please consider citing the following papers:

@article{peng2022spatiotemporal,
  title={Spatiotemporal information conversion machine for time-series forecasting},
  author={Peng, Hao and Chen, Pei and Liu, Rui and Chen, Luonan},
  journal={Fundamental Research},
  year={2022},
  publisher={Elsevier}
}