WaveForM

This is a PyTorch implementation of the paper: WaveForM: Graph Enhanced Wavelet Learning for Long Sequence Forecasting of Multivariate Time Series, published in AAAI 2023.

Requirements

To run the code, you need to install Python(>=3.9.12) and PyTorch(>=1.11.0) at least. The full requirements are specified in requirements.txt.

The pytorch_wavelets package should be installed following the instructions of pytorch_wavelets.

Data

The Electricity, Solar-energy and Traffic datasets can be downloaded from multivariate-time-series-data.

The Weather datasets can be downloaded from Autoformer.

The Temperature datasets (asos) can be downloaded from spacetimeformer.

You should put the xx.csv file into the directory with the datasets' name in dataset directory.

For example, the proper file structure should be like:

dataset
|-- electricity
| |-- electricity.csv
|--solar
| |-- solar.csv
|--temperature
| |-- temperature.csv
|--traffic
| |-- traffic.csv
|--weather
| |-- weather.csv

Running the code

The running script is run.sh, where you can change any arguments which have been declared in run.py.

Contact

If you have any questions, you can raise an issue or send an e-mail to [email protected].

Acknowledgement

Thanks to the following repos for their codes and datasets.

https://github.com/fbcotter/pytorch_wavelets

https://github.com/thuml/Autoformer

https://github.com/nnzhan/MTGNN

https://github.com/laiguokun/multivariate-time-series-data