This repo is the official implementation for Multi-Scale Adaptive Graph Neural Network for Multivariate Time Series Forecasting.

• Python 3.6.12
• PyTorch 1.0.0
• math, sklearn, numpy

To evaluate the performance of MAGNN, we conduct experiments on four public benchmark datasets：Solar-Energy, Traffic, Electricity, and Exchange-Rate.

This dataset contains the collected solar power from the National Renewable Energy Laboratory, which is sampled every 10 minutes from 137 PV plants in Alabama State in 2007.

This dataset contains the road occupancy rates (between 0 and 1) from the California Department of Transportation, which is hourly aggregated from 862 sensors in San Francisco Bay Area from 2015 to 2016.

This dataset contains the electricity consumption from the UCI Machine Learning Repository, which is hourly aggregated from 321 clients from 2012 to 2014.

This dataset contains the exchange rates of eight countries, which is sampled daily from 1990 to 2016.

# Hyper-parameters search with NNI
 nnictl create --config config.yml --port 8080

# Train on Solar-Energy
CUDA_LAUNCH_BLOCKING=1 python train.py --save ./model-solar-1.pt --data solar-energy/solar-energy.txt --num_nodes 137 --batch_size 4 --epochs 50 --horizon 3
# Train on Traffic
CUDA_LAUNCH_BLOCKING=1 python train.py --save ./model-traffic-2.pt --data traffic/traffic.txt --num_nodes 862 --batch_size 4 --epochs 50 --horizon 3
# Train on Electricity
CUDA_LAUNCH_BLOCKING=1 python train.py --save ./model-electricity-3.pt --data electricity/electricity.txt --num_nodes 321 --batch_size 4 --epochs 50 --horizon 3
# Train on Exchange-Rate
CUDA_LAUNCH_BLOCKING=1 python train.py --save ./model-exchange-4.pt --data exchange_rate/exchange_rate.txt --num_nodes 8 --batch_size 4 --epochs 50 --horizon 3

Please cite the following paper if you use the code in your work:

@article{chen2022multi,
  title={Multi-Scale Adaptive Graph Neural Network for Multivariate Time Series Forecasting},
  author={Chen, Ling and Chen, Donghui and Shang, Zongjiang and Zhang, Youdong and Wen, Bo and Yang, Chenghu},
  journal={arXiv preprint arXiv:2201.04828},
  year={2022}
}