PA²MGCN

Official PyTorch implementation of PA²MGCN :Patch Attention Based Adaptive Multi-Head Graph Convolutional Networks for Traffic Flow Forecasting.

Figure 1. Overall architecture of PA²MGCN.

We recommend using Conda.

conda create -n PA2MGCN python=3.9 -y
conda activate PA2MGCN

pip install -r requirements.txt

Supported traffic datasets include:

• METR-LA
• PeMS-Bay
• PEMS04
• PEMS08

Download link:

• Baidu Netdisk: Link
• Google Drive: Link

For the SD dataset source, please refer to: LargeST

⚠️ Note for SD data conversion

You need to modify LargeST's original SD data generation code first. By default, LargeST saves his.npz with z-score normalized flow values. PA²MGCN applies its own normalization internally, so using that file directly will cause double normalization. Remove the standardization step from LargeST's save pipeline before converting the SD dataset for PA²MGCN.

Place dataset files under datasets/ with a structure similar to:

PA2MGCN/
├── datasets/
│   ├── METR-LA/
│   ├── PeMS-Bay/
│   ├── PEMS04/
│   ├── PEMS08/
│   └── SD/
├── scripts/
└── ...

The project provides training scripts in scripts/.

Run commands:

chmod 776 ./scripts/train.sh
./scripts/train.sh

Paper information is coming soon.