GenCast: Traffic Forecasting for Unobserved Regions (official code)

April 20, 2026 ยท View on GitHub

๐Ÿš€ AAAI 2026

What problem?

Forecast traffic in regions WITHOUT sensors

Why hard?

No historical observations โ†’ generalisation issue

What we do?

  • Physics-informed learning
  • Weather-traffic fusion
  • Spatial grouping

Result

โ†“ error 3.1%, โ†‘ Rยฒ 125%

Quick start

This code is based on our previous work STSM STSM Code. Our full paper is available at paper.

Requirements

-pytorch -pandas -numpy -tables -CUDA/12.5.1

The details are in the requirement.txt

Dataset

Google Drive: https://drive.google.com/drive/folders/1_imrTikGhIbZIyrRynG4bXhkaZA9xQgE?usp=share_link

Baidu Drive: https://pan.baidu.com/s/1DtyDp4stCKQ_P-4nox1O8A ๆๅ–็ : cast Due to the dataset is large, we will upload it to google drive and baiduyun for sharing. Or you can download traffic data from STSM and weather data from ERA5.

Putting Dataset dir under the GenCast dir.

Train the model

go to dir GenCast-L or GenCast-H

chmod +x ./metr.sh ./metr.sh