SIGMAformer: A spatiotemporal gaussian mixture correlation transformer for multi-sensor weather fusion forecasting

Proposed SIGMAformer framework

The proposed framework replaces the conventional attention mechanism with the DSTC module, which enhances spatiotemporal weather forecasting by leveraging GMM-based pattern extraction to compute and aggregate weighted temporal and spatial correlations.

Highilights

- Develop SIGMAformer for global weather via multi-station spatiotemporal correlation.

- Gaussian-mixture pattern extraction (GMPE) guides adaptive temporal and spatial attention.

- Attention maps reveal regional and temporal drivers, enhancing interpretability.

- Captures short-term fluctuations, long-term trends, and near/far-station correlations.

- Delivers state-of-the-art forecasts of global temperature and wind speed compared with baselines.

Usage

1. Global weather datasets can be obtained from [Google Drive].

2. Install Pytorch and other necessary dependencies.

pip install -r requirements.txt

3. Train and evaluate model. We provide the experiment scripts under the folder ./scripts/. You can reproduce the experiment results as the following examples:

bash ./scripts/Global_Temp/SIGMAformer.sh
bash ./scripts/Global_Wind/SIGMAformer.sh