Code and processed data for: Physics-guided graph attention networks for multi-horizon PM2.5 forecasting on the northern slope of the Tianshan Mountains

It includes the full analysis pipeline (data cleaning, feature engineering, baseline models, Physics-GAT training, ablations, adaptive prediction-level fusion, and bootstrap significance tests), saved model outputs and weights, the final figures, and the processed hourly dataset cleaned_data.csv for 16 stations.

See README.md in the archive for step-by-step reproduction instructions. A CUDA GPU is recommended for full retraining; all figures and the attention-weight extraction can be regenerated on a CPU-only machine from the saved results.