Physics-Constrained ST-GNNs for High-Performance PM2.5 Prediction

PM2.5 pollution is a serious global health problem. This paper introduces PIST-GNN, a new model for predicting PM2.5 levels in the Beijing-Tianjin-Hebei region that performs very well. Unlike typical machine learning methods, PIST-GNN considers real-world physics. It directly incorporates equations that describe how air particles move into its design as a spatio-temporal graph neural network. The model uses Graph Convolutional Networks to understand relationships between locations and to estimate how PM2.5 moves based on these physics equations. This allows for more accurate predictions over multiple time steps, guided by physical principles. Our experiments demonstrate that PIST-GNN is significantly better. It reduces the prediction error (RMSE) by 47.5% for one-step predictions and by $23.23 \pm 0.32$ for 8-step predictions. These results highlight PIST-GNN’s strong capability for highly accurate PM2.5 forecasting.