This paper introduces PPN (Pollution-Predicting Net for PM2.5), a spatial-temporal deep learning framework for regional PM2.5 concentration prediction. PPN combines preceding PM2.5 observations and numerical weather prediction (NWP) data within an encoder-decoder architecture and uses a weighted loss function to improve forecasting accuracy, particularly during extreme events. Applied to the Beijing-Tianjin-Hebei region, PPN achieved an R² of 0.7 and RMSE of 17.7 µg m⁻³. Comparisons with WRF-Chem show PPN's superior performance, especially within the first 24 hours.
