Regional aerosol forecasts based on deep learning and numerical weather prediction

Abstract Atmospheric chemistry transport models have been extensively applied in aerosol forecasts over recent decades, whereas they are facing challenges from uncertainties emission rates, meteorological data, and over-simplified chemical parameterizations. Here, we developed a spatial-temporal deep learning framework, named PPN (Pollution-Predicting Net for PM 2.5 ), to accurately efficiently predict regional concentrations. It has an encoder-decoder architecture combines the preceding observations numerical weather prediction. Besides, model proposes weighted loss function promote forecasting performance extreme events. We proposed forecast 3-day concentrations Beijing-Tianjin-Hebei region China on three-hour-by-three-hour basis. Overall, showed good with R 2 RMSE values of 0.7 17.7 μg m −3 , respectively. could capture high concentration south relatively low north exhibit better within next 24 h. The use decreased level “high underestimation, overestimation”, while incorporating into encoder phase improved predictive accuracy also compared result that state-of-the-art (WRF-Chem pollutant data assimilation). temporal WRF-Chem were 0.30−0.77 19−45 those 0.42−0.84 15−42 . shows powerful capacity provides efficient accurate tool early warning management pollution