PM_2.5 concentration 7-day prediction in the Beijing-Tianjin-Hebei region using a novel stacking framework

High-precision prediction of near-surface PM_2.5 concentration is a significant theoretical prerequisite for effective monitoring and prevention of air pollution, and also provides guiding suggestions for the prevention and control of PM_2.5-related health risks. It has been acknowledged that existing PM_2.5 prediction models predominantly rely on variables influenced by near-surface factors. This inherent limitation could hinder the comprehensive exploration of the continuous spatio-temporal characteristics associated with PM_2.5. In this study, an optimal 7-day prediction model for PM_2.5 concentration based on the Stacking algorithm was constructed based on multi-source data mainly including atmospheric environment ground monitoring station data, MODIS remote sensing-derived aerosol optical depth (AOD) daily data and meteorological factors. The findings indicated that the PM_2.5 forecasting outcomes derived from this integrated RF-LSTM-Stacking model exhibited a superior fit, with R², RMSE, and MAE values of 0.95, 7.74 µg/m³, and 6.08 µg/m³, correspondingly. This approach enhanced the accuracy of prediction to a degree of approximately 17% in comparison with a solitary machine learning model. The findings of this study demonstrated that the integration of the LSTM-RF model with the fusion-based Stacking algorithm led to a substantial enhancement in the accuracy of PM_2.5 predictions. This model was found to serve as an effective reference for the monitoring of PM_2.5 prediction and early warning systems.