PM 2.5 Concentration 7-days Prediction in the Beijing-Tianjin-Hebei Region Using a Novel Stacking Framework

High-precision prediction of near-surface PM _2.5 concentration is an significant theoretical prerequisite for effective monitoring and prevention of air pollution, and also provides guiding suggestions for PM _2.5 health risk prevention and control. In view of the fact that the control variables of existing PM _2.5 prediction models are mostly dependent on the influencing factors at the near-surface, and it is often difficult to fully explore the continuous spatio-temporal characteristics in PM _2.5 . In this study, MODIS remote sensing-derived Aerosol Optical Depth (AOD) daily data, atmospheric environment ground monitoring station data and meteorological factors are introduced to identify strong correlation factors. A highly robust seven-day prediction model for PM _2.5 concentration is constructed based on the Stacking algorithm combined with various machine learning methods to improve the generalisation ability of the model; the estimation ability of the integrated model is compared and analyzed with LSTM, RF and KNN models. The results demonstrated that the PM _2.5 prediction results on the basis of this integrated RF-LSTM-Stacking model exhibited a better fit, with R², RMSE, and MAE values of 0.95, 7.74 µg/m³, and 6.08 µg/m³, respectively. This approach improved the prediction accuracy by approximately 17% compared to a single machine learning model. Based on this study, it was evident that the LSTM-RF model, integrated with the fusion-based Stacking algorithm, significantly enhanced the PM _2.5 prediction accuracy and provided an effective reference for PM _2.5 predicting and early warning monitoring.