Improving PM2.5 and Visibility Predictions in China Using Machine Learning and Ensemble Forecasting

Abstract Accurate PM2.5 and visibility predictions are essential for effective air quality management in central and eastern China. This study seeks to enhance the regional air quality numerical prediction system (RMAPS‐Chem) by integrating machine learning techniques and a dynamic ensemble algorithm to improve the accuracy of hourly PM2.5 and visibility forecasts across 309 cities in mainland China. Results demonstrate that machine learning methods significantly improve the accuracy of PM2.5 and visibility forecasts. For a 24‐hr lead time, the root mean square error (RMSE) is reduced by an average of 19% for PM2.5 and 30.2% for visibility, while the temporal correlation coefficient (TCC) increases by an average of 4.2% and 22.1%, respectively. Furthermore, we introduce a new ensemble forecasting algorithm, RCHEM‐AI, which capitalizes on dynamic RMSE and TCC weights based on machine learning predictions. RCHEM‐AI outperforms both RMAPS‐Chem and all individual machine learning members, reducing the average RMSE of PM2.5 (visibility) by 33.3% (39.3%) and increasing the average TCC by 16.4% (39.6%). Although the performance of machine learning and ensemble predictions declines with longer forecast time, the improvement in forecast accuracy remains substantial throughout the 1–10 days forecast period. By integrating RMAPS‐Chem, machine learning, and ensemble algorithms, this research provides a powerful tool for earlier and more accurate haze pollution predictions, thereby supporting more effective air pollution control and informed scientific decision‐making.