Development of a data-driven three-dimensional PM2.5 forecast model based on machine learning algorithms

Fine particle matter (PM2.5) pollution is a global environmental problem and has significant impacts on air quality and human health. Accurate prediction is crucial for mitigating PM2.5 pollution and reducing its environmental and health impacts. However, the current data-driven PM2.5 prediction model does not fully consider the vertical distribution pattern and the contribution of source emissions to achieve a broader and more accurate prediction of PM2.5. This study introduces a novel approach to predict three-dimensional (3D) air quality at a high spatialtemporal resolution, with multi-source data and machine learning algorithms. Specifically, we developed a two-stage 3D PM2.5 prediction model by standardizing and integrating meteorology data, anthropogenic emission inventory data, air quality monitoring data, and satellite remote sensing data into a 3D dataset. In the first stage, we used random forest (RF) models to estimate the spatial-temporal distributions of aerosol optical depth (AOD) and ozone (O3) density. In the second stage, we further used these estimations to predict hourly PM2.5 concentrations at both the surface and altitude levels with another RF model. To enhance the prediction performance, dynamic corrections were implemented to the predicted PM2.5 concentrations. Using this model, we predicted PM2.5 concentrations for the next 72 hours and validated the spatial-temporal fluctuations against monitoring data across Shandong Province, China. Furthermore, we assessed the contribution of local emissions and evaluated the air quality improvement resulting from local emission reduction measures. Our findings confirm the capability of the data-driven machine learning model for 3D air quality prediction on a regional scale, emphasizing the importance of regional emission control to improve local air quality.