A ten-year meteorological simulation and optimization in China based on traditional data assimilation and machine learning methods

Abstract

Meteorological conditions are key inputs for chemical transport models and directly impact simulation accuracy. thus, reducing their uncertainties is crucial. To generate long-term meteorological input datasets, we utilizes the WRF model to simulate atmospheric conditions across China over a ten-year period (2014–2023) at a spatial resolution of 27 km. While WRF shows relatively good performance in simulating wind speed, its accuracy in temperature and wind direction remains limited. To further improve the simulation accuracy, the Yangtze River Delta region is selected for a 2023 case study, applying both the traditional 3DVAR data assimilation method and machine learning approaches to optimize these three key variables. The results demonstrate that for non-compliant stations, 3DVAR achieves better optimization in temperature simulation compared to RF and XGBoost, whereas RF and XGBoost outperform 3DVAR in wind field simulation. Among all the methods evaluated, XGBoost delivered the most effective optimization performance.