Enhancing the Simulation Performance of PM2.5 Compositions in the WRF-CMAQ Modeling System Using Machine Learning Techniques

초록

PM2.5 compositions are important indicators for identifying emission sources and formation pathways of particulate matters in the atmosphere. In Korea, the Ministry of Environment has operated Air Quality Research Centers to monitor PM2.5 components continuously. However, relying solely on measurement data has limitations on obtaining temporal and spatial information. A 3-D chemistry-transport model enables us to simulate PM2.5 component concentrations at high spatiotemporal resolutions realistically when the simulated results are accurate. Therefore, this study aims to improve the simulation performance of one of the 3-D chemistry-transport models, Weather Research and Forecasting (WRF)-Community Multiscale Air Quality (CMAQ) model, for PM2.5 and its components using machine learning techniques. The WRF-CMAQ simulation results, including PM2.5 components, meteorology, geography, and emissions, were used as input data in the machine learning models. Measurement data of PM2.5 and its components from Air Quality Research Centers at 10 locations were used as target variables to build the machine learning models. The study period was from January 1st to March 31st, 2022. The best machine learning model showed a correlation coefficient above 0.83 which is quite reasonable to use for PM2.5 and its component simulations. We analyzed the WRF-CMAQ simulation results for PM2.5 episodes occurred nationwide. The machine learning-corrected WRF-CMAQ model results captured nationwide high PM2.5 levels better than the uncorrected WRF-CMAQ model results. It is expected that PM2.5 characteristics in other regions where Air Quality Research Centers do not exist can be accurately provided using the machine learning-corrected WRF-CMAQ model.

키워드