A Development of PM2.5 Forecasting System in South Korea Using Chemical Transport Modeling and Machine Learning

Abstract

Ambient exposure to PM2.5 can adversely affect public health, and forecasting PM2.5 is essential for implementing protection measures in advance. Current PM2.5 forecasting systems are primarily based on the chemical transport model of Community Multiscale Air Quality (CMAQ) modeling systems and the Weather Research and Forecasting (WRF) model. However, the forecasting accuracies of these models are substantially constrained by uncertainties in the input data of anthropogenic emissions and meteorological fields, as well as inherent limitations in the models. The PM2.5 forecasting system developed in this study aimed at overcoming the limitations of CMAQ predictions by utilizing advanced machine learning algorithms. The proposed system was developed using forecast data from CMAQ and WRF, as well as observed PM2.5 concentrations and meteorological variables at monitoring stations in China and South Korea. It was then applied to national PM2.5 forecasting in South Korea. This study focused on developing secondary input data and machine learning models that can reflect the long-range transport in Northeast Asia. The proposed system can forecast 6-h average PM2.5 concentrations up to two days in advance at 19 forecast regions in South Korea. To evaluate the performance of the proposed models, a real-time machine learning-based forecasting system was applied to 19 forecasting regions from January 2020 to April 2021. Herein, the four machine learning algorithms applied, including deep neural network, recurrent neural network, convolutional neural network, and Ensemble, could reduce the over-prediction of the CMAQ forecast by decreasing the normal mean bias and improving the index of agreement. The reduced false alarm rates and high prediction accuracy confirm the feasibility of these models for practical applications.