Short-term prediction of particulate matter (PM10 and PM2.5) in Seoul, South Korea using tree-based machine learning algorithms

Abstract

In this study, highly accurate particulate matter (PM10 and PM2.5) predictions were obtained using meteorological prediction data from the local data assimilation and prediction system (LDAPS) and tree-based machine learning (ML). The study area was Seoul, South Korea, and data from July 2018 to June 2021 as well as LDAPS 36-h predictions with 1-h intervals 4 times a day were used. The predicted PM values were then compared with the observed PM measurements to evaluate the prediction accuracy. The PM prediction performance of the Community Multi-Scale Air Quality (CMAQ)-based chemical transport model (CTM) was compared with that reported by this study. The experimental results report that, among tree-based ML algorithms, light gradient boosting (LGB) is the most suitable for PM prediction. The PM prediction results of the LGB algorithm for the hourly test data were: bias = −0.10 μg/m3, root mean square error (RMSE) = 13.15 μg/m3, and R2 = 0.86 for PM10 and bias = −0.02 μg/m3, RMSE = 7.48 μg/m3, and R2 = 0.83 for PM2.5, and for daily mean were: RMSE ≤1.16 μg/m3 and R2 = 0.996. The relative RMSE (%RMSE) is 21% lower than the results of the CTM model, and R2 is 0.20 higher. Even in the high PM concentration case prediction results, the algorithm showed good predictive performance with %RMSE = 8.91%–20.43% and R2 = 0.89–0.97. Therefore, in addition to the CTM, high-accuracy PM prediction results using ML can also be used for air quality monitoring and improvement.