Determination of Input variables for Artificial Intelligence Models to predict the High PM2.5 concentration events in Seoul, Korea

The concentration of particulate matter (PMs) is governed by complex processes such as long-range transport, vertical diffusion, and local emissions. Therefore, thus it is relatively difficult to accurately forecast high PM concentration events. As the application of artificial intelligence (AI) techniques to air-quality prediction has increased, optimal input variables for AI models have become critical. The purpose of this study was to suggest combined and synoptic variables, in addition to conventional surface meteorological and air quality variables, for AI-based high PM event prediction models. In Seoul and four cities in China, the observed surface meteorological and air quality data, upper air meteorological data, planetary boundary layer height, and temperature gradients between the surface and 850 hPa were tested. The east–west geopotential index (EWGI) and Korean Region Blocking Index (KRBI) have been suggested as regional-scale blocking indices. A concentration-wind (CW) variable was introduced to represent the effects of long-range transport from China. The usefulness of the suggested variables was tested using random forest (RF) and support vector machine (SVM) for 2017–2020. As the forecasting days progressed, the importance of surface variables decreased, whereas those of the EWGI, KRBI, CW, and stability variables increased. The stability variables increased the accuracy, probability of detection, and F1 scores, while decreasing the false alarm rate on the 3‒5 forecasting days. EWGI and KRBI improved the prediction performance after the third forecast day, and CW was important for predicting the 3‒4 forecast days. Newly introduced variables, such as EWGI, KRBI, CW, and stability tended to increase the 1‒4 day forecast hit rate for high PM_2.5 events and were found to be useful input data for machine learning or artificial intelligence-based air quality prediction models.