1164549 - Prediction of PM_2.5 Component Concentrations from Intermittent Measurements Using a Deep Neural Network Model

The concentrations data of fine particulate matter (PM2.5)-bound chemical components are useful information for exploring atmospheric chemical processes and apportioning the source of PM2.5. However, continuous monitoring or quantification is challenging due to time and cost constraints. In this study, the concentrations of PM2.5 chemical components at the unmeasured time were predicted using intermittently measured data using a deep learning model. Through the deep learning model with hyperparameter optimization, it was assumed that air pollution characteristics could be learned, and concentrations could be predicted with only some data. The data was reconstructed to intermittent measurement data from the national intensive measurement data of the Republic of Korea, which was operated continuously, and the prediction accuracy was analyzed. By selecting appropriate input data, it was possible to predict PM2.5 component concentrations for the unmeasured period through 1- or 2-time measurements a week. The accuracy of the determination coefficient was approximately 0.7 or greater depending on the component. In particular, the prediction performance for ion components was higher than those of trace elementals. The result showed that the characteristics of ion species could be effectively learned through the deep learning model using meteorological data and other air pollution information. This was due to a relatively high concentration of ion components and a strong correlation with gaseous air pollutants. Although the prediction accuracy of trace elementals was relatively lower than those of ion components, some trace elementals were sufficiently predictable with a determination coefficient of 0.6 or greater. Therefore, the deep neural network model applied in this study was successful in predicting the concentrations of PM2.5 chemical components by reflecting air pollution characteristics of the region of interest, and ultimately improved the completeness and utilization of the existing datasets with missing values.