(Review) A Study on Deep Learning-Based PM2.5 Forecasting Using Meteorological and Air Quality Data

In this study, we employ artificial intelligence to forecast PM2.5 concentrations, a critical component of air quality. Air pollution caused by fine particulate matter, such as PM10 and PM2.5, poses serious threats to both human health and the environment. Accordingly, the development of accurate and timely monitoring and prediction systems is becoming increasingly important for public health and smart city governance. In this research, we collected public air quality and meteorological data provided by the Ministry of Environment and the Korea Meteorological Administration, focusing on the Seoul metropolitan area. After performing preprocessing steps including time-series alignment, missing-value imputation, and scaling, predictions were conducted using the Canvas platform within AWS SageMaker. The model achieved a root mean squared error (RMSE) of 9.82 µg/m³, a mean absolute error (MAE) of 6.63 µg/m³, and a coefficient of determination (R²) of 0.71 for daily PM2.5 forecasting. In addition, classification of high PM2.5 events exceeding 35 µg/m³ based on the predicted 90th percentile (p90) yielded a precision of 92.3%, accuracy of 91.9%, recall of 66.7%, and an F1 score of 77.4%. The results confirm that forecasting based on the p90 quantile is effective in identifying high-risk air pollution events with high precision. Furthermore, the fan chart visualization between p10 and p90 provides a meaningful representation of forecast uncertainty. This probabilistic approach not only estimates central tendencies but also serves as a valuable tool for issuing early warnings of potential pollution spikes that may impact health policies and public safety.