Prediction of Fine Dust Concentrations in Major South Korean Cities Using a CNN–GRU Deep Learning Model

This study evaluated the applicability of a CNN–GRU hybrid deep learning model for predicting fine particulate matter (PM2.5, PM₁₀) concentrations in eight major metropolitan cities in South Korea. Data on fine particulate matter and air pollutants (SO₂, CO, O₃, NO₂) were collected hourly from 2020 to October 2024 at measurement sites located in areas with the highest population size and density. Basic statistical analysis and correlation analysis using the Pearson correlation coefficient were performed on the collected data. The prediction model was designed using a one-layer CNN for local feature extraction and a two-layer GRU for time-series pattern learning. Model performance was evaluated using root mean square error (RMSE) and symmetric mean absolute percentage error (SMAPE). Results showed that average PM2.5 and PM₁₀ concentrations were highest in Incheon, while PM2.5 was lowest in Busan, and PM₁₀ was lowest in Gwangju. The correlations among air pollutants were prominent in Seoul and Incheon, and the correlation between PM2.5, PM₁₀ and O₃ showed regional differences. The prediction performance of the CNN–GRU model was excellent overall, and particularly, Gwangju showed the best results with RMSE 2.36 in PM2.5 prediction and RMSE 3.44 in PM₁₀ prediction. In terms of SMAPE, Incheon showed the best performance (21.49%) in PM2.5 prediction and Seoul showed the best performance (12.98%) in PM₁₀ prediction. These results showed that the CNN–GRU-based fine particulate matter prediction model has high potential to be used in building a real-time air quality forecasting system. In the future, applicability and model optimization studies for other cities and regions outside the study area are needed.