시계열 결측 자료를 고려한 실내 초미세먼지 예측을 위한 머신러닝 모델 비교

초록

Accurate real-time prediction of fine particulate matter (PM 2.5) in enclosed public transport spaces like subway stations is essential for air quality control and public health. This study developed a machine learning-based model designed to maintain stable predictions even with missing time-series indoor air quality data. Three input datasets were prepared using different methods of incorporating outdoor air quality: data from a single site, averages from multiple sites, and individual values from several sites. Five individual machine learning models and three ensemble models, which do not rely on time-series data, were tested for prediction accuracy over 1-4 hour lead times. The XGBoost-Cubist ensemble model performed best (Kling and Gupta Efficiency = 0.838), showing strong and stable accuracy even at longer lead times. Among the datasets, the one using averaged data from multiple outdoor monitoring sites yielded the most reliable predictions with the least accuracy loss over time. The study highlights that using spatially aggregated outdoor air data enhances the robustness of indoor PM₂.？ forecasts. It also shows the practical value of non-time-series models in dealing with incomplete real-world data, offering insights for future air quality monitoring and alert systems in public transport environments.

키워드