Machine Learning-Based PM2.5 Prediction Using Meteorological Data for Environmental Decision Support

Abstract

Fine particulate matter (PM2.5) remains a critical environmental and public health concern. Accurate prediction of PM2.5 concentrations is essential for air quality monitoring, pollution mitigation, and evidence-based decision-making. Although machine learning has demonstrated promising performance in air quality prediction, many existing studies emphasize predictive accuracy while providing limited comparative evaluation and insufficient interpretation of the variables contributing to model predictions.

This study develops a comparative machine learning framework for predicting daily PM2.5 concentrations using meteorological variables and historical PM2.5 observations. Four models—Linear Regression, Random Forest, Extreme Gradient Boosting (XGBoost), and Stacking Ensemble—were evaluated using MAE, RMSE, MAPE, and R². Explainable Artificial Intelligence was incorporated through SHapley Additive exPlanations (SHAP) to interpret the contributions of meteorological, temporal, and historical PM2.5 features.

The results show that XGBoost achieved the best overall predictive performance, with the lowest RMSE of 6.9639 and the highest R² of 0.7917, while the Stacking Ensemble obtained the lowest MAE and MAPE. SHAP analysis indicated that historical PM2.5 features were the strongest contributors to prediction, while relative humidity was the most influential meteorological variable, followed by temperature and wind speed. These findings demonstrate the potential of combining comparative machine learning and explainable AI to support interpretable PM2.5 forecasting and environmental decision-making.