Forecasting ground-level ozone in Shenyang using interpretable machine learning: Interaction between air pollutants and climate factors

Abstract

Ground-level ozone (GLO), produced through reactions between

and VOCs under sunlight, is a toxic secondary pollutant that adversely impacts ecosystems and human well-being. This study proposes an interpretable, high-resolution forecasting framework based on a weighted ensemble of machine learning algorithms, including LightGBM, XGBoost, and CatBoost. Using hourly observational data from Shenyang, China (2020–2025), the model incorporates meteorological, temporal, and air quality features to capture annual-scale ozone variability. The ensemble outperformed individual learners, achieving strong predictive accuracy ($R^2=0.868$, MAE = $10.3\mu \text{g}/{\text{m}}^3$, RMSE = $14.7\mu \text{g}/{\text{m}}^3$). To enhance interpretability, SHAP analysis was used to reveal nonlinear interactions among meteorological and chemical drivers. High temperatures and moderate humidity were found to promote ozone formation. Seasonal transport patterns identified by concentration-weighted trajectory (CWT) analysis revealed dominant local photochemistry in summer and stronger long-range influence in winter. The proposed framework offers both predictive accuracy and physical interpretability, supporting early warning and targeted ozone control strategies. Its transferable design enables application to other urban regions with complex atmospheric conditions.