Factors influencing surface ozone concentrations in Wuhan: an interpretable machine learning analysis.

Abstract

Ground-level ozone (O₃) has become a major challenge for air quality improvement in China, but its short-term variability is controlled by complex interactions among pollutant persistence, precursor conditions, and meteorological factors. In this study, hourly air quality observations and ERA5 reanalysis meteorological data for an urban monitoring site in Wuhan during 2019-2025 were integrated to evaluate O₃ temporal variability and interpretable machine learning prediction. Three models, namely, deep neural network (DNN), random forest (RF), and extreme gradient boosting (XGBoost), were compared using baseline and lag-enhanced feature sets, and SHapley Additive exPlanations (SHAP) were used to interpret predictive contributions. O₃ remained at relatively high levels during 2019-2025, with exceedances of the national grade II standard in 2019, 2023, and 2024, while most primary pollutants generally declined. O₃ exhibited a unimodal diurnal cycle with afternoon peaks at 14:00-16:00 and a seasonal pattern of summer > spring > autumn > winter. The enhanced XGBoost model achieved the best short-term prediction performance, with R² = 0.9756 and RMSE = 6.35 μg m^-3 on the test set. SHAP analysis showed that the 1-h lagged O₃ concentration was the dominant predictor, indicating strong short-term persistence rather than independent chemical forcing. NO₂ and lagged NO₂ showed negative predictive contributions, consistent with the expected role of NO titration under high-NOₓ conditions, whereas temperature and radiation-related variables showed positive contributions. These results suggest that interpretable machine learning can help identify the statistical conditions associated with O₃ variability, but the absence of direct VOC observations means that O₃ sensitivity and VOC-related control implications should be further verified using VOC measurements or chemical transport/observation-based modeling.