Assessing Meteorological, Local, and Upwind Information for Surface Ozone Through Machine Learning: Implications for Sustainable Urban Air Quality in Hefei, China

Ozone (O3) pollution remains a significant air quality issue in Chinese cities. Identifying which monitoring and modelling information provides reliable incremental value is important for efficient and sustainable urban O3 assessment. This study explored the incremental information provided by meteorological conditions, local pollution, and trajectory-linked upwind observations for the modeling of hourly O3 concentrations in Hefei, China. Hourly observations in 2022 were combined with ERA5 meteorological data and 72 h HYSPLIT backward trajectories. Four random forest models were constructed by sequentially adding temporal variables, meteorological variables, local pollutants, and upwind O3 and PM2.5 indicators. Leave-one-month-out blocked cross-validation showed that meteorological variables substantially improved performance relative to the temporal baseline, while local pollutants produced a further improvement (M2, R2 = 0.753, RMSE = 20.55 µg/m3). Adding the 50 fixed-lag upwind O3 and PM2.5 indicators did not improve annual performance (M3, R2 = 0.746, RMSE = 20.85 µg/m3). SHAP analysis showed that 2 m temperature, NO2, and relative humidity were the leading predictors. M2 also reproduced daily MDA8 O3 variability (R2 = 0.785) but identified only 4 of the 26 observed high O3 days with valid paired estimates. These results show that additional predictors do not necessarily improve model performance and that high O3 conditions remain more difficult to represent than the overall concentration variability. The findings provide a basis for more targeted selection of observational information in urban O3 modelling.