[Prediction of Autumn Ozone Concentration in the Pearl River Delta Based on Machine Learning]

Based on the observation data of the daily maximum 8-hour ozone (O₃) average concentration[MDA8-O₃, ρ(O₃-8h)] and meteorological reanalysis data in the Pearl River Delta Region from 2015 to 2022, four machine learning methods, i.e., support vector machine regression (SVR), random forest (RF), multi-layer perceptron (MLP), and lightweight gradient boosting machine (LG) were used to establish MDA8-O₃ prediction models. The results showed that the SVR model had the best prediction performance on MDA8-O₃ during the whole year, the coefficient of determination (R²) reached 0.86, and the root mean square error (RMSE) and mean absolute error (MAE) were 16.3 μg·m^-3 and 12.3 μg·m^-3, respectively. The prediction performance of the SVR model in autumn was still slightly better than that of LG and MLP, with R²,RMSE,and MAE values of 0.88, 19.8 μg·m^-3,and 16.1 μg·m^-3, respectively. The RF model performed the worst in the autumn prediction. In addition, the models trained by data from the whole year had better prediction ability on autumn MDA8-O₃ than that of those only trained by autumn data, and the R² differed 0.08-0.14.