Integrating Unsupervised and Supervised Machine Learning for Meteorology-Based Ozone Estimation in an Industrial Zone: A Case Study

Fine-scale ozone pollution estimation serves as a crucial prerequisite for promoting sustainable air quality management and facilitating the green industrial transition, especially in petrochemical parks where high concentrations of precursor pollutants are emitted. In this context, focusing on a typical petrochemical industrial park, this study adopts and compares both unsupervised and supervised machine learning methods to evaluate their O3 concentration estimation performance. Six unsupervised algorithms were evaluated using the Silhouette Score (SS) and Calinski–Harabasz Index (CH). The proposed hybrid SK model (SOM + k-means) achieved the best clustering performance (SS = 0.6615, CH = 13,946.60), identifying nine meteorological patterns, two of which were strongly associated with elevated ozone. Among the supervised models, after hyperparameter tuning with 5-fold cross-validation, XGBoost outperformed support vector regression, random forest, and Convolutional Neural Networks on the held-out test set, achieving R2 = 0.87, RMSE = 13.67 μg/m3, and MAE = 10.02 μg/m3. SHAP analysis indicated that northeasterly winds, moderate wind speed (≤4 m/s), low humidity, and temperatures between 10 and 20 °C were positively associated with high ozone levels, whereas southwesterly winds and high humidity showed negative associations. These findings highlight the complex interactions between meteorological factors and ozone dynamics. The proposed meteorology-driven estimation framework provides candidate indicators for ozone pollution risk screening within petrochemical industrial zones, offering preliminary meteorological references for subsequent pollution mechanism research and follow-up validated operational modeling, and contributing to regional environmental sustainability.