Comparison of Machine Learning Methods for Predicting Ground-Level Ozone Pollution in Beijing

Weidong Zhu Zitao Liu ^* Jiansheng Yuan Zhaoxiang Cao Tiantian Cao Shuai Liu Yuelin Xu Xiaoshan Zhang

• College of Marine Sciences, Shanghai Ocean University, Shanghai, China

High ground-level ozone (O₃) concentrations significantly degrade urban air quality and pose threats to human health, highlighting the critical need for accurate and effective prediction of ozone levels to support environmental monitoring and policy formulation. Considering the lagged chemical interactions between ozone (O₃) and nitrogen oxides, this study integrates the historical concentrations of ozone and nitrogen dioxide (NO₂) over the past three hours as lagged feature variables. A Lagged Feature Prediction Model (LFPM) is proposed and evaluated using nine machine learning algorithms, including XGBoost. First, by combining XGBoost with SHAP (Shapley Additive Explanations) for feature selection analysis, 11 critical feature variables are identified. This approach improves computational efficiency by 30% compared to the original feature set while preserving model prediction accuracy. Subsequently, ozone concentration predictions are conducted using six meteorological variables. Results indicate that LSTM-based methods, particularly ED-LSTM, achieve the best performance among meteorological-only models (R² = 0.479). However, predictions relying solely on meteorological variables exhibit limited accuracy. The inclusion of five pollutant variables significantly enhances the predictive performance across all machine learning methods. XGBoost demonstrates the highest accuracy (R² = 0.767, RMSE = 11.35μg/m³), representing a 125% relative improvement in R² compared to predictions using meteorological variables alone. Further implementation of the LFPM model boosts prediction accuracy for all nine machine learning methods, with XGBoost maintaining the optimal performance (R² = 0.873, RMSE = 8.17μg/m³). These findings conclusively demonstrate that integrating lagged feature variables substantially improves ozone prediction accuracy.