Statistical Forecasting of Ozone in Beijing: Evaluating Multiple Time Series Models and the Impact of Meteorological Factors

Ozone pollution has emerged as a critical air quality concern in China, especially in megacity area such as Beijing in recent years. Characterized by its complex, nonlinear interactions among precursor pollutants and significant spatiotemporal variations, ozone poses challenges for numerical models in terms of forecasting accuracy. In contrast, statistical forecasting models offer several advantages, including reduced data requirements, lower computational costs, and enhanced predictive accuracy, making them a viable option for practical ozone forecasting applications.  In this study, we evaluate multiple time series models (such as ARIMA, NNAR, STLF, ETS etc.) for ozone concentration forecasts in Beijing. During the ozone pollution season, ARIMA and NNAR achieved correlation coefficients of approximately 0.65 between predicted and observed values. The ensemble model outperformed these, with a correlation coefficient of around 0.7 and an RMSE of about 45 µg m⁻³. For clear-day pollution events, after accounting for rainfall influence, the ensemble model's correlation coefficient reached approximately 0.9, with an RMSE reduced to about 40 µg m⁻³. The results demonstrate that time series models are effective for both mid-term and short-term ozone forecasting, while the ensemble model based on multiple time series approaches further enhances performance, offering high accuracy, temporal resolution, and spatial universality, particularly during severe pollution episodes. Daily maximum temperature, radiation precipitation are key meteorological factors that significantly influence ozone concentration. Incorporating maximum temperature into a dynamic ARIMA model significantly improved ozone forecasts, raising the correlation coefficient to about 0.75 and reducing RMSE. Future improvements could integrate more meteorological covariates to improve the performance of ozone forecasting models.