Improvement of the ozone forecast over Beijing through combining the chemical transport model with multiple machine learning methods

Abstract

The accuracy of ozone forecasts is influenced by various uncertainty factors, including the meteorological fields, physicochemical parameterization schemes, and emission inventories, which makes it difficult for the chemical transport model (CTM) to comprehensively constrain these factors and leads to inaccurate ozone forecasts. In this study, the CTM is combined with multiple machine-learning approaches including Random Forest (RF), LightGBM and XGBoost to improve the 10-day forecasts of hourly ozone concentrations in Beijing from June to August 2022. The forecasted data from Weather Research and Forecasting (WRF) model and Nested Air Quality Prediction Modeling System (NAQPMS) are adopted as the inputs for training the machine-learning models with sliding intervals. The results demonstrate that all three machine-learning models can significantly enhance the accuracy of 10-day ozone forecasts. Specifically, the reduction in root-mean-square error of ozone forecasts ranges from 35.5% to 39.9% in 24–72 h, 34.3%–36.0% in 72–168 h, and 29.9%–34.4% in 168–240 h, with the error increasing as the forecast time extends. Additionally, the influence brought by the selection of inputs on the accuracy of ozone forecasts is greater than the choice of three machine-learning models, and the weather-forecast data plays a more important role in forecasting improvement than the pollutant-forecast data. Moreover, the machine-learning models effectively improve the occurrence timing of peak and trough and the amplitude of the diurnal variation of ozone, which better matches up to the observation. In conclusion, this study proposes a new approach for improving the 10-day ozone forecasts by combining CTM with machine-learning models.