Analysis of wintertime O variability using a random forest model and high-frequency observations in Zhangjiakou-an area with background pollution level of the North China Plain.

The short-term health effects of ozone (O) have highlighted the need for high-temporal-resolution O observations to accurately assess human exposure to O. Here, we performed 20-s resolution observations of O precursors and meteorological factors to train a random forest model capable of accurately predicting O concentrations. Our model performed well with an average validated R of 0.997. Unlike in typical linear model frameworks, variable dependencies are not clearly modelled by random forest model. Thus, we conducted additional studies to provide insight into the photochemical and atmospheric dynamic processes driving variations in O concentrations. At nitrogen oxides (NO) concentrations of 10-20 ppb, all the other O precursors were in states that increased the production of O. Over a short timescale, nitrogen dioxide (NO) can almost track each high-frequency variation in O. Meteorological factors play a more important role than O precursors do in predicting O concentrations at a high temporal resolution; however, individual meteorological factors are not sufficient to track every high-frequency change in O. Nevertheless, the sharp variations in O related to flow dynamics are often accompanied by steep temperature changes. Our results suggest that high-temporal-resolution observations, both ground-based and vertical profiles, are necessary for the accurate assessment of human exposure to O and the success and accountability of the emission control strategies for improving air quality.