To investigate the interactions among emissions, meteorological phenomena and chemical phenomena, the Weather Research Forecast with Chemistry model (WRF-Chem) is used in the present study. WRF-Chem is a regional online-coupled air quality model that can simultaneously simulate air quality components and meteorological components by using identical transport schemes, grid structures and physical schemes (Grell et al., 2005). The two following model domains are designed: an outer domain (horizontal resolution: 30 km) covering eastern China (20.0–44.5∘ N, 99.0–126.5∘ E) and an inner domain (horizontal resolution: 6 km) covering the YRD region (27.6–32.7∘ N, 116.9–122.4∘ E), as shown in Fig. 1. The Lambert conformal conic projection is applied, with the domain center at 34∘ N, 111∘ E. A total of 31 vertical layers are used, with the model top set at 50 hPa. The simulation period is from 17 August to 6 September 2016, and the first-week simulation is used to spin up the model. Hourly model outputs for 24 August to 6 September are used in the analysis. The gas mechanism CBMZ (Chemical Bond Mechanism Version Z) (Zaveri and Peters, 1999) is used for model simulations. For additional details regarding the model parameterization schemes, please refer to previous studies (e.g., Ni et al., 2018).

The meteorological boundary and initial conditions are determined from the global objective final analysis (FNL) data of the National Centers for Environmental Prediction (Kalnay et al., 1996). The FNL data are mapped to domain 1 (eastern China), and the grid-nudging method (Stauffer et al., 1991) is used to reduce the meteorological integral errors. The chemical initial and boundary conditions are dynamically downscaled from the simulation results of the model for ozone and related chemical tracers, version 4 (MOZART4) (Emmons et al., 2010).

To understand the underlying mechanism of O3 formation, individual physical and chemical processes of O3 formation are investigated by using the integrated process rate (IPR) analysis in the WRF-Chem model (Jeffries and Tonnesen, 1994). The IPR analysis differentiates changes in pollutant concentrations from individual atmospheric processes, which quantitatively elucidates the contributions of each process, mainly including advection, diffusion, emission, deposition, clouds process, and aerosol and gaseous chemistry. The IPR analysis has been widely applied and demonstrated to be an effective tool for investigating the relative importance of individual processes and interpreting O3 concentrations (Gonçalves et al., 2009; Tang et al., 2017; Shu et al., 2016). In the present work, we consider gas chemistry, vertical diffusion, and horizontal and vertical advection as the main atmospheric processes for O3 formation. Other processes, such as cloud process and horizontal diffusion, play minor roles and are thus not considered.

To increase the confidence in interpretations of model results, model outputs should first be evaluated based on observations. Accordingly, the model results derived from domain 2 are compared with hourly surface observational data obtained from 96 air quality monitoring sites in the YRD region (blue dots, Fig. 1b) in this study. These observational data are downloaded from http://www.pm25.in/ (last access: 6 May 2020), and O3 and its precursor NO2 are evaluated in terms of statistical measures, namely the mean fractional bias (MFB), the mean fractional error (MFE) and the correlation coefficient (R), following the recommendation of the US Environmental Protection Agency (US EPA, 2007). Additionally, the meteorological parameters are evaluated based on the observational data, including temperature at 2 m (T2), relative humidity at 2 m (RH2), 10 m wind speed (WS10) and direction (WD10), from the Meteorological Assimilation Data Ingest System (https://madis.noaa.gov/, last access: 6 May 2020). Following the study of Zhang et al. (2014), commonly used mean bias (MB), gross error (GE) and root-mean-square error (RMSE) are calculated as the statistical indicators. All used statistical indicators are summarized in Table 1.

Besides the above evaluation of single-point-based time series results, the vertical spatial distribution of modeled O3 in Hangzhou is also evaluated by comparisons with observed differential absorption lidar (DIAL) data (Su et al., 2017). In the DIAL technique, the mean gas concentration over a certain range interval is determined by analyzing the lidar backscatter signals for laser wavelengths tuned to “on” (λon) and “off” (λoff) in a molecular absorption peak of the gas under investigation (Browell et al., 1998). In the DIAL data of O3, the vertical height available is from 0.3 to 3 km due to the limitations of the signal-to-noise ratio and detection range.

We first evaluate the overall performance of WRF-Chem for the YRD region by incorporating data from the 96 air quality monitoring sites. Specifically, the maximum daily 8 h (MDA8) ozone and daily mean NO2 concentrations at the surface are used. The spatial distributions of MFB and MFE for O3 and NO2 are illustrated in Fig. 2. In general, the model-simulated air pollutant concentrations agree well with the observations, with MFB and MFE at most of the sites meeting the benchmarks (MFB<15 %; MFE<35 %) (US EPA, 2007). A scatter plot of MFB and MFE is shown in Fig. 3, further demonstrating the capability of the present model to reproduce the observations, which is also supported by the high correlation between the model and observations (Fig. 3c, d).

After the above overall evaluation of the present model in the whole YRD region, the site of Hangzhou will be focused on for further analysis. The time series of hourly simulated and observed air pollutants (O3, Fig. 4a; NO2, Fig. 4b) and meteorological factors (T2, Fig. 4c; RH2, Fig. 4d; WS10, Fig. 4e; and WD10, Fig. 4f) at Hangzhou are presented in Fig. 4. It is found that all modeled data are statistically significantly correlated with the observed data at the 95 % level. The MFB and MFE for both O3 and NO2 are well below the benchmarks (MFB and MFE: 15 %/35 %; US EPA, 2007), and the observed diurnal variations are well reproduced. For meteorological parameters, ≤±0.5 ∘C, GE≤2.0 ∘C), 10 m wind speed (MB≤0.5 m s-1, RMSE≤2.0 m s-1) and 10 m wind direction (MB≤±10∘, GE≤30∘). McNally (2009) suggested a relaxed benchmark for 2 m temperature (MB≤±1.0 ∘C). In this study, the 10 m wind speed and wind direction (Fig. 3e, f) results are well within the benchmarks. The GE of 2 m air temperature (1.9 ∘C; Fig. 3c) also satisfies the criteria, but the MB is slightly higher (-1.6 ∘C), which has also been noted in a previous study (Zhang et al., 2014). These comparisons further demonstrate that the present model is able to correctly predict the time series of both meteorological parameters and air pollutants of O3 and NO2 in Hangzhou.

To further evaluate the capability of the model to predict the vertical structure of ozone concentration, the vertical distribution of the modeled O3 in Hangzhou from 24 August to 6 September 2016 is qualitatively compared with the DIAL data, as shown in Fig. 5. It is interesting to find that the present model can successfully predict the spatial–temporal distribution of ozone in Hangzhou. All observed major features of ozone are well captured by the model. This gives us high confidence and lays a solid foundation for further exploring the pollution characteristics and influencing factors of ozone in Hangzhou during the G20 summit.

To discuss spatial–temporal characteristics of O3 pollution in Hangzhou, the whole emission control period can be divided into three stages according to the reduction intensity of the measures. The period from 24 to 27 August 2016 was the first stage (S1), during which industrial and construction emission controls were implemented. During the second stage (S2, 28–31 August), traffic restrictions were further added. The third stage (S3) from 1 to 6 September 2016 was when the emergent VOCs control was further implemented. Figure 4a and b in the above section also present the temporal evolution of O3 and its precursor NO2 in Hangzhou during the emission control period of G20 summit. It is evident that NO2 has been significantly reduced by the emission control measures and that the concentration is well below the national Level II standard of 200 µgm-3. However, the concentration of O3 remains at high levels for the whole 14 d, with 7 d of MDA8 above and 4 d close to the national Level II standard (GB-3095–2012) of 160 µgm-3. This serious O3 pollution indicates that the emission control measures seem to have no obvious effect on ozone, which is consistent with previous observations (Su et al., 2017; Wu et al., 2019). The diurnal variation in O3 is similar for the three stages, with a peak value at around 16:00 LST (local sidereal time) and a valley value at the time around 08:00 LST each day. However, the variation magnitude in Stage 2 is obviously lower than those of other stages, which will be further discussed later.

Figure 5 also clearly shows this diurnal variation in O3 at ground level. However, nocturnal O3-rich mass is observed during certain periods in the upper air (approximately 1 km), such as 25 and 31 August and 3 September, which makes an n-shaped distribution pattern of O3 in the upper air. This kind of spatial distribution of ozone will promote vertical exchange of O3 in the area. In general, high concentrations of O3 appear vertically up to the top of the planetary boundary layer (PBL, approximately <2 km), suggesting the ozone pollution is not a local phenomenon but is instead a regional phenomenon in the whole low-level (from surface to close to the PBL height) region.

Considering that the synoptic circulation is closely related to regional O3 abundance, four representative surface weather charts obtained from the Korea Meteorological Administration are presented in Fig. 6. In the early stage, strong and uniform high-pressure fields covered vast regions of southeastern China, and a tropical cyclone moved northeast over the East China Sea (Fig. 6a). In the middle stage (Fig. 6b), the tropical cyclone approached the YRD region, bringing strong north wind fields to this area. As a result, the long narrow rain band arrived in Hangzhou (red triangle) on 27 August 2016. In the later stage (Fig. 6c), the cyclone continuously moved and eventually hit the land, and the tropical high in the YRD region recovered gradually. Finally, the cyclone faded, and a rainstorm appeared over most of the YRD region (Fig. 6d).

The typical hourly vertical and horizontal O3 distributions in the YRD region are further presented in Fig. 7. The wind fields are also included for better understanding. For stagnation days with weak wind fields and strong radiation before or after the tropical cyclone, meteorological conditions are unfavorable for pollutant dispersion. As a result, O3 pollution is more regional and intense, with an hourly peak O3 concentration of 250 µgm-3 that appeared within the planetary boundary layer in the whole YRD region, as shown in Fig. 7a and c. In these conditions, photochemical reactions dominate the ozone formation and accumulation. This phenomenon is consistent with the satellite-derived tropospheric O3 distribution in the area (Su et al., 2017) and is also supported by the observed ozone data from the 96 sites in the YRD region, as shown in Fig. 3c. During the 14 d emission control period of G20 summit, 52 % of the observed ozone samples from the 96 sites are above the China's national Level II standard (160 µgm-3), suggesting that regional ozone pollution appears in the YRD region during the study period. As the cyclone approached on 27 August, a large belt of O3 mass appeared in the upwind direction and moved toward Hangzhou under a prevailing north wind field (Fig. 7b). Regional pollutant transport may play an important role under this condition. However, because of the rain and cooling effects from the cyclone, the ozone concentration is relatively low in the whole YRD region.

To further investigate the underlying mechanism of O3 pollution, hourly variations in the change rate of low-level O3 resulting from different physical and chemical processes are presented in Fig. 8. It is evident that gas chemistry is the dominant factor for the strong generation of abundant O3 in the entire planetary boundary layer (<2 km) in the daytime but causes a small amount of depletion of O3 at near-surface height (<0.3 km) at nighttime (Fig. 8a). High concentrations of O3 diffuse from the upper layer downward to the ground through vertical diffusion during the whole study period, which is obvious in the daytime (Fig. 8b). However, this effect is relatively weak compared to other processes. Horizontal and vertical advection seem to play more important roles in shaping the near-surface O3, as indicated in Fig. 8c and d. Several interesting dynamic O3 circulations are observed between the near-surface and upper-air regions and indicated by the dashed boxes. During the periods of 24 to 26 August and 31 August to 2 September 2016, the O3-rich mass in the lower layer (<1 km) traveled to Hangzhou through horizontal advection and was then transported upward to the higher layer through vertical advection. At this higher layer, the mass subsequently travels away from Hangzhou to other places through horizontal advection in a circular manner. This phenomenon might be associated with the urban heat island circulation (Lai and Cheng, 2009). However, during the period of 3–6 September 2016, a similar circulation phenomenon is observed, but the flow direction is reversed. The O3-rich mass travels downward to the ground through vertical advection and is then transported to surrounding regions through horizontal advection. This downward circulation is also related to the meteorological conditions after the cyclone. In addition, the horizontal and vertical advection of O3 took on a chaotic status during 27–30 August 2016, suggesting that complicated variable meteorological conditions happened at the time. This is also the reason for the lower magnitude of diurnal variation in Stage 2.

Figure 9 shows the daytime mean change rate of simulated O3 at ground level resulted from various atmospheric processes and the correlation of gas chemistry generation and observed maximum for daily 8 h concentration of O3. As a whole, the main sources of local surface ozone in Hangzhou are from gas chemistry, vertical diffusion and horizontal advection, with mean production rates of 1.9, 3.3 and 6.7 ppb h-1, respectively, from 24 August to 6 September 2016, and the major sink is vertical advection. However, during some days, such as 5–6 September, gas chemistry consumes O3 while vertical advection increases it. In general, strong net horizontal and vertical advection of O3 are observed for most days of the period, except for 27–28 August, during which the strongest cold northwesterly winds (Fig. 4e) occurred and made the net advection of O3 negligible. Similar to Fig. 8, dynamic O3 circulations are observed for the periods of 24–26 August, 31 August to 2 September and 5–6 September. Specifically, the circular direction is reversed during 5–6 September, and the net gas chemistry is to consume ozone due to weak solar radiation during the day, as shown in Fig. 10.

In addition, the variation trend of the daytime mean production rate of gas chemistry is consistent with the observed MDA8 concentration, and the local chemical generation has large positive correlation (Pearson's r=0.77) with the observed MDA8 concentrations (Fig. 8b). This indicates a trade-off effect among vertical diffusion, horizontal advection and vertical advection. High O3 concentrations (i.e., MDA8 on 25 August 2016: 98 ppb) are always accompanied by strong radiation and prolific generation of gas chemical reactions. It is also interesting to find that vertical diffusion may partially compensate for gas chemistry when the chemical reaction rate is relatively low or negative. For example, during 26–27 August and 5–6 September, the vertical diffusion rates are higher than the chemical production rates. The low O3 episode during these periods may result from local chemical consumption.