The Characteristics of Heavy Ozone Pollution Episodes and Identification of the Primary Driving Factors Using a Generalized Additive Model (GAM) in an Industrial Megacity of Northern China

Abstract

Tropospheric ozone is the only normal pollutant with a continuously increasing annual average concentration worldwide. In this study, data were monitored at the Nankai University Air Quality Research Supersite (NKAQRS) (38.99° N, 117.33° E) between 1 April, and 31 August from 2018 to 2020, 33 O₃ episodes from 2018 to 2020 were analyzed to reveal the characteristics of O₃, VOCs and OFP during O₃ episodes and to evaluate the driving factors. The O₃ episodes showed a decreasing trend in terms of pollution frequency, days, heavy pollution duration and peak concentration. Ethane, acetylene, cyclopentane, and methylcyclopentane were the major types in 2020, while 1-hexene was the main component in 2019. The main ozone-contributing species in 2020 were propene cyclopentane methylcyclopentane and ethylene. Alkenes were important contributors to ozone formation. Using generalized additive models (GAMs), the explanatory variables in the study are divided into environmental and meteorological factors, and 16 impact factors are selected as explanatory variables. We found that the influence of these meteorological factors on O₃ pollution was nonlinear and impacted by the interaction between variables. O₃ episodes were mainly driven by meteorological and precursor (NO) factors in 2018, while meteorological conditions (T), followed by precursor (NO₂) were the driving factors in 2019 and 2020, suggesting that O₃ episodes were mainly driven by meteorological conditions.

1. Introduction

Tropospheric ozone is a normal pollutant with a continuously increasing annual average concentration in the mainland China [1]. Ozone affects the healthy growth of plants, leading to a decline in crop yields, and is also hazardous to the human respiratory system, organs, immune system and tissues, which threatens health and even human life [2,3,4,5,6]. Tropospheric ozone (O₃) is mainly produced by the photochemical reaction of its precursors (NOx, VOCs et al.) under favorable meteorological conditions [7,8,9,10]. The concentration of ambient O₃ always shows a nonlinear correlation with these precursors, suggesting a complex physicochemical mechanism in the ozone formation process [11,12,13,14].

As key precursors of O₃ pollution, volatile organic compounds (VOCs) can be transmitted over long distances [15], and they are mostly toxic [16,17] and carcinogenic [16,17]. Under UV irradiation, VOCs react with NOx to generate ozone and enhance atmospheric oxidation [18,19]. They are one of the main contributors to regional air pollution in recent years [20,21]. Meteorological factors also play an important role in ozone formation, dilution, diffusion and transportation [22], including temperature, humidity, solar radiation and wind [8,23,24,25,26]. Studies have shown that high ozone levels are usually related to strong solar radiation, high temperature, low wind speed, low or least rainfall, and low relative humidity [27,28,29].

Due to the nonlinearity between ozone pollution and its precursors, it is difficult to effectively control ozone pollution. It is necessary to evaluate the relationship between O₃ episodes and its influencing factors and to identify the major factors. The common nonlinear methods are the convergent cross mapping (CCM) method [30], the generalized additive model (GAM), and the Observation-Based Model (OBM) [31]. Compared to traditional nonlinear statistical models, the Generalized Additive Model (GAM) can directly deal with the complex nonlinear relationships between multiple response and explanatory variables, can introduce nonlinear and nonmonotonic variables into the analysis [32], and is widely used to identify impact factors and evaluate the nonlinear relationship related to air pollution [33,34], and is a better method for evaluating the nonlinear relationship between influencing factors and air pollution. Gong et al. [35] used GAM to quantify the impact of individual meteorological conditions on O₃ pollution in 16 Chinese cities, emphasizing that the model can explain the relationship between meteorological conditions and O₃ concentration. Kim and Hong [36] used a multiple linear regression model and GAM to study the response of O₃ pollution to meteorological factors (i.e., wind speed, temperature and humidity) in Seoul, South Korea. Although a great deal of research has been conducted, it is still challenging to fully understand the impact of meteorology on ozone concentration. First, previous studies have mainly considered meteorological conditions and the effect of precursors on ozone concentrations, among them, VOCs instead of alkenes, alkanes, aromatics, and acetylene are used as explanatory variables in GAM. Secondly, since different meteorological factors interact closely, and most of them are nonlinear relationships, commonly used correlation analyses may lead to biased results [37]. Therefore, this study refines VOCs into alkenes, alkanes, aromatics and acetylene, and incorporates all factors that affect ozone formation such as meteorological conditions and precursors in the GAM model to quantify the impact of various influencing factors on ozone pollution.

The regional study by Lu et al. [38] pointed out that compared with other industrialized regions in the world, China’s short-term and long-term human exposure to and plant damage by ozone were severe. Ozone pollution becoming the key pollutant in summer of northern China has attracted increasing attention. O₃ episodes often occur in the form of pollution processes that continue for several hours or even days. It is of great value to analyze the process of O₃ episodes to understand the fluctuation law of ozone pollution. Currently, most of these studies have mainly focused on the North China Plain, Yangtze River Delta, and Pearl River Delta regions [30,39,40,41,42]. The studies of the relationships between O₃ episodes and the influencing factors are rarely reported. Tianjin is an important economic center and a densely populated urban agglomeration area in northern China [43,44], therefore we focused on Tianjin in this study. The purposes of this study are as follows: 1—Analyze changes in O₃ episodes between 1 April and 31 August from 2018 to 2020 in Tianjin and the differences in the concentrations of VOCs and OFP in different years. 2—Analyze the main influencing factors (including precursors, meteorological factors, etc.) of O₃ episodes in different years through the GAM model, and determine the main driving factors of O₃ episodes.

2. Methods and Materials

2.1. Site Description

VOCs data were monitored at the Nankai University Air Quality Research Supersite (NKAQRS) (38.99° N, 117.33° E) between 1 April, and 31 August from 2018 to 2020, which is located on the Nankai University campus in the Jinnan Distinct, Tianjin (Figure 1). A student dormitory area is located to the south of the NKAQRS, and a road with relatively low traffic flow is sited 20 m to the north. Easterly and southeasterly winds prevailed during the study period. According to Chinese National Ambient Air Quality Standard and Technical Regulation on Ambient Air Quality Index (AQI) of China (MEE, 2012a), we define an O₃ episode as one day or a set of continuous days longer than 2 days with a maximum daily 8 h average (MDA8) ozone concentration exceeding 160 μg m⁻³. A day with mean ozone values of the maximum daily 8 h average (MDA8) exceeding 160 μg m⁻³ is defined as a high-O₃ day. The mean wind speed (ws), temperature (T) and the relative humidity (RH) were 1.6 ± 1.1 m/s, 23.7 ± 6.5 °C, 64.4 ± 23.1%, respectively during O₃ episodes. The meteorological conditions and pollutants were shown in Table S1.

2.2. Species Monitoring

VOC-species data were continuously monitored at intervals of 1 hr using a GC955 series 611/811 VOC analyzer (Syntech Spectras Inc., Groningen, the Netherlands). In total, 54 VOC species designated as photochemical precursors by the United States Environmental Protection Agency (US EPA) were monitored and used in the premonitoring equipment calibration. The analyses were performed using a photo (PID) and a flame (FID) ionization detector, which ensured high sensitivity and effective identification. The GC955 series 611 and 811 devices are two separated sample and column systems, which measure high (C6–C10) and low (C2–C5) boiling-point VOC species, respectively. For the series 811, the C2–C5 VOC species in ambient air were preconcentrated on Carbosieves SIII at a temperature of −5 °C. The enriched compounds were then thermally desorbed by heating (to 270 °C) and were purged into the separation column. The target compounds were then detected by the PID and an FID. The Series 611 was used to determine C6–C10 VOC-species. Air samples were preconcentrated on Tenax GR at ambient atmospheric temperature (~30 °C). Target compounds were then desorbed at 230 °C, brought into a stripper column, and then an analysis column for separation and a PID for detection. Quality assurance and quality control measures were performed, including routine maintenance of the instrument every week, fortnightly five-point calibration and verification, daily single-point correction, and weekly filter replacement. The analyzer was calibrated using standard gas to determine the retention times and control peak areas; correlation coefficient typically varied between 0.900 and 1.000. The method detection limits (MDL) for VOC species ranged from 0.019 to 0.599 ppbv. In accordance with the detection results, we obtained data for 54 VOC species within four VOC subcategories. Wind roses during the study period are given in Figure S1 in the Supplementary Materials. The data were subject to strict quality control measures, and abnormal data were excluded from the analysis.

SO₂, NO₂, CO data were obtained from SO₂ analyzer (API., Baltimore, ML, USA), NOx analyzer (API., Baltimore, ML, USA), CO analyzer (API., Baltimore, ML, USA) in the Nankai University Air Quality research Supersite (38°59′ N, 117°20′ E) in the Jinnan district of Tianjin, China.

2.3. Methods

2.3.1. Ozone Formation Potential (OFP)

The maximum incremental reactivity (MIR) method proposed by Carter [45] is widely used as a good indicator for comparing the ozone formation potential (OFP) of individual VOC species, and has been used to evaluate the photochemical reactivity of VOCs with OH radicals and for estimation of the contribution of individual organic compounds to ozone formation, and it is defined as follows:

$${\mathrm{OFP}}_{\mathrm{i}}{=\mathrm{MIR}}_{\mathrm{i}}\times {\mathrm{C}}_{\mathrm{i}}$$

(1)

$$\mathrm{OFP}={{\displaystyle \sum }}^{}{\mathrm{OFP}}_{\mathrm{i}}$$

(2)

Here, [VOC]_i is the concentration of VOC species i, OFP_i is defined as the ozone formation potential of individual species i, MIR_i is defined as the maximum incremental reactivity coefficient for individual species i, which is updated by Carter [46]. Additionally, OFP is defined as the ozone formation potential of the total species.

2.3.2. GAM

GAM have a nonlinear relationship between the corresponding variables and predictors using a smooth function,

$$\mathrm{g}\left({\mathrm{u}}_{\mathrm{i}}\right)={\mathrm{a}}_0+{\mathrm{f}}_1\left({\mathrm{x}}_{1\mathrm{i}}\right)+{\mathrm{f}}_2\left({\mathrm{x}}_{2\mathrm{i}}\right)+\dots +{\mathrm{f}}_{\mathrm{k}}\left({\mathrm{x}}_{\mathrm{ki}}\right)+{\mathsf{\epsilon }}_{\mathrm{i}}$$

(3)

where i indicates the ith hour’s observation. k refers to the type of impact factors. f_k(x) are smooth functions of the data. The element g(μ_i) is the “link” function, which specifies the relationship between the linear formulation on the right side of Equation (1) and the response μ_i. Nonlinear functions f_k(x) are used to represent the complex relationship between ozone and impact factors. ε_i is the residual [47].

In this study, we considered hourly O₃ concentration as the response variable, and hourly value of the relevant influencing factors, which were divided into environmental and meteorological factors in this study, as the explanatory variable. The GAM model check is mainly used to evaluate the quality of the proposed optimal model through the gam.check function in the R language mgcv package. In addition, we used the adjusted R² and variance interpretation rate to evaluate the quality of the fitted GAM. The higher the adjusted R² and variance interpretation rate are, the better the model fitting effect are.

3. Results

3.1. Characteristics of Ozone and Its Precursors during O₃ Episodes

O₃ episodes with MAD8 greater than 160 μg/m³ are shown in Table S2. In total, 33 O₃ episodes from 2018 to 2020 were analyzed in this study. The ozone concentration of O₃ episodes in 2020 was 66.4 ppbv, which was higher than 2018 (61.9 ppbv) and lower than 2019 (75.5 ppbv). The frequency of O₃ episodes were 15, 17, 4 and the number of days were 29, 28, and 5 d in 2018, 2019 and 2020, respectively (Figure 2). The longest duration of O₃ episodes was 2 days in 2020, while the longest were 6~8 d in 2018 and 2019. The total hours of O₃ episodes with an hourly value greater than 160 μg/m³ were 209, 191, and 30 h, and the hourly peak concentrations were 344, 258 and 211 μg/m³ in 2018, 2019 and 2020, respectively. The hourly peak concentrations in 2017 was observed (279 μg /m³) in the JJJUA [48]. The MAD8 peak concentration from 2014 to 2018 was observed (297 μg /m³) in Shijiazhuang [49]. It can be seen that although the value of ozone concentration in 2020 was not the lowest, the O₃ episodes showed a decreasing trend in terms of pollution frequency, days, heavy pollution duration and peak concentration. This has a lot to do with the unprecedented reduction in air pollution emissions caused by the various containment measures taken by the Chinese government during the period of COVID-19 outbreak [50,51,52].

The average value of NO₂ were 25, 19, and 15 ppbv during O₃ episodes in 2018, 2019 and 2020, respectively, showing a decreasing trend. From the diurnal variation of pollutants during O₃ episodes (Figure 3), we can see that CO and SO₂ in 2020 were 1.7 ppbv and 3.3 ppbv, respectively, which is significantly higher than previous years.

3.2. Characteristics of VOCs and OFP during O₃ Episode

The highest concentration and proportion of alkanes were ethane (2.8 ppbv, 19.5%), cyclopentane (2.4 ppbv, 16.8%), methylcyclopentane (2.3 ppbv, 15.9%), propane (1.6 ppbv, 11.3%) and n-butane (1.3 ppbv, 8.9%) during O₃ episodes in 2020 (Figure S2). Propane (2.6 ppbv, 29.9%) and ethane (1.9 ppbv, 21.8%) were of the highest content during O₃ episodes in 2018. The concentration of propane (2.7 ppbv) was the highest, followed by cyclopentane, ethane, n-butane, i-pentane, and n-pentane during O₃ episodes in 2019. Ethane and propane mainly come from incomplete combustion of motorcycles and motor vehicles [53] and liquefied petroleum gas/natural gas [54,55], and gasoline volatilization/emissions are characterized by high content of C4 and C5 alkanes [23]. i/n-pentanes were typical tracers for gasoline vehicle exhausts [56]. Cyclohexane can be released from diesel fuel evaporation [57]. Above all, among O₃ episodes, LPG/natural gas emissions were important sources in 2018, and both LPG/natural gas and gasoline volatilization/emissions were important sources in 2019 and 2020.

The concentration and proportion of 1-hexene (0.5 ppbv, 31.2%), ethylene (0.3 ppbv, 22.8%), and propene (0.3 ppbv, 21.9%) were the highest among the alkenes during O₃ episodes in 2020. 1-hexene (3.3 ppbv, 48.9.4%), Styrene (1.5 ppbv, 22.1%), isoprene (1.3 ppbv, 19.7%),had the highest concentrations and proportions during O₃ episodes in 2019. The concentration and proportion of propene (0.4 ppbv, 34.1%), isoprene (0.4 ppbv, 32.8%), ethylene (0.3 ppbv, 24.8%) and cis-2-butene (0.1 ppbv, 13.4%) were the highest during O₃ episodes in 2018. Ethylene and propene are important raw materials and products for rubber, plastics and other chemical manufacturing industries [58,59]. Isoprene was mainly derived from biological sources, indicating that rubber, plastics and other chemical manufacturing industries were important sources, and the impact of biological emissions may show a downward trend.

Toluene (0.7 ppbv, 34.6%), benzene (0.6 ppbv, 29.6%), and m,p-xylene (0.4 ppbv, 20%) were major species among aromatics during O₃ episodes in 2020. BTEX accounted for 91% of aromatics, indicating that they were affected by vehicle emissions [60,61] and solvent sources [62]. The concentrations and proportions of toluene (0.6 ppbv, 37.5%) and benzene (0.5 ppbv, 30.7%) were the highest during O₃ episodes in 2019, when BTEX accounted for 90% of aromatics. Toluene (0.7 ppbv, 42.0%), benzene (0.5 ppbv, 30.3%), m,p-xylene (0.2 ppbv, 12.7%) and ethylbenzene (0.1 ppbv, 8.9%) were major species among aromatics during O₃ episodes in 2018, and BTEX accounted for 94% of aromatics. A toluene/benzene ratio lower than 2 indicates predominantly vehicular emissions, and a value greater than 2 was considered a complex emission source (motor vehicles, industry, solvent use). The toluene/benzene ratios during O₃ episodes from 2018 to 2020 were 1.7 ± 2.0, 1.2 ± 1.5 and 1.5 ± 2.1, respectively, indicating that vehicular emissions were an important source.

The top ten substances of VOCs concentrations are shown in Figure 4. The major VOCs components during O₃ episodes were essentially consistent in 2018 and 2020, and most of the components were low-carbon alkanes. During O₃ episodes, propane, ethane, acetylene and i-pentane were the major species in 2018, and ethane, acetylene, cyclopentane and methylcyclopentane were the major species in 2020. Additionally, 1-hexene was the main component in 2019, followed by propane and cyclopentane.

The value of OFP during O₃ episodes from 2018 to 2020 are shown in Figure S3. The top ten OFP species are shown in Figure S4, where the main ozone-contributing species were propene (4.68 ppbv), isoprene (4.08 ppbv), and ethylene (2.58 ppbv) in 2018. The main ozone-contributing species in 2020 were propene (6.11 ppbv), cyclopentane (4.30 ppbv), methylcyclopentane (3.55 ppbv) and ethylene (3.18 ppbv), and alkenes were the important contributor to ozone formation. The top ten substances of OFP in 2019 were significantly different, and the main ozone-contributing species were 1-hexene (17.72 ppbv), isoprene (11.83 ppbv), propene (6.25 ppbv) and cyclopentane (4.83 ppbv).

During the O₃ episodes in 2018 (Figure 5), the concentration of O₃ was high from 12:00 to 19:00, and the concentration of propane and i-pentane in this period first decreased significantly, then increased sharply after 14:00, and then decreased and leveled off after 16:00. In 2019, the concentration of O₃ was higher from 14:00 to 20:00, and the concentration of 1-hexene and propane first decreased significantly during this time period, then increased after 20:00. In 2020, the concentration of O₃ was higher from 12:00 to 21:00; the concentration of methylcyclopentane first decreased significantly during this time period, and increased sharply after 16:00, then decreased and stabilized after 19:00, and acetylene increased sharply after 18:00. Other major VOCs species did not show an obvious trend during this period.

3.3. Evaluation of the Influencing Factors on O₃ Episodes

Relevant research shows that the concentration of ozone is related to CO, SO₂, NO, NO₂, acetylene, alkenes, aromatics, alkanes, wd(wind direction), ws(wind speed), T(temperature), RH(relative humidity), P(average air pressure), pre(precipitation), vis(visbility), TSRI(Total solar radiation intensity) and other influencing factors. Spearman’s correlation analysis was conducted between hourly ozone and the corresponding influencing factors during O₃ episodes from 2018 to 2020 (Figures S5–S7). The explanatory variables in the study are divided into environmental and meteorological factors. Environmental factors include CO, SO₂, NO, NO₂, acetylene, alkenes, aromatics and alkanes; meteorological factors include wd, ws, T, RH, P, Pre, TSRI, vis. A total of 16 impact factors were selected as explanatory variables, and O₃ concentration as the response variable, whose distributions are all normal (Figure S8). The multifactor correlation analysis was carried out through the GAM model, and the effective data was totaled into 1475 groups.

Multivariate analysis showed that retention factor significantly affected the change of O₃ concentration at the level of p < 0.001, which was statistically significant. Factors that did not pass the significance analysis of TVOC were deleted, and O₃ was taken as the explanatory variable used to reconstruct the multifactor GAM model until all variables passed the significance test. Factors which were included in the final GAM model in different years are listed in Table S3. The R² of the multifactor GAM model during O₃ episodes from 2018 to 2020 were 0.92, 0.90 and 0.98, respectively, and the deviances explained were 94%, 90.8% and 98.9%. The model-fitting effect in Table S3 was better.

The gam.check function was used through the mgcv program package of R version 4.0.2 to evaluate the fitting effect of the multifactor model (Figures S9–S11). From the model residual QQ diagram, the points were approximately distributed in a straight line. From the residual histogram, the residual mean was close to 0, and the frequency distribution was centered at 0. The closer to 0, the higher the frequency, and the two sides were essentially symmetrical. The model residuals were approximately normally distributed. From the scatter plot of residuals and predicted values, the points were essentially randomly distributed, indicating that the residuals were not related to the predicted values. Judging from the scatter plot of the observed values and the fitted values, the two basically had a straight-line distribution of y = x, indicating that the response variable after the model was fitted has a higher degree of matching with the fitted value. In conclusion, the multifactor model proposed in this paper had a good fitting effect.

The F value reflects the relative importance of each explanatory variable of the model to the dependent variable [63]. During O₃ episodes in 2018, vis (16.6), RH (15.3), NO (14.3), and T (13.1) had the highest F value and had the greatest impact on O₃. T (70.7), NO₂ (22.9), wd (22.8), and vis (21.9) had the highest F values in 2019 (Table S3). T (81.8), NO₂ (39.8), vis (24.8), and ws (24.4) had the highest F values in 2020. It can be seen that O₃ episodes were mainly driven by meteorological and precursor (NO) factors in 2018. O₃ episodes were mainly driven by meteorological conditions (T), followed by precursor (NO₂), in 2019 and 2020.

Multifactor correlation analysis was performed through the GAM model to obtain the impact effect map of the influencing factors (Figure 6), and the specific influence of each influencing factor on O₃ concentration were analyzed. T was the factor that had the greatest influence on O₃ during O₃ episodes in 2019 and 2020, which was much higher than other factors. O₃ and T had a mainly nonlinear and positive correlation, with O₃ episodes mainly occurring when T was greater than 20 °C, and ozone concentration increased with the rise of temperature. Further, there was an obvious inflection point, and when the temperature was higher, the increase in ozone concentration was more obvious, while the trend tends to be flat. In 2018, the effect of T on ozone was not obvious at 20–35 °C, and the ozone concentration showed an increasing trend with the increase in temperature after 35 °C. High temperatures can enhance solar radiation and reduce cloud cover, which in turn increases the intensity of photochemical reactions, leading to an increase in ozone concentration [64].

NO₂ was a factor that has a greater impact on O₃ during O₃ episodes in 2019 and 2020, second only to T. O₃ and NO₂ were mainly nonlinear and negatively correlated, and the O₃ concentration decreases with the increase of NO₂ concentration, and the confidence interval of NO₂ concentration was relatively narrow, indicating that there was a significant negative effect. NOx emission controls in China have been motivated mainly by the goal of decreasing nitrate PM₂.₅, and further controls were expected in the future (http://env.people.com.cn/n1/2020/0515/c1010-31710781.html, accessed on 20 August 2021). NO had a greater impact on O₃ during O₃ episodes in 2018 and 2019. When the NO concentration was lower than 10 ppbv, the O₃ concentration showed a decreasing trend, but when the NO concentration was high, the overall impact trend was positive. This may be due to the role of HO₂ in promoting the oxidation of NO to NO₂ in the high-concentration environment [65], and photolysis generating O₃ and increasing its concentration. Under certain conditions, nitrogen oxides, nitric oxide and VOCs generate ozone through photochemical reactions, which increase the ozone concentration and cause higher ozone pollution.

Vis was the factor that had the greatest impact on O₃ during O₃ episodes in 2018, and it also had a greater impact on O₃ in 2019 and 2020. Visibility may be affected by several factors such as rain or thunderstorms, haze, fog and mist. The O₃ concentration and vis were mainly nonlinear and negatively correlated in 2018 and 2019. As the vis increased, the O₃ concentration gradually decreased. Vis showed a nonlinear, positive correlation at 10–20 km in 2020, and a nonlinear, negative correlation after 20 km. RH was a factor that had a great impact on O₃ during O₃ episodes in 2018—it was second only to vis. The O₃ concentration was mainly nonlinearly related to RH. When RH < 60%, the effect of RH on O₃ concentration did not change significantly; when RH > 60%, the O₃ concentration decreased with the increase in RH, and the decrease range was larger. TVOC (alkenes, aromatic hydrocarbons, alkynes, alkanes) had relatively low effects on O₃ concentration, and some had not passed the GAM significance test. Compared with meteorological factors, their impact on ozone was very small.

4. Conclusions

1. O₃ episodes from 2018 to 2020 were analyzed in this study, The ozone concentration of O₃ episodes in 2020 was 66.4 ppbv, which was higher than 2018 (61.9 ppbv) and lower than 2019 (75.5 ppbv), although the value of ozone concentration in 2020 was not the lowest, the O₃ episodes showed a decreasing trend in terms of pollution frequency, days, heavy pollution duration and peak concentration. The average value of NO₂ during O₃ episodes showed a decreasing trend, CO and SO₂ in 2020 were 1.7 ppbv and 3.3 ppbv, respectively, significantly higher than previous years.

2. The major VOCs components during O₃ episodes were essentially consistent in 2018 and 2020, and most of the components were low-carbon alkanes. Propane, ethane, acetylene and i-pentane were the major species in 2018, ethane, acetylene, cyclopentane, and methylcyclopentane were the major species in 2020; 1-hexene was the main component in 2019. Above all, LPG/natural gas emissions were important sources in 2018, and both LPG/natural gas and gasoline volatilization/emissions were important sources in 2019 and 2020.

3. The main ozone-contributing species were propene (4.68 ppbv), isoprene (4.08 ppbv), and ethylene (2.58 ppbv) in 2018. The main ozone-contributing species in 2020 were propene (6.11 ppbv), cyclopentane (4.30 ppbv), methylcyclopentane (3.55 ppbv) and ethylene (3.18 ppbv), and alkenes were important contributors to ozone formation. The top ten substances of OFP in 2019 were significantly different. The main ozone-contributing species were 1-hexene (17.72 ppbv), isoprene (11.83 ppbv), propene (6.25 ppbv) and cyclopentane (4.83 ppbv).

4. Based on the GAM results, O₃ episodes was mainly driven by meteorological and precursor (NO) factors in 2018, while meteorological conditions (T), followed by precursor (NO₂) were the main driving factors in 2019 and 2020. Different factors had different driving impact on the O₃ episodes. T had a nonlinear and positive impact, while NO₂ had a nonlinear and negative impact.