Top-down estimates of benzene and toluene emissions in the Pearl River Delta and Hong Kong , China

Benzene (C6H6) and toluene (C7H8) are toxic to humans and the environment. They are also important precursors of ground-level ozone and secondary organic aerosols and contribute substantially to severe air pollution in urban areas in China. Discrepancies exist between different bottom-up inventories for benzene and toluene emissions in the Pearl River Delta (PRD) and Hong Kong (HK), which are emission hot spots in China. This study provides top-down estimates of benzene and toluene emissions in the PRD and HK using atmospheric measurement data from a rural site in the area, Heshan, an atmospheric transport model, and an inverse modeling method. The model simulations captured the measured mixing ratios during most pollution episodes. For the PRD and HK, the benzene emissions estimated in this study for 2010 were 44 (12–75) and 5 (2–7) Gg yr for the PRD and HK, respectively, and the toluene emissions were 131 (44–218) and 6 (2–9) Gg yr, respectively. Temporal and spatial differences between the inversion estimate and four different bottom-up emission estimates are discussed, and it is proposed that more observations at different sites are urgently needed to better constrain benzene and toluene (and other air pollutant) emissions in the PRD and HK in the future.


Introduction
Benzene and toluene, two volatile organic compounds (VOCs), are toxic to humans and the environment. For example, a sufficiently high exposure of toluene will lead 20 to health issues like intra-uterine growth retardation, premature delivery, congenital malformations, and postnatal developmental retardation (Donald et al., 1991). VOCs, including benzene and toluene, are also important precursors of ground-level ozone that is produced from the reaction between VOCs and NO x in the presence of sunlight (Xue et al., 2014) and contribute to the formation of secondary organic aerosols (Henze zene was one of the most the most abundant VOCs in PRD (Chan et al., 2006). Toluene and benzene, respectively, had the largest and second largest emissions of all anthropogenic VOCs in PRD in 2010(Ou et al., 2015, which highlights the importance of accurately quantifying these emissions. In PRD, the two major sources of benzene are industrial processes and road transport, and those of toluene are industrial solvents 15 and road transport, while minor sources include industrial processes, gasoline evaporation, building paints, biomass burning, etc. (Ou et al., 2015).
Although some bottom-up inventories exist for benzene and toluene emissions in PRD, there are discrepancies among them. For example for benzene emissions in 2010, the Regional Emission inventory in Asia (REAS) v1.1 reference scenario (from 20 here on referred to as REAS REF v1.1) estimates the emissions to be 8 Gg yr −1 (Ohara et al., 2007), while the Multi-resolution Emission Inventory (MEIC) v1.2 (available at http://www.meicmodel.org) estimate is 33 Gg yr −1 , the Representative Concentration Pathways Scenario 3PD (RCP 3PD) estimate is 45 Gg yr −1 (van Vuuren et al., 2007), and the Yin et al. (2015) estimate is 54 Gg yr −1 . Thus, estimates of the total emis-25 sions vary by a factor of approximately seven. For toluene emissions in 2010, the estimates are also quite different: the RCP 3PD and REAS v1.1 REF estimates are 44 and 46 Gg yr −1 , respectively, the Yin et al. (2015) estimate is 64 Gg yr −1 , and the MEIC v1.2 estimate is 181 Gg yr −1 . Atmospheric-measurement-based estimates are needed ber 2010. This study uses these measurement data and an inverse modeling approach to infer benzene and toluene emissions in PRD and HK. This top-down estimate is important to test and improve the existing bottom-up inventories.
2 Methodology 2.1 Measurement data 10 In this study, atmospheric measurements of benzene and toluene at two sites were used, the Heshan site (used for the inversion) and the Mt. Tai Mo Shan (TMS) site (used for validation). The Heshan site (112.929 • E, 22.728 • N) is a rural observatory located on the top of a small hill (∼ 60 m above the surrounding terrain) in Jiangmen (see Fig. 1). The measurement period at Heshan site was from 11 November 2010 15 to 1 December 2010. Data from 1 December 2010 were not used, since we focused on mixing ratios and emissions in November. Detailed information of the measurement system and procedure can be found in Wang et al. (2014). Here we provide only a brief description. Ambient mixing ratios of VOCs were measured using an online automatic gas chromatograph system equipped with a mass spectrometer and a flame ionization 20 detector (GC-MS/FID  -3, 9, and 19-21, November 2010. Detailed information on the sampling time schedule can be found in Guo et al. (2013). After sampling, the VOC canister samples were sent to a laboratory at the University of California, Irvine for chemical analysis. Simpson et al. (2010) provide a full description of the analytical 10 system, which uses a multi-column gas chromatograph (GC) with five column-detector combinations. The measurement detection limit of this system for both benzene and toluene is 0.003 ppb, which is much lower than the typical mixing ratio levels of 0.7 ppb for benzene and 1.6 ppb for toluene during the observation period at the Mt. TMS site.

15
The source-receptor relationships (SRRs, often also called "emission sensitivities") were calculated using the backwards in time mode of the Lagrangian particle dispersion model, FLEXPART (http://www.flexpart.eu) (Stohl et al., 2005(Stohl et al., , 1998 for reaction with OH radicals were expressed for benzene as: (1) expressed for toluene as: where T is the ambient temperature (K). Gridded OH fields (hourly for the period Octo-5 ber to December 2010, at a resolution of 0.5 • × 0.667 • , 47 vertical levels) were derived from the atmospheric chemistry transport model, GEOS-Chem v5 (http://acmg.seas. harvard.edu/geos/). Compared to the reference simulation ran backwards for 20 days with atmospheric chemical loss, Sect. 3.2 shows alternative FLEXAPRT simulations that were run backwards for 10 and 40 days with atmospheric chemical loss, and 20 days without atmospheric chemical loss.

Inverse algorithm
The Bayesian inversion method used in this study is almost the same as described and evaluated by Stohl et al. (2009Stohl et al. ( , 2010, and as used in recent studies of SF 6 emissions (Fang et al., 2014) and HFC-23 emissions (Fang et al., 2015) in East Asia. Briefiy, in 15 this study a Bayesian inversion technique is employed, based on least-squares optimization, to estimate both the spatial distribution and strength of the emissions in the domain to which the measurements are sensitive. The inversion adjusts the emissions to minimize the difierences between the observed and modeled mixing ratios while also considering the deviation of the optimized emissions from an a priori emission 20 field. Observation-model mismatch errors (including transport model errors) are determined as the root mean square error (RMSE) of the model-observation mismatch (Stohl et al., 2009(Stohl et al., , 2010 zero. This is because the backward simulations were run for 20 days and benzene and toluene in the air parcel from emissions occurring prior to this time would have been largely removed from the atmosphere by reaction with OH (typical atmospheric lifetimes of benzene and toluene are ∼ 10 and ∼ 2 days, respectively). Simulated benzene and toluene mixing ratios at the measurement site were obtained by integrating Both monthly inventories of MEIC v1.2 and Yin et al. (2015) were obtained through personal communication with the dataset authors. A priori emission uncertainty in each grid box for benzene and toluene, respectively, was set to 100 and 70 % according to the differences among the bottom-up inventories and was assumed uncorrelated in space.  site ranged from 0.87 to 25.05 ppb and had an average of 5.6 ± 4.15 ppb. The mixing ratios of benzene (0.67 ± 0.21 ppb) and toluene (1.58 ± 1.25 ppb) at the Mt. TMS were only ∼ 30 % of those at the Heshan site. In agreement with previous studies (e.g. Lau et al., 2010;Liu et al., 2008), mixing ratio levels of benzene and toluene in PRD region are overall higher than those in Hong Kong (Table 1), which is most likely due to the 5 fact that Hong Kong often receives clean air masses from over the ocean. Mixing ratios of benzene and toluene in some cities in Europe (e.g. Ait-Helal et al., 2014;Langford et al., 2010) and United States (e.g. USEPA, 1989;Baker et al., 2008) were about 0.5 and 1 ppb (Table 1), respectively, which is about 20 % of the mean observed values in PRD in this study. Mixing ratios of benzene and toluene in Thompson 10 Farm, United State were even 0.08 ± 0.002 and 0.09 ± 0.005 ppb, respectively, which are much lower than the lowest mixing ratios at both Heshan and Mt. TMS sites. Levels of benzene and toluene mixing ratios at different sites mainly reflect the combined infiuence of emission strength, seasonal changes in atmospheric OH concentration and mixing depth.  Figure 2 shows the spatial distribution of average emission sensitivity of benzene and toluene for the Heshan site for 12-30 November 2010. During the observation period, air masses transported to the Heshan site mainly came from easterly and northerly directions. Considering that the major emission sources in PRD are located to the east of 20 the Heshan site ( Fig. 1), this measurement location was ideally situated for constraining emissions from this region for this period and, as the emission sensitivities show, PRD, HK, and neighboring regions, are relatively well constrained by the observations at the Heshan site. Benzene and toluene emissions in PRD and HK are much higher than emissions in neighboring regions Fig. 1 and, consequently, the overall mixing ra-25 tio contributions (the integral of the emission sensitivities multiplied by emissions) from PRD and HK to the observation site comprise more than 80 % of the total simulated mixing ratios. Note that the emission sensitivities for benzene and toluene are different 24847 Introduction

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version

Interactive Discussion
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | because there are differences in the chemical loss of these two compounds during atmospheric transport and in the molecular weight. Specifically, the emission sensitivities for toluene are spatially more confined because of its shorter lifetime. As a sensitivity study, alternative simulations in which FLEXPART was ran backward for 10 days were made. The derived emission sensitivities are almost identical 5 to the reference simulations with 20 days duration (Supplement Fig. S1 for benzene and Fig. S2 for toluene), confirming that 20 day-backward simulations are sufficiently long to account for all benzene and toluene emission sources that can influence the mixing ratios at the Heshan site. Since the lifetime of benzene is ∼ 10 days (much longer than that of toluene), we also made a 40 day-backward simulation from which the emission sensitivities for benzene are also almost identical to the reference simulation (Fig. S3). Without accounting for the loss by reaction with OH in the atmosphere, the emission sensitivities for benzene would only be a little higher (by ∼ 10 % in central PRD) (Fig. S4). On the other hand, the emission sensitivities for toluene would be much higher (by ∼ 50 % in central PRD) (Fig. S5). This indicates that accounting for 15 chemical loss has a relatively small effect for simulating benzene mixing ratios at Heshan, whereas it has a profound effect on toluene mixing ratios. Notice that systematic errors in the simulated chemical loss would lead to smaller errors (i.e., compared to the extreme case of no loss) in the a posteriori emissions retrieved by the inversion. Thus, errors in the retrieved emissions due to errors in chemical loss are marginal for 20 benzene but could be significant for toluene. Figure 3 shows the observed and simulated mixing ratios at the Heshan site. The simulations captured most pollution episodes, and the inversion improved the agreement between the simulations and the observations, as expected. For benzene, root 25 mean square errors between the observed and simulated mixing ratios decreased from 1.53 ppb, using a priori emissions, to 1.26 ppb, using a posteriori emissions, and the mean bias between the simulated mixing ratios and observations decreased 24848 Introduction

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version

Interactive Discussion
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | from 0.96 ppb, using a priori emissions, to 0.41 ppb, using a posteriori emissions. For toluene, the root mean square error between the observed and simulated mixing ratios decreased from 4.77 ppb, using a priori emissions, to 4.30 ppb, using a posteriori emissions and the mean bias between the observed and simulated mixing ratios decreased from 2.35 ppb, using a priori emissions, to 1.99 ppb, using a posteriori emissions. Figure 3 also shows examples of spatial distributions of toluene emission sensitivities for two observed mixing ratios. The toluene mixing ratio at 00:00 UTC on 16 November 2010 was about 2 ppb and the corresponding air mass had not passed over the strong emission sources in the central part of PRD and HK (see the backward emission sensitivities map in Fig. 3c), while the toluene mixing ratio at 00:00 UTC on 24 November 2010 was about 15 ppb and the corresponding air mass had passed over the strong emission sources in the central part of PRD and HK (Fig. 3d). Figure 4 shows the benzene a priori and a posteriori emission fields, their differences and uncertainty reduction. The a priori fields show that emission hot spots are located in Guangzhou, Shenzhen and Hong Kong megacities. Emission changes by the in-15 version are positive in some grid cells and negative in some other grid cells, which shows that the a priori emissions are not systematically lower or higher everywhere than the a posteriori emissions. The biggest emission changes by the inversion occur in two boxes in Guangzhou where the a priori emissions were enhanced by ∼ 50 % in one box and decreased by more than 50 % in the other box. The emission hot spot in 20 Shenzhen did not change much. To test the sensitivity to the a priori emission in this grid cell, we performed an additional inversion in which the a priori emission in this grid box was reduced, and a high a posteriori emission in this grid cell was still found, as in the reference inversion. Figure 5 shows the a priori and a posteriori emissions of toluene and their differ-  Table 2), all produce toluene emission estimates that fall within the uncertainty range of the a posteriori emissions from the reference inversion. Benzene and toluene measurement data at the Mt. TMS site were not used in the inversion but for validating the posterior emissions. For benzene, using the a priori and a posteriori emission fields, respectively, the root mean square errors between the simulated and observed mixing ratios at Mt. TMS site are 0.367 and 0.312 ppb, and the mean bias between the simulated and observed mixing ratios are 0.314 and 0.208 ppb. For toluene, the root mean square error (1.50 ppb) between the observed 15 and simulated mixing ratios using the a posteriori emission fields from the inversion was smaller than that (1.55 ppb) using the a priori field; the mean bias (1.06 ppb) between the observations and simulated mixing ratios using a posteriori emission fields was also smaller than that (1.12 ppb) using the a priori field. Both the root mean square error and mean bias suggest that the a posteriori emissions are more accurate than 20 the a priori emissions. systematically biased. The simulated benzene mixing ratios using the REAS v1.1 inventory are much lower than the observed mixing ratios (Fig. 7). Statistics of the root mean square error, mean bias and squared Pearson correlation coefficients between the simulated and observed mixing ratios show that emission fields obtained from the inversion performed better in simulating the benzene mixing ratios than all four bottom-5 up inventories (See Table S1). For toluene, in most grid cells over the PRD, emissions estimated by RCP 3PD, Yin et al. (2015) and REAS v1.1 REF are lower than the inversion estimates, while MEIC v1.2 emissions are higher than the inversion estimates (Fig. S6). Model simulations show that the simulated mixing ratios using emission estimates from RCP 3PD, Yin  (Fig. S7). The simulated mixing ratios using MEIC v1.2 emission fields are not consistent with some observed pollution peaks (Fig. S7). Statistics of root mean square error and squared Pearson correlation coefficients show that inversion emission fields performed better at simulating toluene mixing ratios at the Heshan site than the 15 four bottom-up emission fields (see Table S1). Table 2  were obtained through personal communication with the dataset authors. Using these ratios, the benzene emissions in PRD and HK for 2010 were estimated to be 44 (12-75) and 5 (2-7) Gg yr −1 , respectively, and the toluene emissions were estimated to be 131 (44-218) and 6 (2-9) Gg yr −1 , respectively.    , which were only ∼ 1/5 of the estimate by Zheng et al. (2009). More studies are available for the year 2010 than for other years.

Benzene and toluene emissions during 2000-2010
For the year 2010, the estimates by RCP 3PD, MEIC v1.2 and Yin et al. (2015) agree with the inversion estimate by this study, which are higher than the estimate by REAS v1.1 REF. According to these estimates (Fig. 8), it is likely that the benzene emissions

Suggestions for more top-down studies
To the best of our knowledge, this study provides the only available top-down estimate for toluene emissions in the PRD and HK regions. All other studies in Fig. 8 are 10 bottom-up estimates. More top-down estimates are needed to validate the bottom-up estimates in the previous years and in the future. In this study, inversions using the Heshan measurement data reduced emission uncertainties in PRD and HK regions. However, the emission uncertainty reductions were not large because there was only one observation site suitable for the inversion (some measurements in urban environ-15 ments are available but not suitable for inverse modeling) and the observation period was not long. Thus, we propose that in the future, observations with better spatial and temporal coverage are urgently needed to better constrain benzene and toluene (and other VOC) emissions in PRD and HK regions. Inversion-oriented observation sites could be situated in rural places outside of the major emission sources located in the 20 central part of PRD and HK regions, and then the major emission sources in PRD and HK regions could be "viewed" from different angles to better constrain the benzene and toluene (and other VOC) emissions.
ACPD 15,2015 Top-down estimates of benzene and toluene emissions in Pearl River Delta and Hong Kong, China

Conclusions
Using atmospheric measurements at the Heshan site, a transport model and an inversion algorithm, this study provides the first top-down estimate of benzene and toluene emissions in the Pearl River Delta (PRD) and Hong Kong (HK) regions, which are emission hot spots in China. According to the measurement data in this study and previous 5 studies, mixing ratio levels of benzene and toluene in PRD region are overall higher than those in Hong Kong, which are much higher than those measured in the United States and Europe. Considering that air masses transported to the Heshan site mainly came from easterly and northerly directions during the observation period and that the major emissions sources in PRD are located to the east of the Heshan site, the Heshan 10 measurement site was ideally situated for constraining emissions from these regions. Based on the measurement data, model simulations and inverse technique, the PRD and HK benzene emissions for 2010 estimated in this study were 44 (12-75) and 5 (2-7) Gg yr −1 , respectively and the PRD and HK toluene emissions for 2010 were 131 (44-218) and 6 (2-9) Gg yr −1 , respectively. We have discussed the spatial distributions 15 of benzene and toluene emissions obtained by inversion in this study in the context of four different existing bottom-up inventories. The discrepancies among these bottomup estimates for the period 2000-2010 are substantial (up to a factor of seven), while this study is the only one available top-down estimate. We propose that in the future, observations with better spatial and temporal coverage are urgently needed to con-20 strain benzene and toluene (and other VOC) emissions in PRD and HK regions more strongly.  , 94, 237-244, doi:10.2307/3431317, 1991. Fang, X., Thompson, R. L., Saito, T., Yokouchi, Y., Kim, J., Li, S., Kim, K. R., Park, S., Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | C 2 -C 10 volatile organic compounds (VOCs), CO 2 , CH 4 , CO, NO, NO 2 , NO y , O 3 and SO 2 , Atmos. Chem. Phys., 10, 11931-11954, doi:10.5194/acp-10-11931-2010, 2010. Song, Y., Shao, M., Liu, Y., Lu, S., Kuster, W., Goldan, P., and Xie, S.: Source Apportionment of Ambient Volatile Organic Compounds in Beijing, Environ. Sci. Technol., 41, 4348-4353, doi:10.1021/es0625982, 2007 Stohl, A., Hittenberger, M., and Wotawa, G.: Validation of the Lagrangian particle dispersion model FLEXPART against large-scale tracer experiment data, Atmos. Environ., 32, 4245-4264, 1998. Stohl, A., Forster, C., Frank, A., Seibert, P., and Wotawa, G.: Technical note: The Lagrangian particle dispersion model FLEXPART version 6.2, Atmos. Chem. Phys., 5, 2461-2474,