Identification of the source of benzene concentrations at Texas City during SHARP using an adjoint neighborhood-scale transport model and a receptor model



[1] The Aerodyne Research Inc. mobile laboratory performed real-time in-situ measurements of volatile organic compounds, nitrogen oxides, and formaldehyde in Texas City, Texas on 7 May 2009 during the Formaldehyde and olefins from Large Industrial Releases experiment of the 2009 Study of Houston Atmospheric Radical Precursors (SHARP) campaign. Our goal was to identify and quantify emission sources within the largest industrial facility in Texas City most likely responsible for measured concentrations of benzene, an important VOC and hazardous air pollutant. The Houston Advanced Research Center inverse air quality model has been used to infer benzene emission rates from all potential source locations that could account for the benzene concentrations measured by the mobile lab in the vicinity of the facility. A Positive Matrix Factorization receptor model was also applied to the concentrations measured by the mobile lab, the results of which strongly supported the source attribution specified by the inverse model. The two independent source apportionment techniques both implicated flare, storage tank, and ultraformer units in the facility as significant contributors to emission plumes of elevated benzene concentrations observed by the mobile lab. The emissions of some of the flare and tank units were found to be greater than reported in emission inventories by about an order of magnitude.

1 Introduction

[2] Many of the volatile organic compounds (VOCs) that lead to ozone formation are also hazardous air pollutants that adversely affect human health. The Houston-Galveston metropolitan area is home to a significant proportion of the United States petrochemical industry. Relatively high levels of both ozone precursors and air toxics are emitted into the atmosphere of residential neighborhoods near petrochemical facilities despite the industry's aggressive efforts to reduce emissions [Knighton et al., 2012].

[3] Eulerian 3D chemical transport models such as CMAQ are widely used to model air quality for regulatory applications for scales ranging from urban to national [Byun and Ching, 1999]. However, due to their coarse resolution (as small as 4 km) as currently implemented, such models cannot capture the high concentration gradients near sources, which are critical in toxic exposure studies and other air quality assessments at neighborhood scales. Recently, Isakov et al. [2009] developed a hybrid approach combining regional-scale photochemical grid and local-scale plume dispersion models to simulate ambient pollutant concentrations resulting from both nearby and distant emission sources. Nevertheless, they emphasized the need to develop a single grid model with sufficiently fine resolution that can be used in exposure assessments.

[4] In addition to inadequate spatial and temporal resolution, most air quality models use emissions which are usually self-reported by industrial facilities based on estimates using standard emission factors rather than measurements. Several studies have shown that emissions of VOCs are either underestimated or not reported in official emission inventories for the Houston area [Mellqvist et al., 2010; Brioude et al., 2011; Kim et al., 2011; de Gouw et al., 2007; Lefer et al., 2010a, 2010b]. Inaccurate emission inventories and the difficulty of simulating large concentration gradients with coarse Eulerian models underline the need to develop finer scale models with source apportionment capabilities. Such models are needed to quantify distinct sources of air pollutants within a large industrial complex and to distinguish between relatively steady fugitive emissions and transient emission events, as well as to evaluate air quality impacts near sources.

[5] Source apportionment methods are used to infer potential sources of air pollutants and to quantify their contribution to measured ambient air concentrations [Hopke, 2003]. Some source apportionment methods combine observed concentrations and an Eulerian air quality model, in many cases using inverse modeling techniques to adjust emissions data used as input to the model [Mendoza-Dominguez and Russell, 2001; Henze et al., 2009]. Inverse models are especially useful in evaluating urban and regional-scale emission inventories and in detecting underreported or unreported source emissions [Mendoza-Dominguez and Russell, 2001]. Recently, Olaguer [2011] developed a neighborhood-scale fine resolution Eulerian grid model (referred to as the Houston Advanced Research Center (HARC) neighborhood air quality model) that can simulate near-source concentrations of long-lived species such as benzene and optimize emissions and other model parameters using the 4D variational (4Dvar) data assimilation technique. Olaguer [2012] then developed a chemical mechanism to simulate near-source concentrations of chemically reactive species, whereas Olaguer [2013] developed the adjoint of the full chemical transport model for 4Dvar inverse modeling of chemically reactive emissions.

[6] Observation-based receptor modeling techniques, such as positive matrix factorization (PMF), have also been widely used by the air quality research community [Paatero and Tapper, 1994]. PMF provides a multivariate statistical analysis of ambient concentration data to determine the number of sources, chemical composition profiles, and contributions of these sources to measured atmospheric mass loadings. Evaluating inverse Eulerian air quality model results with an observation-based statistical model may be useful for improving source apportionment.

[7] During the Formaldehyde and olefins from Large Industrial Releases (FLAIR) experiment in the spring of 2009, the Aerodyne Research Inc. (ARI) mobile laboratory [Kolb et al., 2004] performed in situ measurements of VOCs, nitrogen oxide (NOx), and formaldehyde in the vicinity of the largest industrial complex in Texas City, Texas. The Aerodyne Inverse Modeling System (AIMS), which is a Gaussian puff inverse model developed by ARI based on the adjoint method [Albo et al., 2011], was initially used for the preliminary analysis of the mobile lab data [Stutz et al., 2011]. Benzene concentrations (as high as 205.8 ppbC) measured by the ARI mobile lab on 7 May 2009 from 11 A.M. to 3 P.M. were used as input to AIMS in order to optimize emissions of benzene from a cooling tower identified in the 2006 Texas Commission on Environmental Quality (TCEQ) special point source emission inventory. Concentrations predicted by the AIMS were in general one third or one fourth of the measured concentrations throughout the modeling period [Stutz et al., 2011]. Emission rates from the identified source “cooling tower” deduced by AIMS were up to two to three orders of magnitude higher than the value reported in the 2006 special inventory [Stutz et al., 2011]. However, in the 2009 TCEQ emission inventory (the year of the field campaign), this cooling tower unit was reported to have zero benzene emissions. Furthermore, no emission events were reported to the TCEQ's emission event database from this unit on 7 May 2009. These preliminary results suggested further investigation of the influence of potential sources contributing to benzene concentrations measured by the mobile lab.

[8] In this study, we performed fine horizontal resolution (200 m × 200 m) inverse modeling by running the HARC air quality transport model (i.e., with no chemistry) in adjoint mode. The simulations were conducted with a horizontal domain size of 4 km × 4 km for a 3 hr period (11 A.M. to 3 P.M. local standard time (LST)). Potential point source locations within the petrochemical complex were specified using a high-spatial-resolution digital map of the facility. A positive matrix factorization receptor model was also applied to the concentrations of other compounds measured by the mobile lab in addition to benzene to support the source attribution estimated by the inverse model.

2 Methodology

2.1 Study Location

[9] The modeling study was conducted for the area encompassing the Texas City industrial complex (TCIC), southeast of Houston. The selected modeling domain consists primarily of refineries and chemical plants.

[10] A digital map of the major emission sources in the TCIC was developed from 2009 TCEQ emission inventory to improve the accuracy of source apportionment. The spatial coordinates of the emission units are quality assured using 2006 and 2011 TCEQ emissions inventories, GIS-based permit information, aerial photography, and previous remote-sensing studies conducted at the facility based on the differential absorption lidar technique [TCEQ, 2010; Randall and Coburn, 2010]. Figures 1a–1d show the significant benzene sources selected for inverse modeling as represented by the digital map of the TCIC.

Figure 1.

(a) Final benzene concentrations for the 11:00–12:00 time period. (b) Final benzene concentrations for the 12:00–13:00 time period. (c) Final benzene concentrations for the 13:00–14:00 time period. (d) Final benzene concentrations for the 14:00–15:00 time period.

2.2 Mobile Lab Measurements

[11] The accuracy of source apportionment studies depends heavily on the quality of meteorological data and measurements of ambient concentrations. A rich database of measurements suitable for source apportionment studies was provided by the ARI mobile laboratory, equipped with an extensive set of real-time monitoring devices [Kolb et al., 2004]. A proton transfer reaction mass spectrometer (PTR-MS) onboard the ARI mobile lab performed high temporal and spatial resolution measurements of propylene, benzene, toluene, C2-benzenes, C3-benzenes, styrene, acetaldehyde, and acetone; additional onboard instruments monitored NOx, SO2, CO, CO2, C2H4, and formaldehyde along street level routes in Texas City, Texas during the FLAIR campaign. These measured concentrations were used for source apportionment using the PMF technique. Only the benzene measurements, however, were used for inverse modeling based on the adjoint version of the HARC air quality model. Although high-time resolution meteorological measurements were made onboard the mobile lab, Stutz et al. [2011] found that wind speed resolution from a fixed monitor operated by the TCEQ was of better quality and that wind direction data from the TCEQ monitor was more accurate and consistent for the whole modeling period. Therefore, as stated in section 2.3.2, wind speed and direction data from the fixed monitoring station were used in the HARC model.

2.3 The HARC Air Quality Model

2.3.1 Inverse Modeling Method

[12] This study utilized the passive tracer version of the HARC model originally developed by Olaguer [2011]. We do not require the full chemistry transport model, as benzene has an atmospheric lifetime of about one month. The forward version of the model takes emissions data as input to generate temporally and spatially varying ambient concentrations. The model can also run in adjoint mode to identify and quantify emission sources.

[13] The HARC model run in adjoint mode uses the 4Dvar data assimilation method to minimize a cost function (J), defined as the weighted least squares distance between the 4D temporal and spatial predictions of a forward model and concentration measurements over an assimilation window or time period [Olaguer, 2011]. The cost function in this study includes the emission rates, the horizontal turbulent diffusion coefficient, the forward dispersion model's initial conditions, and the time-dependent concentration predictions.

display math(1)

[14] Kh refers to the horizontal diffusion coefficient, which will be optimized together with the emissions. The quantities with the superscript B are background (or prior) values of the corresponding parameters. Ck is a column vector with dimension equal to the number of grid cells and represents the discretized field of predicted concentrations at time step k. C0,B is the column vector of background initial concentrations, representing the a priori guess concentration field.

[15] The HARC model is first run in forward mode using the background concentrations as the initial conditions. At the end of the forward run, an adjoint final condition is computed as input to the adjoint model. The adjoint final condition quantifies the disagreement between the forward model final concentrations and the observations. The model is then run backward in time using the continuous adjoint of the transport equation [Olaguer, 2011] to correct the emissions, horizontal diffusion coefficient, and background concentration field. The correction runs are conducted iteratively, with several cycles of forward and backward integration over the same assimilation window. In each iteration cycle after the initial one, the background values of the optimized parameters are replaced by their adjusted values from the previous iteration.

2.3.2 Implementation of the HARC Air Quality Model

[16] In this study, the HARC model is implemented over a domain size of 4 km × 4 km × 1 km with spatial resolution of 200 m × 200 m × 50 m. The model time step was set at 20 s. Hourly average wind speed and direction data were acquired from the Continuous Ambient Monitoring Station (CAMS) 147 station located at Texas City Ballpark. Although wind data were collected from the mobile laboratory, wind speed and direction data collected from the TCEQ fixed monitor were more reliable and therefore were used in this study as in Stutz et al. [2011]. Hourly wind speed varied between 5.45 and 6.48 m/s and hourly wind direction ranged from 166° to 170°. Inverse modeling was conducted for four 1 hr periods (11:00–12:00 LST, 12:00–13:00 LST, 13:00–14:00 LST, and 14:00–15:00 LST) on 7 May 2009. A uniform horizontal wind was assumed for each modeling period as in Olaguer [2011]. Mobile lab measurements of benzene were used as input to the HARC adjoint model. The inflow boundary condition for benzene was set at 2 µg/m3. For reference, the 2009 annual average benzene concentration for Texas City monitors was about 2.3 µg/m3 [TCEQ, 2013b]. The background error covariance matrix was assumed to be diagonal, with a magnitude of 2000 µg2/m6. The initial guess horizontal diffusion coefficient during the first modeling period (11:00–12:00 LST) was 50 m2/s. As in (Olaguer, E. P., S. C. Herndon, B. Buzcu-Guven, C. E. Kolb, and A. E. Cuclis, Application of an adjoint neighborhood scale chemistry transport model to the attribution of primary formaldehyde at Texas City during SHARP/FLAIR, J. Geophys. Res., in review), the model-updated diffusion coefficients were used in subsequent modeling periods as initial guess coefficients.

[17] Since the model domain does not have large sources of benzene other than the facilities in the Texas City Industrial Complex [Stutz et al., 2011], only point sources of benzene were included in the modeling simulation; no area or mobile sources were considered. For the source attribution using the HARC model, 37 sources in the TCIC were selected, including 18 storage tanks and 8 flare units. These units had the highest benzene emissions (upper 50%) reported in the 2009 TCEQ emission inventory (Table 1). The total reported benzene emissions from these 37 sources were 2.88 kg/hr and accounted for 83.4% of the total reported emissions from all the benzene sources in the emission inventory. Routine benzene emissions from the 37 sources are displayed in Table 1 and were used as initial guess emissions in the HARC model. A benzene emission event on 7 May from the PX2 Flare unit was reported to TCEQ, allegedly releasing 4.1 kg of benzene over the course of 8 hr. As a sensitivity check, this emission rate (0.511 kg/hr) was used as the initial guess for the PX2 Flare unit instead of the routine emission rate of 2.54 × 10−4 kg/hr. Model results were insensitive to the higher initial emission rate estimate and did not show any significant discrepancies in the predicted concentration fields and the optimized emissions. All the emission sources were assumed to be located within the lowest grid layer, except for the flare units, for which effective release heights exceed 50 m, and therefore were assumed to be located within the second model layer from the ground. The initial estimate for the emissions error variance was 0.1 (g/s)2. The emissions updated by the inverse model are assumed to be averaged over each 1 hr simulation period.

Table 1. Initial Benzene Emission Estimates (kg/hr) for Selected Point Sources Based on 2009 TCEQ Emission Inventory [TCEQ, 2013a]
Selected Emission SourcesIDReported Emissions11:00–12:0012:00–1:001:00–2:002:00–3:00
Env FacilityS30.4320.4240.4010.5430.554
Ultracracker FlareS40.27722.69411.49829.8190.374
Tank 1S60.0841.7570.8672.3270.085
Tank 2S70.0090.8530.4561.1520.518
Tank 3S80.0620.9760.7631.0780.092
Tank 4S90.0101.7521.0822.1610.739
Premium Lead FreeS100.0101.0710.6211.3260.848
Tank 5S110.0221.3240.5241.4690.022
FCCU3 WetgasS120.0080.0080.0080.1660.197
Tank 6S130.0072.5821.5344.1640.027
Aromatic RecoveryS140.0710.0700.0680.1340.186
Distillate Desulfurizer FlareS150.3994.8513.5285.7980.936
Aromatic Unit#2S170.0250.0250.0250.1230.027
Tank 7S190.0210.5540.3830.7090.107
Tank 8S200.0161.9690.9832.8050.017
FCCU3 RegeneratorS210.0140.0140.0140.1970.289
Tank 9S230.0100.0700.0040.0420.067
Tank 10S240.0090.2070.0270.4280.535
Tank 11S250.0092.5671.5893.2990.596
Tank 12S260.0073.4592.0934.6630.429
Residual/Cat Feed FlareS270.0070.0070.0070.0680.011
FCCU No 1S280.0100.0090.0070.0280.029
Tank 13S290.0060.0060.0060.1390.122
Tank 14S300.0060.8800.4961.2530.025
Tank 15S310.0050.0040.0040.0140.016
Tank 16S320.0050.0030.0020.0670.294
Tank 17S330.0050.0000.0040.0120.016
Tank 18S340.0040.0040.0050.0980.501

2.4 Source Apportionment With Receptor Models

[18] In this study, PMF analysis was performed on the mobile lab measurements during the FLAIR campaign. CO, NO2, formaldehyde, and ethylene concentrations were measured by Quantum Cascade Laser (QCL) tunable infrared spectrometers (3 s averages were used for this work); NO and SO2 were measured with commercial chemiluminescent and pulsed ultraviolet fluorescence instruments, respectively. Propylene, benzene, toluene, C2-benzenes (sum of the xylenes and ethylbenzene), styrene, acetaldehyde, acetone, and C3-benzenes (sum of C9H12 isomers) were measured by PTR-MS. Limiting the PMF analysis to the observations on 7 May would produce erroneous results due to the small number of data points. In order to get a more complete picture of the source contributions, PMF was conducted on data collected by PTR-MS and QCL in 3 s intervals from 28 April 2009 to 8 May 2009, and aggregated into 30 min intervals to balance specificity and accuracy.

[19] The weight given to each individual data point is adjusted by PMF according to the measurement uncertainty. In this analysis, 15% of the measured concentrations were assigned to each data point as the measurement uncertainty [Buzcu and Fraser, 2006]. Missing data are substituted by the model with each species' median concentration and assigned an uncertainty of 4 × median concentration [US EPA, 2008].

3 Results and Discussion

3.1 Inverse Model Results

[20] The simulation results for the final benzene concentrations at the surface predicted by the forward model with optimized model parameters at the end of each 1 hr simulation period are displayed in Figures 1a–1d. Figures 1a–1d also show the path of the mobile lab during the corresponding simulation period and the point sources used in the modeling simulation. The peak benzene concentrations occur at the middle of the plume for the first three modeling periods, close to two flare units (ultracracker flare (S4) and distillate desulfurizer flare (S15)), three storage tanks (S13, S25, and S26), and the ultraformer unit (S18). The Paraxylene flare unit (S1) located to the south of the model domain is also one of the high emitters for the first three modeling periods, and despite the fact that the mobile lab route was further away from this source, its effect on benzene concentrations are noticeable in Figures 1a–1c. For the final modeling period, concentrations peaks were simulated near flare units (S5, S15), fugitive units (S16), tanks (S9, S10), and the environmental facility (S3). The locations of these peaks also coincide with the high emitting sources predicted by the HARC model (Table 1). The spatial extent of the concentration isopleths for the three simulation periods is similar; however, the forward model-predicted concentrations show variations in magnitude consistent with the measured concentrations.

[21] Figure 2 compares mobile lab measurements of ambient benzene concentrations averaged over the appropriate model grid cells with corresponding HARC model predictions for the four modeling periods. In general, the HARC model is able to simulate variations in benzene concentrations fairly accurately, with a correlation coefficient of 0.56 (Table 2). According to the two statistical metrics, mean fractional bias (MFB) and mean fractional error (MFE), calculated for model evaluation [Bergin et al., 2007; Baldasano et al., 2010; Tong et al., 2012], model predictions of benzene concentrations showed relatively moderate bias and error relative to observations (Table 2). As a reference, Boylan and Russell [2006] set the model performance goal (best level of accuracy that can be achieved) as MFE < 50% and −30% < MFB < 30%. A model performance criterion (acceptable level of accuracy) is met when MFE < 75% and −60% < MFB < 60% [Boylan and Russell, 2006]. For comparison, Tong et al. [2012] found MFE and MFB values for black carbon in the range of 11.77 to 15.87% and −6.46 to −14.08%, respectively, and Baldasano et al. (2010) calculated MFE and MFB as 29.7% and −29.4% for PM10, respectively, and MFE of 15.6% for NOx in their local-scale model evaluation. Discrepancies between modeled and measured concentrations can be attributed to uncertainties in model assumptions, input concentrations, and meteorological data (uniform wind data), as well as the exclusion of mobile sources.

Figure 2.

Measured and modeled benzene concentrations in µg/m3. (x axis shows the grid cells (x,y) over which the mobile van traveled during the simulation period and for which the concentrations are averaged).

Table 2. HARC Model Performance Measures
Performance MeasuresBenzene
MFB (%)4.52
MFE (%)51.02
Mean Observed Conc. (St Dev) (µg/m3)8.63 (9.76)
Mean Simulated Conc. (St Dev) (µg/m3)8.15 (9.22)
Median Observed Conc. (µg/m3)3.95
Median Simulated Conc. (µg/m3)3.13
Min (Max) Observed Conc. (µg/m3)0.61 (45.13)
Min (Max) Simulated Conc. (µg/m3)2.00 (37.06)

[22] The HARC model-optimized values of the hourly averaged emissions for the selected 37 point sources during the four simulation periods are shown in Figure 3. Total benzene emissions from these 37 sources optimized by the model for the four modeling periods were 63.53, 36.57, 85.11, and 11.07 kg/hr, respectively. For comparison, the total self-reported routine emissions in the 2009 TCEQ emission inventory for the selected benzene sources are 2.88 kg/hr, which is an order of magnitude lower than the model-predicted emissions. In general, emission estimates for the four modeling periods were of the same order of magnitude, with relatively higher estimates in the first and third modeling period for tanks and flares (Figure 3). The variations in predicted emissions were consistent with the variations in measured and modeled concentrations for the four modeling periods (Figure 2).

Figure 3.

The emission rates calculated by the inverse model for all selected point sources.

[23] The HARC model attributed 30% of the total benzene emissions during the first three modeling period to tanks, about 53–55% to flares, 5–8% to the ultraformer units, and 5% to fugitive sources. In the fourth modeling period, these contributions showed some changes; 38% of total benzene emissions were apportioned to tanks, 29% to flares, 5% to an environmental facility, 5% to fluidized catalytic cracking (FCCU) units, 3% to fugitives, and 2% to ultraformer units. The total emissions from flare units were 35 kg/hr, 20 kg/hr, 46 kg/hr, and 3 kg/hr and emissions from storage tanks were 19 kg/hr, 11 kg/hr, 26 kg/hr, and 4 kg/hr for four modeling periods. Emissions from aromatic units and FCCU units increased an order of magnitude in the last two modeling periods. The emissions from the other sources did not change significantly between the three periods but are generally lower in the last modeling period. The optimized values of horizontal diffusion coefficient are 46, 37, 28, and 49 m2/s for the four 1 hr modeling periods, respectively.

3.2 Source Apportionment With PMF

[24] The PMF model runs were optimized based on the analysis of residuals, goodness-of-fit values, calculated versus measured concentrations, and the physical meaningfulness of the factor profiles and their respective contributions. Overall three factors were identified; storage tanks and aromatics units, flares, and ultraformers, which together account for 70.5% of the measured benzene concentrations. The remaining 29.5% of the benzene concentrations is attributed to sources unresolved by PMF (“unknown sources”). The optimized PMF solution was bootstrapped 100 times with a random seed to determine the confidence intervals for each species in each factor profile, and to assess the robustness of the solution. The compositions of the three factors representing the dominant sources of benzene in the TCIC are shown in Figure 4.

Figure 4.

Chemical composition profiles (variability in percentage of species) and associated uncertainties as calculated by PMF. The red stars are the percent contribution of elements to each factor. The box encompasses the 25th and 75th percentiles, and the whiskers are determined as 1.5 times the interquartile range.

[25] The factor compositions in Figure 4 can be interpreted by comparing them to the emission profiles of the major point sources within the TCIC in Table 3. The factor which is dominated by the BTEX (Benzene, Toluene, Ethylbenzene, and Xylenes) compounds and C3-benzenes (trimethylbenzene) represents storage tanks, aromatic recovery units, the environmental facility, and fugitives. This factor accounts for 56.6% of the total benzene concentration explained by the PMF model. The second factor, which is mostly associated with formaldehyde, SO2, ethylene, propylene, and benzene, represents flares and accounts for 36.7% of total benzene. The third factor, represented by CO, NO, ethylene, and propylene, is consistent with the routine emission profiles of ultraformer units in the industrial complex (Table 3). This factor contributes to 6.6% of the total benzene concentration.

Table 3. Emission Chemical Composition Profiles of the Major Selected Sources
Emissions (%)TanksARUDDFFLFL2FL3FL4FL1Res/Cat FLUCFUFUF3UF4EnvFugitive
C-2-Butene 00000.20.90.200    0.0
CO5.9046.380.14.528.034.135.613.936. 0.0
Cyclohexane0.60.12.300. 0.0
Cyclopentane0.40        000  
Ethane000000000000.31.1 0.4
Ethyl benzene2.32.5        00.70.510.112.9
Ethylene000. 1.7
Heptane10.90        00.70.5 0.1
Hexane5.01.30000000000.70.4 0.1
Isobutane0 0.0
m-xylene0000000000   2.84.3
Methane 0        00.20.8 0.1
n-butane8. 0.1
n-hexane9.3 4.700.     
NOx3.8010. 0.0
n-pentane05.00000000000.40.2 0.1
Octane0.2         00.50.5 0.1
o-xylene1.2 00000000   0.32.2
p-xylene2.6 00000000    3.4
Pentane11.7    0.0
Propane0.302. 0.1
Propylene000. 0.1
SO2 04.4088.523.941.725.675. 0.0
Trimethybenzene 1.2        01.41.0 0.0
Trimethybenzene,1,2,42.10           2.70.4

3.3 Comparison of HARC Model and PMF Results

[26] We now compare the model-predicted benzene source apportionment results with the results of the PMF analysis. Predicted percent contributions of each dominant source type from the HARC model and the PMF analysis are shown in Figure 5. Unapportioned mass in the PMF analysis was labeled as “Unknown Sources” and account for 29.5% of the total measured benzene concentration. The apportioned sources accounted for 70.5% of the total concentration; therefore, the percent contribution of the apportioned sources given in the previous paragraph were multiplied by 70.5% to obtain the percent contributions to total measured benzene concentrations measured by the mobile lab in Figure 5.

Figure 5.

Comparison of percent source contributions to benzene concentrations calculated by the PMF method and the HARC inverse model. The acronym UF refers to ultraformer units and FCCU refers to Fluidized Catalytic Cracker Unit.

[27] Figure 5 shows that the results from the two very different and independent source apportionment methods are in reasonable agreement. The HARC model attributes 32% of total benzene emissions to storage tanks, aromatic recovery units, the environmental facility, and fugitives, whereas PMF apportions 40% of the total source contributions to these types of sources. Flare units accounted for 40% of the total benzene emissions in the HARC model results and 26% of total source contributions in the PMF results. The percent source contributions of the ultraformer units were in agreement between the two methods (4 and 5%). The unapportioned sources contributed to 29.5% of the total benzene concentrations in the PMF results, while the HARC model attributed 22% to the other sources in the TCIC, which include cooling towers, FCCU units, loading operations, furnaces, and others. Although the self-reported emissions suggest that the flare units are the dominant sources of benzene in the modeling area, the HARC model and the PMF analysis suggest that storage tanks, an environmental facility, and ultraformer units are also significant sources of the benzene concentrations measured by the mobile laboratory (Figure 5). The unapportioned mass in the PMF analysis likely represents the contribution from sources that could not be distinctively resolved by PMF because of their similarity in chemical speciation of emissions or an insufficient number of species measured in the ambient samples. Despite the uncertainties associated with the models and input data and the differences between the two methods, the source attribution results show strong similarities.

4 Conclusion

[28] Real-time mobile measurements collected in a predominantly industrial area provided an ideal test bed for advanced source attribution. Benzene source apportionment using the HARC air quality model run in adjoint mode, and a receptor modeling technique provided two independent assessments of self-reported emissions from individual point sources in a large industrial complex. Despite their differences and associated uncertainties, the inverse model and PMF source apportion methods yielded comparable findings. The results suggest that self-reported emissions from the industrial facility are underestimated by an order of magnitude. Although the self-reported emissions suggest that the flare units are the dominant sources of benzene in the modeling area, the HARC model and the PMF analysis implicate storage tanks, an environmental facility, and an ultraformer as likewise significant sources of the benzene concentrations measured by the mobile laboratory.


[29] This work was supported by the U.S. Department of the Interior, Fish and Wildlife Service, Coastal Impact Assistance Program through Harris County, Texas.