Flare Event-Comparison of Dispersion Modeling Approaches and Outcomes Regarding Benzene Formation and Dispersion

Paper from the AFRC 2016 conference titled Flare Event-Comparison of Dispersion Modeling Approaches and Outcomes Regarding Benzene Formation and Dispersion

This paper will present a case study utilizing a comparison of multiple state, federal, and international flare dispersion modeling packages including screening models and refined models. The end goal will be to present current flare modeling issues and suggestions for the flare industry to assist in creating better models for industry's use. The paper will use the assessment of a single non-routine flaring event at a petrochemicals facility in Italy performed by using USEPA and Canadian approved models and current available modeling guidelines.

Flare Event - Comparison of Dispersion Modeling Approaches and Outcomes Regarding Benzene Formation and Dispersion Troy Boley, Sage ATC Environmental Consulting, LLC Rishabh Jaishankar, Sage ATC Environmental Consulting, LLC Igor Shnayder, Sage ATC Environmental Consulting, LLC Pooja Sujit, Sage ATC Environmental Consulting, LLC American Flame Research Council 2016 Industrial Combustion Symposium Kauai, Hawaii September 14, 2016 Introduction Air dispersion modeling plays an integral role in the process of understanding the dispersion behavior of pollutants emitted by a combustion source. Various refined and screening dispersion models can be used to predict and assess the ambient air and ground-level concentration of various emissions. Refined models require a significant amount of site-specific data and incorporate a meticulous and detailed data setup approach. This type of modeling uses real-time meteorological data to assess ambient air quality and pollutant dispersion. Screening modeling on the other hand uses simpler algorithms, fewer data inputs, and conservative techniques to estimate the pollutant concentrations in ambient air. Therefore, the concentration results from refined modeling are sometimes much lower than those predicted through the use of screening models. All models share similarities with the inputs required, but they also possess striking differences which ultimately provide a unique insight into the significantly different concentration results that can be derived from each model for the same flaring event. Of interest in this paper, is a comparison of the available modeling packages and their roles in determining the dispersion of pollutants from elevated flares, and particularly benzene. Dispersion modeling of emissions from elevated flares is highly complex, and there are multiple approaches and techniques to consider. This paper aims to highlight the different modeling packages used to analyze both a single non-routine flaring event and a worst case scenario flaring event for a confidential client site located in Italy along a coastal region known for complex meteorology. It also illustrates the differences in results obtained by using the different modeling techniques that rely upon "effective" input parameters. 1 Background Information In typical flare dispersion modeling scenarios, flares are represented and modeled as point sources, in order to represent the flare stack as a ‘true stack'. The dispersion behavior of pollutants emitted from point sources depends on certain model inputs such as gas exit velocity, temperature, stack height, diameter, emission rate, meteorological data (atmospheric conditions), terrain, and building downwash considerations among others. More specifically, modeling point sources requires important stack parameters to be defined: two inputs related to the physical stack (the stack height and diameter), and two related to the gas being emitted (the gas exit temperature and velocity). These stack parameters can either be represented as the actual physical stack parameters of the flare or can be modified or adjusted to better represent the emission of pollutants during a flaring event. Different authors, federal, and state regulatory agencies have historically developed a variety of methodologies to determine these modified, or "effective" flare stack parameters. These different guidelines developed by the United States Environmental Protection Agency (USEPA), various state regulatory agencies of the United States (Texas, Oklahoma, Ohio, Alabama, New Jersey, Iowa); Alberta Department of Environment and Water (Alberta, Canada); and the Practical Guide to Atmospheric Dispersion Modeling by Mr. Bruce Turner were individually used to calculate the modified point source parameters for input into each modeling package used to analyze the non-routine flaring event. While ‘effective' gas exit velocities and temperatures, as well as ‘effective' stack heights and ‘effective' diameters can be evaluated to match those of ‘normal' stacks, typically one to three of the four parameters are assigned certain fixed values (e.g., the gas temperature of 1,000 degrees Celsius and gas exit velocity of 20 meters per second) and the ‘effective values' are calculated for the stack height and stack diameter for input into the model. Since the pollutant concentration predictions for point sources depend on the thermal buoyancy and momentum (dynamic) plume rise, the importance of using appropriate stack parameters is conveyed. Refined Modeling Packages - Single Non-routine Flaring Event In order to assess the single non-routine flaring event and further understand the difference in the modeling results and predictions across each model, a comparison study of various refined models was conducted for a confidential client operating a petrochemical complex in Italy. This significant flare release scenario lasted around six and a half hours and was modeled to determine the dispersion of pollutants after burning in the combustion zone of the flare. In order to model and assess this specific release scenario, various refined modeling techniques were used because real-time meteorological data was available for the specific flaring time period. The modeling assessment involved sophisticated refined air dispersion models (specifically 2 AERMOD, ISC3, and CALPUFF), all of which are capable of accurately evaluating the dispersion of pollutants in the atmosphere under a wide variety of metrological conditions including complex atmospheric phenomena during the specific time period of the flaring event. Furthermore, maximum ground level concentrations (GLCmax) values were determined from the same flare using different refined models with input flare source parameters determined from each state, federal, and international methodology described earlier. Since the main purpose of the event modeling was to compare the predictions of the different models with different modeled stack parameters to each other, a single generic pollutant emission rate was modeled and chosen as an input into each refined model. The assessment produced a range of numerical predictions for GLCs, reflecting differences in modeling algorithms and source parameter evaluation methodologies. The following modeling algorithms were used to model the pollutant dispersion during the specific non-routine flaring event and the GLCmax results were analyzed to determine the difference in the modeling predictions across each model. Ø AERMOD AERMOD is a steady-state Gaussian plume dispersion model designed to predict near field ambient concentrations from multiple emission sources. It can generate daily, monthly, and annual pollutant concentrations and can be configured to assess a multitude of terrain types. The model assumes turbulent horizontal and vertical plume dispersion from the flare tip that results in Gaussian (i.e., bell shaped) concentration distributions. This model is well-suited for short (up to 50 km) distances and steady wind. Beyond 50-km distance, the model predictions become unreliable because the model keeps assuming the meteorological conditions which existed at the time of the release. AERMOD requires processed meteorological data inputs to estimate ambient concentrations. The model contains algorithms that simulate the effects of surrounding buildings and other structures on pollutant concentrations. Other capabilities also exist within AERMOD, including a wide variety of receptor grid configurations, the ability to vary source emission rates, and the ability to make necessary adjustments to account for plume rise, buoyancy, and stack tip downwash. AERMOD can predict concentrations in both stable and convective atmospheric conditions. AERMET is used as the meteorological preprocessor to AERMOD and takes in the real-time wind details and vertical profiles to estimate pollutant concentrations. AERMOD was used for the specific event modeling scenario because this is the most recent refined modeling tool developed by the USEPA and is suitable for accommodating the specific meteorological conditions which were observed during the flaring release event. Ø ISC3 3 The basis of the Industrial Source Complex (ISC) Short Term model is the straight-line, steady-state Gaussian plume equation, which is used with some modifications to model simple point source emissions from stacks, emissions from stacks that experience the effects of aerodynamic downwash due to nearby buildings, isolated vents, multiple vents, storage piles, conveyor belts, and the like. The ISC Short Term model accepts hourly meteorological data records to define the conditions for plume rise, transport, diffusion, and deposition. It estimates the concentration or deposition value for each source and receptor combination for each hour of input meteorology, and calculates user-selected short-term averages. ISC3 is the predecessor model to AERMOD. AERMOD replaced ISC3 in 2006 as the preferred USEPA model for near-field pollutant concentration estimates. ISC3 was used for the specific event modeling scenario because it is a reliable modeling tool that was used by the USEPA for over 25 years as the primary regulatory model and is suitable for accommodating the specific meteorological conditions which were observed during the non-routine flaring event. While this model is typically no longer acceptable for regulatory purposes in the US, it is appropriate for this study to compare results since many historical reviews were conducted using ISCST. Ø CALPUFF CALPUFF is a non-steady state, refined Lagrangian puff dispersion model used for both near and far field applications. The model takes into account the transient, non-steady nature of the flare emissions and simulates the effects of varying meteorological conditions on pollutant transport, transformation, and removal. CALPUFF works efficiently for long-range gas transport (>50km) and is preferred by the USEPA for long-range dispersion applications. CALPUFF can take complex terrain into consideration, and due to supplemental procedures incorporated in CALPUFF ver. 6 and later, it is generally favored for facilities near an ocean/waterfront. Lagrangian puff models like CALPUFF are suitable for long range transport because they generally exhibit higher correlation with observations compared to the steady-state models at longer distances. CALPUFF is better suited within longer distance model domains where the winds vary spatially and where temporal variation in the wind after stack release needs to be considered, as this model allows for time-dependent winds by releasing the plume as a continuous series of puffs. Under low wind conditions, an important effect of non-steady state dispersion is that the puff can change direction with changing winds which results in a meandering trajectory. The ability of CALPUFF to identify and track this change gives the model an advantage relative to AERMOD in replicating concentrations under low wind conditions. CALPUFF was used for the specific event modeling scenario because this program is a refined modeling tool developed by the USEPA and suitable for accommodating the specific meteorological conditions which were observed during the July 2015 release event. Note that in 2015 the USEPA proposed to remove CALPUFF from the list of preferred and 4 recommended models for both near-field and long-range applications, and replace it with AERMOD as the preferred model; however, this model is still used in long-range transport modeling in the US until a final decision is made. Comparison of results - Refined Models Considered In order to accurately determine the GLCmax throughout the entire flaring time period and evaluate the effects post flaring, the concentration results for various averaging time periods was considered. This also served to verify that the pollutant concentration decreases as time passes. The GLCmax results of the non-routine flare event indicated that the highest modeled ground level concentration of pollutants originated from the ISC3 model. It is to be noted that the results from the AERMOD model provided ground level concentration results that were significantly lower than the ISC3 model results. Since benzene is a significant HAP strictly regulated by various air quality standards, a study of the impact of benzene was also conducted for this flaring event in order to assess the potential impacts on public health. It should be considered that benzene can also form during the combustion of hydrocarbons. For an evaluation of benzene formation from hydrocarbons, the total mass of hydrocarbons destroyed in the flame zone of a flare is to be counted. It is estimated that no more than 0.01 mole (vol.) % of total hydrocarbons may be converted to Benzene in the flame zone . A formation benzene analysis was conducted and added to the benzene already present in the fuel to the flare, in order to make a better comparison with the European Union's annual Air Quality Standard (AQS) for benzene. 1 As described earlier in this report, the refined CALPUFF model was also used to determine the ground level concentration of the pollutants dispersed during the non-routine flaring event. The CALPUFF model evaluated impacts using the stack parameters calculated from USEPA flare source parameter methodology only. The CALPUFF predictions were noted to be approximately 50% of the ISC3 predictions that used the comparable USEPA flare modeling parameters for 1hour and 3-hour average concentrations, and are approximately 35% of the ISC3 predictions for 8-hour and 24-hour average concentrations. Therefore, one can confirm that there are significant differences between the predictions from each of the different models, caused mostly by each model's different way of estimating the plume motion and dispersion by turbulence. As a result of a comparison of the model predictions for the three modeling systems (ISC3, AERMOD, and CALPUFF) and flare dispersion source parameter methodologies used, the range of expected ground-level concentration values and locations had been established. For the H.R. Zhuang, et al. Fuel Dependence of Benzene Pathways. Proceedings of the Combustion Institute 32 (2009) pg 377-385 - The University of Utah, Salt Lake City, UT. 1 5 comparable flare parameters, the CALPUFF predictions are 1.8-2.9 times higher than the AERMOD predictions, and ISC3 model predictions and are 1.7-6.5 times higher than the AERMOD predictions, depending on the averaging period. In terms of GLCmax location, the CALPUFF and ISC3 model predictions were also close to each other but differed from the AERMOD predictions. CALPUFF was also noted to take different terrain information into account, since the model predicted the GLCmax results could have occurred in agricultural land terrain, while both the ISC3 and AERMOD models predicted maximum impacts to occur over bodies of water or near rural coastal areas. Therefore, ISC3 when compared with the other refined models, consistently showed higher GLCmax results over the different averaging periods which is a noteworthy conclusion from this comparison analysis. Screening Modeling - Worst-Case Scenario Flaring Event In order to assist the client with determining the most conservative possible ground-level concentration of pollutants, notably benzene, worst-case flaring scenarios were modeled for all flares prevalent at the site. A variety of screening models were used to analyze the worst-case scenario as time and event-specific meteorological data is typically not required in a screening analysis. Screening models require less input data and were considered ideal for modeling of the worst-case scenarios that is not modeled to occur at a specific time period. Worst-case emissions, such as those evaluated in the screening analysis for the client, are typically associated with non-routine, emergency events. These types of emissions do not occur over an entire year, but rather over a few hours or a few days. The conservative factors used to estimate impacts over longer averaging periods likely result in over-predictions of modeling concentrations. In some cases, these concentrations could be greatly over-predicted. The models also assume consistent meteorological conditions along with persistent stream compositions that do not change over time; in reality, these conditions will vary greatly. Nevertheless, this provides a conservative analysis of the GLCmax results from the screening models. The following modeling algorithms were used to model the pollutant dispersion from a possible worst-case flaring event at each flare at the client facility. The modeling was performed using two different modeling systems that have been developed by the USEPA and Softbits Consultants from the United Kingdom (UK). Ø AERSCREEN AERSCREEN is a steady-state, single-source plume screening model that is based on boundary layer equations used by the AERMOD model. It is currently the USEPA's recommended model for use in screening applications. A primary function in AERSCREEN is estimating worst-case 1-hour concentrations for a single source without the need for sitespecific hourly meteorological data. Conversion factors can be applied to the 1-hour concentrations to estimate impacts for other averaging periods, as well. AERSCREEN can evaluate impacts from a wide range of meteorological conditions. Flare modeling in 6 AERSCREEN includes point source inputs, building downwash, terrain influences, wind specifications, and user-specified receptor locations. The AERSCREEN model was utilized with the flare source parameters calculated from the USEPA methodology and the Alberta source parameter methodology. The AERSCREEN results using the source parameter inputs derived from the USEPA methodology consistently derived GLCmax that were higher than the results derived from the Alberta methodology. However, it is important to note that the Alberta source parameter methodology and the USEPA source parameter methodology did not yield massive differences in the GLCmax results. Ø FlareSIM FlareSIM was developed by Softbits Consultants in the UK and is primarily used for designing flare systems. However, it is also capable of flare gas dispersion modeling, and uses Gaussian plume dispersion to determine ground-level impacts of various pollutants at greater distances from the flare. It also uses jet dispersion modeling to determine the details of the flame at the flare tip and concentration of possible flammable gas near the flame. FlareSIM can also be used to evaluate thermal radiation and noise from flares along with the temperatures of exposed surfaces. FlareSIM does not require extensive meteorological data, but relies on the Pasquill-Gifford stability class system and the urban/rural terrain classification to estimate impacts. FlareSIM inputs include fluid type, flare assist system type, stack orientation and dimensions, flare tip descriptions, receptor points and details, and environmental factors like wind speed/direction, radiation, solar radiation, and noise. Comparison of results from the screening models considered Both models do not require extensive meteorological data, but the similarities were noticed to end there. The two models are extremely unique in their modeling features. For instance, the AERSCREEN model was used with the effective stack height and diameter derived from the USEPA and Alberta source parameter methodologies. FlareSIM on the other hand accepts actual physical parameters of the flare and does not typically take in the modified stack parameters. FlareSIM also has the capability to speciate the compounds in the gas being emitted from the flare and derives the concentration results for combusted products and uncombusted flare fluids, while AERSCREEN provides the concentration results as a generic pollutant emitted from the flare. Moreover, FlareSIM significantly differentiates between the GLCmax distances for combustion products such as NO and CO and uncombusted flare fluid products (dependent on the gas composition routed to the flare) which are available for dispersion after combustion in the flare. The model uses the effective release height for dispersion of combustion gases as the end of the flame. It is evident that the model was mainly developed and verified for combustion products but was extended with some assumptions in order to be used for the plume buoyancy of 7 uncombusted flared fluids. As quite a contrast, the AERSCREEN model treats all pollutants as a single generic pollutant being emitted from the flare. The screening models were also run using both the rural and urban settings. The results from the FlareSIM model were very sensitive to the selection of Pasquill Stability Class and the urban or rural terrain specification and relied on these inputs to estimate impacts, which is unlike the AERSCREEN screening model which showed that the rural and urban specification both brought about the exact same GLCmax results. Conclusions Screening Assessment Modeling Results For the four flare screening modeling, Benzene impacts through FlareSIM were conservatively evaluated for Flare 3, as it considers both direct uncombusted Benzene as well as additional Benzene that may theoretically form as a secondary product of hydrocarbons combustion. FlareSIM speciates each compound in the combustion products and/or the uncombusted flare fluid; therefore, GLCmax for Benzene can be estimated. The primary GLCmax for Benzene of 125.8 µg/m is associated with Stability Class A and rural terrain classification. This impact is predicted to occur 2,350 meters downwind of the flare. Model runs with Stability Classes B through F with the Rural dispersion option show negligible concentrations for all these atmospheric stabilities. Stability Class A in the Pasquill-Gifford classification characterizes very turbulent conditions and is infrequent, requiring light (<2.5 m/s) winds, high (> 60 degrees) solar elevation angles, and clear to mostly clear skies that does not screen out incoming solar radiation. For this assessment, a very conservative assumption was applied where clear skies and low wind speeds are observed during all hours of high solar elevation. Ephemeris data show that the maximum number of days in Brindisi when solar elevation exceeds 60 degrees is less than 110 days per year. Additionally, during the summer solstice, the number of hours when solar elevation exceeds 60 degrees is five. If the average number of hours with solar elevation exceeding 60 degrees is three, and assuming it applies for all 110 days of the year, the probability of Class A stability affecting the same receptors in a year is (110 * 3) / 8,760 (hours/yr) is 0.038. Using this conservative factor in combination with the EPA conservative screening multiplier which accounts for frequency of wind directions during a year, the resulting annual GLCmax is less than 0.5 µg/m , or less than 10% of the EU standard. 3 3 The secondary GLCmax for Benzene is associated with urban terrain classification. For Stability Classes A and B the model predicts the same impact of 14.8 µg/m located more than 6,500 meters downwind of the flare. For urban dispersion runs with Stability Classes C through F the model predicts negligible concentrations for all these atmospheric stabilities. Since rural and urban dispersion depend on the wind direction and land features in each direction, only one (urban or rural) characteristic is appropriate for each wind direction. Furthermore, based on the noticed difference in the distance from the flare to the point of GLCmax, the GLCmax from the urban and rural predictions cannot overlay even on an annual basis. Therefore, it is appropriate to separately evaluate the rural and urban predictions for comparison with the annual standards. 3 8 After applying the EPA screening factor of 0.1 for conversion of hourly to annual concentrations (which only accounts for the fact that wind directions are variable and during a year the same receptor is not impacted for more than 10% of the time) to the prediction above, the resulting annual GLCmax for Benzene is less than 1.5 µg/m , or less than 30% of the EU standard. 3 A table summarizing the Benzene impacts evaluated with AERSCREEN and FlareSIM for each flare is shown below. AERSCREEN Flare FLARESIM % comparison Annual GLC of with European Benzene (µg/m ) Union Standard (5 µg/m ) max 3 3 Annual GLC of Benzene (µg/m ) max 3 % comparison with European Union Standard (5 µg/m ) 3 1 0.82 16% - - 2 1.75 35% 1.48 30% 3 1.34 27% - - 4 0.33 7% - - It should also be noted that the modeling domain within FlareSIM is limited to 20km. Stability classes beyond Class A are characterized by moderate to stronger wind speeds that can range from 10 km/hr to speeds exceeding 20 km/hr, in addition to cloudy sky conditions that block solar radiation and sun elevation angles that are less than 60 degrees. It is likely that, particularly for higher stability classes, the plume is traveling at high elevations and not hitting the ground until well beyond the modeling domain used in the FlareSIM simulation, resulting in negligible modeled (i.e. concentrations would be non-zero if the modeling domain could be expanded beyond allowable by the modeling code). When considering the frequency of weather conditions in real-time, concentrations at such larger distances would be much smaller than those predicted by the model for Stability Class A (rural dispersion) and A and B (urban dispersion). This provides another layer of conservatism to the model results, and concludes that no additional modeling is required for comparison of the GLCmax values of Benzene to the EU AQS. Event Assessment Modeling Results General Impacts The highest modeled 1-hour GLCmax impact per 1,000 lb/hr hydrocarbons was 3.50 µg/m for 1hour average concentrations, 1.35 µg/m for 3-hour average concentrations, 0.93 µg/m for 8-hour average concentrations, and 0.34 µg/m for 24-hour average concentrations. All these "highest of 3 3 3 3 9 all" predictions are associated with the ISC3 model and the "Turner" flare parameter calculations. All model results from the AERMOD model were significantly lower than the ISC3 model results. Benzene Impacts Total Benzene emissions from the event were considered from two sources: the potential uncombusted Benzene in the feed stream and Benzene that could have theoretically formed in the flame zone as a product of hydrocarbon combustion, as discussed above. Under these conservative assumptions, the total Benzene considered for the event modeling is equal to 337.4 lb/hr. ISC3 and AERMOD model predictions for the same source are always directly proportional to the emission rate. Using the generic impacts for 1,000 lb/hr of total VOC emissions summarized above, Benzene concentrations can be estimated by using the unit impacts and applying a ratio for Benzene. The maximum 1-hour impact for Benzene is estimated to be 1.18 µg/m = (3.50 µg/m *337.4/1000). Similarly estimated, the GLCmax Benzene impacts from the emissions event are 0.45 µg/m for 3-hour, 0.31 µg/m for 8-hour, and 0.11 µg/m for 24-hour averaging periods. While these impacts cannot be directly compared to the annual EU AQS Benzene standard of 5 µg/m , one can assume that the computed impacts from the event contribute to far less than 0.01% of the annual standard (0.11 µg/m [24-hr impact] / (365 days/yr) / 5 µg/m = 0.006% of the European Union's annual standard). 3 3 3 3 3 3 3 3 CALPUFF Event Modeling Results General Impacts The CALPUFF model evaluated impacts from Flare 3 using the stack parameters calculated from USEPA methodology. Using the 28,600 lb/hr total hydrocarbon emission rate, the GLCmax for VOC for the 1-hour averaging period was predicted to be 14.7 µg/m . Additionally, 3-hour and 8-hour impacts for VOC were predicted to be 5.5 µg/m and 2.1 µg/m , respectively. Since the emissions event lasted less than 8 hours, with remaining hours in the modeling period assumed to be zero or negligible, the 24-hour averaging period was not modeled in CALPUFF. As in common practice, the 24-hour impact was estimated as one-third of the 8-hour impact, or 0.71 µg/m of VOC. 3 3 3 3 For the modeling of 1,000 lb/hr of total hydrocarbon emission rate comparable with the ISC3 and AERMOD, the CALPUFF predictions are approximately 50% of the ISC3 predictions that used the comparable USEPA flare modeling parameters for 1-hour and 3-hour average concentrations, and are approximately 35% of the ISC3 predictions for 8-hour and 24-hour average concentrations. Benzene Impacts 10 In CALPUFF, Benzene impacts were evaluated by directly modeling the 337.4 lb/hr emission rate that was estimated from the weight percent of direct Benzene in the total VOC stream along with the Benzene potentially formed from hydrocarbon combustion kinetics, as discussed above. With this approach, and using USEPA methodology, the 1-hour GLCmax for Benzene was 0.17 µg/m . Predicted 3-hour and 8-hour average concentrations for Benzene were 0.064 µg/m and 0.025 µg/m , respectively. Since the emissions event lasted less than 8 hours, with remaining hours in the modeling period assumed to be zero or negligible, the 24-hour average impact was estimated as one-third of the 8-hour impact, or 0.008 µg/m . These impacts cannot be directly compared to the EU's annual standard of 5 µg/m . However, one can safely assume that the impact will be far less than 0.0005% of the annual standard when extrapolated to consider the entire year. 3 3 3 3 3 This exercise showed that modeling a flaring event with original benzene input plus additional formation benzene uner a variety of modeling approaches produced results of non-impact to the local communities. Naturally, the various model algorithms available do not always produce uniform results. However, all models did predict similar outcomes. Often times, the modeling software used will be dictated by the regulating agency involved, however, in areas without any clear guidance on the matter, the software itself can result in significantly differing results. Thus it is critical to employ the appropriate assumptions and ensure a thourough understanding of each model being considered. 11