Evaluating the NOx Performance of a Steam Generator for Heavy Oil Production: Validation/Uncertainty Quantification in the Field

Paper from the AFRC 2013 conference titled Evaluating the NOx Performance of a Steam Generator for Heavy Oil Production: Validation/Uncertainty Quantification in the Field by Michal Hradisky

Chevron operates approximately 150 steam generators for heavy oil production in Californiaʼs San Joaquin Valley. To meet increasingly stringent NOx regulations, these steam generators were retrofitted with the Fives North American GLE combustion system. In an effort to better understand the combustion environment where NOx is being formed in these systems, researchers at the Institute for Clean and Secure Energy (ICSE) at the University of Utah have teamed with personnel from Chevron U.S.A. and Fives North American Combustion to perform a formal validation/uncertainty quantification (V/UQ) study using high-performance computing Large Eddy Simulation (LES) tools developed at ICSE and field data collected on a Chevron steam generator fired with the GLE system in the San Joaquin Valley. In a formal V/UQ process, the quantity of interest (QOI) is first determined. For this study, the QOI is NOx emissions. Since the NOx formation pathway is a thermal mechanism based on the local temperature and N2/O2 concentrations, these variables, if available, are also of interest. The field data used in this analysis include NOx and O2 concentrations and temperature measurements at various radial and axial locations in the GLE-equipped steam generator. To capture the dynamic motion of the flow field in the steam generator, two simulation tools have been employed. The first, Star-CCM+, is a commercial software tool that has been developed to handle complex geometries in problems involving flow, heat transfer, and stress. The second, ARCHES, couples an LES model with radiation, combustion and NOx chemistry models. This study is specifically focused on the data collected in the near burner region where temperatures are sufficiently high to produce thermal NOx. To fully resolve the turbulent length and time scales that are critical to determining local NOx formation, the computational domain is more than 100 million cells and requires close to 6000 processors for one week to run a single case. Due to the size of this problem and limited computational resources, the parameter space explored by the V/UQ analysis is very limited. This paper will focus on the results of a skeletal V/UQ analysis that utilizes the expertise of team members to reduce the parameter space explored to a couple of variables that have the greatest impact on the QOI (NOx). Particular emphasis will be paid to the difficulties arising from and solutions for performing quantified uncertainty analyses in systems that involve sparse field data and computationally-intensive simulations.

Evaluating the NOx Performance of a Steam Generator for Heavy Oil Production: Validation/ Uncertainty Quantification in the Field Michal Hradisky, Jennifer P. Spinti*, Jeremy N. Thornock,, Philip J. Smith Institute for Clean and Secure Energy, University of Utah, Salt Lake City, UT Nicholas Brancaccio, Beverly K. Coleman, Stein Storslett Chevron Energy Technology Company USA, Inc., Richmond, CA, San Ramon, CA and Bakersfield, CA John Nowakowski, Tom Robertson Fives North American Combustion, Inc., Cleveland, OH *Corresponding author Abstract Chevron operates approximately 150 steam generators for heavy oil production in Californiaʼs San Joaquin Valley. To meet increasingly stringent NOx regulations, these steam generators were retrofitted with the Fives North American GLE combustion system. In an effort to better understand the combustion environment where NOx is being formed in these systems, researchers at the Institute for Clean and Secure Energy (ICSE) at the University of Utah have teamed with personnel from Chevron U.S.A. and Fives North American Combustion to perform a formal validation/uncertainty quantification (V/UQ) study using high-performance computing Large Eddy Simulation (LES) tools developed at ICSE and field data collected on a Chevron steam generator fired with the GLE system in the San Joaquin Valley. In a formal V/UQ process, the quantity of interest (QOI) is first determined. For this study, the QOI is NOx emissions. Since the NOx formation pathway is a thermal mechanism based on the local temperature and N2/O2 concentrations, these variables, if available, are also of interest. The field data used in this analysis include NOx and O2 concentrations and temperature measurements at various radial and axial locations in the GLE-equipped steam generator. To capture the dynamic motion of the flow field in the steam generator, two simulation tools have been employed. The first, Star-CCM+, is a commercial software tool that has been developed to handle complex geometries in problems involving flow, heat transfer, and stress. The second, ARCHES, couples an LES model with radiation, combustion and NOx chemistry models. This study is specifically focused on the data collected in the near burner region where temperatures are sufficiently high to produce thermal NOx. To fully resolve the turbulent length and time scales that are critical to determining local NOx formation, the computational domain is more than 100 million cells and requires close to 6000 processors for one week to run a single case. Due to the size of this problem and limited computational resources, the parameter space explored by the V/UQ analysis is very limited. This paper will focus on the results of a skeletal V/UQ analysis that utilizes the expertise of team members to reduce the parameter space explored to a couple of variables that have the greatest impact on the QOI (NOx). Particular emphasis will be paid to the difficulties arising from and solutions for performing quantified uncertainty analyses in systems that involve sparse field data and computationally-intensive simulations. Introduction In Californiaʼs San Joaquin Valley, steam generators used in heavy oil production operate in a very challenging regulatory environment. Current regulations require that oilfield steam generators operating on gaseous fuel not to exceed 15 parts per million by volume (ppmv) of NOx emitted. Future regulations call for NOx emissions of less than 5 ppmv [1]. This work is a partnership among the Institute for Clean and Secure Energy (ICSE) at the University of Utah, Chevron, which operates approximately 150 steam generators in the San Joaquin Valley, and Fives North American Combustion, the burner manufacturer. The objectives of the project are directed at NOx reduction with either no impact to or increased efficiency of steam generation for Enhanced Oil Recovery (EOR). This project leverages existing Department of Energy funding through the National Nuclear Security Administration under the Advanced Simulation and Computing program and the data-sharing agreement with Chevron, Fives, and the ICSE. Chevron, with significant investment in steam generators currently employed in the field, is interested in retrofit options that balance operational costs and generator performance and efficiency while meeting the requirements of future regulations. Simulation science, involving high performance computing tools tightly coupled with field-scale data, represents an emerging opportunity for examining the impacts of design changes on the outputs of greatest concern, in this case NOx. In the past, NOx emissions from large-scale combustors such as steam generators have been difficult to compute because NOx formation is a spatially- and temporally-varying phenomena that occurs in regions where the temperature and nitrogen and oxygen concentrations are sufficiently high. Additionally, low NOx burners typically have a high degree of geometric complexity that requires computational meshes with very fine resolution. To achieve our goal of producing a simulation tool that can be used to predict impacts of burner and operating design changes on NOx emissions from oilfield steam generators, we propose a loosely coupled approach that utilizes the computational strengths of a commercial Computational Fluid Dynamics (CFD) package, STAR CCM+, and of a highly-paralellized simulation tool, ARCHES, developed by Professors Philip Smith, Jeremy Thornock, and other ICSE researchers. In this paper, we describe this approach and present preliminary results comparing NOx data from an operating steam generator in the field with simulation results. Description of Steam Generator The radiant section of the steam generator has an internal diameter of approximately 3.175 meters or m (10.4 feet) and is approximately 12.2 m (40 feet) in length. Figure 1 shows a typical steam generator with the burner in the foreground. A schematic of the low-NOx GLE burner is shown in Figure 2. Premixed natural gas and air are injected into the primary reaction zone while additional natural gas is introduced into the steam generator through the low NOx injectors. A small quantity of natural gas is also injected through the pipe at the burner center (labeled "Radial gas inlet" in Figure 2). While the steam generator can be operated with flue gas recirculation (FGR), the tests described in this paper do not include any FGR experiments. Figure 1: Steam generator with the burner in the foreground. Figure 2: GLE burner used in steam generator tests. Experimental Dataset Experimental data were collected over the course of two days from the steam generator. Data available for the overall process include the ambient air flow rate and temperature and the fuel flow rate, temperature, and composition. Detailed data collection in the radiant section occurred at four longitudinal locations (0.46 m, 1.37 m, 2.29 m, and 5.94 m) with radial measurements taken every 7.6 cm (3 inches) from the wall to the center of the steam generator. At each location, extractive sampling was used to measure O2, CO, and NO, and NO2 concentrations as well as temperature and pressure. The extracted sample was analyzed with the testo 350 Emission Analyzer. The stated accuracy of the analyzer is ± 2 ppm for NO and ± 5 ppm for NO2. Other sources of sampling error were not analyzed. Infrared thermographic images were also taken of the entire steam generator exterior and of the internals of the radiant section. STAR-CCM+ The commercially-available simulation software Star-CCM+, created by CD-adapco, has been developed to handle complex geometries in problems involving flow (of fluids or solids), heat transfer, and stress. STAR-CCM+ is "unrivalled in its ability to tackle problems involving multi-physics and complex geometries" [2]. Additionally, scaling studies performed by ISCE researchers on Star-CCM+ have shown reasonable scalability up to 768 processes. Because of these capabilities, we determined that Star-CCM+ would be the most appropriate tool for performing a simulation of fluid flow through the low-NOx burner shown in Figure 2. ARCHES ARCHES, a Large Eddy Simulation code, solves coupled mass, momentum and energy conservation equations on a staggered, finite-volume mesh [3]. A dynamic, large eddy turbulence closure model for the momentum and species transport equations is used. The dynamic model accounts for sub-grid velocity and species fluctuations [4]. The gas phase chemistry uses a Rate-Controlled Constrained Equilibrium (RCCE) model that computes the rate of intermediate species that subsequently react instantaneously to equilibrium. The RCCE model is parameterized by two mixture fractions and by heat loss and is preprocessed in tabular form for dynamic table look-up during the course of the LES simulation. Local NOx concentrations are computed from a scalar transport equation using the rate of thermal NOx formation proposed by Bowman [5]. The energy balance includes the effect of radiative heat loss/gains in the IR spectra by solving the radiative intensity equation using a discrete-ordinance solver [6]. ARCHES has shown good scalability up to 262,000 processors, so it is the simulation tool of choice for fluid flow problems that are spatially and temporally varying over many orders of magnitude. Simulations-Scoping Runs The initial scoping runs were performed on local computing resources, which restrict the problem size to 3000 processors. Hence, it was necessary to reduce the size and geometric complexity of the problem. The following assumptions/decisions were made: • Premixed natural gas and air, assumed to be fully mixed, enters the computational domain through annular rings on the inlet face of the primary reaction zone • Velocity of the premixed natural gas/air has a flat profile with a component only in the flow direction • Natural gas is injected on the inlet face of the radiant section in nozzles consisting of a single computational cell (0.5 cm3) each to generate high velocities • Velocity components for each natural gas jet are computed based on information about the jet angle • Computational domain includes the entire primary reaction zone and part of the radiant section - 1 m x 2.12 m x 1 m (3.28 feet x 7.0 feet x 3.28 feet) with the longest dimension in the flow direction • Computational domain excludes the entire burner but includes all of the inlets (center natural gas inlet, annuli with premixed fuel and air, low NOx injectors) at its boundaries. These assumptions resulted in a computational domain with a resolution of 0.5 cm that ran on 2860 processors. The simulations were performed using ARCHES to resolve the requisite length and time scales. Approximately 120 hours of CPU time were required to obtain results sufficient for analysis. Figures 3 and 4 show volume-rendered images of one time slice near the end of the simulation. The volume-rendered image in Figure 3 is of the temperature field with the vorticity field seen on the two planes cutting through the middle of the figure. In Figure 4, vorticity is also shown in the cutting planes but NOx mass fraction is the volume-rendered image. In both images, the primary reaction zone is shown at the right. The low-NOx injectors can be seen as the relatively-cool red regions entering the domain above and below the main combustion plume in Figure 3. These regions are not visible in Figure 4. Figure 3: Volume-rendered temperature field in near-burner region of steam generator with cutting planes showing vorticity. Figure 4: Volume-rendered NOx mass fraction field in near-burner region of steam generator with cutting planes showing vorticity. Based on experimental observation of this burner in the field, the low NOx injectors in the simulation were not producing adequate entrainment rates. Because it was not possible to further reduce the mesh resolution to more fully resolve the burner inlets and/or to include the burner itself in the computational mesh, the decision was made to move to a loosely-coupled approach that utilized the relative strengths of STAR-CCM+ and of ARCHES. Additional concerns from scoping runs focused on the impact of uncertain boundary conditions on the recirculation zones in the near-burner region and the need to completely capture the critical zone for NOx formation on the computational mesh. To remove these sources of uncertainty, the computational mesh needed to be extended to the walls of the steam generator (1.6 m from centerline) and to at least 4.6 m in the flow direction. Simulations-Loosely-coupled approach In the loosely-coupled approach, elements of the burner geometry are fully resolved in STAR-CCM+ and the flows of both natural gas and air through those elements are computed. The output from this simulation is then filtered to match the mesh resolution of the ARCHES computational mesh. Finally, the filtered output is used as the inlet condition for the ARCHES simulation. Each step of this process is described in this section. I. Burner simulation using STAR-CCM+ There are two principle components of the GLE burner that require simulations of the flow through those components: the mixer tube from which the premixed fuel/air stream enters the primary reaction zone and the low NOx injector from which natural gas is injected into the radiant section downstream of the primary reaction zone. Simulation of a mixer tube required 30 million cells and employed the Wall-Adaptive Local-Eddy viscosity (WALE) LES model. The computational mesh of the mixer tube is in shown in Figure 5. Because the simulation included flow of gas through the fuel distributor and through all nozzles, the simulation time step was on the order of 0.00001 seconds. The simulation was run on 600 cores for four days to reach statistical steady state (SSS). After reaching SSS, we began collecting samples of all velocity components, the density distribution, the enthalpy, and the fuel/air mass ratio at the exit plane of the mixer tube for another three days on 600 cores. In total, this simulation required about 100,000 CPU hours to capture a few seconds of mixing occurring inside the mixer tube. The output was then time-averaged over the period of SSS. Figure 5: STAR-CCM+ computational mesh for mixer tube in GLE burner. For the low NOx injector, we also used the WALE LES model with a time step of the same order of magnitude as for the mixer tube simulation. This simulation required approximately three million cells (see Figure 6) and took two days of computer time on 600 cores to reach SSS and to collect time-averaged samples for all velocity components, density, enthalpy, and the fuel/air mass ratio. In total, this simulation required on the order of 30,000 CPU hours to obtain the low NOx injector profiles. As with the mixer tube, this data was filtered as described next. Figure 6: STAR-CCM+ computational mesh for low NOx injector in GLE burner. II. Filtering/handoff operation The next step was to apply a spatial filter to the time-averaged output from the STAR-CCM+ simulation because it was at a much finer resolution than could be achieved on the ARCHES mesh. The spatial filter was applied at a plane just downstream of the mixer tube or low NOx injector tip and involved averaging the fine-scale data over the filter width of the ARCHES computational mesh (7.6 mm). Figure 7 shows the time-averaged output from the mixer tube simulation, including the velocity components, the density (e.g. mass flux), the fuel/air mass ratio (e.g. mixture fraction and eta), and the enthalpy. Figure 8 displays the same set of data after filtering (e.g. coarsening) to the resolution of the ARCHES simulation. Figure 7: Time-averaged output variables from the STAR-CCM+ simulation of the mixer tube. Figure 8: Spatially-filtered output variables from the STAR-CCM+ simulation of the mixer tube to be used as input to ARCHES simulation. The same process was applied to the time-averaged output from the STAR-CCM+ simulation of the low NOx injector. The time-averaged output is shown in Figure 9 with spatially-filtered data displayed in Figure 10. While some fine scale information is lost during the filtering operations, the images in Figure 10 clearly capture the overall distribution of the various velocity components and of the other variables. Figure 9: Time-averaged output variables from the STAR-CCM+ simulation of the low NOx injector. Figure 10: Spatially-filtered output variables from the STAR-CCM+ simulation of the low NOx injector to be used as input to ARCHES simulation. III. ARCHES Simulation with filtered inputs To complete the application of the loosely-coupled approach to the simulation of a steam generator, the data shown in Figures 8 and 10 must be used as an inlet boundary condition for the ARCHES simulation of the primary reaction zone and radiant section. To accomplish this task, the mixer tube and low NOx injector inlets are defined on the boundary of the ARCHES computational mesh. Variable values in the area encompassed by the inlets are then defined by files containing those values at each cell location (e.g. the data shown in Figures 8 and 10). Given the requirements for the mesh size identified by the scoping runs, the new ARCHES computational domain includes the entire primary reaction zone and the first 4.6 m of the radiant section of the steam generator and contains greater than 100 million cells. The simulation was run on an out-of-state computer since it requires nearly 6000 processors. The plot of total kinetic energy from the simulation, as shown in Figure 11, reveals that SSS was reached after 0.5 seconds of simulation time, which required 96 hours of computer time. Additional computational time is required to generate enough data for time-averaging. Figure 11: Total kinetic energy of the ARCHES steam generator simulation as a function of simulation time. The impact of the new inlet boundary conditions was immediately apparent when visualizing the results from this simulation. For example, the mixer tube inlet condition is not perfectly premixed as was assumed in the scoping runs and the velocity profile in each cell has lateral components as well as the component in the flow direction. The mixing pattern in the primary reaction zone produces some regions of high temperature and high NOx formation rates as shown in Figure 12, but NOx formation is a very local phenomena that varies dramatically throughout the primary reaction zone. Figure 12: (left)Temperature field and (right) NOx formation rate in slice near the exit of the primary reaction zone after 0.788 seconds of simulation time. The effect of the low NOx injector boundary condition on the inlet plane of the radiant section was also apparent. These jets now entrain significant flow and exhibit a cohesive structure that penetrates several meters into the radiant section. Figure 13 juxtaposes the temperature field with the NOx mass fraction, providing a visual illustration of the momentum of the low NOx injectors, the interaction between the injector flow and the flow exiting the primary reaction zone, and the effect of the local temperature field on NOx production. Figure 13: (left)Temperature field and (right) NOx mass fraction field in slice through the middle of the steam generator after 0.788 seconds of simulation time. Figure 14 shows volume-rendered images of the NOx mass fraction (top) and the NOx production rate (bottom). These images indicate that the main regions of NOx production are the primary reaction zone and downstream of the low-NOx injectors. Figure 14: Volume-rendered images of (top) NOx mass fraction and (bottom) NOx production rate field in the steam generator after 0.876 seconds of simulation time. Validation/Uncertainty Quantification The quantification of simulation error is achieved by comparing the simulation output to inexact experimental observations, thus constraining simulation accuracy by the accuracy of the experimental measurements used for validation. The V/UQ methodology we use seeks to find the minimum and maximum error in a simulation prediction by propagating the error/uncertainty in the input parameter space (priors) to the simulation output or posterior subject to the constraints of the experimental validation and verification uncertainty [7]. However, due to lack of computational resources, we have been unable to move forward with a V/UQ analysis. A proposal has been submitted requesting additional resources. Comparisons of experimental data from the oilfield steam generator with the NOx, O2, and temperature output from ARCHES scoping simulations using the loosely-coupled approach will be presented at the conference. Conclusions High performance computing tools have been used to simulate the reacting flow environment in an oilfield steam generator. Of particular interest is the generation of NOx and the capability of these computing tools to predict the impact of design changes on NOx emissions. This first phase of the study involves a V/UQ analysis of the simulation output and experimental steam generator data. Using a loosely coupled approach, we have filtered the output of a finely-resolved simulation of the fluid dynamics in the burner to obtain an inlet condition to an ARCHES simulation of the steam generator on a computational mesh of nearly 100 million cells. The next steps are (1) to make adjustments to the ARCHES simulation to improve recirculation near the walls in the near burner region and (2) to check for consistency between simulation results from scoping runs and the available experimental data. Acknowledgments This research was sponsored by the National Nuclear Security Administration under the Accelerating Development of Retrofitable CO2 Capture Technologies through Predictivity program through DOE Cooperative Agreement DE-NA0000740. References 1. S. J. V. A. P. C. District. (2008). Rule 4306, boilers, steam generators, and process heaters-phase 3. http://www.valleyair.org/rules/1ruleslist.htm#reg1, October 2008. 2. STAR-CMM. (n.d.). Available at http://www.cd-adapco.com/products/star-ccm-plus. 3. J. Spinti, J. Thornock, E. Eddings, P. Smith, and A. Sarofim. (2008). Heat transfer to objects in pool fires, volume 20, chapter 3, pages 69-136. Wit Press, Southampton, UK. 4. S. B. Pope. (2000). Turbulent flows. Cambridge Press. 5. C. Bowman. (1975). Kinetics of pollutant formation and destruction on combustion. Progress in Energy and Combustion Science, 1:33-45. 6. G. Krishnamoorthy. (2005, December). Predicting radiative heat transfer in parallel computations of combustion. PhD thesis, University of Utah, Salt Lake City, UT. 7. M. J. Bayarri, J. O. Berger, D. Higdon, M. C. Kennedy, A. Kottas, R. Paulo, J. Sacks, J. A. Cafeo, J. Cavendish, C. H. Lin, and J. Tu. (2002). A framework for validation of computer models. Technical report.