1、OR-05-15-2 (RP-1191) CFD Models and Field-Measured Data from Large Enclosed Locomotive Facilities Liangcai Tan, PhD Member ASHRAE Amy Musser, PhD, PE Associate Member ASHRAE ABSTRACT This paper compares CFD models and field-measured data from four large enclosed locomotive facilities. For each facil
2、ity, CFD models were built using two computationalfluid dynamics programs. These included a commercially available software package using a revised Reynolds Averaged Navier- Stokes (RANS) k-epsilon model and apublic domain large eddy simulation model. This paper discusses the development of the mode
3、ls, compares their predictions with field measurements, and discusses the computing times required. The field data include temperature, NO, and CO concentration. Both CFD programs provided reasonable agreement with the field- measured data in three cases, and uncertainty in specibing the source stre
4、ngth for one case is discussed as the likely cause for lack of agreement in that case. Computing times for the RANS model are faster. The additional time required for the large eddy simulation model varies significantly and is dependent on the details of theproblem and their injuence on thegrid refi
5、ne- ment and number of time steps required. INTRODUCTION This work is part of an ASHRAE research project intended to develop design guidance for ventilation of enclosed locomotive facilities. Prior to that project, available design guidance had been developed primarily based on prac- tical experienc
6、e. With the relatively recent availability of computational fluid dynamics (CFD) modeling, there was a desire to veri and supplement this guidance with a more detailed parameteric study. One of the first steps in developing the parametric study was to veri the approach by comparing computational sol
7、utions with field-measured data from actual facilities. The field measurements were conducted under normal operating conditions in four shops located in the United States (Musser and Tan 2003). This paper compares the predictions of two computational fluid dynamics modeling tools with these data. Be
8、cause the data were collected for a few days at each site, specific operating events had to be selected for modeling. The cases were chosen to cover a variety of venti- lation systems and to allow modeling of both transient and steady-state operation. Use of data collected during occasional instrume
9、nt malfunctions or lapses in field notes was avoided. Each of the four cases was then modeled using two computational fluid dynamics approaches. First, each situa- tion was modeled using a commercially available computa- tional fluid dynamics software package designed to model room airflows. Using t
10、his approach, urbulence was modeled using a revised RANS k-epsilon model. Each simulation was then also modeled using a public domain large eddy simula- tion program that was developed for modeling building fires. Results of both simulations were then compared to the measured data. The approaches ar
11、e evaluated based on this data comparison, computing time, and grid requirements. The computational results are also used to both qualitatively and quantitatively evaluate each ventilation system more thor- oughly than is possible with a limited number of point measurements in the space. NUMERICAL M
12、ETHODS Commercial Software (RANS approach) The commercially available computational fluid dynam- ics software package is designed to predict air flow, heat trans- fer, and contaminant concentrations within rooms or Liangcai Tan is a research assistant professor in the Department of Mechanical Engine
13、ering, University of Nevada Las Vegas. Amy Musser is a principal at Vandemusser Design, LLC, Asheville, NC. 1026 02005 ASHRAE. buildings. This software package offers four options for modeling turbulence, all of which deal with turbulence by solving time averaged forms of the Navier-Stokes equations
14、. They differ in the number of equations that are solved and treatment ofnear-wall conditions. The standard RANS k-epsi- lon model calculates the turbulent viscosity for the fluid cells not immediately adjacent to solid surfaces as a function of two field variables: the kinetic energy of turbulence
15、(k) and its rate of dissipation (epsilon). These two field variables are deter- mined by the solution of two additional differential equations (Launder and Spalding 1974). The revised RANS k-epsilon model used here calculates the turbulent viscosity for the bulk fluid in exactly the same way as the
16、standard RANS k-epsilon model, but it allows the value to vary according to the log law of the wall for those cells close to solid surfaces. This software package has been widely used and exten- sively validated for many types of problems. Its capabilities include automatic grid generation, “on the
17、fly” grid refine- ment, switching between solution techniques during the solu- tion process, monitoring of field values during the solution process, and post-processing graphic capabilities. Public Domain Large Eddy Simulation Software The public domain large eddy simulation software is the NIST Fir
18、e Dynamics Simulator (FDS) (McGrattan et al. 2002). It has been under development for several years and is available to the public at no cost. It was created for modeling building fires but has also been validated for a variety of indoor air quality modeling scenarios (Musser et al. 2001). This code
19、 uses large eddy simulation (LES) to model turbulence. Rather than deal with turbulence through time averaging, LES allows the direct calculation of large scale turbulent motion. Smaller scale motion is simulated using the Smagorinsky sub-grid scale (SGS) model (Smagorinsky 1963), which requires onl
20、y one empirical coefficient. The solution is inherently both three-dimensional and time-dependent. To obtain a steady- state result, the simulation output must be time averaged over a sufficient period to cancel the effects of turbulent fluctua- tions. This can increase the time required for steady-
21、state problems but is ideal for solving transient problems. The public domain large eddy simulation software was specifically created to model building fires and has been vali- dated for situations with high temperatures and velocities, such as those that are experienced near an operating locomo- ti
22、ve. The user interface was designed for use by facility managers and designers to simulate fire scenarios. Text input files are required but are not difficult to generate. The program is accompanied by a graphic post-processor with capabilities exceeding those typically available in public domain pr
23、ograms. This model was developed to provide good results for fire scenarios with coarse grids and fast computing times. One interesting requirement of the fast solver is that the number of grid cells in each direction must be divisible by 2,3, and 5. By default, the grid cells are evenly spaced in e
24、ach direction. However, the user can speci grid transformations to better adapt the grid to the geometry of the project. Transformations can be based on a polynomial function or can be specified in a piecewise fashion that allows the user to speci the number of grid cells to be evenly spaced over va
25、rious portions of the domain. The need for a coarse grid model such as this one arises because much fire modeling is performed for large spaces, where many grid cells are needed even if they have coarse dimensions, on the order of 1 to 5 ft (0.3 to 1.5 m). However, the nature of the model does not e
26、xcuse the user from a grid refinement check, nor does it imply that perfect results will be obtained on very coarse grids. Very coarse grids cannot capture spatial variations as well, and this is most likely to manifest as “smearing” of contaminant profiles. Therefore, it is necessary to check succe
27、ssively finer grids to minimize this effect. BOUNDARY CONDITIONS Locomotives The locomotive itself is an important boundary condition, since it generates heat, airflow, and hot, contaminated exhaust flows. These factors can be significant drivers for the indoor air flow. Locomotives maintained in th
28、e shops come from two primary manufacturers. There are also some regional differ- ences in models, for example, due to the need for third rail capability in the northeastern US. With respect to CFD modeling, there are several impor- tant parameters: the geometry of the locomotive itself; its flow ra
29、te, temperature, and contaminant makeup of the locomotive exhaust; the location, flow rate, and temperature of the radiator fan outlets; and the location and flow rate of air inlets. These parameters are specific to the locomotive model and were obtained from the manufacturers for this effort. Becau
30、se some of this information is considered proprietary, it is not published here. Designers are encouraged to work directly with manufacturers to obtain this information. The diesel exhaust fumes are emitted from the locomotive stacks. The pollutants NO, NO, and CO are introduced as part of the airfl
31、ow from the stacks. The air flow rates of the stacks from the two primary manufacturers vary with throttle settings. The throttle settings we used for the modeling are high idle for both manufacturers locomotives, and the corre- sponding air flow rates of the stacks are 1700 cfm (0.802 m3/ s) and 13
32、00 cfm (0.613 m3/s). For other settings, the actual air flow rates of the stacks are available from the manufacturers. For CFD modeling, the two types of stacks are represented by planar sources (surfaces). The temperature, volumetric flux of airflow rate, and mass fractions of the species (contamin
33、ants) were then designated on those surfaces in both the RANS simulations and the LES simulations. Emissions data reports publish combined NO, values, rather than reporting each NO, component separately. In the model, NO, was modeled as a single contaminant with molec- ASHRAE Transactions: Symposia
34、1027 Figure I Section view of test site 1. dar weight estimated based on the relative quantities of NO, and NO from the field measurements. Weather Weather conditions were monitored during the field measurements because of their potential to influence indoor conditions. Outdoor air temperatures meas
35、ured at the sites at the time ofthe tests were 76.6“F (243C) at Test Site 1,69.1“F (20.6“C) at Test Site 2, 84.2“F (29.0“C) at Site 3, and 47.1“F (8.38“C) at Test Site 4. The effects ofwind were considered as boundary conditions to the model only in cases of high wind conditions or large openings cl
36、ose to the area of interest, when it would have significant influence. Otherwise, openings to the outdoors were considered passive openings for makeup air. The large volume of makeup air needed when the exhaust fans were operating in these facilities supports this assumption. Buildings During all fo
37、ur of the field measurements simulated, indoor and outdoor temperatures were very close to one another. Walls, floors, and ceilings were therefore considered adiabatic surfaces. Open walls and doors were considered passive openings. Platforms and locomotives in these shops were modeled as solid, ine
38、rt obstacles to the flow. Other large objects that may influence airflow, such as enclosed offices, were also modeled. Constant flow rate exhausts were modeled as they appear at each site. FIELD DATA Temperature, velocity, and NO, NO, and CO concentra- tions were all measured and compared with the C
39、FD models; however, only a combined NO, concentration field is shown in the figures for this paper due to space limitations. Also, it was necessary to combine the NO and NO, measurements as an NO, value because the published emissions data used to deter- mine the simulation boundary conditions are g
40、iven as NO,. NO, is currently the “critical contaminant“ used for design of these facilities in most locations, meaning that it is typically present at the highest concentration relative to published limits, such as those by theUS Occupational Safety andHealth Administration (OSHA 2001). Therefore,
41、NO, is both rele- vant and illustrative of the ventilation systems performance. The reader is reminded, however, that regulations change frequently and vary around the world. In some cases, more stringent requirements may be in place for other compounds such as diesel particulate matter. These compo
42、unds may behave differently than the gases analyzed here. Vertical contaminant and temperature arrays were set up inside each shop to obtain profiles that could be used for vali- dation of a computational model. In most cases the array consisted of three sets of sensors that were either hung from th
43、e ceiling structure or mounted on a telescoping pole. These sensors were capable of collecting time series data for up to eight hours at a time. Nitric oxide (NO) and nitrogen dioxide (NO,) were measured using electrochemical sensors. Manu- facturer information for the nitric oxide sensor states a r
44、ange of 0-250 ppm, a resolution of 1 ppm, no cross-reactivity to 300 ppm CO, and motive OE XI Locomotive on 0.5 O O 5 10 15 20 25 Time (min) Figure 12 Comparison of LES and RANS simulation results (NOx concentration) for test site 4 at height = 9.7 3 (3. O m). location but may actually be slightly h
45、igher than the data indi- cated. Figure 12 compares the results of the RANS model with the LES simulation. Contaminant concentrations predicted by the RANS model were lower than those predicted by LES and nearly zero. Upon frther inspection this appeared to be related to localized differences in the
46、 airflow profile in the vicinity of the monitoring points. Computing Time. The RANS simulation shown used 86,l O0 grid cells and 12.45 hours of CPU time on a personal computer with a Pentium III 450 MHz processor and 256 MB of RAM. For comparison, the FDS simulation consumed 2.27 hours of CPU time f
47、or 72,000 grid cells (case i), 7.05 hours for 165,888 grid cells (case 2), and 19.22 hours for 348,920 grid cells (case 3). These computing times were obtained on a personal computer with an Intel Pentium IV processor, 1400 MHz, 256 MB of RAM. DISCUSSION Results Comparison In comparing results from
48、the four shops, several points of interest are raised. First, the magnitude ofpredicted concen- trations is reasonably close to those that were measured in all but one location. In that location, the predicted concentration profiles, as shown in plots of nondimensional concentration, were similar to
49、 the measured profile. This suggests that the emissions data used in the model were too low by a factor of approximately 4. However, the same or similar model loco- motive was simulated in two of the other shops using these emissions data with considerably better agreement. Locomo- tives at that shop were most likely operated at a higher throttle setting, or there may be some other difference of which we were not aware. This result underscores the importance of esti- mating throttle positions and related exhaust emissions correctly when building a model. Another issue that arose
