Académique Documents
Professionnel Documents
Culture Documents
Aerospace Engineering
Jury
President:
Supervisor:
Co-Supervisor:
Member:
October 2009
Acknowledgements
The author would like to thank his Supervisors, Doc. Eng. Nelson Marques for the calm he had
throughout this work, for all his patience and for his kindness in sharing his knowledge with the author
and Professor Jos Carlos Pereira for his availability and support. Without them this work would not
have been possible.
He would also like to thank Eng. Joo Viegas and Eng. Hildebrando Cruz from LNEC, Eng. Cristiano
Neves from Efaflu and Eng. Darko Cuculic and Eng. Joaquim Almeida from LMSA for providing him
with the much needed experimental results and detailed setups and for their availability to answer his
questions. Without their contribution this work would not have been possible. He would also like to
express his gratitude to everyone at blueCape, especially to Eng. Jorge Azevedo, for their invaluable
help.
Finally, a special thank you to Carla Fernandes, Duarte Albuquerque, Lus Domingues, Rodrigo
Taveira, my parents and my brother for their support.
Abstract
Jet fans are nowadays a common option to ventilate enclosed car parks. Those must comply with
strict regulations on ventilation requirements and besides onsite experimental trial of the system itself,
computational fluid mechanics is the only tool available to assess the effectiveness of the system either
during or after the project phase. In this work, a CFD model of a fully functional jet fan was developed.
Despite identifying mesh and domain size influences in the final results, it was the turbulence model
choice that had the biggest impact on results. Jet fans behave like wall jets and k turbulence models
fail to correctly predict the high lateral spread observed, since those models do not correctly simulate
the streamwise vorticity production rates. By using instead a Reynolds Stress Turbulence model it was
possible to fully approach the experimental results provided by LNEC.
This jet fan model was then applied to assess the ventilation effectiveness of a zone of LoureShopping
car park. The use of RST versus k produced different overall flow velocities and a reduction of 10%
on the average age of air for RST. Nevertheless the underlying flow pattern remained the same.
Finally, a software based optimization strategy was tested and it was possible to increase the ventilation efficiency by 10% in terms of the mean age of air, while retaining the ability to maintain pollutants
within the legal boundaries.
In conclusion, it was possible to use a verified CFD calculation process to optimize a real-life jet fan
ventilation system.
Keywords: car park, ventilation, axial flow fans, optimization, CO dispersion.
ii
Resumo
Os ventiladores de impulso so actualmente uma escolha frequente para ventilar parques de estacionamento subterrneos. Estes ltimos esto sujeitos a regras de segurana apertadas e excluindo o teste da
instalao no local, a dinmica de fluidos computacional a nica ferramenta disponvel para avaliar o
sistema de ventilao durante ou aps a fase de projecto.
Neste trabalho, desenvolveu-se um modelo de ventilador de impulso em CFD. Apesar de ter sido
possvel identificar influncias do domnio e da malha nos resultados finais, foi a escolha do modelo de
turbulncia que revelou ter a maior influncia. Os ventiladores de impulso comportam-se como jactos
parietais e modelos da famlia k no conseguem replicar a taxa de abertura lateral destes jactos,
uma vez que estes modelos no simulam correctamente a produo de vorticidade responsvel por este
fenmeno. Utilizando em alternativa um modelo de Reynolds Stress Turbulence, foi possvel replicar os
resultados experimentais fornecidos pelo LNEC.
O modelo desenvolvido foi aplicado ao estudo do sistema de ventilao de parte do estacionamento do
LoureShopping. A utilizao do modelo RST, comparando com k , produziu um campo de velocidades
diferente e permitiu reduzir em 10% o valor da idade mdia do ar das simulaes. Contudo os modelos
produziram padres de escoamento semelhantes.
Finalmente testou-se uma estratgia de optimizao automtica e foi possvel aumentar a eficincia
da ventilao em termos de idade mdia do ar em 10%.
Concluindo, foi possvel utilizar um processo validado de CFD para optimizar a colocao de ventiladores de impulso num parque de estacionamento real.
Palavras Chave: parques de estacionamento, ventilao, ventiladores impulso, CO, optimizao.
iii
Contents
Acknowledgements
Abstract
ii
Resumo
iii
List of Figures
vi
List of Tables
ix
1 Introduction
1.1
Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2
Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.1
Portuguese Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3
Jet Fans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.1
Working Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4
1.5
5
8
1.6
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
2 Methodologies
11
2.1
Chapter Layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
2.2
CFD Methodologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
2.3
11
13
2.3.1
13
2.3.2
15
2.3.3
Continuity Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
2.3.4
SIMPLE Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
2.4
Turbulence Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
2.4.1
17
2.4.2
Wall treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
2.5
Passive Scalar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
23
2.6
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
2.6.1
Space Fillers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
2.6.2
Genetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
2.5.1
iv
2.7
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
28
3.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
3.2
Conceptual Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
3.3
Solid modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
3.3.1
28
3.4
Mesh Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
32
3.4.1
33
3.5
33
3.6
36
38
3.7
LNEC Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
3.8
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
3.6.1
4 LoureShopping case
50
4.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
50
4.2
Solid modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
52
4.3
4.4
4.5
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
4.3.1
53
4.3.2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
4.4.1
Optimization Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
4.4.2
Optimization Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
60
4.4.3
Ventilation Improvements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
5 Conclusion
66
5.1
Results Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
5.2
Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
Bibliography
68
List of Figures
1.1
1.2
1.3
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1
Terminology used. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
2.2
22
2.3
24
2.4
1000 points generated with a Random (left) and a Sobol (right) sequence. Sobol sequence
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
The point A splits this two dimensional objectives space in two zones: the set of the
dominated points is represented by the shaded area. A dominates D, while B, C and C
are non-dominated. [3]
2.6
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
In the left plot 25,000 Random solutions are shown for the function T1 . In the right plot a
MOGA-II run is shown to have reached the real convex Pareto-frontier for the test function
T1 . [3] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
3.1
29
3.2
29
3.3
. . . . . . . . . . . . . . . . . . . .
29
3.4
Model A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
3.5
Model B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
3.6
Model C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
3.7
31
3.8
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
32
3.9
34
3.10 Velocity profiles along a vertical line 10 m away from the jet fan.
. . . . . . . . . . . . .
3.11 Section cut, showing the different mesh base sizes away from the jet fan.
. . . . . . . . .
34
35
3.12 Geometry built to assess boundary conditions, detailing size increases. (top view + half
domain) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
35
3.14 Velocity profiles along a vertical line 4 m away from the jet fan outlet.
. . . . . . . . . .
36
3.15 Velocity profiles along a vertical line 10 m away from the jet fan.
. . . . . . . . . . . . .
36
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37
39
39
39
39
vi
40
40
40
3.24 Jet spread for other Turbulence models tested (left Standard k , right k ). . . . . .
40
41
. . . . . . . . . . . . . . . . . . . . . . . .
41
41
42
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
43
44
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
3.34 Spreading rate at the jet plane for k (left) and RST (right).
. . . . . . . . . . . . . .
45
46
46
46
46
46
46
47
47
47
3.44 Velocity magnitude 3 m Away from jet axis at height of 196 cm. . . . . . . . . . . . . . .
47
47
47
. . . . . . . . . . . . . . . . . . . . . .
48
4.1
. . . . . . . . . . . . . . . . . . . . . . . . . . .
50
4.2
51
4.3
51
4.4
. . . . . . . . . . . . . .
51
4.5
52
4.6
4.7
52
53
4.8
Mesh used. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
4.9
54
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
54
55
55
55
56
56
56
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
vii
58
58
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
. . . . . . . . . . . . . . .
59
. . . . . . . . . . . . . . . . . . . . . . . . . . .
60
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
60
4.26 Area averaged age of air at 1.5 m after several optimization steps. . . . . . . . . . . . . .
62
. . . . . . . . . . . . . . .
62
. . . . . . . . . . . . . . . . . .
63
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
64
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
64
64
64
65
65
4.35 Streamlines for the optimized system, originating at the jet fans. . . . . . . . . . . . . . .
65
viii
List of Tables
1.1
3.1
30
3.2
Mesh characteristics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
32
3.3
33
3.4
42
3.5
45
4.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
52
4.2
53
4.3
59
4.4
4.5
61
63
ix
Chapter 1
Introduction
1.1
Objective
This work aims to assess and optimize the ventilation effectiveness of an underground car park using
computational fluid dynamics (CFD) methodologies. In addition, dispersion and removal effectiveness of
hazardous pollutants will be discussed based on the concept of mean age of air. The first step will be to
validate a CFD model of a jet fan against experimental data: initially, domain, mesh, jet spreading rates
and turbulence models will be checked and then a full scale model will be created and compared against
experimental data from LNEC. This model is to be applied to one fire zone of LoureShopping underground
car park in order to understand how the ventilation performs both under normal operation or under
emergency conditions. Finally, by using modeFrontier, a commercially available optimization software,
the position of jet fans inside LoureShopping underground car park is to be optimized to both reduce the
car park mean age of air and operational costs, while increasing safety and comfort.
1.2
Context
Land in most urban areas is considered an expensive resource, thus it is used mainly for housing
and entertainment purposes. Yet, due to the fact that urban populations have continuously increased
over the last decades, the number of automobiles that everyday flow in or around a city has increased,
consequently creating a great demand for parking places. Due to the available land value, the solution
has been to either build vertical car parks or to reserve the underground floors of buildings for car parks.
Those floors do not receive direct sunlight, hence making them not appreciated for housing or offices. This
solution does look like a win-win situation, since undesired floor space is taken up, while creating parking
places. However, their enclosed nature means they have little interaction with the outside environment
if not properly ventilated.
Cars are one major source of air pollution, as carbon monoxide (CO) is produced whenever there is
an incomplete combustion of hydrocarbons-based fuels. CO is one of the most hazardous pollutants that
needs to be regulated in urban atmospheres due to its toxicity and implications for human health [10].
Inside a car park, people (car owners, workers, etc) and cars have to circulate together, thus there is both
CO production and a need to ensure a breathable environment. This can only be achieved if a proper
ventilation system is running. In addition to this, it has been found that there is a maximum limit of CO
to which people can be exposed [9] whithout suffering poisoning. For instance the average concentration
of CO in the environment ranges form 0.05 PPM to 0.12 PPM [5], whereas in a car park, regulations
allow values up to 200 PPM instantly. Figure 1.2 describes in detail the health effects of CO inhalation.
Fire is also a major concern inside car parks [36], since due to their enclosed nature, a car fire spreads
more easily to nearby cars while abundant smoke production tends to render evacuation routes very
difficult to follow. If a ventilation system properly installed is working inside the car park, it should
be able to extract smoke to a degree that evacuation is possible while reducing the overall ambient
temperature to prevent other cars from igniting.
To create an efficient ventilation system, there is a set of design guidelines and best practices on
mechanical and displacement ventilation available in literature such as ASHRAE technical notes and a
number of regulations which the engineer in charge of the project should adopt. Because, only an efficient
ventilation system can ensure daily operation safety, by actively removing pollutants and by being capable
of removing smoke, in case a fire spreads. In what concerns Portugal, Dirio da Repblica[1] states the
minimal safety conditions the car park must comply to. Yet, those guides and regulations can only provide
some guidance, therefore leaving many aspects dependent upon the engineer expertise and knowledge.
One way to enhance the project quality is to rely on both the advances made in CFD science and in
available computational power. This combination makes it now possible to simulate a car park ventilation
system using previously validated codes. By modeling flow patterns inside a car park it is possible to
adjust it in order to ensure, for instance, that the mean age of air is below a given value or to see
if recirculation of air will occur. In addition, if one is able to combine those codes with some piece of
optimization software, it will be possible to evaluate if changing jet fans location, jet fans power or jet fans
number, one can achieve an optimal solution which increases thermal and environmental comfort while
reducing project costs and operation costs. In this way, developing a correctly validated methodology
would be an invaluable resource for an engineer that needs to design an efficient ventilation system.
1.2.1
Portuguese Law
When building an underground car park it is necessary to obey a certain amount of regulations that
strongly depend on the country where the parking is to be built. Nevertheless, those laws tend to be
supported by best practices in engineering published by organizations like ASHRAE and health criteria
by the World Health Organization which publishes several standards to ensure air quality. Regarding the
Portuguese case, those regulations are published in the Regulamento Geral de Segurana Contra Incndios
em Edifcios which describes safety measures based on building type, occupation levels and total area.
In what concerns car parks, this regulation for instance describes in detail what are the requirements
in terms of evacuation routes, ventilation equipment, ventilation capacity, air quality or maintenance
scheduling. Those regulations were published as Decreto-Lei no 220/2008 de 12 de Novembro and Portaria
no 1532/2008 de 29 de Dezembro, while some other aspects critical for a correct design of a car park
in what concerns its ventilation and energy efficiency can be found in Despacho no 2074/2009 de 15 de
Janeiro, Portaria no 64/2009 de 22 de Janeiro and Decreto-Lei no 78/2006, 79/2006 e 80/2006 de 4 de
Abril.
Relevant for the scope of this work is that CO concentration inside a car park can not exceed 200
PPM instantly and 50 PPM in a 8 hour average. For instance, for other types of buildings this limit is
only 10.9 PPM. In addition, CO monitors have to be placed at 1.5 m height, with at least one for each
400 m2 . Regarding the ventilation, it must be able to remove air at a rate of 300 m3 /h per vehicle space
in case CO concentration reaches 50 PPM and 600 m3 /h per vehicle space if a concentration of 100 PPM
of CO is reached. The ventilation system must also be able to deliver a flow rate of 600 m3 /h per vehicle
space in case of fire 1 . All those are net values, after losses in ducts and fans have been accounted for.
Additionally, the maximum open area of an underground car park should be of 3200 m2 and a person
using the car park must walk no more than 40 m to reach an emergency exit or a maximum of 25 m in
case there is more than one emergency exit nearby. Furthermore, emergency lighting should be enough
to provide 10 lux at 1 m from the floor.
1.3
Jet Fans
One common way of ventilating an enclosed room is with axial jet fans - figure 1.1. As opposed
to conventional ventilation concepts based on transverse ventilation and ducted systems, the jet fan
technology is derived from the longitudinal ventilation systems found in most road tunnels. In a tunnel,
a stream of air is injected into the tunnel by a series of free blowing axial fans that induce the air movement.
For a car park this system requires, however, two components: the jet fan itself plus a number of air inlets
and outlets. Air inlets and outlets are placed in strategic locations for both construction and maintenance
purposes as well as to increase the ventilation efficiency. Those systems are responsible for air insufflation
1 Outdated regulations stated also that air inlets/outlets had to have at least a area of 0.06 m2 per vehicle space and
that if more than 30% diesel cars were expected to use the car park, the owner had to set limits for other pollutants besides
CO.
and extraction from the car park, which is set in motion by the jet fans.
The big advantage of these systems is that they do not require that all the air passes through the
fans in order to set the air in movement, thus reducing losses through ducts and the overall energy
consumption. Since they greatly accelerate the air that passes through them, they are able to set all
the air inside the car park in motion. By installing an adequate number of jet fans in an enclosed
space, a constant air-movement can be created, ensuring that pollutant concentration all over the space
is maintained within regulations. Such kind of system has others advantages:
No ducting required, thus reducing losses in pressure while increasing efficiency.
Saving height, from a construction point of view, since ducts do not need to be factored in.
Since every jet fan can be operated separately, they can be more energy efficient because, in ventilation mode, only those in areas where CO concentration reaches critical levels will be put to work
instead of the whole system.
Some jet fans can operate in reverse mode, so that they can be for instance placed in access ramps
between levels and work as air extractors/insuflators for both levels.
Jet fans can also operate at various speeds, thus allowing for energy efficiency as only the required
power will be used to maintain air within regulations.
1.3.1
Working Principle
Consider for instance a control volume similar to a tunnel (figure 1.3) with an entrance area S, pressure
P0 , inlet velocity V0 and an exit area S, outlet pressure P2 and velocity S2 , where an axial jet fan of
velocity Vs is placed. Through manipulation of the momentum balance equation, one will reach equation
(1.1) [2].
S (P2 P0 ) = Q (VS V0 )
(1.1)
In equation (1.1), the first term represents the force needed to displace air between inlet and outlet
sections and the second one accounts for the jet fan impulse when placed in a flow field with velocity V0 .
Then, the force required to displace a given volume of air equals the jet fan impulse when placed in that
same volume of air.
1.4
Previous sections have already outlined some of the challenges of using jet fans to ventilate an enclosed
car park: on the one hand there is a need to adhere to strict regulations while on the other hand there
is a need to obey to a number of constrains (lack of space, reduced budget, etc) while sometimes lacking
scientific tools to ensure the quality of the proposed jet fan placement. This work aims at offering a new
insight into jet fan ventilation modeling for enclosed car parks, and thus it is important to present an
account on past attempts at solving those problems.
Besides on place experimental validation as performed by [33], the most common tool used to assess
ventilation efficiency has been computational fluids dynamics (CFD) related methodologies, like the ones
used by [31] or [14], either as self-developed codes ([14]) or using commercially available softwares ([13]
or [15]). It is on this last option that this works wants to develop further.
From an historical standpoint regarding car parks, CFD has addressed challenges related with fire
modeling/propagation inside them [36], since they were more critical and evolved then to understand
overall ventilation efficiency and pollutant dispersion ([25] or [35]). Moreover, in recent years, it has
become apparent that CFD can play an important and useful role in fire safety problems [23]. Jet fans
for car park ventilation appeared only at later stages and were based on previous work on ventilation
systems for road tunnels ([29]). Within the scope of this work, it is also important to consider other
situations where ventilation efficiency for other kinds of enclosed spaces is evaluated ([11] or [28]).
CFD is a branch from fluid mechanics where numerical methods and algorithms are used to solve and
analyze problems with fluid flows. The basis of almost all CFD problems are the Navier-Stokes equations.
Those equations are numerically solved with computers that perform the millions of calculations required
to simulate liquids and gases, subject to boundary conditions.
When using CFD methodologies to approach a problem, the same basic procedure is followed, raising
always the same kind or problems/questions. During preprocessing the geometry is defined, the volume
occupied by the fluid is divided into discrete cells (the mesh), the physical modeling is defined and
boundary conditions are defined. The simulation is then started and the equations are solved iteratively
with the specified algorithms. Finally a post processor is used to analyze the results. The two most
common option are either to compare the results against experimental test data ([33], [29]) or to validate
the options taken by applying them to well established and reviewed test cases [24].
From an historical perspective, using CFD to study the ventilation efficiency of underground car parks
can be dated, based on available bibliographic resources, back to the early nineties. Chow ([13]) in 1995,
firstly tried to study ventilation strategies for underground car parks. Very limited by the computation
power available, he used the PHOENICS package with a 3000 cell grid to simulate a car park and was
able to already see some agreement between experimental measurements and numerical results. His
work focused mainly on the assessment of three different options for ventilation and their impact on CO
dispersion, discussing the results against a number of surveyed car parks.
Also in 1995, Gan ([18]) investigated the hypothesis of using CFD to evaluate room air distribution
systems. Choosing k for turbulence model, a SIMPLE algorithm as solver and integrating the
diffusion and convection terms with the QUICK differencing scheme, he applied a grid size of 28080
cells to two different rooms. Despite having reached some interesting results in what concerns thermal
comfort prediction, he outlines that results represent solely one configuration and there could be better
configurations and that an optimized system for cooling tends to poorly perform when heating.
Chow ([14]) in 1998 evaluated several safety systems for underground car parks, under normal operation or in case of fire. He then applied CFD to a particular car park with dimensions of 25x25x5 m,
using a mesh with 10000 cells and a solver he had developed. In this work he was able to develop some
5
correlations between air speed, turbulence intensity and air quality (both in terms of people dissatisfied
and mean age of air). Those correlations were then applied to other car parks and, allowing him to
conclude that those car parks would perform poorly in the event of an emergency.
In 2000, Xue and Ho ([35]) used CFD methodologies to asses both heat and carbon monoxide released
from moving cars. Their work focused on developing numerical strategies to assess ventilation performance, since guidelines were not uniform. Their biggest challenge was to create a model that would
simulate random movements of cars inside car parks so that boundaries and source terms could be well
placed in the computational model. They simulated a car park with 27000 control volumes using the
Standard k turbulence model and the SIMPLER algorithm. They concluded that a random average
model was able to produce significant agreement with experimental results.
A few years later (2003), Papakonstantinou ([25]) applied CFD to evaluate the ventilation efficiency of
an underground car park. He used PHOENICS commercial package to simulate a small car park, divided
into 9880 control volumes, together with a k turbulence model and a SIMPLER solver algorithm. By
taking experimental measurements in the car park under operating conditions, he was able to compare
both experimental and computational results, having identified a good agreement between them, and
once more validating CFD as a tool for ventilation assessment.
Abanto [6] modeled the air flow in a computer room in 2004 using FLUENT software. The advances
in computational power are significant since it already uses an hexahedral mesh with 2 million cells for
an area of 50 m2 . He still continues to use k turbulence model, coupled with a segregated solver
approach. The procedure was checked against experimental results from work with Minibat from NIASL.
The very fine mesh used allowed the investigation to show that boundary conditions and inclusion of
real-world geometries, like diffuser inlets, have a strong influence on overall fluid flow behavior.
Lin ([22]) in 2004 conducted some studies on pollutant dispersion within a public transport interchange. The CFX commercial package was used with RN G k turbulence model and a mesh of 20000
control volumes.
More recently (2006), in a similar work by Chow ([12]), a public transport interchange was simulated
with FLUENT, having approximately 1 million hexahedral control volumes and using Standard k as
turbulence model. An interesting conclusion was the effect of having high speed flows, independently of
the rate of air changes: besides being more efficient for smoke extraction, they promote a better mixing
of the air, reducing pollutants.
Other authors like Zhang ([36] - 2006) also researched smoke movement and fire spread in underground
car parks by using CFD codes. In this specific case, the Fire Dynamic Simulator was used to evaluate a
75x16x3 m car park, using Large Eddy Simulation. He found that numerical results compared well with
experimental ones and that the ventilation scheme used played a major role on the smoke movement
and fire spread - it could even accelerate some stages of fire development, which are critical for rescue
operations.
Going away from car parks, Chow [15] (2001) conduced some studies using CFD on airflows induced
by mechanical ventilation and air conditioning systems, verified against some documented test cases. In
that work, Phoenics software and FLAIR preprocessor were used to simulate several rooms with areas
from 15 m2 to 300 m2 with a constant height of 3 m, split into meshes of 800 cells. Several aspects, like
CO dispersion, temperature profiles and inlet placement were tested, producing reasonable results. Also
in 2001, Papakonstantinou ([26]) used CFD to evaluate CO2 dispersion in an auditorium. He simulated
an auditorium with 20000 control volumes with PHOENICS. Once more he concluded on the applicability
of CFD to address ventilation problems.
CFD has also been used to assess the ventilation efficiency of other kind of enclosed spaces, namely
sports arenas ([28]), in which the ventilation efficiency of a sports hall used during the 2004 summer
Olympics is evaluated in terms of mean number of people dissatisfied. Another kind of building whose
ventilation system has been evaluated by CFD (FLUENT) have been food plants (Chanteloup [11] 2009). This work is also very interesting since it also uses the concept of mean age of air as a tool to
understand how the ventilation performs.
Despite numerous works on car parks, very few have addressed the possibility of using jet fans to
ventilate them, focusing instead on studying the more conventional displacement ventilation schemes. To
find research that includes jet fans there is a need to investigate into road tunnels.
Li [20] in 2003, reviewed several ventilation options for tunnel safety. He split a 100 m long road
tunnel into 127000 volumes and simulated a fire in the middle. Despite raising several concerns regarding
convergence and mesh size, he outlines that all the available systems have their advantages and disadvantages. However, the longitudinal ventilation system, performed by jet fans, is the most simple, having
only some disadvantages in case the tunnel is bidirectional.
A more deep research on the longitudinal ventilation system of a road tunnel was conducted by Vega
([29]) in 2006. Using experimental data from a full scale, 853 m long, road tunnel ventilation system assay,
she produced some interesting conclusions on the possibility of using CFD as a project tool. The road
tunnel (Memorial Tunnel in the USA) was deactivated and a full longitudinal ventilation system, based
on jet fans, was installed along with an array of sensors and cameras. After this installment, a number
of controlled fires was set up, and the ventilation efficiency evaluated accordingly. Using the software
FLUENT, Vega wanted to establish a test case for which CFD codes could be validated against. The
tunnel was divided into some 100000 tetrahedral cells with some localized refinements and turbulence was
modeled with Standard k with wall functions. Jet fans were modeled as simple velocity inlets/outlets.
The comparison of temperature results between both experimental and numerical results showed an
overall good agreement, revealing the accuracy and usefulness of CFD modeling in order to predict the
effectiveness of longitudinal ventilation systems. Finally it is noted that reaction time to the fire might
be critical.
More recently, Maele ([23]) in 2008 investigated the possibility of using RANS and LES field simulations to predict the critical ventilation velocity in longitudinally ventilated horizontal tunnels, since
according to the authors, there should be a shift towards performance-based ventilation design. The idea
is to blow the smoke in one single direction by forced ventilation. As such, the propagation of smoke is
avoided in the other direction and a smoke-free path for evacuation and firefighting purposes is created.
The RANS simulations are performed with FLUENT. The Realizable k model is used in combination
with a buoyancy model based on the generalized gradient diffusion hypothesis. The LES calculations are
performed with the Fire Dynamics Simulator. Comparing to experimental results, LES tended to overpredict whereas RANS underpredicted results, with the addition that LES results can strongly depend
on the mesh, particularly if the mesh is too coarse.
An exception to this rule is the work by Viegas ([33], [32]), who in 2001 and 2002 investigated the
idea of using jet fans to ventilate enclosed car parks. Forum Algarve car park with 90000 m2 was
evaluated along with jet fans from Novenco. The car park was split into several regions, and each one
was divided into 72000 control volumes and simulated for 2400 iterations with a SIMPLER algorithm
with k turbulence model. Longitudinal speeds were validated against test data from Novenco, showing
significant agreement in results. In addition, a fire was simulated in one of the areas to assess how the
flow would behave. Experimental trials in the car park with the jet fans operating produced results
that were in significant agreement with those that were computationally obtained. In 2006, Viegas ([31])
investigated the possibility of using jet fans to ventilate an enclosed car park under normal operation (to
remove pollutants) or in the event of a fire. Using similar procedures to his previous works, he draws
some relevant conclusions for the scope of this work: that the efficiency is highly dependent on the jet fan
placement and air inlets/outlets location; jet fans are able to limit pollutant dispersion; jet fan distance
should be chosen in order for the jets to overlap for the maximum distance, and that entrainment velocity
is high enough to avoid pollutant dispersion. Finally, the need for further research is outlined.
Table 1.1 presents an abridged account of some of those past works. All those past works outlined
challenges, which in most cases are common to any CFD project: mesh base size choice, turbulence
models used, solvers used, codes and software employed, etc. In addition the majority of the previously
mentioned authors referred the importance of developing a reliable tool that would enhance the quality
of the installed ventilation systems. To quote [23], current standards and guidelines are prescriptive and
often the background is not known by practicing engineers.[...] In other words, there should be a shift
towards performance-based approaches. Therefore, physical and numerical tools should be available for
designers and authorities. This work aims at following those directions: by using a commercial software
package it will be verified the possibility of using CFD to assess ventilation efficiency. Specifically, the
use of jet fans to ventilate an enclosed car park. Jet fans advantages have already been presented in an
section 1.3, and from this section it is possible to see that longitudinal ventilation systems are becoming
increasingly popular, thus making them a perfect candidate for efficiency assessment. In addition, jet fans
will be applied to an already built car park, bringing this work scope to the real engineering challenges
of designing a ventilation system for a closed space.
Furthermore, inspired in the possibility of using novel software developed by ESTECO, the idea of
applying optimization algorithms to improve ventilation efficiency will be tried. By coupling optimization
algorithms with CFD, it will be possible to try and build an entirely new ventilation scheme based on
air quality parameters and not only the engineer expertise/awareness. In this way it will be possible to
decide on which performance the system should match and then choose the ventilation option that would
meet those goals.
Finally, as it was already outlined, working with CFD methodologies tends to raise similar challenges,
yet nowadays, the turbulence modeling is the main stake in airflow modeling[19]. Thus, special attention
will be payed to turbulence modeling. A recent work by Launder [16] gave some insight into the challenges
and limitations of turbulence models to correctly predict flows coming from wall jets (from which jet fans
can be considered a case). Knowing this, and having some experimental velocities measurements from jet
fans trials performed by LNEC, this work will try to investigate the shortcomings of the most common
turbulence models and the applicability of models recommended by Launder.
All this together will try to present a new approach to ventilation design using jet fans while increasing
efficiency and reducing pollutant dispersion and fire spread.
1.5
Thesys Layout
So far, this thesis context has been established and a brief account on past work by several authors
was presented. The following chapter will provide a description of equations and models used in this
work, whereas in chapter 3 those models usage will be discussed and compared in order to establish a
model and procedure that can be applied to a real world car park, presented in detail in chapter 4. In
chapter 4, the ventilation scheme installed will be discussed and a new option for the placement of jet
fans, drawn from optimization algorithms, will be introduced. For both options, relevant results will be
presented, compared and discussed as to conclude on the validity of using an optimization strategy as an
auxiliary project tool. Chapter 5 will close the work by summarizing relevant conclusions.
2007
2007
Chang [10]
2003
Papakonstantinou [25]
Vega [29]
2002
Viegas [32]
2006
2000
Xue [35]
Zhang [36]
1995
Gan [18]
2006
1995
Chow [13]
Chow [12]
Year
First Author
Tunel
Memorial
Test data
Experimental data
Grid independece
Experimental data
Experimental data
Experimental data
Validation
CFD-RC
Fluent 6.0
FDS
Fluent 6.1
Phoenics
Own code
Own code
Own code
PHOENICS
Software
Assessment of different
strategies for car park ventilation
Evaluation of room air
distribution systems
Modeling of heat and CO
emitted cars in a Car Park
Isothermal smoke dispersion in an enclosed car
park
Air quality inside car
parks
Assessement of different
ventilation strategies at a
transport interchange
Fire simulation inside car
parks
Ventilation assessment of
a tunel with jet fans during a fire
Simulates CO distribution
in a room and adjacent
balcony
Description
Wall functions
Wall functions
Wall functions
Wall function
k-
Wall functions
Wall Treatment
k Standard
k
Turbulence
Specif.
-
0.15 m
0.35 m
0.2 m
0.4 m
(9880 cells)
0.8 m
0.6 m
0.3 m
0.5 m
Base Size
1.6
Summary
There are both laws and guidelines that apply to the project of underground car parks. Nonetheless,
the outcome and efficiency of a ventilation system still greatly depends on the expertise of the responsible
engineer. Through CFD this dependency on the knowledge and expertise could be reduced, while improving the environmental conditions inside car parks. CFD simulations of car parks have been done for
the past ten years, however due to limitations on hardware available those studies were either of limited
scope (too simple) or only considered small regions of the car park. In addition, given the huge damages
fire may produce in enclosed car parks, most studies have focused on predicting smoke spread. This work,
will deal instead with assessing a ventilation scheme based on jet fans and on its ability to ventilate the
enclosed car park, based on the expected levels of carbon monoxide.The main goals of this work are then:
To create a computational model of a jet fan, discussing model, meshing and turbulence options.
To validate the jet fan model against experimental results from previous work by LNEC.
To apply the jet fan model to a full scale car park, in order to assess the ventilation efficiency.
To develop an optimization strategy that enhances the overall ventilation performance.
10
Chapter 2
Methodologies
2.1
Chapter Layout
In this chapter every physical model will be described as well as how it is used in this work. There
will be a particular emphasis on fluid flow equations and how they are implemented in StarCCM+ rather
than presenting only the underlying physics. Starting from the fluid flow equations, every step needed to
achieve a full computational model will be detailed. The last sections of this chapter will address other
aspects important for this work, while the final section will focus on optimization models.
2.2
CFD Methodologies
This section briefly describes the modeling approach employed in this work, including relevant flow
equations, turbulence models and procedures.
Except for the optimization part, all the work concerning model validations and car park simulations
was done with StarCCM+. StarCCM+ is a commercially available CFD package created by CD-Adapco
which can simulate fluid flows, heat and mass transfer and chemical reactions in generic applications.
This software does all the pre-processing, including mesh generation, processing and post-processing and
only the geometric models were created outside of it using both StarDesign (from CD-Adapco also) and
Autodesk Inventor.
Since this section will only cover essential modeling aspects of this work, further information can be
obtained from [8], [4] or [34].
2.2.1
This work, much like all CFD, is based on the three fundamental governing equations - continuity
(Mass conservation), momentum (Newtons second law) and energy (Energy conservation).
This section will briefly address those equations and the complementary models (i.e. turbulence)
required to fully understand this work.
The equation for conservation of mass, or continuity equation, can be written as (2.1) 1 :
~ ) = Sm
+ (U
t
1 In
~ = (U1 , U2 , U3 )
this work velocity is defined as U
11
(2.1)
(2.1) is valid for both compressible and incompressible flows. The source term (Sm ) is the mass added
to the continuous phase from another phase and/or any user-defined sources.
Then, the equation regarding the transport equation for momentum conservation or Navier-Stokes
equation can be derived from applying Newtons second law to fluid motion, where momentum is the
transported quantity2 (2.2).
~
U
~ U
~
+U
t
!
= p + T + f
(2.2)
Here, p is the pressure, T is the deviatoric stress tensor and f represents body forces (forces per unit
volume). However this equation is still incomplete and one must make hypotheses on the form of T,
that is, there is a need to have a constitutive law for the stress tensor which can be obtained for specific
fluid families; additionally, if the flow is assumed compressible an equation of state will be required,
which will likely further require a conservation of energy formulation. Now, since this work primary and
only fluid will be air which is a Newtonian fluid 3 , the stress tensor is obtained through three additional
assumptions: that the stress tensor is a linear function of the strain rates, the fluid is isotropic and for a
fluid at rest, T must be zero (so that hydrostatic pressure results). All this produces equation (2.3).
Tij =
Ui
Uj
+
xj
xi
~
+ ij U
(2.3)
ij is the Kronecker delta, is the viscosity and is the second coefficient of viscosity (related to bulk
viscosity)4 . In addition, this work will deal with air under constant conditions, implying that (p, T )
and (T ) are constant, while the velocities involved will be within the incompressible range which from
the continuity equation implies that Ui /xi = 0. With all this in mind, it is now possible to write the
Navier-Stokes equations for Newtonian, incompressible and isotermic fluids as (2.4).
~
U
~ U
~
+U
t
!
~ +f
= p + 2 U
(2.4)
Since in StarCCM+ the Segregated Flow approach was used, flow equations (one for each component
of velocity, and one for pressure) will be solved in a segregated, or uncoupled, manner. The linkage
between the momentum and continuity equations is achieved with a predictor-corrector approach. The
complete formulation can be described as using a collocated variable arrangement and a Rhie-and-Chowtype pressure-velocity coupling
[4] combined with a SIMPLE algorithm. This model has its roots in
constant-density flows [4], thus being appropriate for the case being modeled in this study.
Equations (2.1) and (2.4) are then integrated over a given control volume (V ) and the Gauss divergence
theorem is applied resulting in those equations written in an integral form as:
d
dt
I
dV +
~ da = 0
U
(2.5)
2 In
f aces
f ace
f aces
f ace
neighbours
12
d
dt
Z
V
~ dV +
U
~ U
~ da =
U
I
pId da +
Z
T da +
(f )dV
(2.6)
The terms on the left-hand side of (2.6) are the transient term and the convective flux. On the right~ is the
hand side are the pressure gradient term, the viscous flux and the body force terms. T = 2 U
viscous stress tensor from equation (2.3) simplified to fit this work conditions. The body force terms can
account for other terms like the effects of system rotation, buoyancy, porous media or user-defined body
forces. Those equation shall be further manipulated so that they can be used in finite-volume schemes
as will be seen in the next section.
2.3
2.3.1
Numerical Discretization
Finite Volume Discretization
The finite volume method is a method for representing and evaluating partial differential equations,
like (2.1) or (2.2), as algebraic equations and it is based on the control volume formulation of analytical
fluid dynamics. The first step when applying this discretization is to divide the domain into a number
of control volumes where the field of variable values is kept at the centroid of each individual control
volume. Afterwards the differential form of the equations is integrated over each control volume, resulting
in those equations written in an integral form like equation (2.6), shown in the previous section. Variations
between cell centroids are then calculated through interpolation profiles. The most attractive feature of
the control-volume formulation is that the resulting solution would imply that the integral conservation
of quantities such as mass, momentum and energy is exactly satisfied over any group of control volumes
and of course, over the whole calculation domain. This characteristic exists for any number of grid points
[...], thus even the coarse-grid solution exhibits exact integral balances [27]. In this regard, (2.7) which
is a generic discrete form of a transport equation
(2.7)
Diffusion Term
Since this work will deal mostly with steady state solutions, the transient term will not be further
addressed. The source term is approximated solely by the product of the source term evaluated at
the cell centroid with the cell volume. This is the simplest formulation consistent with a second-order
discretization ([4]) which was used for the remainder terms. Regarding the convective term, a second
order upwind scheme was used. This scheme computes the convective flux as:
m
f f,P , for m
f 0
~ a)f = (m)
(U
f =
m
f f,N , for m
f <0
(2.8)
In equation (2.8) f,N and f,P are linearly interpolated from the cell values on either side of the face
as (2.9) or (2.10) and P and N relate to notation from figure 2.1. In those equations both ()r,N and
()r,P are the limited reconstruction gradients in cells N and P, respectively7 .
6 (2.6)
7 More
13
(2.9)
(2.10)
Regarding boundaries, using the second order upwind scheme implies that the flux at boundaries is
obtained from equation (2.11) where f,N is once again interpolated from the cell value using the limited
reconstruction gradients in cell N and f is the face value as set in the boundary condition.
m
f f,N , for m
f 0
(m)
f =
m
f f ,
for m
f <0
(2.11)
Finally the diffusion term is discretized for a general control volume in the form of equation (2.12).
I
da =
A
f f a = Df
|{z}
(p Id a)f +
(T a)
(2.12)
In this equation, A denotes the surface area of the control volume and a denotes the area of a face for
the control volume. In StarCCM+, in order to obtain an accurate second-order expression for an interior
face gradient that implicitly involves the cell values P and N , the decomposition from equation 2.13 is
used, where once more nomenclature is form figure 2.1.
= (N P )~
+ + ( dP~N )
(2.13)
Where, if ~rP and ~rN are position vectors of the centroids of cells P and N respectively, then dP N =
~rP ~rN . In addition
~=
adP N
h
i
Df = f f a = f (N P )
~ a + a dP N
~ a
14
(2.14)
In this equation, and f are suitable face averages8 . The second and third terms in (2.14) represent
the secondary gradient(or cross-diffusion) contribution. They are essential for maintaining accuracy on
non-orthogonal meshes [4]. At boundaries a modified version of (2.14) is used:
h
i
Df = f f a = f (f P )
~ a + P a P dP N
~ a
2.3.2
(2.15)
For reference purposes, equation 2.7 is here presented for the case of momentum transport and writes
as equation 2.169 . The discrete equation for each velocity component may be expressed implicitly as a
linear system.
i
X
X
Xh
d
~ U
~ a =
U
(pId a) +
Ta
(V )0 +
dt
f
f
(2.16)
Regarding this equation, to evaluate the stress tensor T (viscous term), the velocity gradient tensor
~ f ) must be written in terms of the cell velocities for purposes of linearization as equation
at the face (U
2.17, recalling that
~=
a
.
adP N
~ f (U
~f
~ f = U
~
U
~ + U
dP N )
~
2.3.3
(2.17)
Continuity Equation
Regarding the continuity equation (2.1) it is rewritten in a discrete manner as equation 2.18.
X
f
m f =
X
0
(m f + m f ) = 0
(2.18)
The uncorrected face mass flow rate m f is computed after the discrete momentum equations have been
0
solved. The mass flow correction m is required to satisfy continuity. For interior faces and incompressible
f
flows, the uncorrected mass flow rate may be written in terms of the cell variables as equation (2.19).
UP + UN
mf = f a
Gf f
2
(2.19)
solved, Gf is the grid flux10 and f is the Rhie-and-Chow-type dissipation at the face. For cells on
boundaries with a specified velocity, the value of m f is calculated from the known velocity through
0
equation (2.20). Those boundaries use a Neumann condition for the pressure correction (pf = p0 ) and
the mass flux corrections are zero.
m f = f (a Uf Gf )
8 Regarding
(2.20)
15
If those cells have instead a pressure boundary (either a stagnation inlet or a pressure outlet) the
pressure corrections will not be zero. The uncorrected boundary mass flux is given by equation (2.21)
where Uf is the boundary velocity and f is once more the dissipation.
m f = f (a Uf Gf ) f
2.3.4
(2.21)
SIMPLE Algorithm
As previously stated in 2.2.1, this model will make use of a SIMPLE algorithm whose basic steps in
the solution update are as follows:
1. Set the boundary conditions.
2. Compute the gradients of velocity and pressure.
3. Solve the discretized momentum equation to compute the intermediate velocity field.
4. Compute the uncorrected mass fluxes at faces.
5. Solve the pressure correction equation to produce cell values of the pressure correction.
6. Update the pressure field: pk+1 = pk +urf p0 where urf is the under-relaxation factor for pressure.
7. Update the boundary pressure corrections pb0 .
0
avP is the vector of central coefficients for the discretized linear system representing the velocity
equation and V is the cell volume.
10. Update density due to pressure changes
2.4
11
Turbulence Modelling
Turbulent flows are characterized by fluctuating velocity fields. These fluctuations mix transported
quantities such as momentum, energy, and species who might be traveling with the flow, and cause the
transported quantities to fluctuate as well. Since these fluctuations can be of small scale and high frequency, they are too computationally expensive to simulate directly in practical engineering calculations.
Instead, the instantaneous governing equations can be time-averaged, ensemble-averaged, or otherwise
manipulated to remove the small scales, resulting in a modified set of equations that are computationally
less expensive to solve [4]. However, the modified equations contain additional unknown variables, and
turbulence models are needed to determine these variables in terms of known quantities.
In this work, as will be seen later, there was a need to investigate several different turbulence models in
order to obtain a good agreement between theory and experiments. Yet, it is important to stress out that
all the models considered throughout this work and that will be described in the following subsections
are natively available in StarCCM+.
11 This
last step only applies if compressible gases are being considered, which will not be the case in this work. It is only
presented here for completeness.
16
2.4.1
Direct simulation for high Reynolds number turbulent flows in complex geometries with large domains is still not doable [8]. Reynolds-averaging (or ensemble-averaging) can be employed to render the
Navier-Stokes equations tractable so that the small-scale turbulent fluctuations do not have to be directly
simulated. This method introduces additional terms in the governing equations that need to be modeled
in order to achieve a closure for the unknowns.
The Reynolds-Averaged Navier-Stokes (RANS) equations govern the transport of the averaged flow
quantities, with the whole range of the scales of turbulence being modeled. The RANS-based modeling
approach therefore greatly reduces the required computational effort and resources, and is widely adopted
for practical engineering applications as seen in section 1.4.
The basic tool required for the derivation of the RANS equations from the instantaneous NavierStokes equations is the Reynolds decomposition. Reynolds decomposition refers to separation of the flow
~ ) into the mean (ensemble-averaged or time-averaged) component (U
)12 and the
variable (like velocity U
fluctuating component (U 0 ). Thus,
~ (U1 , U2 , U3 , t) = U
(U1 , U2 , U3 ) + U 0 (U1 , U2 , U3 , t)
U
By taking into account that
~
U
i
xj
i
U
xj ,
(2.22)
U
i
~ iU
~j = U
i U
j +U 0 U 0 .
= Ui = 0 from the continuity equation, while U
xj
xj
Uj
= fi
+
Ui Uj
xj
xi
xj
xj
(2.23)
The left hand side of (2.23) represents the change in mean momentum of fluid element due to the
unsteadiness in the mean flow and the convection by the mean flow. This change is balanced by the mean
body
force,
the isotropic stress due to the mean pressure field, the viscous stresses, and apparent stress
Ui0 Uj0 due to the fluctuating velocity field, generally referred to as Reynolds stresses.
However, now the Navier-Stokes equation have turned into an undetermined system, since there is an
additional unknown (Ui0 Uj0 ) and no additional equations. It must be noted that this term is a nonlinear
one, inserted into an averaged equation. This term represent six individual stress components that must
be dealt with, relating them to the mean motion so that this system becomes determined. So, to close
the RANS equations, Reynolds stresses must be modeled and there are several different ways to do that.
In this work, turbulence models based on the eddy viscosity concept and in Reynolds Stress Transport
models were used.
Linear eddy viscosity
The first turbulence model was proposed by Boussinesq; he defined that Ui0 Uj0 = t xUji , where
t is a turbulent viscosity. Under this approximation, xUji Ui0 Uj0 = ef xUji where the effects of the
turbulent flow are treated as an increase in fluid viscosity, given by ef .
Nevertheless, this procedure only shifted the difficulty in calculating Ui0 Uj0 to a new difficult in
calculating ef [8]. To calculate ef , Prandtl suggested a relation based on the kinetic theory of gases,
12 Here,
= limT
U
1
T
R t0 +T
t0
U dt.
17
where is related with both the average molecular velocity and the mean free path, given by t = ut lm
in which both ut and lm are velocity and length scales characteristics of the turbulent flow [8].
Bradshaw further developed this concept and stated that Ui0 Uj0 k, being k = Ui0 Ui0 /2 the turbulent
kinetic energy and solved a transport equation for k. The proportionality between k and Ui0 Uj0 happens
because since the large eddies are the main contributors for the normal Reynolds stresses (the trace from
the Reynolds tensor is Ui0 Uj0 = 2k), they should also be the main contributors to shear Reynolds stresses
[8]. This model still needed to use a length scale, yet some models used in this work were based on this
concept. Next section will present those models in detail.
Standard k Launder suggested a closure model based on two equations, and to overcome some
limitations of a model based on a mixing length, he calculates the transport of a velocity scale and
instead of using a length scale, he uses a time scale that multiplied by velocity produces a result with
units of length [8]. When using the Standard k turbulence model, the velocity scale and time scale
are determined by the following equations:
u=
T =
(2.24)
(2.25)
Therefore, the turbulent viscosity is modelled via (2.26), where C is a model constant with the usual
value of 0.09.
t = C
k2
(2.26)
Equations (2.27) and (2.28) express the transport equations for turbulent kinetic energy (k) and
dissipation rate () respectively. More information regarding this model can be obtained from [4].
(k) +
(kUi ) =
t
xi
xj
() +
(Ui ) =
t
xi
xj
+
t
+
k
k
+ Pk + Pb YM + Sk
xj
2
+ C1 (Pk + C3 Pb ) C2 + S
xj
k
k
(2.27)
(2.28)
(2.29)
t T
Prt xi
(2.30)
Pk = Ui0 Uj0
The effect of buoyancy13 is introduced by
Pb = gi
where P rt is the turbulent Prandtl number for energy and gi is the component of the gravitational
vector in the ith direction. For the Standard k model, the default value of P rt is 0.85.
13 In
this work buoyancy will not be considered, but it is presented here for completeness reasons.
18
(2.31)
p
(k) +
(kUj ) =
t
xj
vj
() +
(Uj ) =
t
xj
xj
t
+
k
k
+ Pk + Pb YM + Sk
xj
t
2
+ C1 C3 Pb + S
+
+ C1 S C2
xj
k
k +
(2.32)
(2.33)
In (2.32) and (2.33), Pk is the generation of turbulence kinetic energy due to the mean velocity
gradients as was defined by (2.29). Pb is the generation of turbulence kinetic energy due to buoyancy and
defined in (2.30). Additional aspects are equated trough, S is is the modulus of the mean rate-of-strain
tensor:
h
i
C1 = max 0.43, +5
,
= S k ,
S=
2Sij Sij
k2
(2.34)
In (2.34) C is no longer constant as with the standard k model, but is instead defined by (2.35):
C =
A0 + As kU
Regarding (2.35)
q
ij
ij
U Sij Sij +
ij = ij 2ijk k
ij = ij ijk k
Where Sij refers to the main strain tensor and the remaining coefficients are
19
(2.35)
1
3
cos
A0 = 4.04, As =
( 6W ), W =
6 cos
p
S = Sij Sij , Sij =
1
2
Uj
xi
Ui
xj
a1 k
max(a1 , SF2 )
(2.36)
k
k
k
+ Uj
= Pk k +
( + k T )
t
xj
xj
xj
(2.37)
1 k
2
2
+ Uj
= S +
( + T )
+ 2(1 F1 )2
t
xj
xj
xj
xi xi
(2.38)
. Such an approach recognizes that the Reynolds stress is really a functional of the velocity; that is,
the stress at a point depends on the velocity everywhere and for all past times, not just at the point in
question and at a particular instant in time [8].
In StarCCM+ the available option is to to use the Reynolds stress transport (RST) model, also known
as second-moment closure models, the transport equations are solved for each component of the Reynolds
stress tensor. This results in these models naturally accounting for effects such as anisotropy due to strong
swirling motion, streamline curvature, rapid changes in strain rate and secondary flows in ducts since it
captures the directional effects of the Reynolds stress fields.
The RST model, however, carries a significant computational overhead. Seven additional equations
must be solved in three dimensions (as opposed to the two equations of a k model): the Reynolds
stress tensor is symmetric, so that only six of the nine components are unique and in addition to the six
RST equations, a model equation is also needed for the isotropic turbulent dissipation rate ()
15
. Apart
from the additional memory and computational time required for these equations to be solved, there is
also likely to be a penalty in the total number of iterations required to obtain a converged solution due
to the numerical stiffness of the RST equations [4].
14 When using this approach, however, due to the non-linear character of the convectiv term, higher level correlations will
appear. There is a need to stop this process at some point, using lower order correlations [8].
15 This is the same equation as the one used in the Standard k model.
20
RST model involves calculation of the individual Reynolds stresses, Ui0 Uj0 , using differential transport
equations. The individual Reynolds stresses are then used to obtain closure of the Reynolds-averaged
momentum equation. The exact transport equations for the transport of the Reynolds stresses, Ui0 Uj0 ,
may be written as (2.39).
i
0 0
h 0 0 0
0 0
0
0
0
0
Ui Uj +
Uk Ui Uj =
Ui Uj Uk + p kj Ui + ik Uj +
Ui Uj
t
xk
xk
xk
xk
Uj
Ui
Ui0 Uk0
+ Uj0 Uk0
gi Uj0 + gj Ui0 +
xk
xk
0
Uj0
Ui
U 0 Uj0
0 0
0
+p
+
2 i
2k Uj0 Um
ikm + Ui Um jkm + Suser
xj
xi
xk xk
(2.39)
or simplifying:
(2.40)
2.4.2
Wall treatment
Turbulent flows are significantly affected by the presence of walls. Obviously, the mean velocity field
is affected through the no-slip condition that has to be satisfied at the wall. However, turbulence also
changes by the presence of the wall in non-trivial ways. Very close to the wall, viscous damping reduces
the tangential velocity fluctuations, while kinematic blocking reduces the normal fluctuations. Toward
the outer part of the near-wall region, however, the turbulence is rapidly augmented by the production
of turbulent kinetic energy due to the large gradients in mean velocity.
The near-wall modeling significantly impacts the fidelity of numerical solutions, inasmuch as walls are
the main source of mean vorticity and turbulence. After all, it is in the near-wall region that the solution
variables have large gradients, and the momentum and other scalar transports occur most vigorously.
Therefore, accurate representation of the flow in the near-wall region determines successful predictions
of wall-bounded turbulent flows [4].
Since this work will deal with high-Reynolds-number flows, the wall function approach was used as
it saves computational resources in comparison to the near wall model approach, because the viscosity21
affected near-wall region, in which the solution variables change most rapidly, does not need to be resolved.
The wall function approach is said to be economical, robust and reasonably accurate. For a majority of
reviewed papers (e.g. [26][12][18]), it is also a practical engineering option.
Under wall functions, velocity at a given distance from the wall is proportional to that same distance
as can be seen in figure 2.2. The wall functions used in this work differ only in their treatment in the
buffer region; the viscous-sublayer and log-layer behaviors are identical. For all those wall functions,
y+ =
yu
u+ =
u
u
(2.41)
where y is the normal distance from the wall to the wall-cell centroid; u is the component of wall-cell
velocity parallel to the wall; is the kinematic viscosity; finally, u is a reference velocity that in actual
practice the is derived from a turbulence quantity specific to the particular turbulence model [4].
It is expected that the flow around the jets will have high-Reynolds number, however, since the car
parks have big areas, the flow around the cars will have a low Reynolds number. This requires the use of
an all y + wall treatment which is an hybrid wall treatment, that attempts to emulate the high y + wall
treatment for coarse meshes and the low y + wall treatment for fine meshes. It is also formulated with
the desirable characteristic of producing reasonable answers for meshes of intermediate resolution (that
is, when the wall-cell centroid falls within the buffer region of the boundary layer[4]).
Eddy viscosity models
Regarding this family of models, the wall function does the following:
For a detailed description of the all y + wall treatment formulation, please refer to [4].
2.5
Passive Scalar
Previous sections have presented the core aspects for CFD modeling used in this work. In the following
sections other aspects that will be used to draw relevant conclusions are explained and clarified. The first
of those aspects is the concept of Passive Scalar. In StarCCM+ a passive scalar is modeled by solving
the transport equation (2.42). This feature will be used for instance to calculate the mean age of air.
I
dV +
~ da =
U
I
[(
A
t
+ ) da +
Z
S dV
(2.42)
In equation (2.42) is the molecular Schmidt number and t is the turbulent Schmidt number. The
molecular Schmidt number is a material property, while the turbulent Schmidt number is assumed to have
a value of 0.9, consistent with the turbulent Prandtl number used for energy [4]. S is a user specified
source term. Note that is assumed to be positive-definite. Since this is a standard transport equation,
it will be discretized in the same way as presented in section 2.3.1.
2.5.1
The local mean age of air (LMA), a concept developed by Sandberg in 1983 [17], is one of the most
important parameters when the removal of contaminants from a room or the ventilation efficiency in a
closed space is considered. At an arbitrary point P of the room LMA is defined as the average time
needed for air to reach the point P since it enters the room. A longer age of air suggests poorer outside
air delivery compared to a shortage of air for the particular location [21]. The "youngest" air is at the
air supply inlets whereas the "oldest" may be at stagnant zones or in the exhaust air.
In StarCCM+, equation (2.42) can be used to implement the calculation of LMA distribution in the
code so that the age of air distribution in an enclosed room can be viewed and the ventilation effectiveness
assessed. Equation (2.42) is a general transport equation for a given variable , which in this case will be
the mean age of air. Since a steady state approach is going to be followed, the first term of it will not be
considered, in addition the source term will be defined as equal to the fluid density (). This approach
uses a passive scalar approach because it considers the local mean age of air a passive quantity that does
not affect the airflow patterns, which is then obtained from the values of fluid velocity and viscosity in
the control volume. This approach is computationally light because it only requires one equation to be
solved.
Regarding boundary conditions, = 0 at the supply inlets where new air enters the domain and
d
dxi
= 0 at the exhaust and walls. The solution of equation (2.42) approaches then the local mean age of
23
and steady in average, may be considered as "quasi-deterministic" and do not constitute a limitation
to the application of the model.
2.6
Optimization
This work aims at developing a new approach for projecting ventilations systems for underground car
parks. In current engineering practice, it is the skill and expertise of the leading engineer that define the
primary placement for the jet fans. This work is then validated at later stages using a CFD approach,
but only whenever required by national regulations. This work wants to create a solution where CFD
is incorporated all the way since the beginning of the project. Having explored in previous section the
formal aspects of CFD modeling, in this section the optimization aspects will be addressed. To help
in performing all the optimization steps, the software modeFrontier from ESTECO was used. Further
information on it can be obtained from [3].
In a basic optimization algorithm, like the one depicted in figure 2.3, after having defined a set of
variables to be optimized, a number of candidate solutions is generated through a random process. Those
solutions are then tested and new ones are generated by modifications to those initial candidate solutions.
This process is iterated until the optimization goals have been met.
Define
Variables
to be Optimized +
Constrains
Start
Verify if
candidate
fits other requirements
Generate
random
populations
Simulate
populations
Generate
new
candidate
solutions
through
genetic
algorithms
evaluate
candidate
solutions
no
Have optimization
goals been
reached?
yes
stop
24
2.6.1
Space Fillers
Genetic algorithms are implemented in a computer simulation in which a population of abstract representations (called chromosomes or the genotype of the genome) of candidate solutions (called individuals,
creatures, or phenotypes) to an optimization
The first step when using a genetic algorithm is to generate an initial population of candidate solutions.
Those candidate solutions represent the initial Design of Experiment table and must respect the problem
constrains. In this work, initial populations were obtained in one of two ways: a purely random value
generator for the variables and a SOBOL approach. The first spreads points uniformly in the design
space and is based on the mathematical theory of random number generation, seeded by current time[3].
SOBOL is a deterministic algorithm that mimics the behavior of the random sequence since it is used to
uniformly sample the design space, however the clustering effects of random sampling are reduced, as can
be seen on figure 2.4. Mathematically speaking, SOBOL generates a low-discrepancy sequence since the
number of points in the sequence falling into an arbitrary set B is close to proportional to the measure of
B, as would happen on average (but not for particular samples) in the case of a uniform distribution. The
points in this sequence are maximally avoiding of each other, so the initial population fills in a uniform
manner the design space.
Figure 2.4: 1000 points generated with a Random (left) and a Sobol (right) sequence. Sobol sequence
fills more uniformly the Design space. [3])
2.6.2
Genetic Algorithms
Optimization of the jet fans placement was then done by using MOGA - II, a multi objective genetic
algorithm implemented in modeFrontier. A genetic algorithm is a search technique used in computing to
find exact or approximate solutions to optimization and search problems. Genetic algorithms are categorized as global search heuristics. Genetic algorithms are a particular class of evolutionary algorithms that
use techniques inspired by evolutionary biology such as inheritance, mutation, selection, and crossover.
Genetic algorithms are implemented as simulation in which a population of candidate solutions to an
optimization problem evolves toward better solutions. They are also a good alternative when there is no
information a priori regarding the solution space.
Since genetic algorithms for multi-objective optimization can find difficulties in converging to the true
Pareto frontier and can get stuck in a local Pareto front, in this work a more efficient algorithm (MOGA
- II) was used[3]. It uses a form of smart multi search elitism. This new elitism operator is able to
preserve some excellent solutions without bringing to premature convergence into local optimal fronts.
For simplicity, MOGA-II requires only very few user-provided parameters, several other parameters are
internally settled in order to provide robustness and efficiency to the optimizer. The algorithm attempts
a total number of evaluations that is equal to the number of points in the Design of Experiment table
25
(the initial population) multiplied by the number of generations. This algorithm was demonstrated in
[3] to reach a Pareto optimal in very timely manner
16
that it only does discard solutions that are fully dominated by newer ones
17
2.5, it is presented a point A in a two-objectives (f1 and f2) optimization problem. The point A defines
two zones: the shaded one (i.e. the left-bottom quadrant) represents the set of the dominated points,
while the complementary area (i.e. the whole of the other three quadrants) represents the set of the
non-dominated points. If A is a point of the previous generation and the actual generation contains the
point B, then the new position is a very favorable one: not only B is non-dominated by A, but even B
dominates A. This kind of evolution is always desirable, and this transition has certainly to be preserved
[3]. If the evolutions brings A to C (or C), the new point is however a non-dominated one: by keeping
non-dominated solutions, it favors the spread of points along the Pareto frontier.
Figure 2.5: The point A splits this two dimensional objectives space in two zones: the set of the dominated
points is represented by the shaded area. A dominates D, while B, C and C are non-dominated. [3]
26
f2 (f1 , g) = g(x)(1
p
f1 /g)
(2.43)
Figure 2.6: In the left plot 25,000 Random solutions are shown for the function T1 . In the right plot a
MOGA-II run is shown to have reached the real convex Pareto-frontier for the test function T1 . [3]
2.7
Summary
In this chapter the most important theoretical aspects, critical to this work, have been covered. The
core of this work will deal with solving the Reynolds Averaged Navier-Stokes equations using either a
Reynolds Stress Transport or a Realizable k turbulence model. Having established the base model, it
will be applied to a car park ventilation model that will be analyzed and optimized. The optimization step
will take advantage of genetic algorithms. Finally, the ventilation effectiveness will be verified through
the mean age of air concept that adds information to the overall model. This verification is targeted
at estimating how effectively the ventilation system removes pollutants like CO and at calculating the
average time that air spends inside the car park.
27
Chapter 3
Introduction
This chapter presents the CFD procedure that was followed to construct and validate a jet fan model,
based on test data from LNEC. This model will, in addition, be checked for grid and domain-size independence. Most geometrical models presented in this section were built with STAR-Design which is developed
by CD-adapco, except for one where, due to its complexity, Autodesk Inventor was used. Computational simulations were performed with StarCCM+, installed in a computer with an Intel XEON 5130
cpu, GeForce 7950 GT graphic card, 8 gb of RAM and a 500 gb hard drive. Simulations were considered
converged if max|n+1 n | < 103 for all dependent variables. This chapter is comprised of three
sections:
The influence of boundary conditions is quantified;
Several turbulence models are presented and their differences discussed;
A real size model is simulated and tested against experimental data obtained from LNEC.
3.2
Conceptual Verification
In order to start studying jet fans a very simple case was modeled using StarCCM+. This test case is
shown in figure 3.1. This case is simply a jet fan placed in a square room without sidewalls, i.e. there is
only a floor and a ceiling. This room has an height of 2.75 m and the jet fan in the center, at an height
of 0.25 cm and is modeled as a cylinder of 3 m length and 0.25 m radius. This model mesh base size and
turbulence model were chosen from table 1.1 as 0.25 cm and Realizable k respectively.
Figure 3.2 shows what was expected: the air comes from one side of the domain to the inlet of the jet
fan and leaves through its outlet, flowing and spreading through the domain ahead of it.
3.3
3.3.1
Solid modeling
Jet Fan Modeling
For the remainder of this work, the jet fan being modeled is model JCR.380.2/4 produced by EFAFLU
and shown in figure 3.3, whose main characteristics are summarized in table 3.1. This particular model
was chosen because it was developed specifically to be used in car parks.
28
Figure 3.3: Jet fan - JCR.380.2/4 model from Efaflu. Source [2].
In this validation process, three different approaches were used to model jets fans:
MODEL A - The jet fan was modeled as not being part of the domain (figure 3.4) and had at
its inlet a boundary condition of velocity inlet and at its outlet a velocity outlet with a velocity
boundary-normal of 18.98 m/s.
MODEL B - In this case the jet fan was modeled as a cylinder shaped region with a momentum
source (figure 3.5) having a thin wall separating it from the rest of the domain. Since from table
3.1 the jet fan produce 51 N of thrust and had a volume of 0.32 m3 , the momemtum source had
160.6 kg/(m2 s2 ).
MODEL C - The third model was similar to A, yet it had a higher level of geometrical detail
(figure 3.6).
Since all those options have advantages and shortcomings, it was important to consider them all
29
Thrust (N)
Velocity (m/s)
51/13
1.9/1.0
21.6/8.25
Rotation
(RPM)
2750/1380
Speed
Diameter
(mm)
380
Length (m)
2.8
Figure 3.7 compares velocities along the jet fan axis for all three models and differences can be noted.
On the one side model C is geometrically more complex, thus requiring a finer mesh to correctly simulate
every aspect, this produces a velocity profile similar to the one that will be seen on the next section
about mesh base size, where mesh size does impact the velocity magnitude along the jet axis. On figure
3.7 it is also noticeable the significant differences in velocity magnitude observed in the first few metes
30
downstream. Those differences are inherent to the modeling approach itself: for both model A and C,
velocity is a prescribed quantity on the face of the fan, whereas model B uses a momentum source. By
using the first approach, an uniform velocity profile is generated whereas in the last option the velocity
profile at the jet fan outlet is not uniform due to the no slip condition on the jet fan walls. If the velocity
magnitude was measured just a few centimeters away from the centerline, both model A and C would
exhibit the same results, whereas model B would have a significantly different velocity profile for the first
few meters downstream.
When assessing all those models, it is also important to verify if any of them will have a significant
impact on lateral jet spread. In section 3.6 it will be seen that turbulence will have a major impact on
jet spread, perhaps the jet fan model could also have some impact on this spread. Figure 3.8 compares
jet spread for all three models and despite being possible to see some slight differences, they are so small
that it can be considered that the jet fan model does not play a major role in the lateral spread of the
jet.
Because differences observed between model C and both A and B are not relevant and it is a model
that requires a finer mesh, it will not be further considered. As for the remaining models, they both have
advantages and disadvantages, and it will be important to consider them both on subsequent sections.
Model A is the model that requires less control volumes to be simulated, and this will be relevant for
31
Maximum
Cell Size
1.0 m
Surface Curvature
36 pts/circle
Surface
Rate
1.3
Growth
Minimum
Size
0.03 m
Target Size
0.125 m
3.4
Mesh Generation
For all the simulations in this work a trimmed mesh with several prism layers was used. This choice
was done because in terms of general accuracy for a given number of cells, the trimmed mesh will always
produce the most accurate solution when compared to a tetrahedral mesh. In addition in terms of solution
quality, it requires approximately five to eight times less cells to produce the same accuracy as other mesh
types. Finally, the trimmer model is not directly dependant on the surface quality of the starting surface
and as such is more likely to produce a good quality mesh for most situations [4].
The trimmer meshing model utilizes a template mesh constructed from hexahedral cells from which
it trims the core mesh based on the starting input surface. The template mesh can contain refinement
based on curvature and proximity, as well as fixed cell sizes based on the boundary surface.
The starting point for the mesh base size used in this work simulations was 0.25 m. This value was
chosen based on references [12], [36] and table 1.1. The remainder options were made accordingly and
are presented in table 3.21 .
In addition, recalling section 2.4.2, table 1.1 and [4], it was decided to use an all y+ wall treatment in
this work. In order to benefit from the wall functions, whenever generating a new mesh, it was necessary
1 This table does not present 0.25 m as base size. This happens because this table is to be representative for the whole
work, and as will be seen later, 0.25 m created a very coarse mesh.
32
Number
Prism Layers
4
Stretching
Thickness
1.1
0.03 m
3.4.1
Starting with Model A and the previously used domain of 120x120 m, it is now important to understand if results are independent from the mesh base size. This study was done by reducing in half
the mesh base size from its original size (0.25 m). Since the reduction in size happened in all three
dimensions, control volumes increased by a factor of 8. Figure 3.9 and 3.10 show both the horizontal and
vertical velocity profiles. From those figures it can be seen that there is a change in results of around
15%, which can be considered significative. Due to this, a further mesh base size reduction was tried and
results are presented in 3.9 and 3.10, represented by the green line. For this last simulation, if the entire
domain base size was halved the number would reach more than 160 million control volumes which was
computationaly not doable, so only a 15x15 m area surrounding the jet fan zone was halved. Comparing
those two simulations, it is possible to see that mesh independence is reached at a base size of 0.0625 m
, if engineering precision of less than 10% in difference is considered acceptable. It must be noted that
the higher differences observed in the first few meters downstream can be explained by effects associated
with numerical diffusity caused by the mesh different sizes themselves. Since finer meshes tend to be less
diffusive, results have a higher difference.
The mesh base size required for fully independent results is rather small, which has a great impact in
the required number of control volumes. Since the computational power available for this work could only
deal with up to about 8 million control volumes, applying such a small base size to the mesh generated
for a car park would not render this work possible. Moreover, the simulation with the finer mesh required
up to 30 hours to reach convergence while the other two needed less than 12 hours. It was then decided to
advance using a mesh base size of 0.125 m, knowing that an approximation error was being introduced. In
addition, differences observed in figure 3.9 in the first few meters downstream are due to effects associated
with numerical diffusity.
3.5
To help understand how the domain size will impact the results, a simple room with a jet fan in the
middle was simulated. The room was modeled with several increasing dimensions, 30x30 m, 60x60 m,
120x120m and 180x180 m. Since the jet fan length is roughly H=3 m, those dimensions were chosen so
that every distance between it and the domains end was approximately 5H, 10H, 20H and 30H. Since the
33
Figure 3.10: Velocity profiles along a vertical line 10 m away from the jet fan.
effects under study were away from the jet fan, for simplification purposes, it was decided to model the
jet fan according to model A. The room was 2,75 m high and the jet fan axis was placed at a height of
0,24 m. The room had top and bottom walls and the sidewalls were treated as pressure inlets with 0 pa
total pressure. Since this section aim is to study domain influences, turbulence was once more modeled
with Realizable k . The advantage of this approach is its simplicity which translates in fewer control
volumes used and less computational time. In what concerns geometry, due to its symmetry, only half
of the domain was built and it was built in an incremental way. First 5H was built and simulated with
a base size of 0.125 m, then, for 10H an add-on with a base size of 0.25 m was applied. It was repeated
for 20H, yet, instead of 40H, 30H was simulated due to computational power available. This incremental
process stopped when 30H was reached, and a section cut of the mesh used for this domain size is shown
in figure 3.11. A schematic of this approach is depicted in figure 3.12 for clarification. This incremental
approach was chosen because on the one hand there was a need to study domain size influence and on
the other hand it would be impossible to study a 30H domain with a base size of 0.125 m since it would
require more than 44 million control volumes which is beyond this work scope. This approach is expected
to produce good enough results at a low computational cost.
Taking velocity profiles along the jet axis (figure 3.13) it is possible to see that except to 5H an 10H
34
Figure 3.11: Section cut, showing the different mesh base sizes away from the jet fan.
Figure 3.12: Geometry built to assess boundary conditions, detailing size increases. (top view + half
domain)
differences between domain sizes are barely noticeable. In addition difference between 5H and 20H are
under 2%, which can be considered under reasonable engineering approximation. If one is to compare
vertical velocity profiles (figure 3.14) for several domain sizes at a distance of 5 m away from the jet fan
outlet it can be seen that the domain size has almost no influence on the results. Differences between
domain sizes are under 10% between 5H and 30H. Comparing results further downstream, 10 m away
from jet fan outlet (figure 3.15), it is possible to reach similar conclusions, since except for 5H which is
close to the domain end, results are in accordance.
In a real car park, jet fans will typically be placed aligned with each others at distances of 10 m to 20
m with the flow from one entering the next one. That is why it was important to understand the domain
size influence. It was verified that the domain size does impact velocity profiles, but this influence is not
35
Figure 3.14: Velocity profiles along a vertical line 4 m away from the jet fan outlet.
Figure 3.15: Velocity profiles along a vertical line 10 m away from the jet fan.
that significant. It was seen that only the smallest domain produced results that were different from the
others. But 10 m downstream the the jet fan, velocity profiles were already within close agreement.
3.6
The final step on the development of the jet fan computational model is to analyze how results are
impacted by the choice of the turbulence model. In section 2.4 several hypothesis on how to calculate the
Reynolds stress tensor have already been introduced, and from table 1.1 it is possible to see that despite
a clear preference for k related models, historically other models have been used.
It is known that a turbulent jet of fluid, discharged from a tube into an expanse of the same fluid
medium at rest, will exhibit a symmetric, linear rate of growth in all directions normal to the jet axis.
If, however, the jet discharge is brought into contact with, or very close to, a plane wall whose surface
is parallel with the jet axis, it is well known that a strikingly different pattern develops; for, the rate of
spread of the shear flow parallel to the wall is between five and nine times as large as that normal to
it [16]. Since jet fans are typically mounted on the ceiling, without space between the casing and the
ceiling itself, they can be considered to behave like a round wall jet. Because the air flow leaving the jet
fan will be responsible for setting the remaining in motion, it is thus critical to correctly model the jet
fan outflow.
36
To better understand this problem, it is important to firstly revisit in more depth the physics pre~ U
~ represents the convective
sented in chapter 2. Recalling equation (2.4) (Navier-Stokes equations), U
acceleration which accounts for the effect of time independent acceleration of a fluid with respect to space.
~ U
~ = (U
~ )U
~ = ( U~ 2 ) U
~ ( U
~ ) it is possible to re-write that equation into equation
Since U
2
(3.1).
~
U
~
~ = 1 pT + 2 U
~ +f
U
t
(3.1)
Dz
=
Dt
W
W
W
+
z
+
x
+ y
x
y
| {zz}
|
{z
}
mean strain bending mean strain stretching
2 (V 2 + U 2 )
2U V
2U V
2 z
2 z
+
+ (
+
)
2
2
2
yx
x
y
x
y 2
|
{z
}
{z
}
|
{z
}
|
turb. normal stress generation turb. shear stress generation viscous diffusion
x (
(3.2)
W
W
U
U
V
V
), y (
), z (
)
z
y
x
z
y
x
For this case, and in accordance with figure 3.16, z is the primary fluid direction. If a laminar three
dimensional wall jet is considered, the Reynolds-stress gradients are absent, the viscous term only diffuses
vorticity and vorticity is generated through the terms on the first line of (3.2). Those terms are responsible
for an arrangement that creates a laterally outward secondary flow for which more information can be
found in [16].
When dealing with a turbulent three dimensional wall jet, which is the relevant situation for this
work, the Reynolds stresses must be approximated in some way. If an eddy viscosity approximation is
chosen, as detailed in section 2.4, then equation (3.3) arises. Hence, (3.4) presents the turbulent stress
2 In
37
terms from equation (3.2) minus small effects due to streamwise velocity gradients. That is to say, the
role of the turbulent stresses is reduced to that of a vorticity diffuser [16].
Ui 0Uj 0 = t
Ui
Ui
+
xj
xj
2
ij k
3
z
z
t
+
t
x
y
y
(3.3)
(3.4)
However, an experimental investigation conducted by Launder and Craft [16] concluded that the only
significant mechanism driving the very high lateral spread of the wall jet is that of the Reynolds stress
field in the (x,y)-plane in providing a source of streamwise vorticity (equation (3.2)). They proceeded even
further to find that it is the first of the Reynolds stress terms in (3.2), involving the spatial variations
of the differences in the normal stresses perpendicular and parallel to the wall, that is predominantly
responsible for the behavior.
Since eddy viscosity models, described in equation (3.3), do not approximate the Reynolds stresses,
there is a need to investigate further options. The option present in StarCCM+ and described in section
2.4 is to make use of the Reynolds stress transport models. Those are second moment closure models based
on transport equations for the Reynolds stresses where those two terms from equation (3.2) are solved to
some extent, thus hopefully making the results closer to reality. The transport equation describing the
transport of the Reynolds stresses Ui Uj was presented in section 2.4 and is presented here modified as
(3.5).
U 0 Uj0
0 0
Uj
Ui
Ui Uj = Ui0 Uk0
+ Uj0 Uk0
+
2 i
t
xk
xk
xk xk
0
i
Ui
h 0 0 0
0 0
0
0
+p
Ui Uj Uk + p kj Ui + ik Uj +
Ui Uj
xj
xk
xk
xk
(3.5)
Craft and Launder [16] also noted that due to the strength of the vorticity source being so sensitive
to the computed normal stress profiles near the wall, the computed behavior is highly dependent on the
model adopted for the calculation of ij . In StarCCM+ this term is approached with the linear model of
Gibson and Launder [4], comprising four terms: the rapid part, the slow part, and their respective wall
reflection terms.
3.6.1
It is clear from the previous section discussion that eddy viscosity models will fail to correctly predict
a turbulent three dimensional wall jet. However, and because this work is intended to have a practical
engineering application, the computational cost of using Reynolds stress transport models can turn into
a disadvantage whenever they are used. Those models require an additional set of seven equations to be
solved and are known [4] to have some shortcomings in what concerns convergence. With this in mind,
the first step to understand how the models implemented in StarCCM+ simulate a three dimensional
wall jet, Abrahamsson experimental data, obtained from the SIG 15 - Turbulence Modeling workshop,
was used. The experimental setup is shown in figure 3.17. The jet was allowed to develop over a surface
3.2 m wide and 2.1 m long. The flow is injected through a circular orifice on the vertical wall with 20 mm
diameter place flush with the horizontal plate. A 1.2 m high vertical wall is placed over the inlet in order
38
to avoid air entrainment. According to Abrahamsson [30], the setup was made so that the flow coming
from the orifice had a uniform velocity profile of 40 m/s with a low turbulence level (0.8%). The wall
jet flow has been shown to be self-preserving and the turbulence structure to be independent of the inlet
Reynolds number which was ReD = 53000. The results from Abrahamsson are therefore thought to be
representative for fully developed wall jets at high Reynolds numbers.
This experiment was simulated with StarCCM+ and a detail of the mesh built for those simulations
is shown on figure 3.18. In those simulations, a trimmed mesh with a number of prism layers as described
in table 3.2 was used. Some further consideration on the meshing options were given in section 3.4. Using
a coarser grid ( 400k control volumes) several runs were performed using k models along with RST
models. Some additional runs were done with more dense grids in order to evaluate grid independence,
which was reached at around 3 million control volumes. Figure 3.19 and 3.20 show that reducing the size
of the control volumes produced results with a slight difference, however those runs who had a coarser
grid took less than 4 hours to reach convergence whereas those with 3 million volumes took more than 24
hours. Further reductions in mesh dimensions were considered but produced similar results. Comparing
both figures, it can be noted that differences in mesh dimensions are almost always under 10% which is
considered to be within engineering tolerance, thus to save computational time, coarser grids were used
except where noted otherwise.
In both 3.19 and 3.20 turbulence models considered were Realizable k which was from table 1.1
a common option among researchers and RST which, in accordance to what was seen earlier, is a second
39
moment closure turbulence model. Comparing both models, RST has a higher reduction in velocity in
the jet axis than Realizable k . This is probably due to the strikingly different pattern observed in
figures 3.21 and 3.22: as predicted RST models produce a much higher jet spread than Realizable k .
Before proceeding further with this analysis, other turbulence models were tried and their results are
presented in figure 3.23. It is possible to see that in terms of axial speeds (figure 3.23) and jet spread
(figure 3.24) they behave much like the Realizable k , failing to approach the relatively high jet spread
of RST. So, a further analysis is only considered for RST and Realizable k .
Figure 3.24: Jet spread for other Turbulence models tested (left Standard k , right k ).
Considering first the vertical spread of the jet, figure 3.25 shows an excellent agreement between
both turbulence models and experimental results from Abrahamsson which is in accordance to what was
40
expected. It is also interesting to compare both computational results at several distances from the jet
outlet. If for figure 3.25 only the most distant plane from the outlet was considered, in both figures
3.26 and 3.27 it is possible to see how the jet develops starting at the jet outlet. It is important to
note that after a few centimeters downstream the results appear to be the same. In [16] some concerns
are raised on whether or not the jet had reached a self similar state, and it is discussed that by 100
diameters downstream the flow has by no means reached full development. For the purpose of this work,
and remembering that it is intended for an engineering application, the results are successful enough, as
will be seen.
If, instead, the lateral variation of velocity is considered, the grave failings of both calculations are
strikingly brought out, Realizable k under predicts and RST over predicts the experimental results,
shown in figure 3.28. And this is the most important part of this analysis since every turbulence model
considered does not correctly approach the velocity profiles. Thinking ahead, it is important to assess this
lack of agreement because in a car park jet fans will be placed in relatively close vicinity, both laterally
and streamwise, thus impacting both the overall air flow and the velocity profile at the jet fans inlet.
Having calculated both vertical and lateral variations of axial velocity, it is now possible to compute
the spreading rates for the jet depending on the turbulence model chosen. Table 3.4 presents results for
computational models with Realizable k and RST and the Abrahamsson experimental results [16].
It is clear that all models have almost no problems in dealing with the streamwise jet spread. And it
is clear that all those models fail to correctly predict the lateral spread of the jet. As seen earlier and
41
Vertical Spread
0.065
0.064
0.062
Abrahamsson[16]
Realizable k
RST
Lateral Spread
0.32
0.076
0.49
Ratio
4.92
1.18
7.90
42
3.7
LNEC Case
In the previous section, the simulation global parameters were established. As a consequence, it is now
possible to advance further and compare those results against experimental data obtained from LNEC.
In LNEC experiment, an EFAFLU model jet fan JCR.380.2/4 was put into normal operation inside a large
enough building. Velocity measurements were done according to figure 3.30 grid, using several Airflow
AV6 anemometers with 30 mm turbines and hot-wire Airlfow TA5 anemometers, that can be seen on
figure 3.31. Measurements were done at several heights: 7 cm, 24 cm, 41 cm, 77 cm and 196 cm.
LNEC experiment was conducted in a pavilion 5 m high and wide enough to avoid any kind of
recirculations. In this simulations, only models A and B were considered. This section test room is
presented in figure 3.33 and 3.32. It is equivalent to previous section 20H model except for the fact that
the roof is higher (5 m instead of 2.75 m). Mesh base size was, from previous section 0.125 m.
Regarding model B, the first step was to check mesh size influence. Using Realizable k turbulence
model, two different mesh sizes were tried. From previous section, a base size of 0.125 m and another
one with 0.0625 m were used, results (not shown) are very similar to those already presented in figure
3.9: it is possible to see that results are not entirely mesh independent, nevertheless using a finer mesh
seemed not doable as it would require more than 20 million control volumes. The meshes used already
took more than 10 and 32 hours to converge, respectively.
Since the only change from the previous section was in the height of the domain, it is expected that
conclusions from the previous sections will hold valid. Due to the increase in height the only thing that
could be expected to change should be some minor increase in the vertical spreading rate. So, comparing
the results from a simulation using Realizable k turbulence model against experimental ones, it is
possible to see that all the past conclusions hold true. From figure 3.35 to figure 3.44 a comparison
43
between results at several heights and distances is made and while the vertical spread of the jet is within
good agreement both in terms of velocity magnitude and velocity profile, the lateral spread is very poorly
approached. Paying special attention to the lateral spread, it is possible to conclude that despite the
velocity profile being similar, velocity magnitudes are too much apart.
Recalling section 3.6, it had already been established that Realizable k would not correctly predict
this lateral spread, and for it to be better approached, it was needed to use a turbulence model that
computed Reynolds stresses. From the options available natively in StarCCM+, the only one that filled
this requirement was Reynolds Stress Transport turbulence model. Applying this turbulence model to the
same simulation produced results that are also displayed in figures 3.35 to 3.44, represented by a red square
with a triangle. Results are remarkably different: except for the first two points, velocity magnitudes are
closely matched for all distances and heights. Since both turbulence models were tested with the jet fan
producing the same amount of momentum, it can be noted that Realizable k mid plane velocities
are all always greater than RST and experimental, while failing to model lateral spread. Instead, RST
produces lower mid plane velocities but closely matches lateral profiles. Since the momentum transfered
to the air is kept constant throughout those simulations, by continuity this had to happen: to have higher
lateral velocities, those at mid plane had to be lower. It must be noted however, that the simple fact of
changing turbulence model, from Realizable k to RST, implied that the simulation took twice the
time to reach convergence ( 20 hours for RST).
A comment has also to be made to the fact that figures displaying results 3 m away from the jet axis
display two sets of experimental results: one that is to the right and another that is to the left of the
jet axis. This happened due to differences in experimental values obtained from LNEC, probably due to
some unbalances in jet fan building or experimental setup aspects. It is important to note that despite
being industrial equipment, jet fans are not precision equipment. So, for the sake of completeness, both
sides were included in those figures.
For all those cases, the spreading rates are computed and calculated and displayed in table 3.53 . If one
recalls results from table 3.4, the similarities are easily brought out. It must be noted that ReD = 53000
for table 3.4 and for this case ReD = 477326 since jet fans have a bigger radius. Despite the Reynolds
number being almost one order of magnitude apart, results are expected to be comparable because
both flows are turbulent[16]. LNEC and Abrahmsson have rather similar results which can function as
3 Left/right
refers to the spreading of the jet to the left/right, due to assimetries in the setup
44
k
RST
Abrahamsson
Experimental (Left)
Experimental (Right)
Vertical Spread
0.0798
0.0652
0.065
0.0581
0.0581
Lateral Spread
0.0682
0.4493
0.32
0.259
0.268
Ratio
0.85
6.8
4.92
4.46
4.60
Figure 3.34: Spreading rate at the jet plane for k (left) and RST (right).
It is also important to understand the behavior of the air entering the jet fan, or stated otherwise,
what is the capacity of the jet fan to attract air. Since the goal of a jet fan is to make the air move,
understanding its capacity to draw air into it is of some importance. However two major details arise:
there are no experimental values of those velocities and those velocities are under 10% of the outlet
velocities. Figures 3.45 and 3.46 show the streamlines and velocity magnitude that enter the jet fan for
the two different turbulence models evaluated, while using model B.
It is interesting to see that velocity magnitudes are rather similar for both turbulence models. This
probably happens because of the low velocities that do not generate high Reynolds stresses. Since jet
45
fans in a car park are usually aligned with each other, the fact that the turbulence model does not greatly
influence the inflow, means that results will be similar, and the capacity of jet fans to draw air into them,
will not be influenced by it.
Finally, all the work from this section was done using model B for the jet fan, however, the next
chapter will deal with establishing an optimization strategy for the placement of the jet fans and, to
ensure that many combinations are tested, there will be a need to reduce mesh size and complexity and
that involves switching from model B to model A for the jet fan since it requires fewer control volumes
46
to produce reasonable results and reach convergence. It is, thus, important to evaluate how those models
compare while using RST and Realizable k turbulence models. In figure 3.47 it is possible to see that
except for the first few meters results tend to be similar for both approaches. It can also be noted that
in what concerns lateral spread, the same pattern, where only RST correctly predicts it, arises. Both
model A and Realizable k turbulence model fail to predict experimental results, nevertheless they
exhibit a trend that is compatible with experimental results. In order to avoid an unnecessary increase
in computational time, they can be considered as possible alternatives. For the same mesh size, changing
both jet fan and turbulence model can result in less than half the time until results converge.
47
Figure 3.47: Flow at jet fan mid plane (left k , right RST).
3.8
Summary
This chapter dealt with validating every aspect of the simulation of a jet fan and understanding which
factors impact the overall result. Despite only being possible to attain mesh independence if a 10%
margin is considered, the remainder aspects of the simulation seem to be clearly understood and under
control.
All three possible approaches for jet fan modeling produced quite similar results. However, the
momentum source model (B), seems to be a more reliable approach as jet fans are factory tested and
calibrated to produce a given momentum instead of an outlet velocity. In addition, by using model B, it
is possible to track soot or some other particles across the jet fan.
Nevertheless it is important to notice that results did not become truly mesh independent for base
sizes of 0.0625 m and domain sizes of less than 5H. Those dimensions are rather small and when applied
to a real car park full with cars and all geometric details would originate such a large number of control
volumes that would render the simulation impossible with the available computer power. To allow the
assessment of an entire car park, 0.125 m will be used instead as base size for the mesh.
The most important aspect coming out of this chapter is the evaluation of the impact that the
turbulence model has on the final results. It was possible to see that for a single jet fan, Realizable k
would fail to completely predict the lateral jet spread whereas RST, which should over predict the results,
produced results that were quite close to the experimental ones.
Despite slightly over predicting the results, RST was the model which showed the best agreement
between experimental and test data, however it was the one that required more computational time
(more than 20 hours versus 10 hours). As stated in section 2.4, the most common k models are in
their formulation isotropic models while RST is a second moment closure model[16]. Studies conducted
with round wall jets [16] showed that such flows have a high degree of anisotropy, caused by second order
induced flows that create high lateral spreading rates. Despite all this, k models are still an interesting
option because they have fewer equations to solve thus being faster, with CPU time requirements being
significantly lower than RST when finer grids are used. They are still able to capture the vertical spreading
rate and are more robust to run, having fewer convergence problems. It remains to be said that it will
be needed to understand how those models impact a full scale car park simulation, and whether they
produce noticeable differences or if those differences tend to fall within engineering tolerance. Differences
in lateral jet spread should not produce changes in flow patterns but rather promote a higher degree of
air mixing, and thus an increase in the rate of air change.
48
Using RST models along with model B, it was possible to simulate and replicate LNEC experimental
results. Those results showed a jet that spreads laterally more quickly than vertically. By modeling
turbulence with Realizable k , one will have a less exact model, as it will fail to model lateral spreading,
but will significantly reduce simulation time. Since the goal of this work is to optimize jet fans placement,
several runs will be needed, thus making it necessary to not only have an accurate model, but also to
have a light one. By saying this, there is a need to do some trade-offs, and by using Realizable k
instead of RST, there is a significant reduction on total simulation time. It must also be understood
that, when testing a jet fan in a big open space building, there are several sources of errors. Jet fans are
not precision material, thus they have performance variations. Turbine anemometers are not precision
equipment either and despite being big, the space is not infinite, thus walls will influence the result.
Finally an open window would be enough to influence the result. This said, the goal of this study was
not to have a perfect match between computational and experimental results, but instead to identify a
trend and to produce results that could be compared.
In short, the most suitable modeling parameters were:
Mesh base size of 0.125 m;
Domain bigger than 15x15 m (5H);
Turbulence model with wall functions - Realizable k for being faster;
Jet fan simulated using momentum source (model B).
49
Chapter 4
LoureShopping case
4.1
Introduction
LoureShopping (Figure 4.1 and 4.2) is modern shopping center owned by Sonae Sierra located in
Loures, nearby Lisboa. It was built in 2004, and the ventilation was projected by LMSA. It has an
underground car park that occupies two underground floors, each having 900 parking spaces as well as
3 levels above ground that have 300 parking spaces in total. For ventilation purposes, each undergound
floor was split into three fire zones, each comprising around 300 parking spaces. In this work, only fire
zone 3 from level -2 will be studied. Figure 4.4 shows the location of fire zone 3 in relation to the car
park.
The goal of this work is to evaluate the performance of the projected ventilation system, identify areas
that may present problems of high mean age of air and optimize the placement of jet fans. The study is
performed for fire zone 3, and the results presented in the following sequence:
1. Solid modeling description and boundary conditions discussion;
2. Air flow analysis including velocity field and flow patterns with only air supply and air exhaust
systems working;
3. Air flow analysis of jet fans in working condition;
4. Optimization procedure discussion;
5. Air flow analysis of jet fans in new configuration;
6. Results Comparison.
50
In all the above simulations (air flow and CO dispersion) the model is considered to be isothermal.
That means that no heat influences through walls or in indoor climate or inlet air temperature will be
taken into account. This results in the assumption that all walls are modeled as physical extensions of
the calculation domain.
51
Base Model
Peugeot 207
Audi A4
BMW 325
Length (m)
4.03
4.70
4.52
Width (m)
1.74
1.82
1.81
Height (m)
1.47
1.43
1.45
4.2
Solid modeling
To create the car park, AutoCAD files were opened in Autodesk Inventor and a solid was created.
From this solid, blocks were taken out to simulate every feature of the car park. All the spaces were filled
with cars, as this would represent the worst case scenario. Figure 4.5 show the overall aspect of the car
park fire zone 3, and figure 4.6 presents a detail of it. In addition, in figure 4.5 boundaries are signaled.
Cars were modeled with three different areas and heights shown in table 4.1. Some cars had more
detail than simply cubic forms. All the cars are detached from both columns and other cars. Figure 4.6
shows some examples of the car models used.
Regarding boundaries, fire zone 3 has two air inlets and two air outlets. In addition, in case a fire
occurs, air can be drawn in from adjacent fire zone 2 and from level -1 through a connection ramp. Table
4.2 describes in detail the dimensions of such boundaries. In the event of a fire, air entrance through
one connection to fire zone 2 and through the street connection is to be prevented and doors in those
places are expected to remain closed. In addition, if a fire occurs, air inlets are supposed to be shut down,
however they still allow air to be drawn in if differences in pressure arise, hence they are treated as special
52
Boundary
Air Inlet
Air Inlet
Air Outlet
Air Outlet
Level -1 Ramp
Fire Zone 2
Condition Applied
Stagnation Inlet
Stagnation Inlet
Pressure Outlet
Pressure Outlet
Stagnation Inlet
Stagnation Inlet
Length (m)
1.47
1.43
1.45
1.45
1.45
1.45
Height (m)
1.74
1.82
1.81
1.81
1.81
1.81
Code
A
B
C
D
E
F
4.3
Having described in the previous section the modeling process, this section will do a first analysis of
the underground car park ventilation system as it is currently installed. For those simulations, based
on insight gained on previous sections (namely conclusions from chapter 3) a trimmed mesh with an
increased base size of 0.5 m was used, refined to a base size of 0.125 nearby jet fans. This mesh also
used three prism layers for the ceiling and only two for the remainder surfaces, to reduce the number of
control volumes used. Figure 4.8 shows a sectional cut of the mesh used.
The first step is then to simulate the car park without jet fans, being air only set in motion due to
the air extractors, placed in positions C and D. The result of this simulation (that modeled turbulence
with Realizable k ) is shown in figure 4.9. From it, it can be seen that this option is not sufficient to
ventilate the entire car park, being it dominated by areas where the air is stagnant or almost stagnant,
causing the mean age of air to reach extremely high values (figure 4.10).
4.3.1
The first approach to simulate the car park with jet fans under operating conditions was done using
Realizable k along with an All y+ wall treatment. From chapter 3 it had been concluded that
such approach would produce fairly good results in respect to longitudinal spreading rates. The mesh
53
Figure 4.10: Mean Age of air without jet fans at an height of 1.5 m.
generated had more than 4 million control volumes and it took more than 30 hours to reach convergence.
Results for this simulation are presented in figures 4.11 to 4.13. The lower plan at an height of 0.5 m is
the height people will be breathing in case of a fire, 1.5 m is the median breathing height, 2.5 m is the
height of the jet fans.
In those figures special attention has to be paid to regions where air is at rest or has very low speeds.
Nevertheless, for the most part of 1.5 m and 2.5 m the air is moving, and apparently the axial fans are
well distributed and should be effective in case of a fire. Perhaps special attention should be paid to the
corners and middle region of the car park as air seems to be more at rest.
For further analysis of the flow field, it is important to look at streamlines originating at every fan
54
outlet and a velocity vector plot to ensure air is actually flowing out of the domain were it is supposed
to. Figure 4.14 shows those streamlines and it can be seen that, apparently, there are no recirculation
spots. From the vector plot on figure 4.15 it is possible to verify that the air flow behaves as expected,
entering and leaving the domain through the defined inlets and outlets respectively.
It is also interesting to analyze where does the air that enters the domain go. Figure 4.16 show the
streamlines that originate at the air entrance points. By inspecting figure 4.16 it is possible to see that
the air coming from the upper level undergoes an impressive recirculation, and instead of heading towards
the extraction point, it returns and goes almost to the zone 2 entrance. On the other hand, air from
55
air inlets goes all the way through the park until it reaches the right extraction point. It is interesting
to observe that this fresh air will travel always in close proximity to the car park wall. The ventilation
scheme should be updated to ensure that fresh air reaches the middle of the car park. Maybe it would
be interesting to include some sort of equipment that would create some downward flow, as it appears
that the air flow adheres to the ceiling and does not reach for lower heights.
Finally, to fully understand how long the air will remain inside the car park, the concept of the mean
age of air was used. By evaluating the mean age of air, it is possible to see which areas have poor
ventilation and which jet fans should have their position modified. The assessment of the mean age or
air at both 1.5 m and 2.5 m is depicted in figures 4.17 and 4.18, where areas of bigger concern are those
with dark blue color. Despite an overall good result and a lack of air accumulations spots, there is a clear
opportunity to improve ventilation on the right area of the car park. This could be achieved by creating
an extra air supply or by changing the jet fan configuration. Also, air from level -1 does not contribute
to properly ventilate the car park. Based on those observations, there is room to improve the jet fan
location in order to have fresher air inside the garage.
4.3.2
One of the major conclusions drawn from chapter 3 was that Realizable k and similar turbulence
models did not replicate in a satisfactory way the lateral spreading of the jet fans. Since the aim of
this work was to assess the ventilation effectiveness, it is important to use the model that best replicates
56
reality. Hence, the complete car park was simulated using a Reynolds Stress Turbulence model. Using
such turbulence model requires some special attention: the mesh surrounding the jet fans must be refined
to a base size of 0.0625 m, it has more equations to solve and the number of iterations needed to reach
result convergence is greater. For this particular case, the car park was split into 8 million cells and
convergence was only reached after 3000 iterations, which is significant if the iteration duration of 95
seconds is taken into account. Figures 4.19 and 4.20 show the velocity magnitude profiles obtained at 1.5
m and 2.5 m, whereas 4.21 and 4.22 show the mean age of air at those same heights. By comparing those
figures to 4.13 and 4.18 it is possible to confirm what was previously expected: more lateral spreading of
the jets with an influence in the mean age of air, that seems to be lower inside the domain.
It is noteworthy that despite differences being noticeable, they appear to be rather small. In table
4.3 those differences have been quantified and it is possible to verify those results for the mean age of air.
The fact that the maximum age of air is 22% lower for the RST model is easily explained by the higher
mixing of the flow due to the differences in jet spread already discussed in previous sections. This higher
degree of mixture and lateral spread, means that the air flow coming from the jet fan can reach more
places and set more air in motion. Nevertheles, those results are in close agreement to what had already
been discussed in chapter 3: for initial engineering analysis the complexity of RST can be avoided by
opting to use Realizable k . Despite some minor differences, jet fans tend to be far away enough so
that differences in lateral spread do not really impact inflows for each other. Moreover, by looking at
streamlines coming out of jet fans (figures 4.14 and 4.23), very similar underlying patterns arise in the
57
flow field. The only significative difference is in the maximum age of air, that is lower for RST. This
happens because this model can promote a higher mixing of the flow. However, the differences in lateral
spread of the jet fans do not change the flow patterns that remain essentially the same, which is what
the engineer is looking for when simulating a ventilation scheme.
58
Figure 4.23: Streamlines originating at jet fans for RST turbulence model.
4.4
4.4.1
Optimization
Optimization Preparation
The optimization part of this work was done using modeFrontier. Since models from previous sections
required more than 20 hours of computational time to reach convergence, there was a need to develop a
new and extremely simplified model. This model was built assuming that the steps from figure 4.24 could
be followed and would produce good results. This is a rather empirical approach, yet, since the model is
expected to run several times before reaching convergence, using the full 8-million volume mesh model
would be beyond the scope of this work both in terms of time and computational resources available.
With all those considerations in mind, a new and simplified model was developed. To reduce the
number of cells on the mesh, cars and columns were removed from the car park model. In addition, jet
fans were modeled according to model A from chapter 3, the mesh base size was increased to 1.0 m and
the prism layer was reduced to two layers in every wall. Only a small region that enclosed each jet fan
was allowed to have a finer mesh with a base size of 0.25 m. Since a few simulations showed that applying
Turbulence Model
RST
Realizable k
59
the boundary conditions from previous models produced some cases of flow entering trough the outlets,
the outlet boundary condition was changed to velocity outlet, with a velocity obtained from reference
data for the air extractors as 9.5 m/s. Realizable k was used to model turbulence and every run was
limited to 160 iterations so that each iteration step lasted approximately 20 minutes. The optimizations
variables used were the maximum age of air at the height of 1.5 m and the area averaged mean age of air
also at 1.5 m, because 1.5 m corresponds to the average breathing height. Those variables were chosen to
be optimized because using one of them alone could produce misleading results and optimizing both is
the main goal of this work. One effective ventilation system is one in which the air that enters the domain
quickly reaches all parts of it. Finally, to possibly fasten the optimization process, the initial position of
axial flow fans was in the nodes of a grid shown in figure 4.25. Each axial fan could then occupy positions
inside an imaginary square whose sides where two meters away from their closest nodes, those possible
positions are given in table (4.4)1 . They had also their angles limited to ensure that the flow would be
directed towards the exhaust system. In addition, their position was checked to avoid them being placed
over or against any column or wall that could exist in the real model. Finally the optimization strategy
(figure 4.24) was assessed for the results it produced.
4.4.2
Optimization Results
This section will show that it is possible to use the methodologies described in the previous section to
optimize a ventilation system. By using modeFrontier it was possible to improve the actual ventilation
system installed by more than 10% in an almost automated way. It must be noted that despite this value
1 In
this table, (XN ; YN ; AAN ) defines the position of a single jet fan.
60
X1
X2
X3
X4
X5
X6
X7
X8
X9
X10
X11
X12
X13
X14
X15
X16
X17
X18
X19
X20
X21
X22
X23
X24
X25
X26
Lower
Bound
(m)
-2
-2
-2
-2
14
14
14
14
30
30
30
46
46
46
30
30
30
50
50
50
50
60
70
70
-1
-1
Upper
Bound
(m)
50
50
50
50
50
50
50
50
80
80
80
80
80
80
80
80
110
98
98
98
112
112
130
130
130
130
Step
(m)
2.6
2.6
2.6
1.9
1.3
1.3
1.3
1.3
1.8
1.8
1.8
1.3
1.3
1.3
1.9
1.9
2.9
1.8
1.8
1.8
2.3
1.9
2.2
2.2
2.4
2.4
Y1
Y2
Y3
Y4
Y5
Y6
Y7
Y8
Y9
Y10
Y11
Y12
Y13
Y14
Y15
Y16
Y17
Y18
Y19
Y20
Y21
Y22
Y23
Y24
Y25
Y26
Lower
Bound
(m)
-5
5
20
55
-5
15
25
30
15
35
15
15
35
55
25
45
25
35
55
75
35
55
35
45
-1
-1
Upper
Bound
(m)
50
50
80
100
50
60
90
90
80
80
85
90
90
85
100
100
95
90
100
110
65
85
65
85
110
110
Step
(m)
Angle
1.8
1.5
2
1.5
1.8
1.5
2.1
2
2.1
1.5
2.3
2.5
1.8
1
2.5
1.8
2.3
1.8
1.5
1.1
1
1
1
1.3
2.6
2.6
AA1
AA2
AA3
AA4
AA5
AA6
AA7
AA8
AA9
AA10
AA11
AA12
AA13
AA14
AA15
AA16
AA17
AA18
AA19
AA20
AA21
AA22
AA23
AA24
AA25
AA26
Lower
Bound
(o )
-15
-15
-15
-15
-15
-15
-15
-15
-15
-90
-90
-90
-60
-60
-90
-60
-180
-180
-180
-180
-180
-180
-180
-180
-180
-180
Upper
Bound
(o )
180
180
180
180
180
180
180
180
180
180
180
180
180
180
180
180
180
180
180
180
180
180
180
180
180
180
Step(o )
5.4
5.4
5.4
5.4
5.4
5.4
5.4
5.4
5.4
7.5
7.5
7.5
6.6
6.6
7.5
6.6
10
10
10
10
10
10
10
10
10
10
61
Figure 4.26: Area averaged age of air at 1.5 m after several optimization steps.
Figure 4.27: Maximum age of air at 1.5 m after several optimization steps.
system. Table 4.5 presents the results of the optimization step. FullModel correspond to a model where
axial fans were simulated according to model B from chapter 3 whereas the remainder used model A.
From this table it is possible see that the mean age of air can improve by at least 10% and the maximum
age of air can be reduced by more than 20%. This result gains special importance because it was reached
in a fully automated way, hence not dependent on the expertise of the engineer in charge. This result
also paves the way to further improvements in the ventilation system, namely to have fewer axial fans
installed.
62
4.4.3
Ventilation Improvements
The optimized ventilation scheme that resulted from the previous section had a very low level of detail.
Recalling figure 4.24, the last section was step (C) and it is now necessary to increase the complexity of
the model in terms of mesh base size and geometrical details, in order to attain the step (D), which is
then comparable to the initial non optimized model, that was already presented. This section will, then,
compare the results of both installed (D) and optimized systems (C) with a high degree of geometrical
detail. In terms of mesh size, it means that the car park will be divided into some 8 million control volumes.
In addition since the differences in turbulence models have already been discussed and compared, in this
section only Realizable k will be used.
Figure 4.29 and 4.30 show the area averaged age of air at both the height 1.5 m and 2.5 m for the
optimized model. Comparing those results to the ones from figure 4.17 and 4.18 it is possible to see
that, overall, the age of air is lower in the first ones. And this happen because a new ventilation scheme
is employed. Figure 4.31 and 4.32 compare side by side the air flow patterns for both cases and the
optimized one is drastically different from the one presented in the first place. Despite seeming very
erratic, the results from the simulation showed that this new ventilation scheme would be able to reduce
the age of air inside the entire car park. By paying special attention to figures 4.33 and 4.34 it can be
seen that the air that enters the domain behaves very differently from the original case, traveling more
defined paths with less backward movements. Also, looking at figure 4.35, that present the streamlines
63
coming from the jet fans, the new flow pattern is outlined. If one looks back at figure 4.16, this new
flow pattern is completely different and even non intuitive, meaning that it would only be reached in a
automated way. In terms of streamlines, this new pattern produces less recirculations and tries to force
the air to follow two main paths towards the outlets and without the backward flows that were observed
in the original scheme, namely on the flow coming from level -1.
64
Figure 4.35: Streamlines for the optimized system, originating at the jet fans.
4.5
Summary
As chapter 3 showed, k performed differently from RST. RST simulations had a higher degree of
air mixing and increased jet spread that affected the mean age of air. Nevertheless, those differences were
only of about 10%, which for most engineering applications does not justify the addedd computational
time.
Regarding the ventilation system installed inside LoureShopping car park, it is effective in what
concerns pollutant removal and should be able to remove CO even in the worst case scenario of a full car
park. By using an optimization software and without a priori knowledge of the installed solution it was
possible to define an entirely new jet fan placement that allowed to decrease the mean age of air inside
the car park and increase pollutant removal, in a reduced amount of time. Those results indicate that it
might be possible for CFD coupled with optimization algorithms to become an extremely important tool
for engineers planning ventilation systems.
65
Chapter 5
Conclusion
5.1
Results Discussion
Since jet fans are located nearby the ceiling, they can be modeled to some extent as wall jets. By
making such approximation it is possible to compare simulation results with experimental data from
several sources. The most cited source are the experimental measurements by Abrahamsson. With this
in mind, several turbulence models were evaluated: Realizable k , Standard k , k and RST.
Despite producing the best agreement between simulations and experimental results, RST is also the
model that is more unstable and that requires the longest per iteration time, thus it is still not the most
suitable to use with finer or more complex meshes like the one required for this car park and in most
engineering applications where differences of up to 10% are considered acceptable. The k family of
turbulence models is the first choice in many past works and it is capable of producing results with good
agreement with experimental ones. From this family, Realizable k was then chosen as the turbulence
model for this work.
Regarding jet fan modeling, Model B (jet fan with a momentum source) produced the best results.
It is also the model that better describes the physics behind the fan, with the advantage that it allows
particle tracking inside the fan, in case one wants so simulate soot propagation. By combining Model B
with RST it was possible to replicate experimental data from LNEC. However those results were obtained
with a high computational cost, that represented a big disadvantage when compared to Realizable k .
Finally, in what concerns LoureShopping car park, the ventilation system installed appears to be efficient
enough and is clearly able to remove pollutants from the studied zone. Yet by using a new approach
that combined CFD with optimization algorithms, it was possible to reach a better solution with a lower
mean age of air. This methodology can clearly be used as an auxiliary in engineering projects.
The proposed goal of this work was clearly achieved: the ventilation effectiveness of a jet fan ventilated
car park was assessed and discussed and it was possible to generate in a fully automatic way a new
placement for the jet fans, capable of increasing comfort and reducing pollutant dispersion.
5.2
Future Work
There were many challenges in this work, one of which was what aspects of the car park ventilation
system should be studied as to obey the time constrains and fulfill the scope of this work. The decision
to study the optimization procedure and analyze the mean age of air inside the car park means that
there are a number of other aspects and options to be tested to further improve the knowledge on jet fan
66
67
Bibliography
[1] Dirio da Repblica I Srie A.
[2] Efaflu Catalogue of Jet Fans.
[3] modeFrontier User Guide.
[4] StarCCm+ User Guide.
[5] WHO - Environmental Health Criteria 213 - Carbon Monoxide.
[6] J. Abanto, D. Barrero, M. Reggio, and B. Ozell. Airflow modelling in a computer room. Building
and Environment, 2004.
[7] J. Balo and P. Le Cloirec. Validating a prediction method of mean residence time spatial distributions. AIChE Journal, 2000.
[8] Vasco Brederode. Fundamentos de Aerodinmica Incompressvel. Vasco de Brederode, Lisboa, 1997.
[9] J. Burnett and M. Chan. Criteria for air quality in enclosed car parks. In Proc. Instn Civ. Engrs.
Transp.
[10] W. Chang and C. Cheng. Carbon monoxide transport in an enclosed room with sources from a water
heater in the adjacent balcony. Building and Environment, 2008.
[11] V. Chanteloup and P. Mirade. Computational fluid dynamics (cfd) modelling of local mean age of
air distribution in forced-ventilation food plants. Journal of Food Engineering, 2008.
[12] T. Chow, Z. Lin, and W. Bai. Assessment of alternative ventilation schemes at public transport
interchange. Transport Research part D, 2006.
[13] W. Chow. On ventilation design for underground car parks. Tunneling and Underground Space
Technology, 1995.
[14] W. Chow. On safety systems for underground car parks. Tunneling and Underground Space Technology, 1998.
[15] W. K. Chow. Numerical studies of air flows induced by mechanical ventilation and air-conditioning
(mvac) systems. Applied Energy, 2001.
[16] T. J. Craft and B. E. Launder. On the spreading mechanism of the three dimensional turbulent wall
jet. J. Fluid Mech., 2001.
[17] L. Davidson and E. Olsson. A numerical investigation of the local age and the local purging flow
rate in two-dimensional ventilated rooms. In Papers - Session 3 - Room Vent 87.
68
[18] G. Gan. Evaluation of room air distribution systems using computational fluid dynamics. Energy
and Building, 1995.
[19] F. Kuznik, G. Rusaoun, and R. Hohota. Experimental and numerical study of a mechanically
ventilated enclosure with thermal effects. Energy and Building, 2006.
[20] J. Li and W. Chow. Numerical studies on performance evaluation of tunnel ventilation safety systems.
Tunnelling and Underground Space Technology, 2003.
[21] Z. Lin, T. Chow, C. Tsang, K. FOng, and L. Chan. Cfd study on effect of the air supply location
on the performance of the displacement ventilation system. Building and Environment, 2005.
[22] Z. Lin, F. Jiang, T. Chow, C. Tsang, and W. Lu. Cfd analysis of ventilation effectiveness in a public
transport interchange. Building and Environment, 2006.
[23] K. Maele and B. Merci. Application of rans and les field simulations to predict the critical ventilation
velocity in longitudinally ventilated horizontal tunnels. Fire Safety Journal, 2008.
[24] Nelson Magalhes. Cfd fire simulation in enclosures. Masters thesis, IST, 2007.
[25] K. Papakonstantinou, A. Chaloulakou, A. Duci, N. Vlachakis, and N. Marka. Air quality in an underground garage: computational and experimental investigation of ventilation effectiveness. Energy
and Building, 2003.
[26] K. Papakonstantinou, C. Kiranoudis, and N. Markatos. Numerical simulation of co2 dispersion in
an auditorium. Energy and Building, 2002.
[27] Suhas V. Patankar. Numerical heat transfer and fluid flow. Taylor & Francis, Inc., 1980.
[28] A. Stamou, I. Katsiris, and A. Schaelin. Evaluation of thermal comfort in galatsi arena of the
olympics athens 2004 using a cfd model. Applied Thermal Engineering, 2008.
[29] M. et al Vega. Numerical 3d simulation of a longitudinal ventilation system: Memorial tunnel case.
Tunnelling and Underground Space Technology, 2007.
[30] B. Venas, H. Abrahamsson, P. Krogstad, and L. Lofdahl. Pulsed hot-wire measurements in two- and
three-dimensional wall jets. Experiments in Fluids, 1999.
[31] J. Viegas. The use of jet fans to improve the air quality in underground car parks. LNEC.
[32] J. Viegas and J. Saraiva. Avaliao com recurso a cfd da aplicao de ventiladores de impulso a
parques de estacionamento cobertos. LNEC.
[33] J. Viegas and J. Saraiva. Cfd study of smoke control inside enclosed car parking using jet fan. LNEC.
[34] Frank M. White. Fluid Mechanics, 4th Edition. McGraw-Hill, Rhode Island, 1998.
[35] H. Xue and J. Ho. Modelling of heat and carbon monoxide emitted from moving cars in an underground car park. Tunneling and Underground Space Technology, 2000.
[36] X. Zhang, Y. Guo, C. Chan, and W. Lin. Numerical simulations on fire spread and smoke movement
in an underground car park. Building and Environment, 2006.
69