Journal of Loss Prevention in the Process Industries 56 (2018) 316–327

Contents lists available at ScienceDirect

Journal of Loss Prevention in the Process Industries

journal homepage: www.elsevier.com/locate/jlp

Numerical simulation of LNG release and dispersion using a multiphase CFD T

model
Tianpei Luoa,b,c, Chuangchuang Yub,c,∗, Ruimin Liub,c, Mao Lib,c, Jiaxian Zhangb,c, Sheng Qub,c
a
State Key Laboratory of Technologies in Space Cryogenic Propellants, Beijing 100028, China
b
Beijing Institute of Aerospace Testing Technology, Beijing 100074, China
c
Beijing Engineering Research Center of Aerospace Propulsion Testing Technology and Equipment, Beijing 100074, China

A R T I C LE I N FO A B S T R A C T

Keywords: Rapid development of Liquefied Natural Gas (LNG) industry and LNG vapor's flammable nature have brought up
LNG dispersion many safety issues. Computational fluid dynamics (CFD) modeling provides a powerful tool for studying the LNG
Multiphase model spill, vapor dispersion, as well as other related safety assessments; however, traditional single phase model is
Falcon test unable to describe the LNG pool formation and phase change, which are essential to risk management. In this
Burro test
paper, we propose an integrated multiphase CFD model to simulate the complete spill process (including LNG
Impoundment
release, LNG-water interaction, LNG evaporation and vapor dispersion) based on Falcon-1 and Burro-8 test. The
computed results are compared with the experimental data and other simulation results. It is demonstrated that
the multiphase CFD model provides better results than other approaches in accuracy and fully describing the
LNG spill process. In addition, the effect of an impoundment on LNG spill and dispersion mitigation is also
investigated. The results show that impoundment can significantly control the vapor's dispersion and shorten the
hazard area, especially on lateral direction.

1. Introduction and motivation
The demand for natural gas supply has been significantly growing
over the past few years due to the intense pursuit of clean energy re-
sources worldwide, leading to a rapid development of the Liquefied
Natural Gas (LNG) industry. Growing interests in LNG and pressing
demand for new LNG terminals to be located close to densely populated
area have caused many safety issues on LNG transportation and re-
gasification (Gavelli et al., 2008). Once the containment of LNG on soil,
concrete or water is failed during transportation or storage, cryogenic
liquid will spread to boil vigorously and form a vapor cloud. The vapor
cloud will stay close to the ground and have limited mixing with am-
bient air by its higher density relative to air (Flates et al., 2016).
Considering the flammable nature of the cloud, studies on its dispersion
characteristic and time-dependent plume flow behavior are crucial for
the risk assessment and management in LNG application.
Early research focused on large field trials. In 1980, eight LNG spill
tests on water were conducted under a variety of meteorological con-
ditions, known as the Burro series (Burro Series Data Report, 1982).
These were followed by 10 combustion, rapid transition and dispersion
tests, known as the Coyote series in 1981 (Coyote Series Data Report,
Oct. 1983). In 1987, the Falcon tests were performed to provide a
database on LNG vapor dispersion involving obstacles and to assess the
efficiency of vapor fences for mitigating dispersion hazards (Falcon
Series Data Report, June 1990). Besides the three most famous trials
above, there are many other tests working on this topic including the
Esso tests (Feldbauer et al., 1972), Shell jettison tests (Kneebone and
Prew, 1974), Avocet tests (Koopman et al., 1979), Maplin Sands tests
(Colenbrander et al., 1984) and the BFTF (Cormier et al., 2009).
However, due to the potential risk and high cost in carrying out such
experiments, using computational methods to model the process of LNG
spill and dispersion is becoming more and more important.
Computational techniques for dense gas cloud dispersion generally fall
into four categories depending on the simplicity of models (Robert,
2012) ―box models, slab models, shallow layer model and CFD model.
Among all the models, CFD model can simulate the dispersion char-
acteristics of the cloud in complex terrain or obstructions, and provide
the most detailed variable fields in 3D space with time. So far, many
scholars have used the experimental data to test the effectiveness of
CFD model, and some good results have been obtained (Gavelli et al.,
2009; Jundika et al., 2014; Qi et al., 2010; Timothy and Harri, 2011).
Besides, Batt (Batt et al., 2016), Blocken (Blocken et al. 2007), Meschini
(Meschini et al, 2014) and Ramponi (Ramponi et al., 2012) et al. have
carried out research on atmospheric boundary layer (ABL) flow

∗
Corresponding author. Beijing Institute of Aerospace Testing Technology, Beijing 100074, China.
E-mail address: chch_yu@foxmail.com (C. Yu).

https://doi.org/10.1016/j.jlp.2018.08.001
Received 25 August 2017; Received in revised form 31 July 2018; Accepted 1 August 2018
Available online 03 August 2018
0950-4230/ © 2018 Elsevier Ltd. All rights reserved.
T. Luo et al.
characteristics and the implementation method of stability classes in
CFD code, which are crucial for the correct prediction of heavy gas'
dispersion. Moreover, Li(Li, X. J. et al., 2016), Derudi (Derudi et al.,
2014), and Tauseef (Tauseef et al., 2011) et al. investigated the effect of
obstacles on gas dispersion and Sun (Sun and Guo, 2013) explored the
mitigation of the vapor dispersion with spray water curtains. However,
the vital problem for theses models is that most of these CFD calcula-
tions are based on the single phase hypothesis without accounting for
the LNG pool formation and phase change, where these complex pro-
cesses are usually simplified by using the water pond surface as a nat-
ural gas source term, setting it to the velocity-inlet or mass-flow-inlet.
This method can obtain some satisfactory results since convective mo-
tion within the water body will help maintain the surface at approxi-
mately constant temperature, and the spill into water will result in an
approximately constant generation rate of LNG vapor. In practical, the
size and evaporation rate of LNG pool will keep changing over time,
and the turbulence induced by the spill and evaporation also effects the
vapor dispersion. Having noticed that fact, two new approaches were
proposed in recent studies. In S.G. Giannissi's paper (Giannissi et al.,
2013), a two-phase jet model was proposed by assuming an isenthalpic
expansion process at the spill pipe's outlet. With this assumption, the
spill LNG is mostly gaseous due to the flash vaporization, which is the
key limitation of this model. Ikealumba and Hongwei (2016) proposed
a new CFD model that can capture the spill process, pool formation, and
dispersion of LNG. However, the multiphase model used and the si-
mulation method for water pool are still not clear.
Therefore, it is highly desired to develop an integrated CFD model,
without using any restrictions and simplification, to directly capture
and simulate the whole process including the spill, LNG pool formation,
interaction between the LNG and water, LNG vaporization and disper-
sion. The primary objective of this study is to build such a model, with
the validation using experimental data from Falcon and Burro test. The
predictions are also benchmarked against the modeling results of the
conventional approaches. The second goal of this paper is to assess the
effectiveness of vapor fences for mitigating hazard gas' dispersion.
Although some precious works have focused on this point (Gavelli
et al., 2008; ), it is still meaningful in the viewpoint of the more ac-
curate prediction of the flow field within the impoundment. In present
study, Fluent 15.0, the most commonly-used commercial CFD package
is used as the CFD solver together with a user-defined function (UDF) to
deal with the vaporization process and related thermal effects. This
paper is organized as follows: The Falcon series tests are described in
Section 2. In Section 3, the computation methods are presented in three
sub-sections. Computation zone and mesh method are presented in the
first section; in Section 3.2, detailed model and its implementation in
Fluent are discussed; in Section 3.3, the boundary conditions and initial
conditions in this study are illustrated. In Section 4, the results and
discussion are shown in three sub-sections. Validations of different
models by Falcon-1 test data and corresponding analysis are presented
in the first section; in Section 4.2, additional validation simulation case
of Burro-8 test is presented to further test the multiphase model's re-
liability; in Section 4.3, the impoundment effect on mitigating gas
dispersion is discussed. And in the last Section of 5 are the conclusions.
2. Falcon Series test
The Falcon series test consists of five large-scale LNG spill tests,
conducted by Lawrence Livermore National Laboratories (LLNL) at
Frenchman Flat in Nevada in 1987, sponsored by the Gas Research
Institute and the Department of Transportation. LNG was spilled onto a
rectangular water pond (60 × 40 × 0.76 m) equipped with a water
recirculation system so that the water temperature can be maintained
stable, and introduced through a multi-exit spill “spider” consisting of
four arms (each was approximately 11.6 m long) spaced at 90° intervals
to provide uniform LNG distribution on the pond surface to balance the
total mass vaporization rate with the spill rate (Fig. 1). The ends of the
317
Journal of Loss Prevention in the Process Industries 56 (2018) 316–327
spider arms were made with a 90° turn to point downward and fitted
with a square plate flush on the water level respectively to control the
development of LNG pool and prevent LNG sinking into water (Gavelli
et al., 2009) (Fig. 2). The pond was enclosed by a vapor fence structure
(88 × 44 × 8.7 m). And a “billboard” structure (13.3 m tall and 17.1 m
wide) was located upwind of the pond to simulate the effect of a storage
tank inside the fence on the vapor cloud dispersion.
The spill rate ranged from 8.7 to 30.3 m3/min, spill duration was
from 78 to 301 s, volume ranged from 20.6 to 66.4 m3, average wind
speed at 2 m height from 1.7 to 5.2 m/s and the atmospheric stability
was from neutral to highly stable (Pasquill-Gifford classes D-G). In the
Falcon series, all necessary physical parameters were measured, in-
cluding the air temperature, pressure and humidity, wind speed and
direction, turbulence intensity, heat flux from the ground, and gas
concentration. A group of 77 gas concentration sensors were employed
to record the behavior of the vapor cloud, located at 50, 150, and 250 m
downwind of the fence trailing edge and at 1, 5, 11, and 17 m height. In
this study, Falcon-1 test was selected to validate the model, which
combined the most stable atmosphere condition among the entire
Falcon Series with other challenging conditions: largest spill volume
and spill flow rate, as demonstrated in Table 1.
3. Computation methods
3.1. Computational zone and mesh
The domain of the simulation is oriented in such a way that the x-
direction is horizontal and parallel to the wind, pointing downwind, the
y-direction is horizontal and perpendicular to the wind, and the z-di-
rection is vertical, extending 500 m, 500 m, and 50 m in the x, y, and z
directions, respectively (as shown in Fig. 3). The origin of the domain is
at ground level, in the middle of the downwind fence. In the test, an
angle of −9.3°of average wind direction was measured with respect to
the alignment of the fence. However in this work, wind direction was
assumed parallel to the fence and the x-z plane was the plane of sym-
metry to save computational time. The domain was carefully dis-
credited with hexahedral elements, which are much more computa-
tional efficient than tetrahedral elements, by using finer mesh in the
regions of high flow gradients such as pond, ground and billboard. The
minimum cell size was 0.018 m in the horizontal direction and 0.030 m
in the vertical direction at the LNG source term location to capture flow
details. In summary, the computational domain contained a total of
1 495 058 elements. The above mentioned grid provided grid-in-
dependent results.
3.2. Mathematical formulation
3.2.1. Volume fraction equation
The VOF model, a fixed grid surface-tracking technique, was used to
simulate the multiphase flow in this study. This model solves a single
set of momentum equations shared by the fluids and tracks the volume
fraction of fluids in each computational cell throughout the domain to
model two or more immiscible fluids. In this study, three phases were
included―the water, LNG and mixture vapor. Tracking of the interface
is accomplished by solving a continuity equation for the volume frac-
tion of one phase. The equation for the qth phase has the following
form:
n
1 ∂
[ (αq ρq ) + ∇ . (αq ρq→
vq ) = Sαq + ∑ (m˙ pq − m˙ qp)],
ρq ∂t p=1 (1)
Where Sαq is the mass source term of the qth phase and can be user-
→
defined, vq is the three dimensional velocity vector, ṁpq is the mass
transfer from the pth phase to the qth phase, m ⋅ is the mass transfer
qp
from the qth phase to the pth phase. αq, the qth fluid's volume fraction,
can have the following values:
T. Luo et al. Journal of Loss Prevention in the Process Industries 56 (2018) 316–327

Fig. 1. Schematic of the spill facility, vapor fence, barrier and wind direction (Gavelli et al., 2008).

Fig. 2. Schematic of square plate and water pond (Gavelli et al., 2009).

Table 1
Falcon-1 trial release and weather conditions.
Spill rate (m3/min) 28.7

Spill duration (s) 138

Spill volume (m3) 66.4
Wind Speed (m/s) 1.7 (@ 2 m)
Atmospheric Stability class G
Ambient Temperature (lower-upper/K) 305.35–307.25
Ambient pressure (pa) 90890
Water temperature (pre/post-K) 301.55/295.55

αq = 0, if the cell is empty of the qth fluid

αq = 1, if the cell is full of the qth fluid
0 < αq < 1, if the cell contains the interface between the qth fluid
and other fluids

The volume fraction for the primary-phase is computed based on the

following constraint:
Fig. 3. Computational mesh and its local refinements at water pool and LNG
N
release points.
∑ αq = 1,
q=1 (2)
n

Properties such as density and viscosity in the transport equations

ρ= ∑ αq ρq ,
q=1 (3)
are determined from the volume fraction of the component phases in
each cell. For example, the volume fraction averaged density is com-
puted through the following manner:

318
T. Luo et al. Journal of Loss Prevention in the Process Industries 56 (2018) 316–327

3.2.2. Momentum equation Tsat − Tv

m˙ vl = coeff ⋅α v ρv ,
Tsat (13)
∂ →
(ρ v ) + ∇ . (ρ→→
v v ) = −∇p + ∇ . [μ (∇→
v + ∇→
T →+→
v )] + ρg F,
∂t (4) Where ṁlv and ṁ vl are the mass transfer rates due to evaporation and
→ condensation, respectively, with units of kg/(s·m3). coeff, which can be
Where p is pressure, μ is the dynamic viscosity, ρg is the gravitational
→ interpreted as a relaxation time, is 0.1 for both evaporation and con-
body force, and F is the sum of body forces including the surface densation in this study. Tsat is the saturation temperature. Then, the
tension. source term for the energy equation can be obtained by multiplying the
rate of mass transfer by the latent heat.
3.2.3. Energy equation
∂ 3.3. Boundary conditions and initial conditions
(ρE ) + ∇ (→
v (ρE + p)) = ∇ . (keff ∇T ) + Sh,
∂t (5)
The inflow wind velocity profile was defined using the power law
Where energy E and temperature T are treated as mass-averaged vari- correlation, and the values for turbulence kinetic energy and dissipation
ables and keff is effective thermal conductivity shared by the phases, as were calculated as a function of height by the following equations
shown below. Sh is energy source term and can be user-defined. (14)–(16), which have been declared to be suitable for wind environ-
n ment around buildings (Architectural Institute of Japan, 2004;
∑qa1 αq ρq Eq
E= N
, Yoshihide et al., 2008). The inflow wind temperature was set to
∑qa1 αq ρq (6) 305.35 K.
N vz z
= ( )α ,
k eff = ∑ α q k q, v0 z0 (14)
qa1 (7)
k = 1. 5⋅(vz⋅Iz )2 , (15)
3 3
3.2.4. Species equation ε = 0. 09 4 ⋅k 2 /L, (16)
∂ (ρYi ) →
+ ∇⋅(ρ→
v Yi ) = −∇⋅ Jj + Ri + SI , Where vz is the wind speed at a given height (z), v0 is the wind speed at
∂x (8) the reference height of z0, α is a dimensionless parameter whose value
→ μ ∇T depends upon atmospheric stability category and surface roughness. In
Jj = −(ρDi, m + t ) ∇Yi − DT , i , this study, α was set to 0.16. Iz is the turbulence intensity and L is the
Sct T (9)
turbulence integral scale.
→
Where Yi is the local mass fraction of species i, Jj is the diffusion flux of Since flow details were not known prior to the solution, the outlet
species i, Ri is the net rate of production of species i by chemical re- boundary was set as an outflow boundary, which means that a fully-
action, Si is the source term, Sct is the turbulent Schmidt number, and developed flow is assumed at the boundary and a zero diffusion flux
Di,m and DT,i are the mass diffusion and thermal (Soret) diffusion condition for all flow variables was applied at the boundary cells. The
coefficients for species i, respectively. side and top boundaries were set as symmetrical boundaries, and this is
appropriate with the fact that these boundaries are located far from the
3.2.5. Turbulence equations region and the gradients of flow variables across them would be nearly
Turbulence was modeled using the realizable k–ε model, which is zero. The wall boundary condition was applied to the ground and the
reported to be excellent at modeling heavy gas dispersion in the pre- vapor fence. The source of LNG vapor was modeled with two different
sence of obstacles (Tauseef et al., 2011). The model is presented below. approaches: as a constant area evaporating pool (Estimated pool model)
and direct simulation without any simplifications (multiphase model).
∂ (ρk ) ∂ (ρkuj ) ∂ μ ∂k In the estimated pool model, it was assumed that the LNG evapo-
+ = [(μ + t ) ] + Gk + Gb − ρε − YM + Sk ,
∂t ∂x j ∂x j σk ∂x j rated instantly and covered the entire water pond surface area as soon
(10) as it released onto the water pond. The LNG vapor, assumed to be sa-
turated, entered the domain over the water pond area with a vertical
∂ (ρε ) ∂ (ρεuj ) ∂ μ ∂ε ε2 velocity of 0.053 m/s. This velocity was calculated from the given spill
+ = [(μ + t ) ] + ρC1 Sε − ρC2
∂t ∂x j ∂x j σε ∂x j k+ vε rate (28.7 m3/min), the density of LNG (422.5 kg/m3), the density of
ε saturated vapor methane (1.59 kg/m3) and the assumed pool area
+ C1ε C3ε Gb + Sε ,
k (11) (2400 m2). The temperature of the LNG vapor was set to 110.35 K, the
boiling temperature of LNG at 90 890 pa. In the VOF multiphase flow
In the equations, k is the turbulence kinetic energy, ε is the dis-
simulation, the saturated LNG was released on water through the four
sipation rate, μt is the eddy viscosity, σk and σε are the turbulent Prandtl
pipes with the given total mass rate 202.096 kg/s. The turbulence
numbers for k and ε, respectively. Gk and Gb are the generation of
parameters of the LNG inlet boundary can be determined using in-
turbulence kinetic energy due to mean velocity gradients and buoy-
tensity and hydraulic diameter method. The turbulence intensity of a
ancy, respectively. Sk and Sε are user-defined source terms.
fully developed duct flow can be estimated from the following equation.
Hydraulic diameter DH was set to the inlet pipe diameter.
3.2.6. Mass transfer equations
In this study, the mass transfer between the LNG and its vapor was (
I=0. 16⋅ Re DH )−1/8, (17a)
modeled using the Lee model, a mechanistic model with a physical
basis. In the model, the mass transfer based on the temperature regimes Where ReDH is the Reynolds number of the inlet flow.
is described as follows:
If Tl > Tsat (evaporation): 4. Results and discussion
Tl − Tsat
m˙ lv = coeff⋅αl ρl , 4.1. Validations of different models and analysis
Tsat (12)

If Tv < Tsat (condensation): Fig. 4 shows the predicted methane concentration (v/v) time

319
T. Luo et al. Journal of Loss Prevention in the Process Industries 56 (2018) 316–327

Fig. 4. Measured vs. Predicted gas concentration time histories at 50 m, 150 m, 250 m downwind of impoundment and 1 m above ground.

Fig. 5. Measured vs. Predicted temperature time histories at 50 m and 250 m downwind of impoundment and 1 m above ground.

histories against the experimental data for three sensors at locations 50,
150 and 250 m downwind of the impoundment on the plane y = 0 and
at 1 m above the ground surface. The red line represents the computed
result of multiphase model, while the blue line is the result of estimated
pool model, and the green line is the result of two-phase jet model from
ref (Giannissi et al., 2013). It can be observed that for both the esti-
mated pool model and two-phase jet model the predictions under-
estimate the peak value of vapor concentration with respect to the
experiment. In contrast, the calculation results of multiphase model are
in better agreement to the experimental at all sensors. Similar results
can be seen in Fig. 5 where the results of multiphase model are sig-
nificantly close to the experimental results for the prediction of
temperature.
Effect of turbulence is the main reason for this improvement. Fig. 6
shows the comparison of turbulent kinetic energy between the multi-
phase model and estimated pool model in the impoundment on the
plane z = 4 at time 20s, 60s and 100s. It can be seen that the turbulent
kinetic energy predicted by multiphase model is much larger than that
from the estimated pool model. In the Falcon tests, the impoundment
shields the LNG pool from the ambient wind, so the turbulence is
mainly introduced by the LNG injection, LNG-water interaction and
LNG evaporation process, which allows the dense gas to fully mix with
air, grow in size and ultimately overflow the impoundment. Previous
work (Gavelli et al., 2009, 2008) indicates that the inlet's turbulent

320
T. Luo et al.
Fig. 6. The predicted turbulent kinetic energy contours at 20 s, 60 s, 100 s in the case of multiphase model (left) in comparison with estimated pool model (right).
kinetic energy affects the result of estimated pool model. They quan-
tified the turbulent kinetic energy by analyzing video recordings, and
got some satisfied results by using the approximate turbulent kinetic
energy. However, this method cannot capture the real vortex brought
about by the spill process, and it is impossible to simulate the real
mixing process between the LNG vapor and air in the impoundment and
properties changing occurred earlier in the simulation. Fig. 7 shows the
321
Journal of Loss Prevention in the Process Industries 56 (2018) 316–327
iso-surface of LFL at 100s. It is found that the vapor cloud simulated by
the estimated pool model is limited in the impoundment at this mo-
ment, while the cloud simulated by the multiphase model has been
dispersed out of the impoundment and spread downwind.
Another important reason for this is the temperature effect. LNG
absorbs energy when evaporates. As mentioned in 3.3, the energy
transferred during the phase change can be evaluated through mass
T. Luo et al.
Fig. 7. The predicted LFL (5% v/v) iso-surface colored by turbulent kinetic energy at 100 s in the case of multiphase model (left) in comparison with estimated pool
model (right).
transferred multiplied by the latent heat. During the test, 66 m3 LNG
was spilled, and a total of 33 616 KJ was needed if all of the LNG
evaporated. Such large amount of heat loss will lead to a decrease of
temperature in the impoundment. In the multiphase model, the energy
transferred is considered and added to the energy equation through the
source term. While in the estimated pool model, the inlet temperature is
set directly to the temperature at which LNG is evaporated, and the
effect of phase transition on the surrounding environment is not taken
into account, which leaded to a significant difference in the tempera-
ture prediction at the impoundment exit (top) between the two models.
Fig. 8 shows the temperature variations at three points of impoundment
exit over time. It is seen that the temperature predicted by the multi-
phase model is lower than the estimated model, which makes the vapor
cloud calculated by multiphase model sink more effectively. Hence, the
estimated pool model couldn't accurately predict the diffusion law of
vapor cloud in the vertical direction compared to the multiphase model.
The measurement location in Figs. 4 and 5 located 1 m above the
ground, while the height of impoundment was 8.7 m. Therefore the
estimated pool model underestimated the peak value of vapor.
In addition, the evaporation rate of LNG was changing during the
whole spill process. Fig. 9 shows the variations of LNG's volume in the
computation zone. It can be seen that the amount of LNG is increasing
in the first 60s, indicating the evaporation rate is decreasing. From 60s
to 131s, the amount changed small and reached to its maximum at 131s
when release stopped, the evaporation process ended until about 170s.
Therefore, the release rate can be considered to be substantially equal
to the evaporation rate only from 60s to 131s. In the estimated pool
model, a consistent value throughout the process and closed source
term at 131s were used in the computation, which resulted in a clear
deviation in the calculated results.
The multiphase model have provided an accurate prediction on the
peak value of vapor concentration, nevertheless it failed to exactly si-
mulate the arrival time and departure time of vapor cloud as shown in
Figs. 4 and 5. One possible reason for this is the fact that the water
recirculation system was not simulated since there are no corre-
sponding detail parameters in the experimental report. So the water
kept stable until LNG started to leak. Although lacking of the simulation
of water recirculation, the phenomenon of icing couldn't occur due to
the large heat capacity and thermal conductivity of water. Fig. 10
shows the temperature distribution of iso-surface of water volume
fraction equaled 0.995 at the end of LNG spilling process, which can be
considered as the water surface at that time. It can be observed that the
lowest temperature on the iso-surface is 281 K, while the area-weighted
average temperature is 301.5 K. The water surface's temperature was
still higher than the freezing point, so the icing phenomenon couldn't
occur during the spill. The similar conclusion can be found in the
Woodward's book (Woodward and Pitblado, 2010). In spite of this fact,
the stirring effect of water recirculation system to increase the local
turbulence is ignored. And turbulences caused by the stirring will affect
the subsequent development of vapor cloud. So it could be the first
reason of the deviation. Another possible reason is the error of the
turbulence model itself. In this study, the realizable k–ε model, used the
322
Journal of Loss Prevention in the Process Industries 56 (2018) 316–327
Boussinesq hypothesis, was employed to model the turbulence. This
model tends to over-predict the generation of turbulent kinetic energy
in a normal boundary layer, where wind impinges on a structure. The
excess turbulent kinetic energy over-estimates the degree of mixing
downwind (Gavelli et al., 2008). The RSM model, which involves cal-
culation of the individual Reynolds stresses, is more preferable. How-
ever, this model requires more computation source.
From the above results, it can be seen that the multiphase model can
significantly improve the simulation accuracy compared with the esti-
mated pool model and the two-phase jet model, since it can correctly
predict the turbulent kinetic energy and temperature change during the
LNG spill process. In addition, it is clear that the multiphase model can
be used more widely. For example, when a large mount of LNG is re-
leased to the soil or concrete, the evaporation rate of LNG will decrease
rapidly and a liquid pool will be formed as the ground surface's tem-
perature decreased. So the other two models can't be applied as they are
based on the assumption of LNG rapid evaporation, and this assumption
is only proper when LNG released to the water surface. At the same
time, as the multiphase model is more complex, it is relatively less
robust and the amount of computation will be increased.
4.2. Further validation of multiphase model by Burro-8 test
Although the comparison with the Falcon-1 test has indicated the
better accuracy of multiphase model, just one case study is not enough
to ensure its reliability. So a further validation case of Burro-8 test is
shown here. In this simulation case, the computational domain con-
tained a total of 1 502 204 hexahedral elements with the size of
1000 m × 400 m × 50 m (As shown in Fig. 11). All computation
methods are consistent with the former case expect the inlet boundary
condition. In the Falcon test's simulation, the most effective factor on
the gas diffusion is the vortex behind the impoundment rather than
atmospheric boundary layer as all the measuring points are located at
the back of impoundment (see Section 4.3 for more detail), therefore,
the power law correlation is applied due to its simplicity and suitability
for wind environment around buildings or obstacles. However, when
simulating experiments without obstacles such as Burro or Coyote, the
effect of atmospheric boundary layer and its stability is much more
significant. Therefore, the logarithmic law coordinated with Monin-
Obukhov similarity theory, a better method for maintaining the velocity
profile (Dyer, 1974), is applied in the Burro test simulation to accu-
rately simulate atmospheric boundary layer and its stability. The inflow
wind velocity, temperature profile, values for turbulence kinetic energy
and dissipation were calculated by the following equations (Pontiggia
et al., 2011):
u∗ z z
u (z ) = +( )⋅[ln( ) − ψm ( )]
κ z0 L (17b)
T∗ z z
T (z ) = T0 + ( )⋅[ln( ) − ψh ( )]
κ z0 L (18)
T. Luo et al. Journal of Loss Prevention in the Process Industries 56 (2018) 316–327

Fig. 8. Predicted temperature time histories of two models at three points of impoundment's outlet.

1 2 u∗3
k= u∗ ε=
cμ1/2 (19) κ⋅z (20)

Where for stable conditions (L ≥ 0):

323
T. Luo et al.
Fig. 9. Predicted volume of LNG time histories.
Fig. 10. The temperature distribution of the water's surface at t = 131s.
Fig. 11. Computational zone of Burro-8 test.
z z z seen in the results.
ψm ( ) = ψh ( ) = −5⋅
L L L (21)
In the equations, u* is the friction velocity, z0 is the roughness
height, ψm and ψh are used to account for the effects of atmospheric
stability. T0 and T* are the surface temperature and dynamical tem-
perature, respectively. Where u* = 0.075 m/s, κ = 0.41, z0 = 0.0002 m,
T0 = 310.76 K, T* = 0.029 K and cμ = 0.09 (Ikealumba and Hongwei
(2016)).
Fig. 12 show the predicted methane concentration (v/v) and tem-
perature time histories against the experimental data at location 140 m
downwind of the pond on the plane y = 0 and at 1 m above the ground
surface, respectively. Fig. 13 presents a comparison of vapor cloud
concentration contours at 1 m height and 100s between the simulation
and test. And in Fig. 14, the predicted vapor concentration contour on
the crosswind plane x = 57 m at 100s is compared with the respective
experimental data.
It can be seen that the overall agreements between multiphase
324
Journal of Loss Prevention in the Process Industries 56 (2018) 316–327
model prediction and the experiment are fairly good, so the reliability
and universality of multiphase model have been verified once again.
The biggest deviation between the two results is the vapor con-
centration's distribution in the vertical direction (Fig. 14). The possible
reason for this is that the sloping terrain is not considered for some
former studies ignoring it too (Anay et al., 2007; Ikealumba and
Hongwei, 2016; Xiaobin et al., 2015). Further study considering the
sloping terrain can be carried out in the future.
4.3. Impoundment effect analysis
To analyze the influence of impoundment on vapor dispersion, spill
simulation of Falcon-1 test was repeated without the impoundment and
billboard. Fig. 15 demonstrates the LFL iso-surface of two cases which
colored by height at different moment (20 s, 50 s, 100 s and 120 s). It
can be seen that the vapor cloud simulated by the case without im-
poundment dispersed upwind at 20s, in comparison, there is a smaller
dispersion zone with impoundment. It is observed that the cloud spread
upwind is close to the ground, where the wind velocity is smallest and
its motion is mainly driven by the concentration difference. Therefore
in spite of the existence of wind, the vapor can still disperse upwind,
and the vapor cloud will be blocked by the billboard which is simulated
with impoundment, followed by a smaller diffusion zone. At 50 s, it is
seen that the vapor cloud within the impoundment is beginning to
creep over the impoundment, while the vapor cloud without im-
poundment has formed a great lateral and downwind spread already. At
100s, it can be seen that the cloud has spread out the impoundment,
and the maximum height reached 35 m, while the case without im-
poundment is below 10 m. This is caused by two vertical vortex formed
before and after the billboard in the case with impoundment (Fig. 15)
where the vertical vortex rolled the vapor cloud to a high altitude.
However in the case without impoundment, the high atmospheric sta-
bility environment will weaken the effect of wind dynamics, and the
vertical turbulent mixing will also be reduced, allowing the dense cloud
to travel close to the ground. At 120 s, the vapor cloud extends to 270 m
in the case without impoundment, compared to 118 m that with im-
poundment. And it can also be seen that the impoundment has more
important effect on mitigating the lateral diffusion of vapor cloud. The
lateral diffusion distance of two cases was 215 m (without impound-
ment) and 30 m (with impoundment). This is caused by the lateral
diffusion which is mainly concentrated in the low-speed boundary layer
closed to the ground and the motion of vapor cloud is mainly driven by
the concentration difference. In the case with impoundment, there is a
vortex formed in the low pressure zone after the impoundment
(Fig. 16), and this vortex will enhance transport of vapor's momentum
and energy in the vertical direction. The effect is more obvious for the
boundary layer closed to the ground. Overall, the effect of impound-
ment on vapor cloud diffusion and mitigating LNG spill can be clearly
5. Conclusions
In this work, the Fluent 15.0 software was used to simulate the
Falcon-1 test, firstly. Two different source modeling methods were
applied and evaluated: a) conventional estimated pool model, b) new
multiphase model. Predictions of the two models as well as a two-phase
jet model were compared with the experimental results.
Both the estimated pool model and two-phase jet model under-
estimate the vapor's peak concentration with respect to the experiment.
In contrast, the calculation results of multiphase model are in better
agreement with the experimental at all sensors. The significant im-
provements are obtained due to the more accuracy predictions on the
turbulence effect and temperature variations brought by the phase
change and other processes in the multiphase model. Nevertheless, this
model fails to estimate the arrival time and departure time of the vapor
cloud. One possible reason is that the simplification of recirculation
T. Luo et al. Journal of Loss Prevention in the Process Industries 56 (2018) 316–327

Fig. 12. Measured vs. Predicted gas concentration (a) and temperature (b) time histories at 140 m downwind of the pond on the plane y = 0 and at 1 m height.

Fig. 13. Measured vs. Predicted vapor cloud concentration contours at 1 m height and 100s.

Fig. 14. Measured vs. Predicted vapor concentration contour on the crosswind plane x = 57 m at 100s.

system made in present study, so the turbulence caused by the stirring
the pool couldn't be simulated. Another possible reason is the error of
the turbulence model itself.
Besides, a simulation of Burro-8 test used multiphase model was
also carried out to test the model's reliability and universality. And
results proved the accuracy of multiphase model once again.
At last, the effect of an impoundment on controlling an LNG spill
and dispersion was also investigated. Results showed that the im-
poundment had great effect on vapor cloud's dispersion mitigation both
in the downwind direction and lateral direction, and especially on
lateral direction. It is worth mention that the cloud will spread higher
for the vertical vortex caused by the impoundment and billboard.
These results show that the multiphase model method is a prefer-
able method to study the LNG spill and dispersion processes. The

325
T. Luo et al. Journal of Loss Prevention in the Process Industries 56 (2018) 316–327

Fig. 15. The predicted LFL (5% v/v) iso-surface colored by height 20 s, 50 s, 100 s, 120s after the start of the spill (left = baseline case, right = no-barrier case).

326
T. Luo et al. Journal of Loss Prevention in the Process Industries 56 (2018) 316–327

atmospheric boundary layers with commercial CFD software for use in risk assess-
ment. Chemical Engineering Transactions 48, 61–66.
Blocken, B., Carmeliet, J., Stathopoulos, T., 2007. CFD evaluation of wind speed condi-
tions in passages between parallel buildings – effect of wall-function roughness
modifications for the atmospheric boundary layer flow. J. Wind Eng. Ind. Aerod. 95
(9–11), 941–962.
Burro Series Data Report, 1982. LLNL/NWC Report No.UCID-19075 v.1 2 Lawrence
Livermore National Laboratory, Berkeley, CA.
Coyote Series Data Report, Oct. 1983. LLNL/NWC, UCID-19953 1 2.
Colenbrander, G.W., Evans, A., Puttock, J.S., May, 1984. Spill Tests of LNG and
Refrigerated Liquid Propane on the Sea’ Maplin Sands 1980, Dispersion Data Digest,
Trial 27. Shell Research Ltd, Thornton Research Center Report TNER. 84.028.
Cormier, B.R., Qi, R., Yun, G.W., Zhang, Y., Mannan, M.S., 2009. Application of com-
putational fluid dynamics for LNG vapor dispersion modeling: a study of key para-
meters. J. Loss Prev. Process. Ind. 22, 332–352.
Dyer, A.J., 1974. A review of flux-profile relationships. Boundary-Layer Meteorol. 7,
363–372.
Fig. 16. Schematic vectors on centerline. Derudi, M., Bovolenta, D., Busini, V., Rota, R., 2014. Heavy gas dispersion in presence of
large obstacles: selection of modeling tools. Ind. Eng. Chem. Res. 53, 9303–9310.
Feldbauer, G.F., Heigl, J.J., McQueen, W., Whipp, E.H., May, W.G., 1972. Spills of LNG on
analysis also shows the effect of an impoundment on controlling LNG Water: Vaporization and Downwind Drift of Combustible Mixtures. ESSO Research
spill and vapor dispersion, which will be of great importance for the and Engg. Co Report no EE61E-72.
construction of LNG terminals. Falcon Series Data Report, June 1990. Gas Research Institute, 1987 LNG Barrier
Verification Field Trials. GRI Report No.89/0138, Chicago, IL. .
Flates, J., Santos, R.R.C., Neto, F.F., Francesconi, A.Z., Simoes, V., Savio, S.V., 2016. An
Acknowledgments alternative CFD tool for gas dispersion modeling of heavy gas. J. Loss Prev. Process.
Ind. 44, 583–593.
Gavelli, F., Bullister, E., Kytomaa, H., 2008. Application of CFD (Fluent) to LNG spills into
The fund of the State Key Laboratory of Technologies in Space geometrically complex environments. J. Hazard Mater. 159, 158–168.
Cryogenic Propellants, the People's Republic of China, supported this Gavelli, F., Chernovsky, M.K., Bullister, E., Kytomaa, H.K., 2009. Quantification of
work. (No. SKLTSCP1514). source-level turbulence during LNG spills onto water. J. Loss Prev. Process. Ind. 22,
809–819.
Giannissi, S.G., Venetsanos, A.G., Markatos, N., Bartzis, J.G., 2013. Numerical simulation
Nomenclature of LNG dispersion under two-phase release conditions. J. Loss Prev. Process. Ind. 26,
245–254.
Ikealumba, W.C., Hongwei, W., 2016. Modeling of liquefied natural gas release and
αq the qth fluid's volume fraction
dispersion: incorporating a direct computational fluid dynamics simulation method
t time(s) for LNG spill and pool formation. Ind. Eng. Chem. Res. 55, 1778–1787.
ρ density(kg·m−3) Jundika, C.K., Agus, P.S., Arun, S.M., 2014. CFD simulation of methane dispersion and
innovative methane management in underground mining faces. Applied
p pressure(pa)
Mathematical Modeling 38, 3467–3484.
μ dynamic viscosity(Pa·s) Kneebone, A., Prew, L.R., 1974. Shipboard jettison test of LNG onto the sea. In:
→
F sum of body forces including the surface tension(N) Proceedings on the 4th International Conference on LNG, pp. 1–25 Algiers.
→
v velocity vector(m·s−1) Koopman, R.P., Bowman, B.R., Ermak, D.L., 1979. Data and Calculation of Dispersion on
5m3 LNG Spill Tests. Lawrence Livermore Laboratory.
E total energy(J) Meschini, D., Busini, V., Ratingen, S.W., Rota, R., 2014. Modeling of pollutant dispersion
T temperature(K) in street canyon by means of CFD. J. Chem. Ind. Eng. 58 (8), 1967–1972.
keff effective thermal conductivity(W·m−2·K−1) Pontiggia, M., Labriele, G., Busini, V., Derudi, M., Aiba, M., Scaioni, M., 2011. CFD model
simulation of LPG dispersion in urban areas. Atmos. Environ. 45 (24), 3913–3923.
Yi mass fraction for species i Qi, R.F., Ng, D., Cormier, B.R., Mannan, M.S., 2010. Numerical simulations of LNG vapor
Tsat saturation temperature(K) dispersion in Brayton Fire Training field tests with ANSYS CFX. J. Hazard Mater. 183,
Tl liquid temperature(K) 51–61.
Ramponi, R., Blocken, B., 2012. CFD simulation of cross-ventilation for a generic isolated
ṁlv mass transfer rates from liquid to gas phase(kg·s−1) building: impact of computational parameters. Build. Environ. 53, 34–48.
ṁ vl mass transfer rates from gas to liquid phase(kg·s−1) Robert, N.M., 2012. CFD modeling of dense gas cloud dispersion over irregular terrain.
Sαq user-defined mass source term of the qth phase J. Wind Eng. Ind. Aerod. 104–106, 500–508.
Sun, B., Guo, K., 2013. LNG accident dynamic simulation: application for hazardous
LFL lower flame limit
consequence reduction. J. Loss Prev. Process. Ind. 26, 1246–1256.
u* friction velocity(m·s−1) Timothy, L.M., Harri, K.K., 2011. The effect of turbulence on the rate of evaporation of
z0 roughness height(m) LNG on water. J. Loss Prev. Process. Ind. 24, 791–797.
Tauseef, S.M., Rashtchian, D., Abbasi, S.A., 2011. CFD-based simulation of dense gas
T0 surface temperature(K)
dispersion in presence of obstacles. J. Loss Prev. Process. Ind. 24, 371–376 2011.
T* dynamical temperature(K) Woodward, J.L., Pitblado, R.M., 2010. LNG risk Based Safety: Modeling and Consequence
Analysis. John Wiley & Sons, Inc., Hoboken, New Jersey Published simultaneously in
References Canada: AIChE.
Xiaobin, Z., Jingfeng, L., Jiakai, Z., Limin, Q., 2015. Computational fluid dynamics study
on liquefied natural gas dispersion with phase change of water. Int. J. Heat Mass
Anay, L.H., Ronald, P.K., Donald, L.E., 2007. On the application of computational fluid Tran. 91, 347–354.
dynamics codes for liquefied natural gas dispersion. J. Hazard Mater. 140, 504–517. Yoshihide, T., Akashi, M., Ryuichiro, Y., 2008. AIJ guidelines for practical applications of
Architectural Institute of Japan, 2004. AIJ Recommendations for Loads on Buildings. CFD to pedestrian wind environment around buildings. J. Wind Eng. Ind. Aerod. 96
Architectural Institute of Japan, Tokyo, Japan. (10), 1749–1961.
Batt, R., Gant, S.E., Lacome, J.M., Truchot, B., 2016. Modelling of stably-stratified

327