CFD modeling of hydrogen dispersion under cryogenic release conditions

ww.sciencedirect.com

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 3

Available online at w

ScienceDirect

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

CFD modeling of hydrogen dispersion undercryogenic release conditions

S.G. Giannissi a,b,*, A.G. Venetsanos a, N. Markatos b, J.G. Bartzis c

a Environmental Research Laboratory, National Center for Scientific Research Demokritos, Aghia Paraskevi, Athens,

15310, Greeceb National Technical University of Athens, School of Chemical Engineering, Department of Process Analysis and Plant

Design, Heroon Polytechniou 9, 15780, Athens, Greecec Department of Mechanical Engineering, University of Western Macedonia, Parko Agiou Dimitriou,

West Macedonia, 50100 Kozani, Greece

a r t i c l e i n f o

Article history:

Received 24 January 2014

Received in revised form

7 July 2014

Accepted 10 July 2014

Available online 15 August 2014

Keywords:

LH2 dispersion

Two phase jet

Humidity

Slip velocity

Wind direction

ADREA-HF

* Corresponding author. Tel.: þ30 210 650341E-mail addresses: [email protected]

(N. Markatos), [email protected] (J.G. Bartzishttp://dx.doi.org/10.1016/j.ijhydene.2014.07.00360-3199/Copyright © 2014, Hydrogen Ener

a b s t r a c t

The use of hydrogen as a fuel should always be accompanied by a safety assessment

concerning the case of an accidental release. To evaluate the potential hazards in a spill

accident both experiments and simulations are performed. In the present work, the CFD

code, ADREA-HF, is used to simulate the liquefied hydrogen (LH2) spill experiments (test 5,

6, 7) conducted by the Health Safety Laboratory (HSL). Two horizontal releases, the one

along the ground and the other one at a distance above the ground, and one vertical release

are examined with spill rate 60 lt/min. The main focus of this study is on the presence of

humidity in the atmosphere and its effect on the vapor dispersion. When humidity is

present is cooled, condenses and freezes due to the low prevailing temperature (~20 K near

the release), and releases heat. In addition, during the release hydrogen droplets are

formed due to mechanical and flashing break up, and water droplets and ice crystals due to

humidity phase change. Therefore, two models are tested: the hydrodynamic equilibrium

model, which assumes that the phases are in thermodynamic and kinematic equilibrium

and the non hydrodynamic equilibrium model (slip model), which assumed that the

phases are in thermodynamic equilibrium but they can obtain different velocities. The

fluctuating wind direction was also taken into account, since it greatly affects the hydrogen

dispersion. The computational results are compared with the experimental measure-

ments, and it is concluded that humidity along with the slip effect influences the buoyancy

of the cloud to a great extent. The best simulation case (humidity and slip effect) is

consistent with the experiment for all three tests for the majority of the sensors.

Copyright © 2014, Hydrogen Energy Publications, LLC. Published by Elsevier Ltd. All rights

reserved.

6.ritos.gr (S.G. Giannissi), [email protected] (A.G. Venetsanos), [email protected]).42gy Publications, LLC. Published by Elsevier Ltd. All rights reserved.

Nomenclature

Xi Cartesian j co-ordinate (m)

ui i component of velocity (m s�1)

p pressure (Pa)

gi gravity acceleration in the i-direction (m s�2)

qk mass fraction of k component (dimensionless)

t time (s)

H Enthalpy (J kg�1)

Sct,Prt turbulent Schmidt and Prandtl number

(dimensionless)

N particles' number (dimensionless)

D;D particle's diameter , mean diameter (m)

N0 Marshal Palmer constant (m�4)

fdrag drag function (dimensionless)

Greek

dij Kronecker delta

ε turbulent energy dissipation rate (m2 s�3)

m,mt laminar and turbulent viscosity (kg m�1 s�1)

r mixture density

d depth (m)

a thermal diffusivity (m2 s�1)

Subscripts

nv non vapor

l liquid

sl slip

i n t e rn a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 315852

Introduction

Hydrogen is a competitive fuel in the energy market due to its

high energy carrier and low emissions. However, its wide

flammability range brings up safety issues. A practice for

hydrogen storage and handling is its liquefaction under

pressure and low temperature. In case of a fracture in the

cryogenic tank, LH2 is spilled forming a cryogenic pool and a

dense vapor cloud. The jet release is two phase and prediction

of both vapor dispersion and liquid pool evaporation and

spreading is required. Computational Fluid Dynamics (CFD)

codes are usually used to simulate such complicated cases.

In the past, several experiments have been performed

related to hydrogen dispersion under cryogenic release con-

ditions in open environment [1e7]. The most recent related

experiments are the HSL experiments [7] that were conducted

by the Health and Safety Laboratory in 2010. In the present

work, these experiments (HSL experiments) have been simu-

lated with the help of ADREA-HF code.

ADREA-HF is a CFD code that has been validated against

hydrogen dispersion and other liquefied fuels such as natural

gas, or denser than air gases [8e18], and it is considered to be a

useful and reliable tool for such applications.

During HSL experiments LH2 was released above a concrete

pad at a fixed rate of 60 lt/min. Four tests were performed with

different release directions and duration. Wind speed and di-

rection were measured at the edge of the pad 2.5 m from the

ground. Sensors thatmeasured the temperaturewere deployed

in line with the release at several distances and heights. The

hydrogen concentration was derived by the temperature data

assuming adiabatic mixing. This approach is considered valid,

sinceonce the cloud is liftedoff the ground theair andhydrogen

mixing is adiabatic within 1% [1]. Thermocouples inside the

groundmeasured also the underground temperature.

The parameters that influence the hydrogen dispersion are

various, especially when the release takes place in an open

environment. Apart from the release conditions significant

role play the weather conditions (wind speed and direction,

ambient humidity), the atmospheric conditions (neutral, sta-

ble, unstable), the ground terrain, etc.

Previous work has simulated tests 6 and 7 of the HSL ex-

periments using the CFD software FLACS [19]. In that work,

the condensation or solidification of the air (nitrogen and

oxygen) was examined. In their computations of the two

phase jet the dispersed and continuous phases were assumed

to be in thermodynamic and hydrodynamic equilibrium. At

the source they tested four different flashed hydrogen volume

fractions and the case with pure gas flow, and it was

concluded that for both tests the more coherent results were

obtained with flashed volume fraction equal to 99%. The

condensation and solidification of air affected the flow field in

test 6, by releasing heat close to the ground, generated an

upward velocity and made the cloud more buoyant. In test 7

the effect was not significant, because the area where

condensation/solidification is occurred is located close to the

release and seems not to influence the flow away from it.

When humidity is present in the atmosphere, apart from the

air phase change, thewater is also condensed and solidified due

to thevery lowprevailing temperature.Moreover, theareawhere

humidity phase change is occurred is much more extended

than the area where air phase change is occurred. Therefore,

the present work focuses on the effect of ambient humidity on

vapor dispersion and on the computational results. Test 5, 6

and 7 of the HSL experiments are chosen to be simulated.

Moreover, during the release hydrogen droplets are formed

due to mechanical and flashing break up. The humidity

condensation and solidification produces also water droplets

and ice crystals.These “particles” canbeassumedeither tohave

the same velocity as the vapor mixture or to obtain different

velocities, because of the gravitational acceleration. The differ-

ence between the phases' velocity is called slip velocity.

In the simulation process two cases were examined: one

case that solves assuming both thermodynamic and hydro-

dynamic equilibrium and one case that solves assuming

thermodynamic equilibrium but non hydrodynamic equilib-

rium. With the non hydrodynamic equilibrium model the

vapor and non vapor phase of both hydrogen and humidity

are allowed to develop different velocities. The effect of the

slip velocity on the flow field is studied.

In all simulation cases the fluctuating wind direction was

taken into account, since it was observed that it results in

many hydrogen concentration oscillations.

Description of the experiments

The HSL experiments were conducted by the Health and

Safety Laboratory in 2010. During these experiments LH2 was

spilled above a concrete pad in open environment. A number

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 3 15853

of tests were performed in which LH2 was released either

horizontally along the ground or at 860 mm above it, or,

vertically downwards 100 mm above ground. The nominal

storage pressure in the tankwas 1 bar and the nominal release

pressure was 0.2 bars for all test cases.

The release rate was 60 lt/min in each test and the release

duration was ranged from 265 to 561 sec. The wind speed

varied between 1.4 and 3.35 m/s at 2.5 m height. The wind

direction fluctuated around the average wind direction, which

was approximately in line with the release for each test. The

relative atmospheric humidity varied from 64 to 87%.

For test 5, 6, and 7 sensors were placed in line with the

release at 1.5, 3, 4.5, 6, 7.5 m downwind the spill point, and

0.25, 0.75, 1.25, 1.75, 2.25, and 2.75m above the ground. For test

10, thirty sensors were positioned at a range of heights and

distances around the spill point.

Simulation process

ADREA-HF code and mathematical equations

ADREA-HF code is a three dimensional time dependent finite

volume CFD code that has been successfully used for predic-

tion of pollutants and hazardous gases dispersion in open and

closed environments. The main focus has been given to pre-

diction of hydrogen dispersion for both cryogenic and com-

pressed releases [11e14].

ADREA-HF software solves the Navier-Stokes equations for

the mixture using homogeneous equilibrium model. Coupled

with the Navier-Stokes equations the mass and energy con-

servation equations are solved. There is also the option to

solve assuming thermodynamic equilibrium but non hydro-

dynamic equilibrium permitting the non vapor phase and the

vapor phase to obtain different velocity by using extra slip

terms in the conservation equations. It is assumed that only

w-component (along z-direction) of slip velocity and only in

the vertical direction (z-direction) is significant, due to the

gravitational acceleration, and therefore only the respective

terms are added in the equations.

Themixture conservation equations are the following. The

last term on the right side of the equations is the slip term:

vr

vtþ vrui

vxi¼ 0 (1)

vrui

vtþ vrujui

vxj¼ �vp

vxiþ v

vxj

�ðmþ mtÞ

�vui

vxjþ vuj

vxi

��þ rgi

� d3iv

vz

�rqnvws

�1� qnv

�wsl

�(2)

vrqk

vtþ vrujqk

vxj¼ v

vxj

�mt

Sct

vqk

vxj

�� vrqnvwsl

�1� qk

�vz

(3)

vrHvt

þ vrujH

vxj¼ v

vxj

�mt

Prt

vHvxj

�þ Dp

Dtþ v

vz

�rqnvwslðH� HnvÞ

�

� qnvwsl

�1� r

rnv

�vpvz

(4)

For the turbulence modeling there are available the RANS

turbulence models of zero equation (e.g. mixing length), one

equation and two equations (e.g. standard k-ε, RNG k-ε), and

the LES model [20]. In the present work, k-ε model [21] is used

with extra source terms for buoyancy [8], [22]. This model is

chosen because of its credibility, its simplicity and its limited

computational requirements. The k-ε model is a widely used

and validated turbulence model. Moreover, the model with

the extra buoyancy terms is suitable for environmental flows,

such as pollutant dispersions.

The ADREA-HF code has several slip models incorporated

in order to calculate the vertical slip velocity. In the present

study, the slip model that is used is based on the Ogura and

Takahashi empirical relation for water [23], which is suitable

for rain precipitation. However, it has been modified in order

to take into account all the flow regimes (laminar, turbulent,

transition) and all liquids.

Analytically, the slip velocity is calculated assuming

spherical particles (droplets or ice crystals) distributed in

diameter according to Marshall Palmer distribution:

N ¼ N0e�DD (5)

where N0 ¼ 8$106m�4.

Themean diameter is calculated with the help of the liquid

mass, solving the following equationwith respect of themean

diameter:

rql

rl¼

Z∞

0

Np

6D3dD ¼ pN0D

4(6)

The vertical liquid flux is derived by the following

relationship:

rqlwsl ¼Z∞

0

Nrlp

6D3wsldD (7)

The integral in the right side of the above equation is

calculated numerically. The slip velocity in the integral is

calculated with the help of the particle's diameter. The regime

can be either laminar or turbulent dependent on diameter

size. The equation is:

wsl ¼ 118mfdrag

ðrl � rÞgD2 (8)

The drag function, fdrag, defines the flow regime, and its

value is given by Ref. [24]:

fdrag ¼�1þ 0:15Re0:687

p ; Rep � 10000:01833Rep ; Rep > 1000

(9)

Using this model there is no need in estimating the parti-

cle's size which is very important, in cases where there are no

such data available.

Modeling approach

Test 5, 6 and 7 of theHSL experiments are simulated. In all tests

the source is modeled as a two phase jet. The percentage of

vapor phase is calculated assuming isenthalpic expansion from

the storage pressure to the release pressure, i.e. from 2 bars to

1.2 bars (absolute pressure). The spill velocity is calculated

using the source area and themixture density of the two phase

i n t e rn a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 315854

jet, and it is approximately 6 m/s for each test. The vapor vol-

ume fractionwas equal to 71%. This is less than the vapormass

fraction (99%) that was concluded to be the best case in Refs.

[19], but since no accurate data are available the isenthalpic

expansion assumption appeared to be more consistent to use.

During cryogenic releases both vapor and liquid hydrogen

coexist in the domain. The hydrogen evaporates as it absorbs

heat by its surrounding. The main heat sources are the

ambient air and the ground. The computational code solves

for the mixture and not for each phase separately. Therefore,

for the phase distribution in the mixture is assumed that the

liquid phase appears when the mixture temperature is equal

or lower than the dew temperature of the mixture [25], which

is calculated using the Raoult's law. In case that the mixture

temperature drops below the freezing point of a component,

then the solid phase of that component appears.

The liquid and solid phases are dispersed with the vapor

phase in each cell. If liquid hydrogen is calculated in a bound-

ary cell (cell on the ground), it is assumed that the area of that

cell projected to the ground is the area of the liquid pool. In this

way the pool formation and spreading is computed.

The presence of atmospheric humidity makes the cloud

more buoyant due to the heat liberation by the water'scondensation and solidification. The effect of the ambient

humidity on the vapor dispersion is examined in the present

study. The air humidity is taken into account by solving an

additional mass fraction conservation equation for the water.

The initial watermass fraction of the air is calculated from the

measured relative humidity, and it is 0.00529 for test 5 and 6

and 0.00532 for test 7.

In these experiments the fluctuating wind direction seems

to affect to a great extent the hydrogen concentration, which

exhibits many oscillations. Therefore, a transient wind di-

rection was imposed as initial and inflow condition. More

specifically, a transient v-component (along y-axis) of wind

velocity was imposed. The v-component is computed by

multiplying thewind speedmagnitude by the sine of the angle

of the wind direction in each time.

Analytically, in the simulation process the wind field was

calculated as following: One dimensional steady problem was

solved, in order to obtain the initial wind profile based on the

average wind speed and direction. The momentum equation

for the horizontal velocity component (u-component) is

solved and for the turbulence the k-εmodel is used. On the top

domain a given value boundary condition for the horizontal

velocity component is set. Several boundary values were

tested until the average measured wind speed at 2.5 m height

Table 1 e Release and weather conditions for test 5, 6, and 7.

Te

Spill diameter (mm) 26.6

Source height (mm) 3.36

Source direction Horizontal

Release rate (kg/sec) 0.07

Release duration (sec) 248

Average wind speed (@ 2.5 m) (± Standard deviation) 2.675 ± 0.0

Average wind direction (@ 2.5 m) (± Standard deviation) 291 ± 15.5

Average ambient temperature (@2.5 m) 283

Ambient relative humidity (%) 68

(Table 1) is predicted. The calculated profile was set as initial

wind field for the whole domain and as inflow values for all

variables, except for the v-component of velocity (the

component in the y-direction). The v-component of velocity

was transient and uniform along z-direction and its values

were set on boundaries based on the filteredmeasured values.

Specifically, the experimental values of v-component were

filtered using the Savitsky-Golay smoothing filter. Savitsky-

Golay smoothing filter can be thought as a generalized mov-

ing average and its advantage is that it preserves higher mo-

ments. Details about the method can be found in Ref. [26]. In

the present study, a second order polynomial fits each data

point with the help of 10 neighbors before and 10 neighbors

after the point and replaces its value with the value that the

polynomial calculates. This value is the imposed on the

boundary value at the respective time. In Fig. 1 the measured

and the filtered v-component for each test are depicted.

Finally, for temporal discretization the first order fully

implicit scheme was used. For the convective terms the first

order upwind scheme was used. The initial time step was

10�4, and then gradually increased using a Courant number

(CFL) restriction.

Computational domain and grid

The computational domain extended 15 m in the x-direction

(from 5 m upwind to 10 m downwind of the release point),

20 m in the y-direction (crosswind) and 10m in the z-direction

in the three tests. The grid is Cartesian and consists of 146

034 cells (61 � 63 � 38), 140 034 cells (61 � 63 � 38) and

180 621 cells (61 � 63 � 47) in test 5, 6, and 7 respectively.

Refinement points are close to the release and on the ground.

The minimum cell size was equal to 0.05 m at the adjacent to

the source cells, and was kept approximately constant for a

small area near and downwind the release. Then, the grid

expanded with a factor of 1.12. Denser grids were tested with

no significant effect on the results but with demand of higher

computational time. Therefore, the abovementioned grid

resolution was used which provides independent results and

with relatively small computational cost. In that way it is

possible to simulate and compare several different cases

within logical time limits.

Boundary conditions

The west domain (upwind the release) was the inlet boundary.

At the inlet boundary inflow (Dirichlet) conditionswere imposed

st 5 Test 6 Test 7

26.6 26.6

100 860

(x-direction) Vertical (z-direction) Horizontal (x-direction)

0.07 0.07

556 305

9 3.35 ± 0.95 3.07 ± 0.82

294.63 ± 13.4 294.48 ± 14

283 284

68 64

measurementfiltered

-3

-2

-1

0

1

2

3

244 294 344 394 444 494 544 594time (sec)

)s/m(

yticolev-v

-3

-2

-1

0

1

2

3

400 500 600 700 800 900 1000 1100 1200time (sec)

)s/m(

ytico lev- v

-3

-2

-1

0

1

2

3

400 450 500 550 600 650 700 750 800 850time (sec)

v-ve

loci

ty (m

/s)

Fig. 1 e The measured and the computed filtered v-

component of the wind velocity for test 5 (top), test 6

(center) and test 7 (bottom).

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 3 15855

with values provided by the 1D steady state problem for all

variables, except for the v-component of the velocity whose

value was imposed as described in Section Modeling approach.

The east domain (downwind the release) was outlet

boundary. The outlet boundaries were prescribed by applying

zero gradient for all variables except components' mass frac-

tion and temperature, for which either a zero gradient bound-

ary condition was applied if outflow occurs or a given value

boundary condition (equal to the initial value) if inflow occurs.

The boundary condition for themass fraction and temperature

is applied in order to ensure that there will be no hydrogen flow

back in the domain. The zero gradient boundary condition is a

common practice for external flows, and when the boundary is

located away from geometrical disturbances, since the flow

reaches a fully developed state [27], which is our case.

The top is free boundary, since the domain was extended

enough; therefore, symmetry boundary condition was applied.

The normal component of velocitywas set equal to zero and for

all other variables zero gradient boundary condition was

imposed. The components' mass fraction and temperature

boundary conditions were the same as the outlet boundary.

The north and south side boundaries were either inflow or

outflow boundaries depending on the wind direction, and the

respective boundary conditions were imposed during the

simulation process.

On the bottom domain (ground) a wall boundary condition

was applied with roughness length 1 mm. The temperature at

the bottom boundary is calculated by solving a transient one

dimensional temperature equation in the underground. The

depth of the underground should be taken deep enough to be

able to assume that the temperature in the bottom of the

underground is constant in time equal to the initial temper-

ature. The minimum depth which the above assumption is

applicable is calculated by the following relationship:

d ¼ 4ffiffiffiffiffiat

p; (10)

In this work, the underground consists of concrete 20 cm

thick. No experimental data for the concrete properties were

available. The properties of the concrete used in the simulation

were density r ¼ 2371 kg/m3, hat capacity cp ¼ 880J/kg K and

thermal conductivity l¼ 1.13W/mK.Using these properties the

minimumdepth required is 4.6 cm, 7 cm and 5 cm for the test 5,

test 6 and test 7 respectively. So, the assumption of constant

temperature at the bottom of the concrete pad is consistent.

Results and discussion

Three cases weremodeled and compared to each other and to

the measurements for all tests. Table 2 shows these cases.

For the sake of brevity, from this point each case will be

mentioned with its number, instead of using the whole

expression.

Fig. 2 depicts the predicted temperature contours of test 5, 6

and 7, 12 sec after start of the release and on plane y ¼ 0. The

prediction is for the case with dry air (case 1). The area where

the temperature is below the freezing point of thewater (yellow

area) is much more extended than the area which is below the

boiling point of oxygen and nitrogen. Therefore, the effect of

humidity freezing on the dispersion is expected to be higher

than the air freezing. Moreover, even though the mass fraction

of oxygen and nitrogen in air is much higher than the mass

fraction of humidity, the water specific latent heat of vapor-

ization (2260 kj/kg) is approximately 11 times the nitrogen

specific latent heat of vaporization (198.57 kj/kg) and the oxy-

gen one (213.125 kj/kg). These two factors combined lead to the

conclusion that the effect of humidity phase change would be

more significant than the air‘s phase change, and, therefore,

only its effect on vapor dispersion is examined in this study.

The comparison is presented in three different ways. First

the time histories of the hydrogen concentration are displayed.

Table 2 e Modeled cases.

No case Description

1 Dry air e thermodynamic and hydrodynamic

equilibrium model

2 Humid air e thermodynamic and hydrodynamic

equilibrium model

3 Humid air e thermodynamic and non hydrodynamic

equilibrium model (slip model)

Fig. 2 e The temperature contours on plane y ¼ 0 and 12 sec after start of the release for test 5 (left), test 6 (center) and test 7

(right) for case 1.

i n t e rn a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 315856

However, the presence of fluctuating wind direction in

conjunctionwith the limited informationabout thewindprofile

in order to simulate it accurately, made the comparison based

only on the time histories rather difficult. Therefore, additional

comparisons were deemed necessary. Specifically, the time

averaged hydrogen concentration along with the maximum

hydrogen concentration downwind the release and at different

heights from the ground are considered more suitable for

comparison. The time averaged concentration is defined as the

average of the total dosage from the arrival time to the leaving

timeof thecloud.Total dosage is the totalmassofhydrogenthat

reaches each sensor, while the arrival and the leaving time

define the interval timeafter the releasewhen5%and95%of the

total dosage respectively reaches the sensor [28].

Further evaluation of the predictions was made by per-

forming statistical analysis. Two statistical indicators were

used in this work, the geometric mean bias (MG) and geo-

metric mean variance (VG). MG measures the relative mean

0.0

0.1

0.2

0.3

0.4

240 290 340 390 440 490time (sec)

)v/v(noitartnecnoc

nego rdyh

experiment case1 (dry air)

Sensor (7.5,0,0.25)

Fig. 3 e The predicted and measured time histories for test 5 at

The exact location of the sensors (x, y, z coordinates (m), in respe

ground level at z ¼ 0) is shown in the parenthesis.

bias, whilst VG measures the relative scatter, and they both

have ideal value of 1. MG values greater than 1 show over-

prediction of the model.

The equations for the above measures are the followings.

The overbar denotes the average over the entire dataset.

MG ¼ exp

24ln

�Cp

Co

�35 (11)

VG ¼ exp

24ln

�Cp

Co

�235 (12)

Test 5

Fig. 3 shows the measured and predicted time histories on

plane y ¼ 0 at 7.5 m downwind the release for all modeled

case2 (humid air) case2 (humid air-slip model)

0.00

0.04

0.08

0.12

240 290 340 390 440 490time (sec)

)v/v(noitartnecnoc

negordyh

Sensor (7.5,0,2.25)

7.5 m downwind the release and at two different heights.

ct with the release point located at x ¼ 0 and y ¼ 0, and the

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 3 15857

cases. The concentration oscillations are due to the wind di-

rection oscillations, as mentioned in Section Modeling

approach. The fluctuating wind direction affects the vapor

dispersion to a great extent, and Fig. 3 shows the significance

in modeling the wind direction. Even though the average

measured directionwas in line with the release, the cloudwas

able to follow the instantaneous direction and deviate from

the centreline direction.

The case with the non hydrodynamic equilibrium model

(case 3) is in better agreement with the experiments than the

other two cases. The case with the dry air (case 1) and the case

with humid air and the hydrodynamic equilibrium model

(case 2) highly overestimate the concentration at the lower

sensor, while they both underpredict it at the high sensor. The

concentration oscillations may not fall exactly on the

measured ones, but one should keep in mind the difficulty of

reproducing the exact wind profile, especially in case with

high wind variations and with limited available weather data.

More weather sensors would provide valuable information,

and consequently could improve the performance of the

simulation.

The average hydrogen concentration, as was defined pre-

viously and the peak hydrogen concentration are depicted in

Fig. 4. In these diagrams it is shown that the cases 1 and 2

overpredict both average and peak concentration at most of

the lower sensors (at 0.25 m height) and underpredict it at the

higher sensors (at 1.25 m height). Case 3 overpedicts the

average and peak concentration near the release at the lower

sensors and far from the release at the higher sensors, whilst

the prediction at the rest sensors is consistent with the

0

10

20

30

40

50

0 2 4 6 8downwind distance (m)

)v/v%(

noitartnecnocegarevA

0.25 m

(a)

0

20

40

60

80

0 2 4 6 8downwind distance (m)

)v/v%(

noitartnecnockae

P

0.25 m

experiment case1 (dry air)

(a)

Fig. 4 e The predicted and measured average (top) and peak (bo

the release and at height above the ground, a) 0.25 m and b) 1.2

experiment. At 1.25 m height the prediction with case 3 cap-

tures the experimental trend quite well.

As it can be observed in Fig. 4 the cloud with the humid air

(case 2) is slightly more buoyant than the cloud with dry air,

since the condensation and solidification of air humidity lib-

erates heat. However, the effect is small, and still in both cases

(1 and 2) the model underpredicts the vapor concentration at

most of the high sensors. This is due to the fact that when

humidity condenses and solidifies the mixture density in-

creases resulting in a heavy cloud. The hydrodynamic equi-

librium model that was assumed is not able to describe fully

and accurately the physical phenomenon.

Only when humidity is modeled with the non hydrody-

namic equilibrium model (case 3) the predicted cloud is lifted

more, and the prediction is improved. The non hydrodynamic

equilibrium model allows the vapor and the non vapor phase

to obtain different vertical velocities. The non vapor phase

falls faster to the ground due to gravity, and leaves a lighter

and more buoyant cloud enriched with the vapor phase of the

components. Moreover, when the non vapor phase falls down,

it is accumulated to the ground near the cold release (~20 K),

and extracts more heat to the mixture.

Possible reason for the overprediction near the release in

case 3 is the underestimation of the heat flux from the sur-

roundings to the cloud. More heat input either from the

ground or from the air freezing that occurs near the release

point could contribute to the buoyancy of the cloud. The cloud

would be liftedmore and the concentration at the low sensors

near the release would decrease, while the concentration at

the high sensors near the release would increase.

0

2

4

6

8

10

0 2 4 6 8downwind distance (m)

)v/v%(

noitar tnecnocegarevA

(b)

1.25 m

0

10

20

30

40

0 2 4 6 8downwind distance (m)

)v/v%(

noitartn ec nockae

P

1.25 m

case2 (humid air) case2 (humid air-slip model)

(b)

ttom) concentration on plane y ¼ 0, at distances downwind

5 m for test 5.

Table 3 e Evaluation of the statistical performancemeasures for the average and peak concentration levels for all modeledcases for test 5.

Ideal value Case 1 Case 2 Case 3

Average Peak Average Peak Average Peak

MG 1 0.018 0.0107 0.027 0.015 0.25 0.11

ln(VG) 0 40.2 42.5 36.6 38.9 18.9 19

i n t e rn a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 315858

In Table 3 the statistical measures for the average and peak

concentration at all available sensors and for all modeled

cases are displayed. According to the MG values all cases

overall underpredict both average and peak concentration.

However, case 3 exhibits better performance and its pre-

dictions have the lowest scatter.

Test 6

The measured and predicted hydrogen concentration time

histories on plane y ¼ 0 at 7.5 m downwind the release are

depicted in Fig. 5. Similar remarkswith test 5 can also bemade

for this test too. The computational results for case 3 are in

better agreement with the experiment compared to the other

cases. Here, as in test 5, the arrival times of the concentration

peaks are not exactly the same as in the experiment and it

seems that more peaks are predicted.

Fig. 6 compares the measured and predicted average and

peak hydrogen concentration at distances downwind the

release at two different heights for all three modeled cases.

Case 1 and case 2 results are similar. Case 2 shows very good

agreement with the experiment at the sensors far from the

release and at 1.25 m concerning the average concentration.

Case 3 overpredicts both average and peak concentration levels

at the low sensors near the release, while it is consistent with

the experiment at the low sensors far from the release and at

the high sensors. In general, it exhibits better agreement with

the measurements than the two other cases regarding both

average and peak concentration at most of the sensors.

The overprediction at the closer downwind the release and

lower sensor (Fig. 6a) could beattributed to theunderestimation

of the heat flux near the release as mentioned in test 5. The air

)52.0,0,5.7(rosneS

0.00

0.10

0.20

0.30

450 550 650 750 850 950 1050time (sec)

)v/v(noitartnecnoc

negor dyh

experiment case1 (dry air)

Fig. 5 e The predicted and measured time histories for test 6 at

The exact location of the sensors (x, y, z coordinates (m), in respe

ground level at z ¼ 0) is shown in the parenthesis.

condensation effect near the release is reported in Ref. [19]

which showed that air condensation generates an upward ve-

locity near the release, and makes the cloud more buoyant in

this region.

According to the statistical performance (Table 4), all cases

underestimate both average and peak hydrogen concentra-

tion. However, MG value in case 3 is closer to the ideal value

and has less scatter. Therefore, it can be concluded that

overall the performance of case 3 is in better agreement with

the experiment than the other two cases.

Test 7

The predicted and experimental time histories for test 7 are

illustrated in Fig. 7. The predicted and measured average and

peak hydrogen concentration are shown in Fig. 8. According to

these figures in the two first cases the model underestimates

the hydrogen concentration at the higher sensors. Case 3 gives

an improved prediction with increase of the concentration

levels at those sensors. In general, in case 3 with the non

hydrodynamic equilibrium model the cloud is more buoyant

compared to the other cases, and the results are closer to the

experiment. However, it overestimates the average concen-

tration at most of the sensors.

In Fig. 9 the temperature contours on plane y ¼ 0 and the

experimental cloud 170 sec after start of the release are

illustrated. The cloud in the experiment seems to fall sharply

to the ground at some distance from the release point. This

steep fall is not that obvious in the simulation at that time

snapshot, because at 170 sec after the release the wind di-

rection and consequently the hydrogen cloud is not in line

with the release, so the comparison on the plane y ¼ 0 is not

)52.1,0,5.7(rosneS

0.00

0.05

0.10

0.15

0.20

450 550 650 750 850 950 1050

time (sec)

)v/v(noitartnecnoc

negordyh

case2 (humid air) case2 (humid air-slip model)

7.5 m downwind the release and at two different heights.

ct with the release point located at x ¼ 0 and y ¼ 0, and the

0

20

40

60

0 2 4 6 8downwind distance (m)

)v/v%(

noitartnecnocegarevA

0.25 m

(a)

(a)

0

2

4

6

0 2 4 6 8downwind distance (m)

)v/v%(

noita rtne cnoce gar evA

1.25 m

experiment case1 (dry air) case2 (humid air) case2 (humid air-slip model)

(b)

(b)0

20

40

60

0 2 4 6 8downwind distance (m)

)v/v%(

noitartnecnockae

P

0.25 m

0

10

20

30

40

0 2 4 6 8

downwind distance (m)

)v/v%(

noit art necnockae

P

1.25 m

Fig. 6 e The predicted and measured average (top) and peak (bottom) concentration on plane y ¼ 0, at distances downwind

the release and at height above the ground, a) 0.25 m and b) 1.25 m for test 6.

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 3 15859

very accurate. Therefore, the next available time that the

hydrogen cloud was in line with the release is also chosen for

reproducing the temperature contours (see Fig. 9-bottom). The

figures are comparable although there are at different times,

because we are interested in the qualitative behavior of the

cloud. At that snapshot the steep fall of the cloud is observed

in the simulation too.

In the experiment close to the release the hydrogen and the

formed by the humidity phase change droplets and ice crys-

tals are influenced by the initial jet momentum and follow the

main stream (see Fig. 9). As they traveled further downstream

the “heavy” cloud falls to the ground. Similarly, in the simu-

lation with case 3 the temperature contours reveal a hori-

zontal movement of the cloud near the release and a rapid fall

approximately 2 m downwind the release. Some meters

further the cloud absorbs heat by the ground and rises again.

This behavior shows that the slip model that was used is

capable to capture the cloud behavior qualitatively, and it is

shown that it is of great importance the use of the non hy-

drodynamic equilibrium model in such cases.

Table 4 e Evaluation of the statistical performance measures focases for test 6.

Ideal value Case 1

Average Peak

MG 1 0.05 0.016

ln(VG) 0 24.5 31.12

In Table 5 the statistical indicators for test 7 are displayed.

According to the MG values the predictions with case 1 and 2

underestimate the average and peak hydrogen concentration,

whilst the prediction with case 3 overestimates the average

concentration and underestimates the peak concentration.

Overall discussion

The modeling of the dispersion of liquefied hydrogen in open

environment was studied with the help of three tests per-

formed by HSL. The parameter that was examined is the effect

of ambient humidity.

In the presence of ambient humidity the cloud becomes

more buoyant, due to the heat liberation by the condensation

and solidification of the water. Therefore, when humidity was

modeled the predicted concentrations at the higher sensors

increased compared to the case without humidity. However,

the effect was not as significant as it was expected. This

behavior is due to the fact that when the humidity is

condensed and/or solidified leads to an increase of the

r the average and peak concentration levels for all modeled

Case 2 Case 3

Average Peak Average Peak

0.07 0.023 0.27 0.1

22.5 28.36 12.8 14.4

0.00

0.05

0.10

0.15

0.20

420 460 500 540 580 620 660 700 740 780

time (sec)

)v/v(noitartnecnoc

negordyh

0.00

0.02

0.04

0.06

0.08

0.10

420 460 500 540 580 620 660 700 740 780

time (sec)

)v/v(noitar tnecno c

n egordy h

experiment case1 (dry air) case2 (humid air) case2 (humid air-slip model)

)m52.2,0,5.7(rosneS)m52.1,0,5.7(rosneS

Fig. 7 e Time histories of the predicted hydrogen concentration for test 7 for all simulation cases compared with the

experiment at 7.5 m downwind the release and at two different heights. The exact location of the sensors (x, y, z coordinates

(m), in respect with the release point located at x ¼ 0 and y ¼ 0, and the ground level at z ¼ 0) is shown in the parenthesis.

i n t e rn a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 315860

mixture density, and provides a heavier cloud. Therefore, the

buoyancy effect of the released heat by the phase change

process is counterbalanced by the higher mixture densities

that are produced.

To improve the prediction, the non hydrodynamic equi-

librium model was applied and tested. With the non hydro-

dynamic equilibrium model the condensed and/or solidified

0

4

8

12

16

20

0 2 4 6 8downwind distance (m)

)v/v%(

noitartnecnocegarevA

m52.0

(a)

0

20

40

60

80

0 2 4 6 8downwind distance (m)

)v/v%(

noitartnecnockae

P

(a)

0.25 m

experiment case1 (dry air)

Fig. 8 e The predicted and measured average (top) and peak (bo

the release and at height above the ground, a) 0.25 m, and b) 2

humidity and the liquid hydrogen are allowed to obtain

different velocity from the vapor phase. Therefore, they fall

faster to the ground leaving a lighter cloud. The difference

between the vapor velocity and the non vapor velocity is

called slip velocity. In addition, as the condensedwater and/or

the formed ice crystals are accumulated at high concentration

levels near the ground more heat is transferred to the cold

0

0.5

1

1.5

2

2.5

0 2 4 6 8downwind distance (m)

)v/v%(

noit ar tne cnocega revA

m52.2

(b)

0

5

10

15

20

0 2 4 6 8downwind distance (m)

)v/v%(

noitartnecnockae

P

(b)

2.25 m

case2 (humid air) case2 (humid air-slip model)

ttom) concentration on plane y ¼ 0, at distances downwind

.25 m for test 7.

Fig. 9 e The predicted temperature contours on xz plane (y¼ 0) 170 sec after start of the release (top) and 220 sec after start of

the release (bottom) for case 3 and the experimental cloud [19] 170 sec after the start of the release. The contour images are

the symmetric images with respect to the z-axis, in order to have the same viewpoint as in the experiment.

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 3 15861

cloud. The consideration of the slip velocity in the modeling

has a great impact on the vapor hydrogen dispersion. The

prediction with the slip model has been improved compared

with the case with hydrodynamic equilibrium model

assumption for all the examined tests, by predicting higher

concentration levels at the higher sensors, as was observed in

the experiment. In general, the predictionswith the slipmodel

captured better the cloud behavior and the experimental

trend in terms of both average and peak concentration.

For the best simulation case (humid air and slip model) the

prediction is consistentwith the experiment for all three tests.

However, in the tests where the release is close to the ground

(test 5 and 6) the model overpredicts both time average and

peak concentration at the lowest sensor near the release. The

reason for this overprediction may be attributed to underes-

timation of the heat flux in this region. Possible heat contri-

bution could be from the air condensation that occurs near the

release which is not modeled in this study. In the test where

Table 5 e Evaluation of the statistical performance measures focases for test 7.

Ideal value Case 1

Average Peak

MG 1 0.73 0.15

ln(VG) 0 5.69 6.65

the release is high above the ground (test 7) themodel tends to

overpredict the average concentration at the lower sensors

and at the higher sensors far from the release, while it tends to

underestimate the average concentration at the higher sen-

sors near the release. However, regarding the peak concen-

tration the model is in good agreement with experiment at

most of the sensors.

Conclusions and future work

Three LH2 release experiments on flat ground (60 lt/min,

26.6mmdiameter) performed in 2010 by the Health and Safety

Laboratory (HSL tests 5, 6 and 7) were simulated with the CFD

code ADREA-HF taking into account the measured fluctuating

wind direction and the presence of humidity in the ambient

air. Humidity was modeled both with and without the hy-

drodynamic equilibrium assumption. In the second case a

r the average and peak concentration levels for all modeled

Case 2 Case 3

Average Peak Average Peak

0.98 0.22 1.43 0.42

4.77 4.85 4.44 2.92

i n t e rn a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 315862

modified slip velocity model originally developed for rain was

used to provide the difference between vapor and non vapor

phases. Simulations were also performed without humidity

for comparison.

It was found that in all three experiments accounting for

humidity and using the presented slip model (humid-slip)

results in better qualitative and quantitative agreement with

the experimental data than the other two cases (humid-no-

slip, dry) regarding both time averaged and peak

concentrations.

A statistical analysis was also performed using the MG and

VG measures. According to the analysis the predictions in all

modeled cases overall underestimated the time averaged and

peak concentrations in all tests, except for test 7 with the case

humid-slip (case 3) which overall overestimated the time

averaged concentration. In general, case 3 exhibited better

performance than the other two cases with MG values for the

time averaged concentration 0.25, 0.27 and 1.43 for test 5, 6

and 7 respectively.

Further research in the future could include more insight

on the effect of air (N2, O2) solidification, more insight into the

non hydrodynamic equilibrium assumption and other slip

velocity models and more LH2 experiments under more

“controlled” external/ambient conditions to provide more

data for model validation.

Acknowledgments

The work presented here is supported by the H2FC project

(“Integrating European Infrastructure to support science and

development of Hydrogen- and Fuel Cell Technologies to-

wards European Strategy for Sustainable, Competitive and

Secure Energy”, Grant agreement no.: FP7-284522).

The authors would also like to acknowledge Philip Hooker

and his group for the helpful discussion and for providing the

experimental data.

r e f e r e n c e s

[1] Witkofski RD, Chirivella JE. Experimental and analyticalanalyses of the mechanisms governing the dispersion offlammable clouds formed by liquid hydrogen spills.Hydrogen Energy J 1984;9(5):425e35.

[2] Chirivella JE, Witkofski RD. Experimental results from fast1500 gallon LH2 spills. Inst Chem Eng Symp Ser 1986;82(28).

[3] Marinescu-Pasoi L, Sturm B. Messung der Ausbreitung einerWasserstoff- und Propangaswolke in bebauten Gel€ande undGasspezifische Ausbreitungversuche. BattelleIngenieurtechnik GmbH; 1994. Reports R-68.202 and R-68.264.

[4] Schmidichen U, Marinescu-Pasoi L, Verfodern K, Nickel V,Sturm B, Dienhart B. Simulation of accidental spills ofcryogenic hydrogen in a residential area. Cryog J1994;34:401e4 (ICEC 15 Supplement).

[5] Hijikata T. Research and development of International CleanEnergy Network using hydrogen energy (WE-NET). HydrogenEnergy J 2002;27:115e29.

[6] Chitose K, Takeno K, Yamada Y, Hayashi K, Hishida M.Activities on hydrogen safety for the WE-NET project e

experiment and simulation of the hydrogen dispersion. In:Proceedings of WHEC-14, Toronto; 2002.

[7] Hooker P, Willoughby DB, Royle M. Experimental releases ofliquid hydrogen, 4th International Conference on HydrogenSafety; 2011. San Francisco, Paper 160.

[8] Venetsanos AG, Papanikolaou E, Bartzis JG. The ADREA-HFCFD code for consequence assessment of hydrogenapplications. Hydrogen Energy J 2010;35:3908e18.

[9] Bartzis JG. ADREA-HF: a three dimensional finite volumecode for vapour cloud dispersion in complex terrain. Rep Eur1991;13580.

[10] Wurtz J, Bartzis JG, Venetsanos AG, Andronopoulos S,Statharas J, Nijsing R. A dense vapour dispersion codepackage for applications in the chemical and processindustry. Hazard Mater J 1996;46(2e3):273e84.

[11] Statharas JC, Venetsanos AG, Bartzis JG, Wurtz J,Schmidtchen U. Analysis of data from spilling experimentsperformed with liquid hydrogen. Hazard Mater J2000;77(1e3):57e75.

[12] Venetsanos AG, Baraldi D, Adams P, Heggem PS,Wilkening H. CFD modelling of hydrogen release, dispersionand combustion for automotive scenarios. J Loss PrevProcess Ind 2008;21:162e84.

[13] Baraldi D, Venetsanos AG, Papanikolaou E, Heitsch M,Dallas V. Numerical analysis of release, dispersion andcombustion of liquid hydrogen in a mock-up hydrogenrefuelling station. J Loss Prev Process Ind 2009;22:303e15.

[14] Venetsanos AG, Bartzis JG. CFD modeling of large-scale LH2spills in open environment. Hydrogen Energy J2007;32:2171e7.

[15] Giannissi SG, Venetsanos AG, Markatos N, Bartzis JG.Numerical simulation of LNG dispersion under two-phaserelease conditions. J Loss Prev Process Ind2013;26(1):245e54.

[16] Statharas JC, Bartzis JG, Venetsanos A, Wurtz J. Prediction ofammonia releases using ADREA-HF code. Process Saf Prog J1993;12(2):118e22.

[17] Andronopoulos S, Bartzis JG, Wurtz J, Asimakopoulos D.Modelling the effects of obstacles on the dispersion ofdenser-than-air gases. Hazard Mater J 1994;37(2):327e52.

[18] Andronopoulos S, Stratharas JC, Deligiannis P, Bartzis JG.Evaluation of the predictions of the ADREA-HF code fordense gas dispersion with real scale ammonia releaseexperiments, 4th International Conference on Air Pollution;1996. p. 81e6. Toulouse, Code 45563.

[19] Ichard M, Hansen OR, Middha P, Willoughby D. CFDcomputations of liquid hydrogen releases. J Hydrogen Energy2012;37:17380e9.

[20] Koutsourakis N, Venetsanos AG, Bartzis JG. LES modelling ofhydrogen release and accumulation within a non-ventilatedambient pressure garage using the ADREA-HF CFD code. JHydrogen Energy 2012;37:17426e35.

[21] Launder BE, Spalding DB. The numerical computation ofturbulent flow. J Comput Methods Appl Mech Eng1974;3(2):269e89.

[22] Markatos NC, Pericleous KA. Laminar and turbulent naturalconvection in an enclosed cavity. Int J Heat Mass Transf1984;27(No. 5):755e72.

[23] Ogura Y, Takahashi T. Numerical simulation of the lifecycle of a thunderstorm cell. Mon Weather Rev J1971;99:895e911.

[24] Manninen M, Taivassalo V, Kallio S. On the mixture modelfor multiphase flow. VTT Publications, 288, TechnicalResearch Centre of Finland; 1996.

[25] Giannissi SG, Venetsanos AG, Bartzis JG, Markatos N,Willoughby DB, Royle M. CFD modeling of LH2 dispersionusing the ADREA-HF code, 4th International Conference onHydrogen Safety; 2011. San Francisco, Paper 196.

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n en e r g y 3 9 ( 2 0 1 4 ) 1 5 8 5 1e1 5 8 6 3 15863

[26] Press W, Flannery BP, Teukolsky SA, Vetterling WT.Numerical recipes in C: the art of scientific computing.Cambridge University Press; 1988e1992.

[27] Versteeg HK, Malalasekera W. Introduction tocomputational fluid dynamics. New York: Longman

Scientific & Technical, John Wiley & Sons Inc.; 1995. 605Third Avenue.

[28] Harms F, Leitl B, Schatzmann M, Patnaik G. Validating LES-based flow and dispersion models. J Wind Eng Ind Aerodyn2011;99:289e95.