Professional Documents
Culture Documents
Asimplifiedstatistic Basedprocedureforgas
Asimplifiedstatistic Basedprocedureforgas
net/publication/321691219
CITATIONS READS
15 351
6 authors, including:
Jingde Li
Curtin University
42 PUBLICATIONS 315 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Dynamic performance of precast concrete beam and joint under impact loads View project
Precast Segmental Concrete Beams Prestressed with FRP Tendons under Static Loads View project
All content following this page was uploaded by Jihao Shi on 03 May 2018.
Jihao Shi a,b , Jingde Li b , Yuan Zhu a,∗ , Hong Hao b , Guoming Chen a , Bin Xie c
a Center for Offshore Engineering and Safety Technology, China University of Petroleum, Qingdao 266580, China
b Centre for Infrastructural Monitoring and Protection, School of Civil and Mechanical Engineering, Curtin University,
WA 6102, Australia
c GexCon China, Shanghai 200135, China
a r t i c l e i n f o a b s t r a c t
Article history: In explosion risk analysis, Frozen Cloud Approach (FCA) and Dimensionless Response Sur-
Received 20 February 2017 face Method (DRSM) are both commonly used to achieve a balance between simulation
Received in revised form 20 October workloads and accurate results. However, the drawbacks of these two approaches are obvi-
2017 ous. FCA is not reliable for risk study of fuel-dominated regions. Whereas DRSM usually
Accepted 1 December 2017 couples the dimensionless parameters and generates a large numbers of correlations to
Available online 8 December 2017 predict the flammable cloud size, which brings a heavy computation burden for engineers.
Therefore, this paper aims to propose a simplified procedure which can quickly and accu-
Keywords: rately provide a large number of non-simulation data based on limited CFD simulation data.
Automatically Selected Model Full Factorial Design of Experiment (FFDOE) based RSM is adopted. Codification is applied
Technology to couple all the dimensional parameters into a single correlation. Automatically Selected
Computation cost reduction Model Technology (ASMT) is used to easily determine the suitable structure of correlation.
Explosion risk analysis Compared to the conventional procedures, the simplified procedure is proven to be more
Frozen Cloud Approach robust. For subsequent Explosion risk analyses (ERAs) in the fuel-dominated regions, the
Dimensionless respond surface simplified procedure becomes a superior alternative.
method © 2017 Published by Elsevier B.V. on behalf of Institution of Chemical Engineers.
Fuel-dominated region
1. Introduction different leakage and wind conditions. The representative gas clouds
are subsequently used in gas explosion simulations to obtain DALs.
Explosion risk analyses (ERAs) are widely used to derive the dimen- Due to the fact that performing a large number of dispersion simu-
sioning accidental loads (DALs) for design of offshore topside facilities. lations by using CFD tools to derive DALs is inefficient and expensive,
ERAs can predict explosion loads in detail, including overpressures, Frozen Cloud Approach (FCA) and Dimensionless Response Surface
differential pressure, and drag loads (FABIG, TN-08). Loads with return- Method (DRSM) have been proposed by others to increase the effi-
ing frequencies of 1 × 10−04 per year are then adopted as DALs and are ciency of the risk analysis and to obtain acceptable results based
incorporated with standards and legislations such as NORSOK Z013 on a limited number of CFD simulations. DNV initially proposed the
(NORSOK, 2010) and ISO19901-3 (2015) in safety study. commonly-used Frozen Cloud Approach (FCA) (GexCon, 2015). In this
In a standard ERA, Computational fluid dynamic (CFD) simulation approach, linear relations amongst gas concentration, leak rate and
is usually employed. Gas dispersion simulation, which plays an impor- the wind speed for each leak scenario are firstly assumed and the gas
tant role in ERAs, is performed before gas explosion simulation. The cloud sizes from the non-numerically-simulated scenarios are then
major aim of conducting gas dispersion simulation is to identify cred- obtained. However, the FCA can exclusively provides satisfactory pre-
ible gas cloud sizes, gas concentrations and cloud locations under diction results for leakages in ventilation dominated regions while
poor estimations are expected for leakages in fuel dominated regions
(GexCon, 2015).
∗
Corresponding author.
E-mail addresses: shi jihao@163.com (J. Shi), jingde.li@curtin.edu.au (J. Li), zhy3323@163.com (Y. Zhu), Hong.Hao@curtin.edu.au (H.
Hao), offshore@126.com (G. Chen), xie.bin@gexcon.com (B. Xie).
https://doi.org/10.1016/j.psep.2017.12.002
0957-5820/© 2017 Published by Elsevier B.V. on behalf of Institution of Chemical Engineers.
Process Safety and Environmental Protection 1 1 4 ( 2 0 1 8 ) 48–63 49
Afterwards, Cleaver et al. (1999) developed a set of simple correla- S is the energy term, xi is the coordinate in i-direction, ui is
tions to predict the flammable gas cloud size for different gas release the velocity component in i-direction. The equation consists of
scenarios. Based on these correlations, Huser and Kvernvold (2000) four distinct terms and the first in the left represents transient
proposed the concept of DRSM. The DRSM calculates the flammable term while the second takes into account the convection. The
gas cloud size by using several response surfaces with a large number
terms in the right represent diffusion term and the last is the
of simulations. Compared to other old approaches, the DRSM indeed
source term.
improves the accuracy of gas cloud estimation. Nevertheless, the uti-
It is critical to model the Reynolds stress tensor as follows to
lization of these equations incurs heavier computation requirements
for engineers (Ferreira and Vianna, 2014). In order to decrease the describe the turbulence of dispersion process for the leakage
number of those correlations, Qiao and Zhang (2010) adopted a very gas.
conservative way combining the DRSM with the FCA, which however Reynolds stress tensor:
leads to large safety redundancy and large cost of offshore platform
construction. Later, Ferreira and Vianna developed the DRSM to couple ∂ui ∂uj
the most relevant dimensionless parameters into a single mathemati- −ui uj = ueff + − 2kıij /3 (2)
∂xj ∂xi
cal model, which though is exclusively suitable for early design phase
(Ferreira and Vianna, 2014).
Precisely, DRSM has a significant drawback even though it improves
where xj is the coordinate in j-direction, uj is the velocity com-
the accuracy of gas cloud estimation. As RSM is a collection of math- ponent in j-direction, k is turbulent kinetic energy, ıij is stress
ematical and statistical techniques for analyzing the relationship tensor.
between different parameters and response (Bezerra et al., 2008), the Following Boussinesq eddy viscosity assumption, the eddy
derived correlations estimate data for the ERAs without providing the viscosity models are used to model Reynolds stress tensor and
physical meaning of gas dispersion. The coupled dimensionless param- solve k–ε model as follows:
eters by DRSM result in a variation of two coupled parameters at a fixed Turbulent kinetic energy and one for dissipation of turbu-
ratio which will not influence the final gas cloud size calculation. In lent kinetic energy:
other words, the gas cloud size will be significantly under/over pre-
dicted if the coupled parameters change simultaneously. Alternatively,
∂(ˇv k) ∂(ˇi kui ) ∂
∂k
eff
if the variations of individual dimensionless parameters are all consid- + = ˇi + ˇv Pk − ˇv ε (3)
∂t ∂xi ∂xi ck ∂xi
ered in a correlation, several correlations need to be regenerated, which
may cause extra amount of calculation burden (Ferreira and Vianna,
∂(ˇv ε) ∂(ˇi εui ) ∂
∂ε ε2
2014). eff
+ = ˇi + ˇv Pε − C2ε ˇv (4)
Therefore, this study aims to propose a simplified procedure to ∂t ∂xi ∂xi cε ∂xi k
quickly generate thousands of non-simulation data with acceptable
interval for further ERAs based on a limited number of simulations. where t is time, Pk is the production of turbulent kinetic energy,
Compared to conventional procedures, the new simplified proposed Pε is the production of dissipation, C2 is constant in FLACS and
procedure is more robust and user-friendly. It can also reduce the equal to 1.92, ˇi is area porosity in the i-direction, ˇv is volume
computation cost. Moreover, for the fuel-dominated region of offshore porosity, ck and cε is Prandtl–Schmidt number.
platform, the simplified procedure becomes a better alternative than
FCA. In combination with stochastic simulation technique (e.g. Monte
2.2. Wind boundary model
Carlo simulation), the simplified procedure can be eventually used for
further ERAs.
Monin–Obukhov length is adopted by FLACS to describe the
properties of the atmospheric boundary layer close to Earth’s
2. Numerical dispersion model
surface and explain the buoyancy effect on the atmospheric
boundary layer. The characteristic length scale L is defined
In this study, FLACS (Flame Acceleration Simulator) is used
based on Pasquill classes as follows:
to model gas dispersion. FLACS is a 3D CFD software vali-
dated over the last 40 years against numerous experiments 1 z0
for a extensive range of different dispersion scenarios (Middha L= log (5)
Ls zs
et al., 2009, 2010; Hansen et al., 2010; Bleyer et al., 2012). The
software has been widely used in the oil and gas industry to where z0 is an atmospheric roughness length, zs and Ls are
simulate gas dispersion and vapor cloud explosion in both 1.0 m and 0 m respectively under Pasquill class D. With the
offshore and onshore facilities (Vianna and Cant, 2012; Qiao Pasquill class, the wind velocity profile can also be calculated
and Zhang, 2010; Huser and Kvernvold, 2000; Patankar, 1980; as follows:
Hansen et al., 2013).
∗ (z − zd ) + z0
u(z) = ln − u (z) If z0 > 0 (6)
2.1. Turbulent model a z0
FLACS solves the compressible RANs equations on a 3D Carte- u(z) = 0 If z0 < 0 (7)
sian grid using a finite volume method. The conservation
equations for mass, momentum, enthalpy, and mass fraction where z is the height above the ground, and zd is the canopy
of species, closed by the ideal gas law, are included. The con- height, a is Von Karman constant, typically equal to 0.41. u0 is
servation equations can be represented in general as: average wind velocity, z0 is an atmospheric roughness length,
* is the friction velocity, which is defined as follows for the
∂( ) ∂(ui ) ∂ ueff ∂ neutral and non Pasquill class:
+ = ( )+S (1)
∂t ∂xi ∂xi ∂xi
u0 a
∗ = (8)
In this equation, represents the dependent variable. ueff is (zref −zd )+z0
ln z0 − u (zref )
the effective turbulence viscosity, is the gas mixture density,
50 Process Safety and Environmental Protection 1 1 4 ( 2 0 1 8 ) 48–63
Fig. 1 – 3-Parameter-2-level FFDOE (three independent parameters A B C and two levels 1, 2 of each parameter result in 8
combinations) (Croarkin et al., 2002).
Table 1 – Configuration of FFDOE of 2 levels of leak rate, wind speed, wind direction and 3 levels of leak direction with
codification method.
Std Leak rate Wind speed Wind direction Leak direction Std Leak rate Wind speed Wind direction Leak direction
1 −1 −1 −1 [1 0] 13 −1 −1 1 [0 1]
2 1 −1 −1 [1 0] 14 1 −1 1 [0 1]
3 −1 1 −1 [1 0] 15 −1 1 1 [0 1]
4 1 1 −1 [1 0] 16 1 1 1 [0 1]
5 −1 −1 1 [1 0] 17 −1 −1 −1 [–1 –1]
6 1 −1 1 [1 0] 18 1 −1 −1 [–1 –1]
7 −1 1 1 [1 0] 19 −1 1 −1 [–1 –1]
8 1 1 1 [1 0] 20 1 1 −1 [–1 –1]
9 −1 −1 −1 [0 1] 21 −1 −1 1 [–1 –1]
10 1 −1 −1 [0 1] 22 1 −1 1 [–1 –1]
11 −1 1 −1 [0 1] 23 −1 1 1 [–1 –1]
12 1 1 −1 [0 1] 24 1 1 1 [–1 –1]
After codification, configuration of FFDOE of different levels The single and interactive effect of different parameters on
of varied affecting parameters can be conducted. Table 1 takes dispersion results are then derived from ANOVA results.
2 levels of leak rate, wind speed, wind direction and 3 levels Step 4: Further diagnosis analysis of mathematical model
of leak direction as an example to illustrate the configuration Diagnosis analysis is conducted to further determine the
of FFDOE based on codification method. As shown in Table 1, most suitable transformation of correlation, e.g. Ln trans-
FFDOE is orthogonal so that each parameter can be evaluated formation, Exponent transformation etc. Internally studentized
independently of all the other parameters. Moreover, the inter- residuals and Box-Cox plot are used to diagnosis the out-
active effect of varied parameters can be completely analyzed, liers from varied transformation and to determine the correct
which make the correlation derived more accurate and robust power law transformation.
(Box et al., 2005). Step 5: Dispersion results output: 2D contour plots, 3D sur-
Step 3: Mathematical modeling of dispersion simulation faces and vapor cloud volume prediction
results After deriving the suitable correlation, 3D surfaces and
The most suitable and robust polynomial order is deter- 2-D predicted lines with 95% confident intervals and predic-
mined in this step. ASMT, namely F-test, coefficient of multiple tion intervals are obtained. Those surfaces can also be used
determination R2 , Adjusted-R2 and Predicted-R2 is adapted for ERAs in combination with stochastic probability analysis.
to select the polynomial order. F-value, R2 , Adjusted-R2 and However, it should be noted that this procedure are useful for
Predicted-R2 can be calculated as below: statistical estimation rather than deterministic estimation. In
2 m 2 other words for further ERAs, not only the predicted results
SSreg /(p − 1) m −
F value = = ŷj − y / yj − ŷj but also the interval’s upper and lower limits should be used
SS/ res (p − 1) j j together with stochastic probability method.
(12)
m − 2
to FLACS. A 3D view of the central part of the offshore platform
+ ŷj − y (15) is shown in Fig. 3.
j The overall length, width and height of the offshore plat-
form are 25 m (X), 60 m (Y), 40 m (Z), respectively. The platform
where m is total number of simulation results in the design; p consists of process module and accommodation module. The
is number of parameter of model; yj is the simulation results process module (central part), which is the focus in this study,
for the level j; ŷj is estimated value by the model for the level has three deck levels, i.e. the lower deck level, the middle deck
−
j; y is the average value of m simulation results; SSreg is the level and the upper deck level. Most of the gas equipment
sum of the square due to regression and SSres is the sum of locate on the lower and middle deck level. The lower deck
the square due to residuals. It is worth noting that Predicted- level of the process module is occupied by accommodation
R2 is calculated by systematically removing each output (yi,−i
ˆ ) modules on +Y side and the middle deck level is occupied by
from the data set, which can be used to estimate the regres- the steel plates and accommodation module on −X and +Y
sion equation, and determine how well the model predicts the side, respectively. The length, width and height of the process
removed output. module are 25 m, 24 m, and 10 m, respectively.
52 Process Safety and Environmental Protection 1 1 4 ( 2 0 1 8 ) 48–63
Fig. 3 – Geometric model of a typical fixed and naturally ventilated offshore platform.
Fig. 5 – Ventilation conditions inside the facility with 5 m/s wind speed and wind direction is 270◦ . The upper figure shows
the wind vector at XY plane at Z = 1.8 m (the middle deck level). The lower figure shows the wind vector at XZ plane at
Z = 2.8 m (the lower deck level). The wind vectors are shown as m/s.
a) The middle deck level. b) The lower deck level.
2010), the leak point (115,110,29.5) as an example, is set at the Leak rates used in this study cover the range of all critical
center of the middle deck level. A typical composition of natu- hole sizes. 4 levels of leak rates are considered. The leak rate
ral gas is adopted for gas dispersion simulation. Table 2 shows can be calculated by (Spouge, 1999):
the composition of natural gas. Table 3 shows the leak property
of critical segment.
M
2 + 1 −1
−1 P >P ( 2 ) (16)
Q0 = Cd Av P0 0 a +1
RT0 +1
54 Process Safety and Environmental Protection 1 1 4 ( 2 0 1 8 ) 48–63
Fig. 6 – 3D plot of flammable cloud distribution when leak rate is 20 kg/s, wind speed is 5 m/s and leak directions are −X &
+X, wind direction is 270◦ .
a) 3D plot of flammable cloud distribution when leak rate is 20 kg/s, wind speed is 5 m/s and leak direction is –X, wind
direction is 270 at time 4 s after leakage. b) 3D plot of flammable cloud distribution when leak rate is 20 kg/s, wind speed is
5 m/s and leak direction is –X, wind direction is 270 at time about 70 s after. c) 3D plot of flammable cloud distribution when
leak rate is 20 kg/s, wind speed is 5 m/s and leak direction is +X, wind direction is 270 at the time about 70 s after leakage.
The other criterion is combination results of R2 , Adjusted R2 Parametric sensitivity study on simulation results of Q9 is
and Predicted R2 . Fig. 8 shows the coefficients of determination also conducted based on F-test. The smaller P-values indicate
results. As shown in Fig. 8, R2 , Adjusted R2 and continuously the higher significance of the corresponding parameter. As
increase by increasing the polynomial order. However, at a cer- shown in Table 7, the terms of leak rate A, wind speed B, leak
tain point of Cubic order, the Predict R2 starts to decrease. It is direction D, leak rate and wind speed AB, leak rate and leak
seen that the Predict R2 at Quartic order is smaller than that at direction AD, wind speed and leak direction BD play vital roles
Cubic order, which means the Quartic correlation is not robust. in affecting Q9 since the corresponding P-values are less than
Therefore, in combination with the criterion 1, the Cubic order 0.05.
is chosen for the mathematical model. Based on the results derived by ASMT, Cubic model is cho-
sen and Ln is used to transform the response values of Q9. In
56 Process Safety and Environmental Protection 1 1 4 ( 2 0 1 8 ) 48–63
Fig. 7 – Equivalent stoichiometric cloud volumes versus time of varied leak rates under different wind speeds when leak
direction is −X and wind direction is 270◦ .
a) Simulation results of varied leak rates under 1.5 wind speed. b) Simulation results of varied leak rates under 5 wind
speed. c) Simulation results of varied leak rates under 7 wind speed. d) Simulation results of varied leak rates under 12
wind speed. e) Q9 versus different leak rates under different wind speed.
Fig. 9 – Comparison of internally studentized residuals between transformed correlation and original correlation.
a) Normal plot of residuals from the original. b) Normal plot of residuals from the transformed. c) Residuals versus the
experimental run order of the original. d) Residuals versus the experimental run order from the transformed.
order to simplify the correlation, stepwise multiple regression correlation. Meanwhile the transformed one (Ln(cubic)) is com-
method (Draper and Smith, 2014) is used to delete the terms pared with the original one (cubic). Fig. 9 shows comparison
that have less effect on the prediction ability. Final equation results of diagnosis analysis of internally studentized residu-
of actual factors in terms of leak direction −X is determined als between original and transformed correlation. As can be
as below: seen, both the residuals from the transformed and original
one fall on a straight line, which means both residuals are nor-
Ln(Q9)=+7.52986+0.023671 × A − 0.063698 × B − 6.48637E − mally distributed. However, Ln(cubic) indicates better residuals
normal distribution as less outliers occur compared with the
003 × C+3.44187E − 003 × A × B+4.54538E − 005 × A × C −
cubic. Fig. 9c) and d) presents residuals versus the experimen-
4.21516E − 004 × B × C − 9.41053E − 004 × A 2+5.75984E tal run order from both correlations. The random pattern of
residuals indicates the suitability of both correlations. As can
−005 × C 2 − 3.34552E − 005 × A 2 × B − 1.38770E be seen, Ln(cubic) is relatively more robust as less experimental
points are above the 95% confident intervals.
−007 × A × C 2+1.96656E − 006 × B × C 2+7.94896E In order to effectively determine the correct power law
transformation, i.e. b of Ln(cubic + b), Box-Cox plot for power
−006 ∗ A 3 − 1.28618E − 007 × C 3 (17) transform is used. As can be seen in Fig. 10, pink color line
indicates the upper (0.17) and the lower (0) limits of the 95%
confident intervals for the Box-Cox lambda value, the blue
(overlapping with the pink color line) and green lines represent
5.3.4. Diagnosis analysis on correlation transformation the current (0) and the best (0.09) lambda values, respectively.
Diagnosis analysis is to check if the determined correlation The data in Fig. 10 indicates that b = 0 is suitable for the trans-
satisfies the normal assumption initially made. However, the formed correlation as green line locates between the upper
variances of the raw residuals at different input samples may and lower limits of 95% confident intervals.
differ, which leads to the raw residuals belong to different nor- Fig. 11 shows comparison results between Ln(cubic) and
mal populations or normal distributions even if the variances (cubic) in terms of predicted results versus simulation results.
of the errors are equal at these different samples. Therefore, As can be seen, the majority of predicted results by Ln(cubic)
internally studentized residuals should be used to map all the are within 50% intervals, which further verify the suitability
different normal distributions into a single standard normal of transformation compared with (cubic), as lots of predicted
distribution. results by (cubic) locate beyond the acceptable range.
As correlation transformation is conducted in above sec- The fitness and accuracy of the Ln-Cubic mathematical
tion, the part is to check the feasibility of transformed model are verified in the above diagnosis analysis. Therefore,
58 Process Safety and Environmental Protection 1 1 4 ( 2 0 1 8 ) 48–63
Fig. 11 – Comparison between the transformed and original correlation in terms of predicted results versus simulation
results.
a) Is the original one (cubic). b) Is the transformed one Ln (cubic).
Process Safety and Environmental Protection 1 1 4 ( 2 0 1 8 ) 48–63 59
Fig. 12 – 3D results of predicted flammable cloud size of leak rate and wind speed when the leak direction is aligned to −X
and wind direction is 270◦ (red dots present the simulation results, contour are curves of constant response drawn in the
leak rate and wind speed plane keeping all other parameters fixed. Each contour corresponds to a particular value of Q9).
(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
Fig. 13 – Predicted lines with corresponding lines of 95% confident intervals and 95% predicted intervals under varied leak
rates and wind speeds when the leak direction is aligned to −X and wind direction is 270◦ .
a) 1.5 m/s wind speed. b) 5 m/s wind speed. c) 7 m/s wind speed. d) 12 m/s wind speed.
60 Process Safety and Environmental Protection 1 1 4 ( 2 0 1 8 ) 48–63
FCA when wind speed is 1.5 m/s for small leak rate. Fig. 15b)
also shows the comparison results under leak rates 20 kg/s and
60 kg/s when wind speed is 3 m/s, wind direction is 270◦ and
leak direction is −X. For FCA, the result under leak rate 20 kg/s
is predicted by using given scenario data when leak rate is
10 kg/s (a factor of 2 is used for the multiplication). While the
result under leak rate 60 kg/s is calculated by multiplying given
scenario (i.e. leak rate is 40 kg/s) data by a factor of 1.5. From
it, one can see the results predicted by the single correlation
are closer to the simulation results compared to those of FCA.
Fig. 16 shows comparison results among the simplified pro-
cedure, FCA and CFD of 864 simulation scenarios. As can be
seen in Fig. 16a), the results predicted by the simplified pro-
cedure agree well with those of CFD. The majority of these
results are within +50% confident intervals. Whereas more
results derived by FCA in Fig. 16b) are smaller than those esti-
Fig. 14 – Single effect of leak direction when leak rate is mated by using CFD. Fig. 16c) shows the exceedance frequency
80 kg/s, wind speed is 12 m/s and wind direction is from curves for all simulation results. CFD simulation data is used
South to North (270◦ ). as a benchmark.
The leak frequency used for those curves is calculated by
C2 , the Q92 is calculated by multiplying the Q91 by a factor of using DNV LEAK software based on the HSE database. These
A2 C1 /A1 C2 (GexCon, 2015). databases provide leak frequency depending on the size or
Extra non-simulation dispersion results are predicted by type of equipment and piping elements. The wind frequency
both of the simplified procedure and FCA, where the gas dis- is derived based on the statistics of wind conditions. Fig. 16c)
persion conditions are the same. Fig. 15 shows the comparison indicates that the curve derived from FFDOE-based RSM is
results between the simulation results and predicted results closer to the CFD curve than that of FCA. In other words, the
from these two methods when wind speed is 3 m/s, wind simplified procedure is more reliable than FCA in ERA when
direction is 270◦ and leak direction is −X. The results from RSM wind condition is not dominant.
are predicted by using a single correlation. For cases with wind
speed of 3 m/s and varied leak rates of 5 kg/s, 20 kg/s, 60 kg/s
and 80 kg/s, the FCA results are predicted by using given dis- 5.6. Comparison between FFDOE-based RSM and
persion scenarios’ data when wind speed is 1.5 m/s (multiplied DRSM
by a factor of 0.5 as the ventilation rate is 0.013 s−1 ) and wind
speed is 5 m/s (multiplied by a factor of 1.6 as the ventilation The disadvantages of conventional DRSM are shown in the
rate is 0.027 s−1 ). As can be seen Fig. 15a), there is a signif- comparison of the predicted results between FFDOE based
icant difference between the simulation results and that of RSM and conventional DRSM.
Fig. 15 – Comparison results between FFDOE-based RSM, FCA and simulation results when wind speed is 3 m/s, wind
direction is 270◦ and leak direction is −X.
a) Results for varied leak rates. b) Results when leak rates are 20 kg/s and 60 kg/s. c) Simulation results of 3 m/s by CFD. d)
Predicted simulation result of 3 m/s by FCA when wind speed is 1.5 m/s.
Process Safety and Environmental Protection 1 1 4 ( 2 0 1 8 ) 48–63 61
Fig. 16 – Comparison results between FFDOE-based RSM, FCA and CFD under 846 simulation scenarios. Horizontal axis
shows all the CFD results.
a) Predicted results by this procedure versus CFD results. b) Predicted results by FCA versus CFD results. c) Exceedance
frequency curve versus equivalent stoichiometric cloud volume.
Fig. 17 – Dimensionless data predicted by FFDOE-based RSM (the dots present the simulation dimensionless data, the lines
present the dimensionless results generated by FFDOE-based RSM, Vf presents the maximum flammable cloud volume and
V presents the volume of the process module, R is the ratio of the volume flow of flammable gas Qg and the volume of air Qa .
a) Results of wind speed 1.5 m/s. b) Results of wind speed 5 m/s. c) Results of wind speed 3 m/s and this dimensional
simulation results are shown in Fig. 15c).
62 Process Safety and Environmental Protection 1 1 4 ( 2 0 1 8 ) 48–63
through probabilistic consequence modeling. Proceedings of Middha, P., Hansen, O.R., Storvik, I.E., 2009. Validation of
the 7th Global Congress on Process Safety, 16. CFD-model for hydrogen dispersion. J. Loss Prev. Process Ind.
Draper, N.R., Smith, H., 2014. Applied Regression analysis. John 22, 1034–1038.
Wiley & Sons. Middha, P., Hansen, O.R., Grune, J., Kotchourko, A., 2010. CFD
FABIG. Protection of Piping Systems Subject to Fires and calculations of gas leak dispersion and subsequent gas
Explosions: (FABIG Technical Note 8). Steel Construction explosions: validation against ignited impinging hydrogen jet
Institute (SCI), TN-08. experiments. J. Hazard. Mater. 179, 84–94.
Ferreira, T.D., Vianna, S.S., 2014. A novel coupled response NORSOK, 2010. Risk and Emergency Preparedness Analysis,
surface for flammable gas cloud volume prediction. Int. J. NORSOK Standard Z-013, Oslo.
Model. Simul. Pet. Ind. 8. Patankar, S.V., 1980. Numerical Heat Transfer and Fluid Flow:
GexCon, 2015. FLACS v10.4 User’s Manual. Doxygen, Norway. Computational Methods in Mechanics and Thermal Science.
Hansen, O.R., Gavelli, F., Ichard, M., Davis, S.G., 2010. Validation of Hemisphere, New York.
FLACS against experimental data sets from the model Qiao, A., Zhang, S., 2010. Advanced CFD modeling on vapor
evaluation database for LNG vapor dispersion. J. Loss Prev. dispersion and vapor cloud explosion. J. Loss Prev. Process
Process Ind. 23, 857–877. Ind. 23, 843–848.
Hansen, O.R., Gavelli, F., Davis, S.G., Middha, P., 2013. Equivalent Spouge, J., 1999. A guide to quantitative risk assessment for
cloud methods used for explosion risk and consequence offshore installations, CMPT, Aberdeen, SD.
studies. J. Loss Prev. Process Ind. 26, 511–527. Vianna, S.S.V., Cant, R.S., 2012. Explosion pressure prediction via
Huser, A., Kvernvold, O., 2000. Explosion risk polynomial mathematical correlation based on advanced CFD
analysis—Development of a general method for gas modelling. J. Loss Prev. Process Ind. 25, 81–89.
dispersion analyses on offshore platforms.
ISO19901-3, 2015. Petroleum and Natural Gas Industries e Specific
Requirements of Offshore Structures Part 3: Topside Structure.
International Organization for Standardization.