On the use of consequence models for accident ...

Paper 5b

On the use of consequence models for accident investigations

Prankul Middha and Kees van Wingerden GexCon AS P. O. Box 6015 Postterminalen NO-5892 Bergen Norway Phone: +47-55-57-4145 [email protected]

[This work is a property of GexCon AS and may not be reproduced without permission]

Prepared for Presentation at American Institute of Chemical Engineers 2009 Spring National Meeting 43rd Annual Loss Prevention Symposium Tampa, Florida April 26-30, 2009

UNPUBLISHED

AIChE shall not be responsible for statements or opinions contained in papers or printed in its publications

51

LPS 2009 ____________________________________________________________ Paper 5b

On the use of consequence models for accident investigations Prankul Middha and Kees van Wingerden GexCon AS, P.O. Box 6015 Postterminalen, Bergen, NO-5892, Norway

Abstract Accident investigations are performed to understand the cause of an accident to avoid it from happening again. Consequence models can contribute understanding the chain of events that led to the accident such as where the ignition occurred (including indicating hat the ignition source was), how much fuel was involved in the accident, can explain damage that was seen, help identifying the source of leak of flammable material released and estimating the potential consequences of an event. The models can also be used to identify and design protection measures after an accident. In the present paper several examples of use of a consequence model (FLACS) for accident investigations performed are given. The examples give an insight of e.g. sub-models that needed to be developed and implemented to be able to verify observations made and to identify possible related causes of the accident, they show how to represent geometrical details including failing structural components and discuss necessary validation of sub-models used/implemented. The accidents that were considered include among others an aeroplane crash (TWA 800), a tank explosion (Sløvåg), a coal mine explosion (Sago) and a chlorine release (Festus). The paper focuses on the developments, geometrical representation and validation in these accidents. The accidents themselves are presented only briefly.

1. Introduction It is a well-known that accidents are found “everywhere” in the process and transport industry. Several years of research in safety has reduced the incidence of fatal accidents but they still occur with alarming frequency. Many accidents have occurred in the past 30 years that have caused severe damage be it in the petrochemical industry (Piper Alpha), chemical industry (Flixborough), mining industry (e.g. Sago mine), transport industry (TWA 800), and many more. It is essential to learn from past accidents in order to understand the mechanisms of accidents and to develop accident prevention and control strategies. In order to fulfil this aim, the root causes have to be revealed and corrective actions implemented. This is because among these accidents many are due to the repetition of the same/similar faults (Kletz, 1991a; Kletz, 1991b). However, industries (both process and transport) are generally reluctant in revealing what had happened and have a tendency to underplay their mistakes. This aspect has been discussed by several authors (e.g. Lees, 1996). Therefore, thorough accident investigations are a necessity. Investigations will very often highlight flaws in manufacturing procedures or in new work equipment and processes and should therefore be taken seriously. Further, it is also 52

2

LPS 2009 ____________________________________________________________ Paper 5b

important to investigate all “near misses” since it is impossible to predict which time it could have escalated to have serious consequences. Consequence models can contribute understanding the chain of events that led to the accident (or resulted in a serious accident being avoided even if the initiating event was present) such as where the ignition occurred (including indicating what the ignition source was), how much fuel was involved in the accident, can explain damage that was seen, help identifying the source of leak of flammable material released and estimating the potential consequences of an event. The models can also be used to identify and design protection measures after an accident. They are thus a very valuable resource in identifying the cause or causes of the accident and establishing recommendations to avoid a repeat. These consequence models can be of several kinds, starting from simple engineering correlations to complex Computational Fluid Dynamics (CFD) tools. The CFD tools provide the possibility to model the underlying physics involved in safety analyses. It is possible to take account of effects of buildings, mitigation measures, piping and vessel arrangements, etc. which have been found to have a strong influence on the consequences of any accident or unwanted incident. The effect of accident scenario parameters such as release location, rate, direction, ignition source location, cloud size, cloud location, etc. can also be modelled. Due to these advantages, CFD has increasingly been used in recent years as a part of risk analyses in the oil and gas and associated chemical industry, (NORSOK, 2001; Hermann, 2007). In the present paper, several examples of use of a CFD-based consequence model (FLACS) for accident investigations performed are given. The examples give an insight of sub-models that needed to be developed and implemented to be able to verify observations made and to identify possible related causes of the accident, they show how to represent geometrical details including failing structural components and discuss necessary validation of sub-models used/implemented. The accidents that were considered include among others an aeroplane crash (TWA 800), a tank explosion (Sløvåg), a coal mine explosion (Sago) and a chlorine release (Festus). The paper focuses on the developments, geometrical representation and validation in these accidents. The accidents themselves are presented only briefly.

2. Brief Description of the CFD tool The FLACS tool used by the authors is a CFD code solving the compressible Navier-Stokes equations on a 3-D Cartesian grid using a finite volume method. The conservation equations for mass, impulse, enthalpy, turbulence and species, closed by the ideal gas law are included. It has been developed in-house since 1980 in order to simulate the dispersion and subsequent ignition in process areas, such as oil platforms. The SIMPLE pressure correction algorithm is applied (Patankar, 1980), and extended to handle compressible flows with additional source terms for the compression work in the enthalpy equation. Hjertager (1985,1986) describes the basic equations used in the FLACS model, and Hjertager, Bjørkhaug & Fuhre (1988a,b) present the results of explosion experiments to develop and validate FLACS initially. The numerical model uses a second order scheme for resolving diffusive fluxes and a secondorder “kappa” scheme (hybrid scheme with weighting between 2nd order upwind and 2nd order central difference, with delimiters for some equations) to resolve the convective fluxes. Second order schemes in time have been implemented, but are generally not used due to short 53

3

LPS 2009 ____________________________________________________________ Paper 5b

time steps. FLACS uses a standard k- model for turbulence (See, e.g. Harlow and Nakayama, 1967). with certain modifications, e.g., a model for generation of turbulence behind sub-grid objects, a model for build-up of proper turbulence behind objects of a particular size (about one control volume) for which the discretization produces too little turbulence, and turbulent wall functions (with no slip condition) for adding production terms to the relevant control volume across the boundary layer. The FLACS code uses a “distributed porosity concept” which enables the detailed representation of complex geometries using a Cartesian grid. Large objects and walls are represented on-grid, and smaller objects are represented sub-grid. This enables geometrical details to be characterized while maintaining reasonable simulation times. This is necessary as small details of “obstacles” can have a significant impact on flame acceleration, and hence explosion pressures (Bjerketvedt, Bakke & van Wingerden, 1997). This approach represents geometrical details as porosities (opposite of blockage) for each control volume. The porosity concept models the blockage, drag formulation, sub-grid turbulence generation and flame folding coefficients to obtain good simulation results despite coarse grid resolutions. Sub-grid objects contribute to flow resistance, turbulence generation and flame folding in the simulation as it is important to model the turbulence correctly for partly porous and “subgrid” objects to obtain good results. In case of small objects, the flow kinetic energy lost due to drag is compensated as a source term for turbulent energy. The flame folding contribution is very important for explosion calculations. More details of this concept are given in Hjertager (1985,1986) and Arntzen (1998). FLACS contains a combustion model that assumes that the flame in an explosion can be regarded as a collection of flamelets. One-step reaction kinetics is assumed, with the laminar burning velocity being a measure of the reactivity of a given mixture. The model consists of two parts: a three-step burning velocity model and a flame model. The flame model gives the flame a constant flame thickness (equal to 3-5 grid cells) and assures that the flame propagates into the reactant with the specified velocity. To represent flame folding around subgrid obstacles, a flame folding model has also been implemented. The burning velocity model consists of the following three models: (a) A laminar burning velocity model that describes the laminar burning velocity as a function of gas mixture, concentration, temperature, pressure, oxygen concentration in air and amount of inert diluents, (b) A model describing quasi-laminar combustion in the first phases of flame propagation after ignition. Due to flame instabilities, the observed burning velocity increases as the flame propagates away from ignition (due to flame wrinkling), and (c) A model that describes turbulent burning velocity as a function of turbulence parameters (intensity and length scale). The model is based on a broad range of experimental data (Abdel-Gayed, Bradley & Lawes, 1987; Bray, 1990).

3. Examples of Accident Investigations 3.1. TWA 800 Explosion Investigation General Summary The TWA flight 800 accident occurred on July 17, 1996 just outside New York City shortly after take-off. The airplane blew into pieces as a result of an “explosion” and 230 people were

54

4

LPS 2009 ____________________________________________________________ Paper 5b

killed. The investigation concentrated on the possibility of a gas explosion in the Centre Wing Fuel Tank (CWT). The hypothesis was that the heating of fuel in the tank by the air conditioning system was the cause of the flammable vapor concentration (temperature above flash-point). The explosion investigation used laboratory experiments, scale-model tests, and numerical simulations to examine the explosion of Jet-A (aviation kerosene) mixtures with air under conditions simulating the center wing tank environment at the time of the accident. Work was carried out over a period of four years to determine the chemical and physical properties of Jet A, particularly the flammability limits, combustion behavior, and the propagation of flames through the compartmentalized structure of the center wing tank. The CFD tool FLACS was adapted and validated against scale-model experiments. It was then used in full-scale simulations to explore the effects of various parameters and assumptions, especially ignition locations within the tank. All of this information was integrated through a rule-based system to attempt to narrow down the number of plausible ignition locations that would be consistent with the observed damage as deduced from the recovered wreckage. The propagation of an explosion in the multi-compartment center fuel tank was investigated using scale-model experiments and numerical methods for simulating fluid motion, heat transfer and propagating flames. The problem of quenching or flame extinction was identified as an issue and addressed through experiments and modeling. Modelling Details The question our part of the investigation team received was: where was the ignition source? At that time it was clear that the source of the accident had to be found in the Central Wing Tank. Figure 1 shows a schematic view of the Central Wing Section. It was divided laterally into compartments by six beams, including (from farthest aft to forward) the rear spar, spanwise beam (SWB)-1, the mid spar, SWB-2, SWB-3, and the front spar. It had been established that due to heat given off by the air conditioning system located underneath the CWT, the remaining fuel (Jet-A) had been evaporated causing a flammable fuel-air mixture in the tank. Now it had to be determined how the observed failure of the fuel tank would have been possible by a gas explosion in the tank. On the basis of the analysis of the various parts of the Central Wing Tank (recovered from the bottom of the ocean) it appeared that only one of the separation walls (in fact the front wall) had failed. This failure led to a progressive failure of the whole fuselage resulting in the plane breaking in two. It was found that a butane-air mixture with an equivalence ratio (ER) of 0.79 at an initial temperature of 50 °C was able to represent the Jet-A fuel (that is a complex mixture of over a 100 hydrocarbons) reasonably well (at the correct operating conditions). The burning velocity of this mixture was 43 cm/s. Standard tests in the ¼-scale single bay geometry revealed that the reactivity of the butane-air mixture agreed very well with the Jet-A fuel while the maximum overpressure was slightly under-predicted. Similarly, it was found that a Jet-A mixture at 40 °C could be represented by an ER = 0.62 butane-air mixture. This mixture had a laminar burning velocity of 21.6 cm/s. Scale-model experiments and numerical simulations revealed that the flame front could propagate rapidly between the compartments of the tank once the flame reaches the passageways and vent stringers connecting the compartments. In some cases, the flame was 55

5

LPS 2009 ____________________________________________________________ Paper 5b

quenched as it passed through these openings or connections, especially when ignition occurred far away from the opening. Due to the pressure build-up inside a compartment a strong turbulent field is generated in the wake of the opening. When the flame arrives at such an opening the flame was quenched by the turbulence.

Figure 1. A schematic of the Centre Wing Tank

At the time of the investigation, a flame quenching criteria was not available in FLACS. This was developed on the basis of small-scale experiments carried out at the University of Bergen (Larsen, 1998). On the basis of experiments a quenching criterion was developed and inserted 56

6

LPS 2009 ____________________________________________________________ Paper 5b

into FLACS (these were independent from the Jet-A experiments performed in the ¼-scale CWT facility). The criterion was based on the Karlovitz number K. The Karlovitz number is a ratio of the combustion time scale (  / S L ) to the turbulence time scale ( u '/  ). This has been used by Abdel-Gayad et al. (1987) to express conditions at which quenching due to flame stretch would occur. From the calculated velocities and observed flame transmission phenomena, a critical Karlovitz number was established as a function of the opening diameter. For opening diameters larger than 7 mm (the minimum size in the full scale CWT), the critical Karlovitz number was found to be K = 140150. This criterion was further consolidated by comparing to the results of ¼-scale experiments. On the basis of these results, the criterion derived from the experiments performed by Larsen (1998) was revised into a criterion representing a likelihood of transmission as a function of Karlovitz number. This indicated e.g. a quenching probability of 20 % for K = 100200 and a quenching probability of 100 % for K > 300.

To verify whether FLACS would be able to perform such calculations the code was validated against experiments performed on 1:4 scale. These experiments were initially performed with a fuel representing the reactivity of Jet-A (consisting of 1.4 % v/v propane and 7 % v/v hydrogen.). Later, experiments were also performed with Jet-A. The laboratory and scalemodel experiments with Jet-A were carried out under conditions (temperatures of 4050 °C and a pressure of 0.585 bar) simulating the CWT environment at the event altitude. Additional experiments were carried out with simulated jet fuel in the scale model tests. The experiments with Jet-A demonstrated that the mixture was flammable over the range of temperatures (4060 °C) and fuel loading (50100 gallons) present in the tank. The elevated temperatures inside the tank increased the explosion hazard over a cool tank by substantially increasing the amount of fuel vapour and decreasing the ignition energy. While the amount of liquid fuel inside the tank was limited and had been in the tank since the previous flight from Athens, there was sufficient fuel vapour to create an explosion hazard. The measured peak pressure rises were between 1.5 and 4 barg, sufficient to cause failure of structural components. The majority of the ¼-scale tests in which simulant fuel was used were simulated in order to investigate and qualify the properties of the code regarding predicting the outcomes of explosions in both the ¼-scale experimental set-up and the real Centre Wing Tank. Figure 2 shows how the ¼-scale rig was represented in FLACS. Dimensions and possible other properties of openings, stringers, vents and partitions were based on information made available by the California Institute of Technology (1997). Figure 2 also shows several moments of flame propagation of an explosion ignited in bay 3. The simulations revealed reasonable agreement with experimentally observed overpressures. It can be seen from Figure 2 how the flame shoots from bay to bay via the connecting holes after a relatively slow initial phase causing turbulence and considerably faster explosions in these other bays. For some bays (especially bays 3, 4, 5 and 6) flames can enter a bay from more than one direction due to the presence of several surrounding bays. This could for some scenarios lead to a very big sensitivity of change of ignition position. Sandia also carried out some simulations for this investigation with their tool varying the ignition location all over the ¼-scale set-up to investigate both the sensitivity of the ignition source location and to find locations where the observed damage agrees with observed predicted loadings. The two tools were also validated against each other. Thus ignition locations could be established at which the pressure differential across each separation wall was able to explain the damage seen.

57

7

LPS 2009 ____________________________________________________________ Paper 5b

Figure 2. Moments of flame propagation in the ¼-scale rig as described by FLACS.

After the completion of the validation against ¼-scale experiments, simulations for the real CWT were carried out. The ignition position was limited to the seven fuel probe positions and the position of the compensator, thus giving a total of 8 ignition locations (large tanks had 2 ignition positions, and smaller tanks had 1). The different scenarios were created by parametrically varying the ignition location (8 points), fuel vapor concentration (2 levels), and time-delay (16 or 40 ms) between failure of SWB-3 and the panel in SWB-2. Therefore, a total of 32 scenarios were considered. The results of the simulations considered the loading of each partition as a function of time and the possibility of flame quenching in the various orifices upon flame arrival. The question of probable ignition location was addressed by combining the results of numerical explosion simulations with structural failure estimates to predict damage caused by the pressure differences within the wing tank structure. The predicted and observed damages were folded into a rule-based system to evaluate the consistency between internal explosion scenarios and the observed damages associated with the early events. This system was developed by Combustion Dynamics, Inc. in order to evaluate the explosion scenarios as a function of various parameters (Thibault, 2000). The rules and structural failure thresholds were based on a consensus developed through discussions with some members of the Accident Sequence group and Boeing. The rule-based system evaluation of these numerical simulations found ignition locations that would produce propagating flame fronts within the tank volume and pressure differences on the structural components that were consistent with the observed damages. A single ignition location could not be established but the results were such that one could conclude that a single ignition source could have ignited the fuel and the resulting gas explosion could have caused the damage seen and the accident as it occurred.

3.2. Sago Mine Explosion

Several thousands of people die globally in mine accidents every year. A significant fraction of the accidents are caused by explosions of methane and sometimes coal dust. Most of these take place in China, quite often due to low safety standards in sometimes illegal coal mines. In recent years there have also been very severe accidents in USA and the EU. In January 2006, USA experienced its worst coal mine explosion accident for decades when 12 workers died in the Sago mine. Poland also had a very severe accident in 2006 with 23 people killed.

58

8

LPS 2009 ____________________________________________________________ Paper 5b

The gas causing these explosions, methane, is continuously generated from the rock. During mining operations proper ventilation systems will generally make sure that the gas concentration is far below flammability limits. When miners leave an area, this area is usually sealed with a thick barrier of concrete. Behind this seal, the gas concentration will gradually build up, and may for shorter or longer periods become explosive. An ignition source, which could be smoldering dust or sometimes lightning, may initiate an explosion. If the seals are leaking, this could also lead to explosive regions in the vicinity of the seals. The purpose of the seals is to contain possible explosions inside the sealed area. In 1997 FLACS was used to blindly predict methane explosions in two different mine tunnel test geometries. The South-African research institute, CSIR, thereafter performed the experiments and evaluated the blind predictions. Good results were seen for FLACS blind predictions for both scenarios. As a consequence of the Sago Mine accident in 2006, the US National Institute for Occupational Safety and Health, NIOSH (previously the US Bureau of Mines), started an activity evaluating the design of seals. One observation from the Sago accident was that the concrete seals applied were far from strong enough to withstand the explosion. In the research initiated by NIOSH, one activity included the simulation of existing NIOSH experiments (performed in the Lake Lynn experimental mine) to evaluate the ability to predict or reproduce existing experiments. Thereafter potential mine explosions at larger scales were simulated to study the required seal strength as a function of separation distance of seals. The team also visited several countries to learn from their practices and experiences in constructing seals. Experimental mine explosions can generally only study comparatively small volumes of explosive mix. Most experiments worldwide fill less than 20 m (65 ft) of tunnel with methane-air mix, although a few tests have filled as much as 58 m (190 ft) of tunnel with explosive mix. In this study, gas explosion simulations were used to extrapolate small-volume gas explosion data to larger gas explosions typical of what could happen in a coal mine. The first step was model validation. Herein, FLACS calculations were performed to duplicate measured pressure-time histories from six tests done in the Lake Lynn Experimental Mine (LLEM). Six different experiments with gas cloud sizes ranging from about 3.718.3 m in two different mine geometry configurations were simulated. Figure 3 (left) shows the test and model geometry for three multiple-tunnel experiments in A, B, and C drifts of the LLEM, and Figure 3 (right) shows the same for three single-tunnel experiments in D drift. The FLACS simulations were carried out prior to receiving information about the test results. Still very good results were obtained, with 8 out of 12 maximum overpressures (6 tests and 2 observation locations) predicted within 10 % of observed values, and the worst prediction of the 12 with a deviation of only 24 %. The results for all 6 tests for both sets of drifts are summarized in Tables 1a and 1b. Figure 4 shows measured versus computed pressure-time histories for three different locations. It can be seen that the experimental and simulated overpressures agree well with each other. The magnitudes of the peak pressures seen in the simulations, as well as the shape of the pressure waves, compare well with the observations. However, some discrepancy is seen in the arrival time of the pressure waves. The first arrival of the calculated pressure wave is slower than that seen in the experiments. This difference can be explained by the difference in the ignition source model used as compared to the actual ignition source. The FLACS simulations used a single-point ignition, whereas in the actual tests, an electric match that emitted a shower of sparks started the explosion simultaneously in 59

9

LPS 2009 ____________________________________________________________ Paper 5b

many different locations. In summary, despite the difference in timing, the gas explosion models reproduced the measured experimental data well.

Figure 3. Layout of calibration models in the LLEM. Left: A,B,C-drift calibration tests; right: D-drift calibration tests.

Table 1a. Comparison of calculated to measured LLEM experimental gas explosion overpressures in single-tunnel (D-drift) tests

Test No.

Length of methane zone (m)

Approximate methane volume (m3)

468

3.66

4.25

469

8.23

9.91

470

12.2

15.21

Gauge No.

Measured peak pressure (kPa)

D1 D10 D1 D10 D1 D10

21.53 18.59 58.87 48.13 75.03 74.80

60

Simulations Calculated pressure (kPa) 18.70 17.60 57.50 51.90 76.30 71.00

Percent difference -13.14 -5.33 -2.33 7.83 1.69 -5.08

10

LPS 2009 ____________________________________________________________ Paper 5b Table 1b. Comparison of calculated to measured LLEM experimental gas explosion overpressures in multiple-tunnel (A-, B-, and C-drift) tests

Test No.

Length of methane zone (m)

Approximate methane volume (m3)

484

12.2

16.14

485

18.3

23.64

486

18.3

23.64

Gauge No.

Measured peak pressure (kPa)

B-10 B-526 B-10 B-526 B-10 B-526

83.89 29.43 101.30 34.14 48.12 24.88

Simulations Calculated pressure (kPa) 71.20 28.66 97.64 42.34 57.06 23.10

Percent difference -15.13 -2.62 -3.61 24.02 18.58 -7.15

After the successful model validation, the next set of simulations examined larger volumes of a completely confined explosive mixture. In this case, no experiments were carried out but the simulation results were used as an input to improved design of seals. The model geometry was based on the same LLEM model employed earlier. “Infinitely strong” seals were placed 41, 71, 161, 228, or 300 m from the end of B drift for the purpose of simulations. A stoichiometric methane-air mixture filled the entire volume, and ignition occurred at the end of B drift. Figure 5 shows the computed pressure-time history at one of the seals (seal B). With the 41-m cloud, the pressure rises to a pressure level that is close to the 908-kPa constant volume (CV) explosion pressure over 0.5 s and then remains at that level as expected. The computed pressure-time history shows some reflections, but their magnitude is small. With the 71-m cloud, the pressure rises to about 1.0 MPa and then stabilizes again at the same pressure level as that for the 41-m cloud (close to the CV explosion pressure). When larger clouds (161, 228, and 300 m) are present, sharp pressure peaks are seen where the pressure rises to 2–3 MPa in less than 0.1 s, but then equilibrates to CV explosion pressure as expected. These very high pressures may not be accurate since detonation may have occurred, and models to capture DDT or detonation are not included in FLACS. However, the models are likely correct in indicating that very high pressures have developed. Further, in complementary validation tests with FLACS for hydrogen systems, it has been seen that FLACS can predict explosion pressures relatively accurately even in cases where a transition to detonation occurs. These calculations also suggested that gas clouds with run-up lengths less than 50 m may not develop pressures much beyond 1.0 MPa and may be less likely to detonate unless there are turbulence-intensifying obstructions, such as mine timbers, standing supports or other forms of roughness. Further simulations considered an explosive mixture that forms directly behind a seal due to air leakage. This explosive mixture is only partially confined and able to vent freely into inert atmosphere deeper into the sealed area. The model geometry is again based on the LLEM. The model has infinitely strong seals in the A, B, and C drifts at 228 m from the beginning of B drift. A stoichiometric methane-air mixture filled the volume for 15, 30, or 60 m behind the seals. The ignition point is right behind the B-drift seal, which is the worst possible case. In this case, computed pressures at the B seal range from 100500 kPa. As a result of this work, a public report has been issued on explosion pressure based design criteria for new seals in US coal mines (Zipf, Sapko & Brune, 2007). Thus, FLACS simulations played an important role in confirming the reasons for the accident and in 61

11

LPS 2009 ____________________________________________________________ Paper 5b

evaluating and developing measures in order to limit the damage in case of any future incident.

Figure 4. Comparison of experimental data with simulation predictions. Top: Lake Lynn Experimental Mine calibration data; bottom, calculations from FLACS model.

62

12

LPS 2009 ____________________________________________________________ Paper 5b

Figure 5. Calculated pressure-time histories at seal for large-volume explosions using FLACS.

3.3. Festus Chlorine Release Accident

A chlorine release accident occurred at a typical medium-sized industrial site in Festus, Missouri on August 14, 2002. We were not asked to contribute to accident investigations. This work was carried out to demonstrate the possibilities that FLACS could offer to assist in emergency response and planning in case of accidental or deliberate toxic gas releases. In recent years, FLACS has been validated against several large-scale tracer gas release experiments, including those carried out at Kit Fox, MUST, and Prairie Grass. This work has involved model improvements and comparison against experimental data. The results have been presented in Dharmavaram, et al. (2005) and Hanna, et al. (2004). FLACS has also been applied to blind predict tracer gas concentrations following a release in Manhattan (Hanna, et al., 2006). Thus, the capabilities of FLACS to analyze such an incident can be established. A description of the Festus accident and the analysis work performed is given in the following. More details can be found in Hanna, et al. (2009). At the time of the incident, chlorine was being offloaded from a railcar parked at the site. The tank containing chlorine had a pressure of approximately 12 barg. No information on the temperature was available. However, it could be assumed that the tank temperature was the same as the ambient temperature that was around 20 °C. The underlying cause of the incident was the rupture of a 2.5 cm diameter hose near the railcar at a height of 3.5 m. However, due to the sequence of the automatic closing of various safety valves in the system, the actual release occurred from a pipe at a height of 3 m. Weather conditions at the time of the accident 63

13

LPS 2009 ____________________________________________________________ Paper 5b

were overcast with some drizzle, which meant a neutral atmospheric stability (Pasquill class D). There was no direct information on wind speeds at the accident site. However, wind speeds were observed to be 3 m/s at the St. Louis airport, which is about 60 km north of Festus. It was observed that the chlorine jet impinged on the side of the railcar 2 or 3 m away following the release. The release continued for over 3 hours before it was stopped by emergency personnel and in that time, approximately 22,000 kg of liquid chlorine was released. The release rate was nearly steady (about 2.02 kg/s). Most of the jet was deflected underneath the railcar and could be seen coming out from the other side. Videos and photos taken during that time show a large visible chlorine gas cloud, of depth about 1 m and width about 20 or 30 m around the railcar. However, no information exists about the observed arrival times of the visible gas cloud from photographs and videos. Because there were no concentration observations taken during the release period of the Festus accident, it is not possible to carry out quantitative evaluations of the FLACS concentration predictions. FLACS simulations were used to determine whether the simulations yielded similar behavior as was observed. A more important motivation was to see whether CFD simulations can be a valuable tool for accident investigations and for predicting the consequences of an accidental or deliberate toxic gas release. For this purpose, a continuous chlorine source emission rate of 2.02 kg/s was used, assuming that the release was at a height of 3 m above ground. Based on the observations, an inclined jet pointed 45 degrees downwards hitting the cylindrical body of the railcar was used. The wind speed assumed as input to FLACS was 2 m/s (at 10 m elevation), since it was assumed that the local wind speeds at Festus were likely to be slightly lower than at the flat St Louis airport, where the official wind speed of 3 m/s was recorded. Due to various objects in the general area, surface roughness length could be considered to be 20 cm. The process immediately following the release was not directly modelled. A utility program FLASH (Salvesen, 1995) was applied to estimate conditions in the gas jet at the distance from the source where all the flashing chlorine would have evaporated and some air entrainment had taken place. This is done prior to the actual CFD simulations. The FLASH utility estimated the initial jet as a mixture of chlorine gas and air with a mass rate of 6.83 kg/s (2.02 kg/s chlorine), a temperature of 71 °C, and a velocity of 76 m/s at the given distance (1.0 m). This was used as an input to FLACS. Thus, the actual release in FLACS was already a chlorineair mixture. A grid resolution of 0.20 m  0.20 m  0.20 m was used very close to the source in order to properly resolve the dense jet. A slightly coarser resolution of 1 m  1 m  0.20 m was used in the area within about 50 m of the source, where the videos showed a shallow layer of dense gas. The grid was gradually stretched (made coarser) further away and above the source location. Since only visual data on the chlorine gas cloud is available, it is not possible to compare the simulated results with observations directly. Further, it is essential to determine the concentration that represents the edge of the visible cloud in order to carry out comparisons with the observations. According to a chlorine safe work practices report, chlorine is supposed to be invisible below 1000 ppm. However, direct comparison is complicated because the visible cloud is thought to contain chlorine and water droplets in addition to chlorine gas. Therefore, after comparing many of the videos with the FLACS concentration patterns, it was decided that a simulated chlorine concentration of 2000 ppm seemed to agree better than 1000 ppm to the observed cloud edge. 64

14

LPS 2009 ____________________________________________________________ Paper 5b

The FLACS results are presented in Figure 6 in side-by-side format with observations as available in the videos from Fox News. The view angles and the tanks and buildings are approximately the same in each pair of figures. It is seen that the FLACS-simulated 2000 ppm cloud compares well in a qualitative manner to the observed chlorine cloud. The horizontal dimensions are similar and the simulated cloud is also seen to be bounded by obstacles and tree lines. The initial jet shape following the impingement also agrees reasonably well with observations. The cloud in the simulations is also seen to flow underneath the railcar.

Figure 6. Comparison of frames from Fox News videos on the left and FLACS simulations on the right (where the 2000 ppmv level is shown) for two different views (used with permission from M. Ichard).

The vertical dimension of the cloud obtained in FLACS was around 1 m (The top panel of Figure 6 shows the simulated cloud contour touching the base of the tank car, which is 1.2 m above the ground.). This agreed well with observations and is typical of the known behaviour

65

15

LPS 2009 ____________________________________________________________ Paper 5b

of dense gas clouds. The 1 m estimate of the observed depth of the visible cloud is partly based on videos and photographs. It is not possible to compare observed and simulated arrival and departure times of the visible chlorine cloud at Festus from photographs and videos, since they are not available from those periods of the accident. The cloud is seen to be more or less steady state in the available visual records. In the FLACS simulations, where a constant source emission rate is assumed, the visible gas cloud is simulated to reach a steady state in the domain around 10 minutes after the release starts. Thus, it can be concluded that FLACS can be a valuable tool to aid investigations of such toxic gas releases (deliberate or accidental) and can also be used to predict the flow and concentration behavior following a release.

3.4. Sløvåg Tank Explosion Accident

This section describes the accident investigation performed by GexCon following the tank explosion (in tank T3) and subsequent fire that occurred at Vest Tank AS at Sløvåg Industrial Area in Western Norway on 24th May 2007. An atmospheric storage tank exploded, and the subsequent fire destroyed the remaining tanks in the tank farm, as well as some buildings and trucks near the tank farm. The extent of the work included revealing the direct causes of the accident, as well as evaluating plausible courses of events in light of the extent of the accident, witness observations, and a literature study on similar events. The analysis revealed that it was highly unlikely that a “physical” explosion was the cause of the accident and it seemed reasonable to assume that it was a chemical explosion that led to the accident. FLACS simulations were used to evaluate the possibility of formation of explosive atmosphere inside the tank. There was considerable uncertainty about the actual chemical composition of the mixture present in the tank prior to the explosion. The accident occurred one day after the process of adding hydrochloric acid to the tank was started in order to dissolve the precipitated solids containing high amounts of sulphur-containing coumpounds (mainly methyl mercaptan) at the bottom of the tank formed after many months of treatment of coker gasoline (in order to reduce the sulphur content). It may be concluded that the primary contribution to the explosive atmosphere came from the release of methyl mercaptan, and possibly also other sulphur containing components. This was due to their reduced solubility in the alkaline solution following the addition of hydrochloric acid (reduced pH). Buoyancy due to a vertical temperature gradient (exothermal reactions presumably raised the temperature of the solution somewhat), diffusion due to a vertical concentration gradient, and/or convection due to the motion of the liquid surface (circulation during the addition of hydrochloric acid), may have promoted the transport and mixing of air and flammable gases released from the liquid surface. Simulations with the CFD-code FLACS were performed to explore the process of mixing flammable vapour and air inside the tank. The purpose of this work was to explore reasonable locations for potential ignition sources, based on the extent of the flammable mixture. The analysis assumed an ambient temperature of 12 oC, a total mass flow rate of 1 kg s-1 evenly distributed over the liquid surface, and temperatures of the released vapours from the liquid surface of 35, 50, or 70 oC. A release rate of 1 kg/s corresponds to a hypothetical release scenario involving boiling pentane. However, the released vapour was modelled as propane, with a molecular weight of 44 kg/kmol, since this 66

16

LPS 2009 ____________________________________________________________ Paper 5b

corresponded reasonably well with both the lighter components in coker gasoline and the most relevant sulphur containing compounds: methyl mercaptan and hydrogen sulphide have molecular weights 48 and 34 kg/kmol, respectively. Figure 7 illustrates the simulated concentration gradient after 71 minutes in a vertical crosssection of the tank, for a release of vapour at 35 oC from a liquid surface. The flammable mixture has reached the top of the tank, while the concentration close to the liquid surface is above the upper flammability limit. The flow velocities resulting from buoyancy are approximately 45 cm s-1. Figure 8 shows how the simulated fuel concentration develops in the monitor points along the centre of the tank for the same simulation. Figure 9 illustrates the effect of decreasing the density of released vapour (by increasing the temperature, but reduced molecular weight would in principle have given similar results), on the simulated mole fractions of fuel vapour near the roof of the tank. Reduced density increases the buoyancy, and the flammable mixtures reach the top of the tank faster.

Figure 7. Vertical cross-section of tank T3 illustrating the simulated mole fraction of fuel vapour and velocity vectors after a 71 minutes release from a liquid surface at elevated temperature; total release rate 1 kg/s; release temperature 35 °C. The drop in concentrations near the left edge of the tank is a visualization artefact.

67

17

LPS 2009 ____________________________________________________________ Paper 5b

Mole fraction of fuel vapour_

1,000

0,100

M 01 M 20 M 30 M 40

0,010

M 50 M 60 M 70

0,001 0

1000

2000

3000

4000

5000

6000

Time (s)

Figure 8. Simulated mole fractions of fuel vapour as a function of time at various elevations inside the tank; total release rate 1 kg/s, release temperature 35 °C; monitor points shown in Figure 7.

Mole fraction of fuel vapour_

1,000 T = 35 °C T = 50 °C 0,100 T = 70 °C

0,010

0,001 0

2000

4000

6000

8000

Time (s)

Figure 9. Simulated mole fractions of fuel vapour near the roof of the tank as a function of time; monitor point M 70 in Figure 7; release rate 1 kg/s, and liquid temperatures of 35, 50, and 70 oC, respectively.

The simulations suggest that the entire free volume of the tank at some point in time would contain a flammable mixture. In the early stages, the explosive atmosphere would be limited to a zone just above the liquid surface. However, over time the zone with explosive mixture would move upwards, and probably reach the roof inside the tank within hours, while the fuel concentration near the liquid surface would exceed the upper flammability limit. Hence, it is not likely that the explosive mixture occupied the entire free volume of the tank at the same time, but this could nevertheless occur due to buoyancy driven flow or convective mixing, or by other means. Further simulations were carried out to explore the possibility of releasing flammable mixtures through the openings in the tank, and thereby creating explosive atmospheres that could reach potential ignition sources outside the tank. A geometry model of the tank facility

68

18

LPS 2009 ____________________________________________________________ Paper 5b

at Sløvåg was implemented in FLACS. To ensure conservative estimates of the simulated flammable gas clouds, the analysis assumed that pure fuel vapour (simulated as pentane) had replaced all the air in the tank, and the corresponding release rates are therefore high. The first simulation involved a leak with release rate 1 kg/s, pointing downwards from the position of the exit from the air filter, and wind from the south at a speed of 6 m/s. This simulation suggested that no traces of fuel reached the monitor point upstream of the tank, and that the concentrations in the monitor points downstream of the release are of the order 0.011 ppm, i.e. significantly lower than the lower flammability limit for the relevant fuels. Additional sensitivity calculations were carried out including one with a higher release rate, and one with a lower wind speed, among others. The simulation results suggested that it is highly unlikely that an ignition source outside the tank ignited a release of flammable mixture from tank T3 on 24th May 2007. Even though the simulations involved quite conservative assumptions, i.e. high release rates or low wind speeds, the extent of the explosive atmosphere was found to be very limited. Thus, FLACS calculations were able to provide a direct estimate of the possibility of flammable clouds in various regions, showing that the only possible ignition source can have been self-ignition of activated carbon in the air filter on top of the tank (this was further supported by the finding of whitish ash-like material amongst the activated carbon granules).

4. Final Remarks Through the work presented in this article, it has been shown that FLACS can be a powerful tool in order to carry out accident investigations and to propose measures to reduce consequences of any future incident. Four diverse examples are chosen to demonstrate the possibilities of the use of FLACS. These include an example from the public transport industry (TWA 800 crash), mining industry (Sago mine accident), chemical transport industry (toxic gas release), and petrochemical industry (Sløvåg tank accident). Some of the investigations involved further model development while others involved applications in new areas. The strategy was to carry out validation against available experiments in order to gain confidence in the predictive capabilities of FLACS and then apply it to the problem of interest. CFD simulations with the FLACS tool offer the possibility to simulate realistic physics and should be a part of the tool-box of any investigator.

5. References Abdel-Gayed, R.G., Bradley, D., & Lawes, M., 1987. Turbulent burning velocities: a general correlation in terms of straining rates, Proc. R. Soc. Lond. A, 414, 389–413. Arntzen, B.A., 1998. Modelling of turbulence and combustion for simulation of gas explosions in complex geometries, PhD Thesis, NTNU, Trondheim, Norway, ISBN 82-4710358-3. 69

19

LPS 2009 ____________________________________________________________ Paper 5b

Bjerketvedt, D., Bakke, J.R. & van Wingerden, C.J.M., 1997. Gas Explosion Handbook. J. Haz. Mat., 52 (1), 1-150. Bray, K.N.C., 1990. Studies of the turbulent burning velocity, Proc. R. Soc. Lond. A, 431, 315–335. Dharmavaram, S., Hanna, S. R. & Hansen, O. R., 2005. Consequence analysis – using a CFD model for industrial sites, Process Safety Progress, 24, 316–327. Hanna, S. R., Hansen, O. R. & Dharmavaram, S., 2004. FLACS air quality CFD model performance evaluation with Kit Fox, MUST, Prairie Grass, and EMU observations, Atmos. Environ. 38, 4675–4687. Hanna, S. R., Brown, M. J., Camelli, F. E., Chan, S., Coirier, W. J., Hansen, O. R., Huber, A.H., Kim, S. & Reynolds, R. M., 2006. Detailed simulations of atmospheric flow and dispersion in urban downtown areas by Computational Fluid Dynamics (CFD) models – an application of five CFD models to Manhattan. Bull. Am. Meteorol. Soc., 87, 1713–1726. Hanna, S. R., Hansen, O. R., Ichard, M. & Strimaitis, D., 2009, CFD model simulation of dispersion from chlorine railcar releases in industrial and urban areas. Atmos. Environ., 43, 262270. Harlow, F.H. & Nakayama, P.I., 1967. Turbulence Transport Equations. Phys. Fluids, 10, 2323-2332. Herrmann, D., 2007. Why Use CFD for Explosion Studies? Proceedings of the 41st Annual Loss Prevention Symposium, Houston, TX. Hjertager, B.H., 1985. Computer simulation of turbulent reactive gas dynamics. J. Model. Identif. Control 5, 211–236. Hjertager, B.H., 1986. Three-dimensional modeling of flow, heat transfer, and combustion. Handbook of Heat and Mass Transfer. Gulf Publishing Company, Houston, Texas, pp. 304– 350 Chapter 41 Hjertager, B.H., Bjørkhaug, M., & Fuhre, K., 1988a. Gas explosion experiments in 1:33 scale and 1:5 scale; offshore separator and compressor modules using stoichiometric homogeneous fuel–air clouds. J. Loss. Prev. Process Ind., 1, 197–205. Hjertager, B.H., Bjørkhaug, M., & Fuhre, K., 1988b. Explosion propagation of nonhomogeneous methane-air clouds inside an obstructed 50 m3 vented vessel. J. Haz. Mater., 19, 139–153. Kletz, T.A., 1991. The Philips explosion—this is where I came in. Chemical Eng. 493, 47–53. Kletz, T.A., 1991. Process safety—an engineering achievement. Proc. Inst. Mech. Engr., 205, 11–23.

70

20

LPS 2009 ____________________________________________________________ Paper 5b

Larsen, Ø., 1998. A study of the critical dimensions of holes for transmission of gas explosions and development and testing of a Schlieren System for studying jets of hot combustion products, M.Sc. Thesis, University of Bergen. Lees, F. P., 1996. Loss prevention in chemical process industries. London: Butterworth. NORSOK Standard Z-013, 2001. Risk and emergency preparedness analysis, Norwegian Technology Center, Oslo, Norway Rev. 2. Available from www.standard.no/pronorm3/data/f/0/01/50/3_10704_0/Z-013.pdf. Patankar, S.V., 1980. Numerical Heat Transfer and Fluid Flow, Hemisphere Pub, ISBN: 0070487405. Salvesen, H.C., 1995. Modeling Release of Liquefied Gas under High Pressure, Report Nr. CMR-95-F20062, Christian Michelsen Research. Thibault, P.A. (2000) A Rule-Based System for the Evaluation of Explosion Scenarios. Combustion Dynamics Ltd Report CDL-00-XXXX. Development of rule-based system and application to ¼-scale and full-scale simulations. Zipf, R. K., Kr., Sapko, M. J. & Brune, J. F., 2007. Explosion Pressure Design Criteria for New Seals in U.S. Coal Mines. Available from http://www.cdc.gov/ niosh/mining/pubs/pdfs/2007-144.pdf.

71

21