Anda di halaman 1dari 28

12

Fire Technology 2012 Springer Science+Business Media New York. Manufactured in The United States DOI: 10.1007/s10694-012-0305-6

Advanced Methods for Determining the Origin of Vapor Cloud Explosions Case Study: The 2006 Danvers Explosion Investigation
Scott G. Davis*, Derek Engel, Filippo Gavelli, Peter Hinze and Olav R. Hansen, GexCon US, 4833 Rugby Ave, Ste 100, Bethesda, MD 20814, USA
Received: 31 May 2011/Accepted: 25 October 2012

Abstract. Gas dispersion and explosion dynamics can be very complex. Blast eects in the far eld are less sensitive to the local dynamics of an explosion event, and simplied techniques have been developed to roughly estimate the amount of fuel involved (i.e., energy released) in an explosion based on the observed damage in the far eld. However, these simple techniques are not suitable to predict the complex explosion dynamics that occur in the near eld. The phenomena that drive explosions and resulting damage in the near eld are complex in nature, as they depend on nonlinear interactions between multiple variables (e.g., ignition location, ame accelerations due to object interaction, fuel reactivity, geometry of the facility regarding connement and interconnected rooms, vent areas, etc.). Furthermore, the damage created in the near eld by overpressure development, blast wave reections and focusing, pressure impulse, pressure piling, blast wind and ensuing drag forces, may be dicult to interpret using simple methods. The origin of an explosion may, in fact, dier from the intuitive epicenter determined from oversimplied assumptions and may lead to incorrect conclusions regarding the cause and origin or initiating events of the accident. State-of-the-art 3D modeling tools, such as FLACS, are typically required to understand and evaluate complex explosions. FLACS was specically developed to predict the consequences associated with complex explosions, especially in the near eld, and has been extensively validated against hundreds of full-scale experiments. Therefore, FLACS simulations can be used to investigate the chain of events of the explosion and provide a more complete understanding of the evidence, including near-eld blast damage. This paper presents the ndings of how FLACS was used to help determine the explosion origin in the explosion that occurred at an ink and paint manufacturing facility in Danvers, Massachusetts, on November 22, 2006. This paper is not intended to provide complete investigation ndings and will only provide the necessary background material to follow the explosion origin analysis.
Keywords: FLACS, Explosion modeling, Vapor cloud explosions, Fire and explosion investigation, Cause and origin

* Correspondence should be addressed to: Scott G. Davis, E-mail: sgdavis@gexcon.com

Fire Technology 2012

1. Introduction
At approximately 2:46 a.m. on November 22, 2006, the largest explosion in the history of Massachusetts occurred in Danvers at an ink and paint manufacturing facility. The facility was a shared space between coatings, adhesives, and inks (CAI) and Arnel Company, Inc. (Arnel). The explosion was caused by an accumulation of gas or vapors within the facility that was subsequently ignited. The blast completely destroyed the 1,400 m2 (15,000 ft2) facility and caused signicant damage to the surrounding property and structures. Broken windows were observed up to 1.6 km (1 mile) from the facility. Based on the extent of far-eld damage, all investigative groups concluded that almost the entire facility needed to be lled with a ammable fuelair mixture [1, 2]. The ensuing re burned in the facility for hours after the explosion. In addition, residences immediately to the north and commercial structures to the west were severely damaged by the blast. Blast eects in the far eld are less sensitive to the local dynamics of an explosion event and simplied techniques have been developed to roughly estimate the amount of fuel that participated in the explosion. These simplications, commonly used in accident investigation, are typically not suitable to evaluate damage caused by complex explosion dynamics in the near eld. For the Danvers explosion, some key observations were made during extensive site inspections regarding the near-eld blast damage, which include: (1) the walls of a production room were observed to have been blown outward during the explosion which must have been due to a signicant overpressure early in the explosion, (2) directional damage within the facility that included lifted and bent mezzanine grating, and (3) damage to the neighboring structures outside the facility. In order to understand and evaluate complex near-eld explosion phenomena, state-of-the-art 3D modeling tools, such as FLACS, are typically required. These modern methods, along with detailed investigation techniques (e.g., relevant data collection concerning the blast, etc.), may allow the investigator to model complex scenarios and reconcile the available evidence. This paper presents an advanced analysis into determining the origin of the explosion using the computational uid dynamics (CFD) software FLACS. A brief description of the facility is rst provided, followed by the sequence of events leading up to the explosion. Next, a summary of the inspection ndings is provided. Finally, explosion analyses are conducted by igniting a ammable mixture within the facility at various locations and comparing the development of the deagrating fuelair mixture with the observed blast damage. A more detailed analysis regarding GexCons investigation can be found in a previous publication 1. This paper is not intended to address dierences in the studies conducted on the incident but to demonstrate how advanced CFD tools such as FLACS can provide valuable data for explosion investigations.

2. Background
CAI manufactured solvent-based inks and used an assortment of solvents (alcohols, aliphatic hydrocarbons, glycols, and esters), pigments, and resins in the facility.

Advanced Methods for Determining the Origin of Vapor Cloud Explosions Arnels portion of the plant was equipped to produce solvent- and water-based paints and coatings, and therefore also contained various solvents (alcohols, ketones, aromatic hydrocarbons, and esters), pigments, and resins. Both CAI and Arnel kept solvents in three underground storage tanks (USTs), and both had industrial grade nitrocellulose in trailers adjacent to the building. The layout of the facility is shown in Figure 1. The employee oces and lab facilities were on the south side of the plant. The warehouse and manufacturing areas were located on the west and north sides of the plant, respectively. The facilitys interior and exterior walls were primarily constructed of unreinforced concrete masonry units (CMUs or cinder blocks). The manufacturing section was divided into two rooms: Room E and Arnels storage and packing area (shown in Figure 2). Figure 2 also shows the locations of the underground tanks where the solvents were stored and the trailers where the industrial grade nitrocellulose was kept. The manufacturing section of the facility contained Arnels storage and packing area and Room E, as shown in Figure 3. There were two re doors in Room E: one on the west wall leading to the warehouse section and one on the east wall leading to Arnels storage and packing area. Room E had a roof with two dierent ceiling heights and housed one T-1250 (V12) mixing tank, eight 1,893-l (500gallon) totes used for solvent storage, and one 3,785-l (1,000-gallon) mixing tank. The USTs were piped directly into each companys respective pump and distribution piping manifold. Arnels 3,785-l mixing tank did not have a motor rated for use in environments having ammable gas and vapors. The motor was located just outside the facility, on the lower roof section east of the tank, in a motor housing.

Fig. 1.

Layout of the CAI/Arnel facility.

Fire Technology 2012

Fig. 2.

CAI/Arnel plant layout.

The warehouse section of the facility contained both CAIs production area and Arnels storage area (shown in Figure 4). CAI had ve closed-top, atmospheric pressure mixing tanks in the manufacturing section. The rst mixing tank (V1) was not in use. The remaining four mixing tanks, from west to east, were 03-6000 (V2), 03-1103 (V3), 01-1038 (V4) and 03-8600 (V5); these four tanks were being used to manufacture ink vehicles. Access to the top of the mixing tanks and controls was possible via a steel mezzanine. The mezzanine consisted of a 4.8 mm (3/1600 ) diamond plate welded to the top bar grating on the entire mezzanine. Arnel stored hundreds of 208-l (55-gallon) drums and thousands of pounds of powdered raw material in the south area of the warehouse section of the facility. Arnel also stored solvents in 46 1,325-l (350-gallon) portable totes stacked on the south wall.

2.1. Accident
At approximately 2:46 a.m. on November 22, 2006, multiple witnesses reported that they were woken up by a loud explosion. Many witnesses also reported hearing a second, slightly less severe explosion within 1 min after the initial blast. At the time of the incident, outside temperatures were between 3C (27F) and 2C (29F) and the wind was from the north at approximately 2.7 m/s (6 mph). Video footage from the Fox 25 News 5:00 a.m. program showed signicant burning in

Advanced Methods for Determining the Origin of Vapor Cloud Explosions

Fig. 3. Manufacturing section of CAI/Arnel facility, showing various pieces of equipment.

Fig. 4.

Perspective view of warehouse section of the CAI/Arnel facility.

Fire Technology 2012 various areas of the facility, including some tanks and other areas of the plant. The blast destroyed the CAI/Arnel facility and caused signicant damage to the surrounding properties and structures.

2.2. Site Inspections


Figure 5 is an overhead image of the remains of the CAI/Arnel facility taken 2 days after the explosion. It is clear from the image that the facility was destroyed and many objects were blown away from the building. Signicant damage was observed to the surrounding structures. A manhole cover, just north of the facility, was blown o during the explosion and found 22.9 m (75 feet) from its original location. The nitrocellulose drums stored in the trailers all appeared to have burned as a result of the ensuing re. Signicant heat and re damage were observed in the south section of the warehouse, where large quantities of raw materials were stored. In contrast, the northern wall had minimal heat and re damage. The CAI mixing tanks were still standing and uncompromised after the blast. All four mixing tanks had signicant thermal damage on the sides facing the Arnel storage racks. Despite the entire facility being destroyed, the inspection revealed that Room E had evidence of an internal overpressure condition early in the explosion event. Figure 9 shows a representative image of the original position and post-blast movement of the major components that were in Room E as viewed from the north. CAIs T-1250 and the eight 1,893-l (500-gallon) totes were still standing after the blast. The re door to the west was found blown into the warehouse area 6.1 m

Fig. 5.

Blast damage at the CAI/Arnel plant.

Advanced Methods for Determining the Origin of Vapor Cloud Explosions (20 feet) away. The east re door blew further east into the manufacturing area and was found 5.5 m (18 feet) away, covered with 20.3-cm (8-in.) thick cinder blocks. The 3,785-l mixing tank motor was blown over the totes to the east (see Figures 6, 7, 8). Inspection of Room E further revealed that the cinder block walls (30.5-cm or 12-in. thick) supporting the high ceiling section (west wall) were blown west into the manufacturing area, while the cinder block walls (20.3-cm or 8-in. thick) supporting the lower section ceiling (east wall) were blown to the east. To demonstrate the movement of Room E walls, the base of the 12-in. cinder block wall was painted green and 12-in. cinder block remnants blown further west into the warehouse were also painted green, while the base of the 8-in. cinder block wall was painted blue along with the 8-in. cinder block remnants blown further east into Arnels Storage and Packing. Inspection also revealed that there were practically no cinder blocks remaining in Room E after the explosion. Figure 9 shows a summary of the directional blast damage observed in Room E. Just to the west of Room E, directional damage from east to the west (originating from the direction of Room E) was observed in the warehouse, and consisted of lifted and deformed grating on the mezzanine (Figure 10), displaced steel trusses, and displaced railing and piping (Figure 11). The 4.8 mm (3/1600 ) diamond plating that was welded to the top of the bar grating was lifted up from the mezzanine structure and observed to be deformed away from Room E (see upper images in Figures 10, 12). This damage was consistent with pressure and drag eects associated with a blast wind originating from the direction of Room E.

Fig. 6.

Room E after the blast.

Fire Technology 2012

Fig. 7. Movement of the green-painted base blocks of the west wall of Room E, as well as the 12-in. cinder blocks blown further west into the warehouse. Note the pedestal of the supporting roof truss structure was blown to the west.

Residences and buildings immediately surrounding the facility were severely damaged by the explosion, suggesting a strong explosion (overpressures >70 kPa, or 10 psi) [3, 4] in the near eld.

3. The FLACS CFD Model


FLACS, developed and maintained by GexCon AS in Norway, is a CFD tool to model ventilation (i.e., natural or mechanical air ow), gas dispersion, gas/vapor cloud explosions and blast propagation in three-dimensional geometries, such as complex process areas. Because of its capabilities, FLACS is widely used for the quantication and management of explosion risks in the oshore petroleum industry and onshore chemical industries [5, 6]; and in some cases is explicitly required by the regulations (e.g., NORSOK standard [7]). FLACS solves the compressible Reynolds-Averaged NavierStokes (RANS) equations (conservation equations for mass, momentum, enthalpy and species) on a 3D Cartesian grid using the nite volume method [8]. The RANS equations are closed using the k-e turbulence model with the standard set of constants taken from Launder and Spalding [9]. The SIMPLE pressure correction method [10] is applied, and extended for compressible ows with source terms for the compression work in the enthalpy equation. FLACS contains a amelet-based combustion

Advanced Methods for Determining the Origin of Vapor Cloud Explosions

Fig. 8. Movement of blue-painted base blocks of east wall of Room E, as well as the 8-in. cinder blocks blown further east into Arnels Storage and Packing (Color gure online).

Fig. 9.

Directional analysis of Room E.

model with one-step reaction kinetics, where the laminar burning velocity is one important measure of the reactivity of a given mixture. A model for ame growth that describes how the local reactivity changes with parameters like concentration, temperature, pressure, turbulence, etc. is implemented to model combustion and

Fire Technology 2012

Fig. 10. ages.

Inspection photos showing thermal and directional dam-

Fig. 11.

Directional blast damage from east to west.

explosion dynamics. The real ame area is then properly described and corrected for curvature at smaller scales than the grid, using sub-grid models. Flame acceleration due to ame instability, ame-folding by obstacles and turbulent mixing is included [11, 12]. Since even small objects can have a signicant impact on gas dispersion and even more soon ame acceleration, the accurate representation of the 3D geometry has always been a focal point in the development of FLACS. A so-called distributed porosity concept [13] was developed to allow the ecient handling of complex geometries using a Cartesian grid: large objects and walls are represented on-grid, while smaller objects are represented as sub-grid elements.

Advanced Methods for Determining the Origin of Vapor Cloud Explosions

Fig. 12. Blast damage to Arnels 1,000-gallon tank and crushed partially lled 50-gallon drums in Room E (left) and thermal damage to CAIs mixing tanks in the warehouse (right).

Sub-grid elements contribute to ow resistance, turbulence generation and ame folding in the simulation, which allows the eect of small obstacles to be accounted for accurately, while maintaining reasonable simulation times [14]. Validation of FLACS for methane, propane, ethane, ethylene, and heptane gas deagrations was performed in the 1990s and included several hundred experiments [15]. Additional test series, which focused on natural gas, but included other gases such as propane, ethylene and hydrogen, were used for validation in the 1990s and 2000s and some of the signicant eorts included:  Six MERGE geometries 46 m3 and 365 m3 unconned explosions, blockage ratios of 0.05 to 0.20 for methane, propane, ethylene [16].  British Gas (Advantica) 180 m3 box [17], ve congestion levels, dierent vent sizes, geometries, dry or with water deluge (nozzles/pressure), dierent gas concentrations and more (almost 100 experiments).  CMR (GexCon) 3D-corner (27 m3), 3D pipe arrays, methane/propane, almost 10 geometries with volume blockage ratios from 0.1 to 0.5.  CMR (GexCon) M24 module (50 m3) [18, 19] variation in congestion, vent conguration, gas type and concentration, ignition location, deluge, non-homogenous clouds, pre-ignition turbulence, etc. (100s of tests conducted).  Shell SOLVEX chambers (2.5 m3, 550 m3), four congestion levels for methane, propane, ethylene (small scale only).  British Gas (Advantica) BFETS (1,600 m3 tests) [20]/HSE Phase 3A [21] (2,600 m3 tests)/Phase 3B full-scale tests [22] (1,600 to 2,700 m3 tests), which varied congestion and connement levels, ignition location, water deluge, gas concentration, cloud size variation, vent area, realistic releases, etc. (almost 100 full-scale tests).  Six tunnel tests with up to 150 m3 and 220 m3 gas cloud sizes, dierent tunnel designs [23]. In addition, signicant validation work for explosion studies for scenarios relevant to hydrogen safety has been carried out where results obtained with FLACS

Fire Technology 2012 have been compared with results from hundreds of experiments [24, 25]. This includes tests in tunnels, channels, unconned geometries (balloons, refuelling station), highly congested rigs, smooth and obstructed geometries, vented geometries, jet ignition scenarios, and more. This extensive set of experiments evaluated the eects of congestion (i.e., obstacle density) in simple symmetric and more complex geometries (see Figure 13), along with various connement levels (with or without walls). The experiments also considered: geometries that were completely lled with ammable gas versus ones that were only partially lled; blast eects due to varying the ignition location; and the eect of water sprays (deluge) on the consequences of explosions. Figure 14 shows explosion pressures for the BFETS test 7. This gure demonstrates the complexity and asymmetric nature of actual blast propagation, unlike that predicted by simple models (e.g., TNO-Multienergy [26] or Baker-StrehlowTang [27]). Figure 15 shows comparisons between experimentally recorded pressures (diamonds) and those predicted from FLACS simulations (circles) in both the near and far elds. Explosion patterns in such geometries can be quite complex and directional, where the experimental overpressure can vary by a factor of 5 to 10 for a given distance yet dierent direction from the structure. The FLACS simulations very closely reproduced the observed asymmetric overpressures inside the test geometry as well as the far-eld overpressures. Figure 15 also shows the results of the more simplistic TNO-Multienergy method. Due to the directional and asymmetry of the explosion events, it is not possible to accurately predict the explosion pressures solely with one explosion curve. This further demonstrates the failure of such methods for the prediction of near eld pressure and for geometries that are far from ideal. Grid sensitivity studies and comparisons with simpler scenarios (e.g., shock-tube blast propagation, transport of passive scalars, turbulence generation behind an object with diering grids, etc.) were other important validation eorts. As a result of the validation and grid sensitivity work, a set of meshing guidelines was developed for FLACS simulations. The parameters pertinent to this study are the following:  Cubical cells must be dened inside gas cloud and vicinity. Maintain the grid spacing all the way to the boundary in the direction of the external blast targets.  The maximum acceptable grid size in each direction will be to resolve the gas cloud (or congested region, whichever is smaller) by $8 cells if cloud is conned on both sides $10 cells if cloud is conned on one side (e.g., ground) $13 cells if the cloud is unconned  The grid size may be gradually increased towards the boundary in directions other than those towards the external blast targets. Maximum stretch factor from one cell to the next should be approximately 1.2  Main conning walls should be on grid lines

Advanced Methods for Determining the Origin of Vapor Cloud Explosions

Fig. 13. Illustration of BFETS geometries, the geometry with low obstruction density is shown in the upper picture and the high obstruction density in the lower picture. Note, the front wall has been removed on the picture to visualize the inside geometry details.

Similar studies have been carried out to ensure adequate performance of FLACS in other areas, including far-eld blast propagation from deagrations [28, 29], dispersion of ammable gas, and ignition of non-homogeneous gas clouds [6].

4. Analysis of the Danvers Explosion


Energy is released during a chemical explosion causing an overpressure and a blast (pressure) wave to propagate away from the source or epicenter of the explosion. After all the combustion energy is released, the pressure wave decreases in

Fire Technology 2012

Fig. 14. Simulated maximum pressures for BFETS test 7 (upper plot) and 3D pressure distribution just after the exit of the ame (lower plot). The non-symmetrical nature of such explosion scenarios is clearly illustrated.

strength as it propagates away from the epicenter. The blast wave can cause damage to nearby structures. Evaluating the damage caused by the blast wave allows one to approximate the energy released in an explosion, and hence the amount of fuel involved. Previous analyses [1, 2] determined that almost the entire facility needed to be lled with a ammable mixture near ideal (stoichiometric) conditions. However, sophisticated computational tools, which are capable of modeling ame propagation and resulting overpressures, need to be used to advance the investigation so that the likely fuelair mixtures and the ignition location can be determined by comparing the predicted and observed blast damage.

Advanced Methods for Determining the Origin of Vapor Cloud Explosions

Fig. 15. Sample BFETS simulations (circles) compared to experiments (diamonds). Low congestion density and end ignition (left plot), high congestion density and end ignition with activated water deluge 17 l/m2/min (right plot).

The CFD tool FLACS was used to model the explosion itself and to compare the predicted overpressures and dynamics of the burning fuel mixture with the observed structural response and blast damage within the facility. While the neareld blast damage can provide valuable information in determining the location of the ignition source and the dynamics of the initiating event, the far-eld overpressure damage is less sensitive to the local dynamics of the explosion. Explosion simulations were conducted in order to reconcile both the near- and far-eld blast damage with the likely location of the ignition source. Some key observations regarding the blast damage include:  Observed far eld blast damage estimates  Walls of Room E blown outward by the explosion (Figure 9)  Directional damage in the warehouse including the lifted and bent grating (Figure 10) The rst step in modeling the explosion is to construct a geometry model of the facility. Flame acceleration and pressure buildup within the facility are very sensitive to both connement and congestion (i.e., railing, support structures, various types of piping, hoses and ducts, small containers, etc.). An exact replica of the facility was not possible because it had been completely destroyed by the blast, and there were a limited number of photographs showing the interior of the building prior to the incident. Therefore, the approach followed in this investigation was to start from a 3D geometry model that would accurately describe the major elements of congestion, such as walls, tanks, etc., and then add congestion to approximate the likely conditions inside the building. First a detailed 3D laser scan of the building footprint and surrounding area was performed. An example of the point-cloud, which is the compilation of all the points in space recorded from multiple scans, is shown in Figure 16. This data provided accurate three-dimensional

Fire Technology 2012

Fig. 16.

Point cloud of the facility footprint looking from the North.

information regarding the building layout (including wall position, support pedestals, piping, drains, oor elevation) as well as the surrounding area (including topography, property location). The data was accurate to within 1 cm and was used to construct a model of the facility prior to the explosion. Figure 17 shows the geometry model used for FLACS simulations. Based on available photographs and the 3D laser scan, the position of walls, doors and windows could be specied accurately. The degree of congestion from this model was obviously less than observed in the available photographs (see Figure 18), and did not include many of the smaller objects. To evaluate the sensitivity of explosion pressure to the degree of congestion, a limited amount of congestion was added based on the available photographs. The congestion added to the warehouse area is visible in Figure 19 and includes three forklifts, tanks, mixers, mills, stairs, miscellaneous equipment, some railing and stationary piping. For certain ignition scenarios, the newly added congestion increased local explosion pressures approximately by a factor of three (from 25 psig to 75 psig) in the warehouse section, as shown in Figure 20. The sensitivity of the results to the level of congestion demonstrated that the explosion overpressures could be signicantly underpredicted in certain areas, such as the manufacturing area, unless the remaining items were included (e.g., exhaust ducting for all of the stationary tanks, some solvent piping for the tanks and mixers, water piping to the mixers and mills, refrigerant lines to the mills and associated jacketed tanks, smaller steam piping, sprinkler heads and associated suppression systems, lighting and associated electric conduits, control elements, miscellaneous support structures, etc.). Accounting for all sources of congestion can be quite challenging and oftentimes not possible due to the lack of information. Therefore, representative congestion (RCM) was applied to the preliminary model [30]. The addition of objects at a given congestion density, which is a

Advanced Methods for Determining the Origin of Vapor Cloud Explosions

Fig. 17. Preliminary geometry model of the CAI/Arnel facility in FLACS. The surrounding houses were modeled as simple boxes.

Fig. 18. (right).

Preliminary geometry (left), actual facility congestion

Fig. 19. Preliminary geometry (left) compared to the model with limited additional items (right).

method commonly used in explosion safety risk analyses, has been shown to more accurately represent explosion consequences for areas within facilities where the geometry model is incomplete [30]. For the present analysis, a coarse array of 6-in. diameter beams with horizontal spacing of 2 m (6.6 feet) was placed within the initial geometry model to account for some of the missing elements. The added obstructions were not meant to replicate the exact position of all the missing elements in the plant, but to bring the

Fire Technology 2012

Fig. 20. Results of congruent explosions with differing levels of geometry detail. Blue is the preliminary geometry model; green is the preliminary model plus limited additional items; red is preliminary model plus representative congestion (Color gure online).

overall congestion level closer to the actual equipment density within the plant on the day of the incident. Figure 20 shows the resulting overpressures for similar scenarios for the three geometries (preliminary, preliminary + limited additional items, preliminary + representative congestion). The trend suggests that the actual maximum (local) overpressures within the facility could have been even higher than the 890 kPa (130 psig) calculated using FLACS with the addition of the coarse congestion levels. These high pressures were noted close to the corner of the warehouse where pressure reections and focusing likely occurred. However, once strong near-eld overpressures (>70 kPa or 10 psi) are obtained, the far-eld blast propagation tends not to vary greatly; therefore, additional accuracy in the model would not translate into improved correlation with blast damage indicators in the far-eld. For the present study, both geometries preliminary + limited additional items and RCM were included as a sensitivity to the near-eld explosion development and blast damage. The temporal development of an explosion is very complicated; the goal of forensic explosion modeling is not to understand the exact failure of each element within the facility, but to capture sucient details to understand the dynamics of the explosion and to reconcile the modeling results with the observed damage. The yielding of walls and other barriers is simulated in FLACS by assigning a failure pressure and other dynamic parameters such as inertia and post-failure venting area to walls, ceiling elements, doors and windows. Windows were assumed to fail at approximately 2.0 kPa to 2.8 kPa (0.3 psi to 0.4 psi), doors at approximately 4.8 kPa (0.7 psi), and walls and ceilings at approximately 6.9 kPa to 20.7 kPa (1 psi to 3 psi) depending on the type of construction [3]. When a roof (concrete or wood) failed during the explosion, it was assumed that the available eective venting area was approximately 25% to 50%. Simulations were also performed assuming weaker roofs (50% lower opening pressure and twice as high eective

Advanced Methods for Determining the Origin of Vapor Cloud Explosions vent area) to conrm that observed trends presented are not sensitive to this parameter. For the exterior envelope of the present structure, the failing roof represented almost 65% of the available vent area. The east (manufacturing) and west (warehouse) walls of the structure each had a large overhead door and two windows that were allowed to fail and vent at 100% area. Regarding the northern wall of the warehouse: (1) the lower four feet of this wall were below grade and prevented venting; (2) various equipment and trailers abutted much of the aboveground section of this wall, further inhibiting initial venting; and (3) ve windows and one door were allowed to fail and vent at 100% area. The northern wall of the manufacturing section had three windows and one door allowed to fail and vent at 100% area, as well as the upper section of the wall in Room E. The southern wall of the facility was partially supported by the oce roof structure, as well as other external trailers and equipment, which could inhibit initial venting. The southern wall contained four windows and two doors that were allowed to fail and vent at 100% area. Therefore, the failure of internal walls and some of the external walls was ignored in the modeling, as it was assumed to have a negligible impact on the initial explosion development. Figure 21 shows an example of the ame and pressure dynamics as viewed from the north for a deagrating stoichiometric fuelair mixture, which was ignited in Arnels storage and packing area. For this scenario, the ame accelerates away from the ignition source and vents out of the building and into Room E, causing a signicant overpressure in Room E. The ame front then continues to the warehouse and signicant overpressures are observed in this area of the facility. As the ame vents from the facility, a blast wave is seen propagating away from the building. Figure 22 shows an example of the calculated peak blast pressures measured at various locations within the facility and at the surrounding structures (residences and buildings). These peak blast pressures can be used to evaluate the likelihood of various explosion scenarios. In order to determine the location of the ignition source and resulting building damage during the explosion, it is necessary to understand the dynamics of an explosion. FLACS was used to evaluate the ignition of a fuelair cloud, the propagation and acceleration of the ame front around obstacles and congestion, and the resulting overpressures. The goal of this study was not to identify which of the various hydrocarbons identied as potential fuel sources within and outside the facility participated in the explosion, but to understand the ignition origin of the resulting fuelair cloud. Many hydrocarbons found in the facility, including n-heptane, have very similar burning characteristics to propane (Figure 23). Therefore, explosion simulations were conducted using propane as the fuel, but comparable results would be expected for heptane or similar hydrocarbons found at the facility [31]. Methane and natural gas have very similar burning characteristics to each other, yet are distinctly dierent from propane and other hydrocarbons. As such, methane and natural gas explosion scenarios were modeled using methane. The explosion dynamics and far-eld loads for methane deagration scenarios were comparable to those for propane.

Fire Technology 2012

Fig. 21. North).

Flame and overpressure front progression (as viewed from

Fig. 22. facility.

Example of peak blast pressure distribution around the

Based on the damage observed in the far eld, earlier studies concluded that the energy released in the explosion would require the facility to be almost entirely lled with a ammable mixture [1, 2]. In addition, residences and buildings immediately surrounding the facility were severely damaged by the explosion, suggesting a strong explosion (overpressures >70 kPa, or 10 psi) in the facility. In order to achieve strong explosion overpressures, the mixture throughout the facility would

Advanced Methods for Determining the Origin of Vapor Cloud Explosions

Fig. 23. [32].

Laminar burning velocity curves for various hydrocarbons

Fig. 24. Expected blast pressures from explosions with equivalence ratios of a 0.6, b 0.9, c 1.1, d 1.3, and e 1.6. Red indicates overpressures 70 kPa (10 psi) and blue pressures = 14 kPa (2 psi) (Fuel = propane).

have to be near stoichiometric or ideal mixtures. To verify this hypothesis, scenarios were run with dierent fuelair ratios, ranging from fuel-lean to fuel-rich mixtures (equivalence ratio = 0.6, 0.9, 1.1, 1.3, 1.6). Figure 24 compares the peak overpressures obtained from one such set of scenarios for propane-air and conrms that only the near-stoichiometric clouds (equivalence ratio = 0.9, 1.1, 1.3) can produce near-eld overpressures that are consistent with the observations of a strong explosion (overpressures >70 kPa, or 10 psi) in the facility.

Fire Technology 2012 Explosion dynamics are not only sensitive to the fuelair mixture concentration but also to the ignition location. Therefore, eight scenarios were simulated assuming dierent ignition locations for the same near-stoichiometric cloud lling the entire production area, as shown in Figure 25. Generic ignition locations were chosen to investigate ame and pressure propagations from sources in dierent areas of origin and not specic ignition sources (e.g. motors or other electrical equipment). This was done to identify potential areas of origin and not a specic source. All ignition locations were chosen near the oor level. Locations #31 to #34 are within the warehouse, #21 to #22 within Room E, and #11 to #12 in Arnels storage and packing east of Room E. Both the preliminary geometry with limited additions and that with RCM were evaluated. These explosion simulations provided interesting results regarding the dynamics of the explosion. One nding was that the blast waves emanating from the facility were somewhat directional and depended on the location of the ignition source (Figures 25, 26). It became immediately evident from these explosion scenarios that signicant blast damage to the west (bakery) and north (houses) could be obtained from most scenarios except the two most western ignition locations (#31, #33). In addition, the additional congestion levels found in the RCM geometry resulted in pressures somewhat higher on the bakery due to the directional blast eects associated with the higher near-eld overpressures. Regardless, explosions using the preliminary geometry with limited additions resulted in elevated pressures (4 psig and 7 psig) for these scenarios. While the location of the ignition source within the facility had an observable eect on the dynamics of the explosion, sensitivity studies shows that the exact height of the ignition source has little eect on the outcome of the explosion. Figure 27 shows that ignition sources within 1.5 m from the ground (1.5 m, 1.0 m and 0.5 m) cause very similar near-eld directional and far-eld pressures. Based on the inspection of the facility, the walls of Room E were blown outward during the explosion, which can only be explained by a signicant overpressure in Room E early in the explosion. The explosion simulations revealed that:

Fig. 25. Ignition locations evaluated (left) and variation in side-on pressure at the severely damaged bakery WNW of the facility (right).

Advanced Methods for Determining the Origin of Vapor Cloud Explosions

Fig. 26. Illustrations of peak blast pressure distribution (psig) for the eight ignition scenarios.

Fig. 27. Comparison of explosion results for ignition heights of a 0.5 m, b 1 m, and c 1.5 m.

 Ignitions outside of Room E resulted in strong ame accelerations venting into Room E, causing signicant overpressures in Room E sucient to blow out its walls. Hence, the ignition must have occurred within a limited distance from the doorway to Room E in order for the ame to reach Room E prior to any signicant pressure buildup in the room of ignition (#11, #12, #32, #34).  The westernmost ignition locations in the warehouse (#31, #33) resulted in signicant overpressures in the warehouse before the ame entered Room E. This would have blown the wall that separates the warehouse and Room E in the opposite direction as that observed after the explosion.  Ideal stoichiometric mixtures ignited in Room E (#21, #22) resulted in moderately high overpressures 21 kPa and 69 kPa (3 psi and 10 psi) in Room E. However, if the mixture deviated from near-ideal conditions, signicant overpressures would have occurred outside of Room E prior to signicant overpressures in Room E, causing the walls in Room E to fail inward. Based on the above analyses, it appears that ignition locations well within the warehouse (#31, #33) can be ruled out, whereas ignitions just inside the warehouse and close to Room E (#32, #34) or in Arnels storage and packing area (#11, #12) yielded the highest overpressures in Room E. While ignition in Room E (#21, #22) cannot be ruled out, these scenarios appear less likely as near ideal conditions would be required.

Fire Technology 2012 As mentioned earlier, directional damage from east to the west was observed in the warehouse, and consisted of lifted and deformed grating on the mezzanine (Figure 10), displaced steel trusses, and displaced railing and piping. Diamond plating originally covered the mezzanine grating. This damage appears to be consistent with signicant forces associated with a directional blast wind from east to west, as well as an initial pressure dierence across the mezzanine (higher pressure from below), which had 4.8 mm (3/1600 ) diamond plates welded to the top of the bar grating. Sensor points were placed within the FLACS model to monitor the dynamic pressure (qu2), or the forces associated with drag loads, and pressure both above and below the plating of the grated mezzanine for the six remaining scenarios (Figure 28). Figure 29 shows the simulation results for the dynamic pressure loads (from east to west) and the pressure dierence across the mezzanine (Pbelow - Pabove) for the various ignition scenarios. The explosion simulations predicted an initial pressure dierence across the mezzanine (higher pressure below the plated mezzanine) followed very closely by drag (dynamic pressure) forces acting from east to west. The RCM model simulations suggest that the dynamic pressure loads (from east to west) for the ignition scenarios within the warehouse (#32, #34) were too low compared with the observed damage (see Figure 18). Simulations also revealed that the dynamic pressure associated with ignition locations in Arnels storage and packing (#11, #12) resulted in larger dynamic pressures on the mezzanine grating in the location where they were lifted and bent towards the west as compared to ignitions within Room E (#21, #22). Furthermore, the dynamic pressures from ignition locations #11, #12 were accompanied by static pressure dierences of approximately 27.6 kPa to 34.5 kPa (4 psi to 5 psi) across the mezzanine surface for ignition positions #11 and #12, which are also consistent with the lifting of the grating. Lower pressure dierences across the mezzanine were observed for ignition positions #21 and #22 (within Room E). The preliminary model with limited additions revealed that ignition scenarios within Arnels storage and packing (#11, #12) were the only ones consistent with the lifted and bent grating. In addition, the explosion trends exhibited by the geometry with limited additions are consistent, yet lower in magnitude than those

Fig. 28. Monitoring panel and points placed to measure the observed pressures and drag forces on the mezzanine.

Advanced Methods for Determining the Origin of Vapor Cloud Explosions

Fig. 29. Dynamic pressure (upper) from east to west and pressure difference (lower) measured across the mezzanine (Pbelow 2 Pabove).

obtained using the RCM. This is not surprising, as the geometry with limited additions is known to not contain all of the known congestion at the facility. Both geometries adequately predict the governing dynamics in the explosion, which demonstrate that when the RCM method is properly employed it can account for missing congestion in the model to more closely predict the explosion overpressures while maintaining predictions of the dominant explosion dynamics. The results from both geometries demonstrate that the explosion most likely originated in Arnels storage and packing area. These ndings are consistent with a previous study, which evaluated the migration of ammable vapors within the facility at night, when the ventilation system was turned o. With the ventilation system o, a release of gases or vapors originating in the manufacturing and solvent storage areas of the facility would be quickly mixed to near homogeneous conditions by the sole heater in operation located in the manufacturing section. This would allow the cloud to mix, get larger and gradually migrate to the remote storage and packing area of the facility. At this time, the cloud would have lled almost the entire facility (determined necessary to cause the far-eld blast damage) and ultimately found an unidentied ignition source in the remote storage and packing area of the facility.

5. Conclusion
Considering the damage observed inside the facility, it would be intuitive to deduce that the explosion initiated in Room E. The directional analysis of the

Fire Technology 2012 blast (shown in Figure 9) makes clear that the walls, re doors, and associated equipment were all blown outwards from Room E. Detailed simulations show that this scenario, however, is unlikely, as the damage in other parts of the facility cannot be reconciled by ignition inside Room E. While ignition in Room E cannot be entirely ruled out, simulation results indicate that Arnels storage and packing area is the most likely area of origin. Ignition locations in Arnels storage and packing were capable of reconciling the observed blast damage in the facility, including: (1) signicant overpressure in Room E causing the walls to blow outward, (2) the blast venting into the warehouse causing the observed drag eects associated with a blast wind in this section of the buildings, and (3) directional damage to surrounding structures outside the facility. This suggests that this area of the facility likely contained an unidentied piece of equipment that was either not rated for use in ammable environments or compromised, and was a key nding in the investigation. Gas dispersion and explosion dynamics can be very complex. Oversimplication of the driving factors of an explosion may result in incorrect determinations of accident causes. The evidence provided by detailed inspections may be compared with simulations using state-of-the-art 3D modeling tools such as FLACS to better understand complex explosions. Exploring the balance of assumptions and details in the presented case study demonstrates the possibility of counter-intuitive results that are discovered by coupling investigation ndings with FLACS simulations.

References
1. Davis SG, Hansen OR (2010) New investigation ndings on the 2006 Danvers MA explosion. J Loss Prev Process Ind 23(2):194210 2. CSB Investigation Report (2008). 2007-03-I-MA, May 2008 3. Lees FP (1996) Loss prevention in the process industries, 2nd edn. Elsevier, Butterworth 4. Harris RJ (1983) The investigation and control of gas explosions in buildings and heating plant. British Gas Corporation, London 5. Hansen OR, Renoult J, Bakke JR (2001) Explosion risk assessment: how the results vary with the approach chosen. In: Proceedings of Mary Kay O Connor Process Safety Symposium, pp 395410 6. Hansen OR, Talberg O, Bakke JR (1999) CFD-based methodology for quantitative gas explosion risk assessment in congested process areas: examples and validation status. International Conference and Workshop on Modelling the Consequences of Accidental Releases of Hazardous Materials. AIChE, Centre of Chemical Process Safety, pp 457477 7. NORSOK Z-013 (2010) Risk and emergency preparedness analysis. Norsok standard. Available from Standard Norge, Postboks 242, N-1326 Lysaker, Norway 8. Hjertager BH (1985) Computer simulation of turbulent reactive gas dynamics. J Model Identif Control 5:211236 9. Launder BE, Spalding DB (1974) The numerical computation of turbulent ows. 2. s.l. Comput Methods Appl Mech Eng 3:269289

Advanced Methods for Determining the Origin of Vapor Cloud Explosions


10. Patankar SV (1980) Numerical heat transfer and uid ow. Hemisphere Publications, New York 11. Abdel-Gayed RG, Bradley D, Lawes M (1987) Turbulent burning velocities: a general correlation in terms of straining rates. Proc R Soc Lond A 414:389413 12. Bray KNC (1990) Studies of the turbulent burning velocity. Proc R Soc Lond A 431:315335 13. Arntzen BJ (1998) Modeling of turbulence and combustion for simulation of gas explosions in complex geometries. PhD thesis, NTNU, Trondheim, Norway. ISBN 82-4710358-3 14. Bjerketvedt D, Bakke JR, van Wingerden K (1997) Gas explosion handbook. J Hazard Mater 52:1150 15. Hansen OR, Storvik I, van Wingerden K (1999) Validation of CFD-models for gas explosions, where FLACS is used as example: model description and experiences and recommendations for model evaluation. In: Proceedings of the European Meeting on Chemical Industry and Environment III Krakow, Poland, pp. 365382 16. Mercx WPM (1994) Modelling and experimental research into gas explosions. Overall report of the MERGE project, CEC contract: STEP-CT-0111 (SSMA) 17. Catlin CA, Gregory CAJ, Johnson DM, Walker DG (1993) Explosion mitigation in oshore modules by general area deluge. Trans IChemE 71, Part B 18. Hjertager BH, Bjrkhaug M, Fuhre K (1988) Explosion propagation of non-homogeneous methane-air clouds inside an obstructed 50 m3 vented vessel. J Hazard Mater 19:139153 19. Hjertager BH, Fuhre K, Bjrkhaug M (1988) Gas explosion experiments in 1:33 and 1:5 scale oshore separator and compressor modules using stoichiometric homogeneous fuel/air clouds. J Loss Prev Process Ind 1:197205 20. Selby C, Burgan B (1998) Blast and re engineering for topside structures, phase 2, nal summary report. Steel Construction Institute, UK. Publication number 253 21. Al-Hassan T, Johnson DM (1998) Gas explosions in large scale oshore module geometries: overpressures, mitigation and repeatability. Presented at OMAE-98, Lisbon, Portugal 22. Johnson DM, Cleaver RP, Puttock JS, van Wingerden CJM (2002) Investigation of gas dispersion and explosions in oshore modules. OTC paper 14134, Houston 23. Zipf RK, Sapko MJ, Brune JF (2007) Explosion pressure design criteria for new seals in U.S. coal mines. U.S. Department of Health and Human Services, Public Health Service, Centers for Disease Control and Prevention, National Institute for Occupational Safety and Health. DHHS (NIOSH) Publication No. 2007-144, Information Circular 9500, pp 176 24. Middha P, Hansen OR (2009) Using computational uid dynamics as a tool for hydrogen safety studies. J Loss Prev Process Ind 22:295302 25. Middha P, Hansen OR (2008) Predicting deagration to detonation transition in hydrogen explosions. Process Saf Prog 27:192204 26. van den Berg AC (1985) The multi-energy method: a framework for vapour cloud explosion blast prediction. J Hazard Mater 12:110 27. Pierorazio AJ, Thomas JK, Baker QA, Ketchum DE (2005) An update to the BakerStrehlow-Tang vapor cloud explosion prediction methodology ame speed table. Process Saf Prog 24(1):5965 28. van Wingerden CJM, Hansen OR, Teigland R (1995) Prediction of the strength of blast waves in the surroundings of vented oshore modules. In: Fourth international conference of oshore structures-hazards, safety and engineering working in the new era, London, England, 1213 December

Fire Technology 2012


29. van Wingerden K, Hansen OR, Foisselon P (1999) Predicting blast over pressures caused by vapor cloud explosions in the vicinity of control rooms. Proc Saf Prog 18:1417 30. Hansen OR, Hinze P, Engel D, Davis S (2010) Using computational uid dynamics (CFD) for blast wave predictions. J Loss Prev Process Ind 23:885906 31. Hansen OR, Wilkins B (2003) Oil mist explosions in a test channel. In 37th Annual Loss Prevention, symposium, New Orleans 32. Davis SG, Law CK (1998) Determination of and fuel structure eects on laminar ame speeds of C1 to C8 hydrocarbons. Combust Sci Technol 140:427499

Anda mungkin juga menyukai