Evaluating Vapor Dispersion Models for SafetyAnalysis of LNG Facilities

Research Project

Technical report

Prepared by:

J. IvingsF. Jagger

J. LeaM. Webber

Health Safety Laboratory

(i)CID

THEFIRE PROTECTIONRESEARCH FOUNDATION

.. ....

.... r

.... -- .",.- .....-... -.. --

--.w

........--- . - .... - - " ---.. - &. --... ---.. --...-

JI,

----. ". "'-

THE FIRE PROTECTIONRESEARCH FOUNDATION

ONE BATTERYMARCH PARKQUINCY, MASSACHUSETTS , U.S.A. 02169

MAil: Foundation~NFPAorgWEB: www.nfpa.org/Foundation

19 Copyright The Fire Protection Research FoundationApril 2007

FOREWORD

NFPA 59A Standard for the Production , Storage , and Handling ofLiquefied Natural Gas (LNG), provides limited guidance on the use ofvapor dispersion models for the analysis of safety features for LNGfacilities. Due to recent developments , and new laboratory andmodeling research , the NFPA 59A Committee anticipates receivingproposals requiring NFPA 59A to reference specific vapor dispersionmodels. These models are complex; hence , the Committee is in needof some guidance on their application to large LNG spill scenarios.The Committee is also seeking evaluation tools/criteria to assist themin their decision making process for referencing models in thestandard.

As a result the Fire Protection Research Foundation of the NFPAhave let a contract with the Health & Safety Laboratory, the researchagency of the United Kingdom Health & Safety Executive , to examinemodeling of dispersion of LNG spills on land , and provide them withguidelines for assessment of available dispersion models.

The Research Foundation expresses gratitude to the report authorsJ. Ivings , S.F. Jagger, C.J. Lea and D.M. Webber of Health &

Safety Laboratory; and to the Project Technical Panelists andsponsors listed on the following page.

The content, opinions and conclusions contained in this report aresolely those of the authors.

Evaluating Vapor Dispersion Models for SafetyAnalysis of LNG Facilities

Technical Panel

Anay Luketa-Hanlin , Sandia National Laboratories

Andrew Kohout , FERC

Robert Morrill , Wrentham Fire Department

Davis Parsons , BWD Consulting

David Butler , City of Everett Fire Department

Frank Licari , U. S. Department of Transportation

James Lewis , PTL Associates , Div. of ICF International

Kevin Ritz , Baltimore Gas & Electric Co.

Ted Lemoff, NFPA

Paul Croce , Consultant

Phani Raj, Technology & Management Systems , Inc.

Richard Hoffmann , Hoffmann & Feige

Robert Meroney, Colorado State University

Thomas Felleisen , U.S. Coast Guard

SponsorsAmerican Gas Association

BP America

Distrigas of Massachusetts

ExxonMobil

Southern LNG

Weavers Cove Energy, LLC

CB&I

Harpur Hill , BuxtonDerbyshire , SK17 9JNT: +44 (0)1298 218000F: +44 (0)1298 218590

W: www. hsi.gov.

Evaluating Vapor Dispersion Models for SafetyAnalysis of LNG Facilities

MSU/2007/04

Project Leader: Dr M.J. Ivings

HEALTH & SAFETYLABORATORY

Author(s): Dr M.J. Ivings, Dr S.F. Jagger, Dr C.J. Leaand Dr D.M. Webber

Science Group: Health Improvement

DISTRIBUTION

K. Almand Fire Protection Research Foundation

D. MorganA. Curran

Mathematical Sciences Unit Head, HSLDirector, Health Improvement Group, HSL

PRIVACY MARKING:

RESTRICTED: COMMERCIALThe work described in this report was undertaken by the Health and Safety Laboratory undercontract to Fire Protection Research Foundation. Its contents, including any opinions and/orconclusions expressed or recommendations made, do not necessarily reflect policy or views ofthe Health and Safety Executive.

HSL report approval:Date of issue:Job number:Registry file:Electronic file name:

Dr A.D. Curran21 February 2007

JC7300025026305LNG-model-evaluation.pdf

iC) Crown copyright (2007)

ACKNOWLEDGEMENTS

The authors would like to thank Dave Huckaby, US Dept of Energy, National EnergyTechnology Laboratory, Jerry Havens and Tom Spicer, University of Arkansas for the hugeamount of effort that they have put into providing us with all of the information that we haveneeded to apply the MEP to their models

We would also like to thank the Fire Protection Research Foundation, the Project TechnicalPanel and the NFPA Committee 59A for their support and useful feedback during the course ofthis project.

The authors would also like to that Nijs Jan Duijm, Riso National Laboratory, and Joseph CChang, George Mason University, for helping us acquire the REDIPHEM database, and theModelers Data Archive and supporting documentation, respectively.

iii

1.4

2.4

3.4

CONTENTS

INTRODUCTION ..................................................................................... 1

Background..............................................................................................Outline of report.......................................................................................Liquefied Natural Gas (LNG) ................................................................... 2Model evaluation......................................................................................

LNG DISPERSION MODEL EVALUATION PROTOCOL.....................Overview................................................................................................Scientific assessment ............................................................................Verification.............................................................................................Validation...............................................................................................Evaluation criteria ..................................................................................

CLASSIFICATION OF MODELS

..........................................................

Introduction............................................................................................Workbooks Correlation ........................................................................ 42Integral models ......................................................................................Shallow layer models.............................................................................CFD models...........................................................................................

MODEL REVIEWS

................................................................................

DEGADIS ..............................................................................................FEM3A...................................................................................................FLUENT.................................................................................................

GUIDANCE ON MODEL APPLICATION

..............................................

CONCLUSIONS

...................................................... ..............................

Recommendations.................................................................................

REFERENCES

......................................................................................

GLOSSARY...........................................................................................

APPENDIX A.........................................................................................

Example contents of a validation database ........................................... 69

10 APPENDIX B.........................................................................................

10. Questionnaire ........................................................................................10.2 DEGADIS MER......................................................................................10.3 FEM3A MER.......................................................................................... 7610.4 FLUENT MER........................................................................................

EXECUTIVE SUMMARY

Background

NFPA 59A Standard for the Production, Storage, and Handling of Liquefied Natural Gas(LNG), provides limited guidance on the use of vapor dispersion models for the analysis ofsafety features for LNG facilities. Due to recent developments , and new laboratory andmodeling research, the NFPA 59A Committee anticipates receiving proposals requiring NFPA59A to reference specific vapor dispersion models. These models are complex; hence, theCommittee is in need of some guidance on their application to large LNG spill scenarios. TheCommittee is also seeking evaluation tools/criteria to assist them in their decision makingprocess for referencing models in the standard.

As a result the Fire Protection Research Foundation of the NFPA have let a contract with theHealth & Safety Laboratory, the research agency of the United Kingdom Health & SafetyExecutive, to examine modeling of dispersion of LNG spills on land, and provide them withguidelines for assessment of available dispersion models.

Objectives

The objectives of this project were:

To develop a tool dedicated to the evaluation of predictive models of dispersion of LNGspills on land and assist the decision process in referencing models in NFP A 59A.

To make recommendations to the NFPA regarding criteria for model evaluation andtranslate these into a form suitable for use by the Association for model selection.

To review appropriate versions of the DEGADIS , FEM3A and FLUENT models forLNG dispersion by reference to their scientific basis, their user interface and uncertaintyand validity of results in the range of applications relevant to situations of concern.

To provide guidance to the Association on the application of models to large LNGspills.

Main Findings

This project has led to the development of a Model Evaluation Protocol (MEP) that can be usedto assess the suitability of dispersion models for predicting hazard ranges associated with largespills of LNG. The protocol is based on that developed by the EU SMEDIS project (Carissimoet al. 2001 , Daish et al. 2000) for dense gas dispersion but has been modified to make itspecifically applicable to the dispersion of LNG on land.

The MEP is based on three distinct phases: scientific assessment, model verification and modelvalidation. The scientific assessment is carried out by obtaining detailed information on a modelfrom its current developer using a specifically designed questionnaire and with the aid of otherpapers, reports and user guides. The scientific assessment examines the various aspects of amodel including its physical , mathematical and numerical basis, as well as user oriented aspects.This assessment allows the model to be evaluated against II proposed qualitative assessmentcriteria. The outcome of the scientific assessment is recorded in a Model Evaluation Report(MER), along with the outcomes of the verification and validation stages. The template for theMER has been designed to aid the reviewer to extract all of the necessary information tocomplete the scientific assessment.

The verification stage of the protocol is treated passively as in the original application of theSMEDIS protocol. This means that instead of carrying out a specific exercise to verify that themodel has been implemented correctly and accurately, evidence of model verification is soughtfrom the model developer and this is then assessed and reported in the MER. The validationstage of the MEP involves applying the model against a database of experimental test casesincluding both wind tunnel experiments and large-scale field trials. The aim of the validationstage is therefore to quantify the performance of a model by comparison of its predictions withmeasurements. Specific datasets and validation cases against which the model can be comparedhave been identified. A number of physical comparison parameters and statistical performancemeasures have been defmed which allow the model to be assessed via a number of proposedquantitative assessment criteria. The content of a validation database has been defmed togetherwith initial guidance on its construction and use. However, the validation database has yet to beconstructed and so this part of the MEP cannot currently be applied and is outside the scope ofthis work.

The MEP has been applied, excluding the full validation stage, to "DEGADIS Version 2.FEM3A February 2007 version" and "DOE-NETL LNG Dispersion Module for FLUENT2/6. . A full scientific assessment of the three models has been undertaken and the MER'

have been included in full in the Appendices. All three models met all of the qualitativeassessment criteria. For each model a general description of the model has been given alongwith the scientific basis of the model. The limits of applicability of each model are describedand an assessment of previous validation of the model is given.

The full application of the MEP will help the NFP A to make decisions on the appropriateness ofdispersion models for predicting hazard ranges for large LNG spills. However, like the modelsthemselves, the MEP is subject to uncertainty and, although we have made the best possible useof previous work on model evaluation, the MEP would benefit ftom refinement following theapplication of the MEP in full to a number of models.

How a model is used is at least as important as the choice of model itself. Therefore someguidance has been given on the application of dispersion models for assessing the hazards fromLNG spills. The importance of the source model is discussed and concerns are raised forsituations where the ' source ' model includes a model for the initial dispersion of the vapor. Incases such as these we recommend that the LNG dispersion MEP is applied to the ' dispersionpart of the source model.

Recommendations

Based on the work that we have carried out we make the following recommendations:

To complete the MEP , the validation database should be constructed and based on thepresent work.

A number of models should then be subjected to the full MEP , including the validationexercise. The MEP should then be refmed in the light of this new information.

The need for a separate model evaluation exercise to assess the appropriateness ofsource models for LNG dispersion, including guidance on their use, should be givenserious consideration.

INTRODUCTION

BACKGROUND

The 2006 edition of the NFPA Standard 59A for the Production, Storage and Handling ofLiquefied Natural Gas (LNG) concerns the safe design, siting, construction and operation ofLNG facilities including loading and unloading facilities, vaporizing plant and processequipment. In the United States this Standard has been incorporated into the Department ofTransportation (DOT) Siting Regulation 49 CFR 193 which the Federal Energy RegulatoryCommission (FERC) relies on for guidance and determination of siting requirements.

Regulation 49 CFR 193 and the standard NFPA 59A require exclusion zones defmed as areascontrolled by the terminal operator or the government. The exclusion zones must not extendbeyond the area controlled by the operator or the government. One exclusion zone concernsvapor dispersion and extends to the distance nom the release at which the mean concentrationof vapor falls to half the Lower Flammability Limit (LFL).

The standard prescribes the events or LNG releases which must be considered and then sets outmethods which must be used to predict the resulting hazard zones and conditions under whichthe hazard zones must be derived. There is a requirement that either the model DEGADIS(Havens and Spicer, 1985 , 1988 , 1990) is used or a model that meets a number of specifiedcriteria and is approved by the appropriate authority.

However, there have been a number of recent developments which have led the NFPA toreconsider this situation:

The increased interest in LNG as a fuel is very likely to involve an expansion in thenumber of LNG facilities and, as a result, proposals to use a wider range of models forthe prediction of exclusion zones. The NFP A are therefore seeking advice onmethodologies for assessing the validity and quality of LNG dispersion models.

Recent research work has questioned the existing methodologies. Wind tunnel modelingstudies at the University of Arkansas (Havens and Spicer, 2006) has suggesteddeficiencies in the SOURCE 5 model (which is often used to compute the inputconditions for the DEGADIS model) for the specific case of an LNG release nom atank or pipe into an impoundment area. This work also led to further development of theFEM3A model, producing a version tailored to the prediction of LNG dispersion, as

well as information which the DOE NETL is now using to assess FLUENT as analternative Computational Fluid Dynamics (CFD) model to FEM3A.

Recent research, and in particular that described in a recent special issue of the Journalof Hazardous Materials (Volume 140, 2007), has highlighted deficiencies anduncertainties in our knowledge in the area of LNG dispersion.

As a result the Fire Protection Research Foundation of the NFPA have let a contract with theHealth & Safety Laboratory, the research agency of the United Kingdom Health & SafetyExecutive, to examine modeling of dispersion of LNG spills on land , and provide them withguidelines for assessment of available dispersion models.

OUTLINE OF REPORT

The remainder of the Introduction provides some background information on LNG hazards andan introduction to model evaluation. Further background information can be found in the

reviews by Luketa-Hanlin (2006) and Cleaver (2007) and the references cited therein. Volume140 (2007) of the Journal of Hazardous Materials also provides much information on thehazards associated with LNG.

Section 2 of this report introduces a Model Evaluation Protocol (MEP) for LNG dispersionmodels and describes in detail how the information is gathered and then evaluated using anumber of qualitative and quantitative assessment criteria. A review of the different types ofdispersion models is presented in Section 3. Application of the protocol, excluding activevalidation of the model, is presented in Section 4 as applied to three models: DEGADIS Version

, FEM3A February 2007 version and DOE-NETL LNG Dispersion Module for FLUENT2/6.

Section 5 of the report provides guidance on the application of models to large LNG spills andthe conclusions from this project are presented in Section 6. References are provided in Section

A Glossary of key terms used within this report is provided in Section 8.

LIQUEFIED NATURAL GAS (LNG)

LNG as methane

Liquefied Natural Gas (LNG) is mainly methane (CH4) with a small admixture of higher, lessvolatile hydrocarbons. For the purposes of hazard assessment, LNG is usually consideredeffectively to be methane to a very good fIrst approximation and this will be adopted throughoutthis report.

Properties of methane

Table l.l below provides the properties of methane (Reid et a!. , 1987). We use Sf units, exceptfor the adoption of kmol instead of mol to make molecular weights look more familiar, and barinstead of Pascals to make pressures more manageable (I bar = 10' Pa). The vapor pressurecurve up to the critical point is shown in Figure l.l and the section focusing on a range aroundthe normal boiling point is presented in Figure 1.2.

Table 1.1 Physical properties of methane

Molecular weight: 16.04 kglkmolFreezing point: 90.Boilingpoint: 111.7 Liquid density at B.P. 425 kg/mCritical temperature:Critical pressure:

190.446. bar

~ 25

e. 2.

Vapour pressure

110 130 150 170 190 210 230 250 270 290

T (K)

Figure 1. Vapor pressure of methane

Vapour pressure

100 105 110 115 120 125 130

T(K)

Figure 1. Vapor pressure of methane around its boiling point

The critical temperature of 190.4 K means that methane cannot be liquefied by pressure atambient temperature. Rather it would have to be cooled significantly below this temperature.In practice, therefore, liquefaction is achieved at ambient pressure by cooling to the boilingpoint at II!.? K. This is quite different trOll LPG (liquefied petroleum gas which is largelypropane and butane) which is liquefied under a pressure of several bar at ambient temperature.

The liquid density means that a large tank of LNG - say 30 m high - would have a liquid headof around 1. 3 bar. This gives a measure of the sort of pressures one has to pump against. Theyare significantly lower than those involved in LPG storage.

An increase of temperature by only a few degrees , corresponds with an increase of saturatedvapor pressure comparable with the head of liquid.

The low molecular weight of methane (compared with air at around 29 kg/kmol) means that atambient temperature methane is lighter than air. However methane at its boiling point issignificantly denser (typically by about a factor of 1.5) than ambient temperature air, and LNGspills are therefore likely to result in heavy gas clouds.

The source of the hazard

Liquid spills

If LNG escapes from its containment a cold, flammable gas cloud will result. A breach in pipe-work or the side of a tank may result in a boiling liquid pool.

The source term for a heavy gas cloud dispersion calculation is crucially dependent on the areaof the pool and its rate of vaporization. There are broadly two modes of vaporization:

Boiling: the liquid temperature is the boiling point, and the rate of vaporization is controlled bythe rate at which heat is transferred from the surroundings (primarily the substrate in mostcases) to supply the heat of vaporization. The methane gas concentration immediately abovethe pool surface is 100%.

Evaporation: the liquid is well below the boiling point and the rate of vaporization is controlledby the rate at which air flow above the pool can carry the vapor away. The methane gasconcentration immediately above the pool surface is governed by its partial pressure and issignificantly less than 100%.

In reality vaporization is a combination of both modes. The heat balance is governed by anequation schematically of the form:

Tt=Qi" Q",p l.l

where Qi" is proportional to the heat transfer rate to the pool, and Q",p is proportional to the heatof vaporization multiplied by the vaporization rate. In the boiling scenario the temperature Tisconstant at the boiling point and

Q",p Qi" determines the vaporization rate. As heat is

extracted from the surroundings and the ground coolsQi" may drop to the point where the air

steam can remove material faster than the incoming heat can vaporize it. At this point Q",p

Qi" and the pool will start to cool. In this case there will thus be a gradual transition to theevaporation mode" described above. Whether or not this happens may depend on the detailed

spill scenario: for example a spreading pool can continually reach warm ground and the boilingregime may be prolonged. Whether or not it is significant may depend on whether the initialrapid boiling was enough to initiate a disaster.

Roll over

Another hazard which should be considered is escape of gas through the roof of the tank. Thiscan happen following "rollover : if one considers liquid at the top of a high tank to be at apressure of I bar, then the liquid at the bottom also experiences the head of liquid above it andmay be at 2 bar or so from our observations above. A glance at the vapor pressure reveals that itcan exist there in liquid form at a temperature higher by quite a few degrees K. If this isallowed to happen , then the equilibrium in the tank can become unstable, and a sudden rollovercan send liquid from the bottom of the tank to the top. Its higher temperature means that itshigher vapor pressure is applied to the roof of the tank, which may then fail , leading to the

escape of a significant gas cloud. The quantity of gas created will depend on the temperatureand heat capacity of the liquid. It is not clear whether hazard analyses routinely consider thisscenario , though it is known to have happened.

Rapid phase transition (RPT)

Also of interest is the possibility of rapid phase transitions. If heat transfer to the liquid LNG israpid enough then an RPT may occur - a rapid vaporization causing a shock wave. Predictingwhen they occur is difficult, and they are not generally, as far as we know, considered in hazardanalysis, presumably under the implicit assumption that the flammable hazard is greater.Cleaver et al (2007) include a summary of possible effects. Interestingly this is one area wherethe higher hydrocarbon content of the LNG may be important in determining a propensity forRPT. Furthermore a hazard which should perhaps be considered is an RPT arising ITom a spillfrom a ship on to water, involving rapid mixing with water (providing the high heat transfer)and causing a shock wave in the water which may increase the damage to the ship. Luketa-Hanlin (2006) also reviews rapid phase transitions and notes that the enhanced vaporization rateshould an RPT occur, can lead to significantly longer hazard ranges. Rollover is considered byCleaver et al (2007).

3.4 LNG Dispersion and modeling

General considerations

LNG hazards are usually analyzed in three phases: source term (usually covering thedevelopment and vaporization of a pool), dispersion (the transport of the gas) and effect(radiation ITom fire or pressure wave from explosion).

Initially the pool will boil very rapidly, and the vaporization rate is controlled mainly by theheat flux into the pool from the ground. If the pool is bunded, the ground beneath it will cooland the heat flux will diminish with time, leaving a still very hazardous pool though vaporizingmore slowly. If the pool is not bunded then it will be able to spread on to new warm ground andrapid boiling may continue. The rate of production of gas also increases with increasing surfacearea of the pool. An LNG cloud formed in this way is cold, concentrated and flammable, andrequires a dispersion calculation to estimate the hazard range.

As noted above, the result of vaporization will be a heavy cloud - the heaviness being caused bythe coldness. It will initially both slump and disperse, and a heavy gas model is required in orderto make predictions.

Sometimes it may become passive before it reaches the minimum concentration of interest(often half LFL), and in this regime classical methods of analyzing passive dispersion will apply

Thermodynamic considerations

Because the initial heaviness of the cloud is associated with its low temperature, it may bereasonable a priori to expect an appropriate dispersion model to be able to handlethermodynamic effects as well as heavy, passive and even (as ambient temperature methane islighter than air) buoyant dispersion. Very rapid boiling may also throw up drops of liquid intothe cloud , and so perhaps aerosol dispersion should be considered.

It is generally believed that any methane droplets will be large and fall back rapidly into thepool, and that any water condensation from the atmosphere (which typically renders cold cloudsopaque) will not affect the dispersion process significantly. With only some reservations

(expressed below) we share this general belief but are unsure whether it has ever been testedconvincingly.

There is also a limit to which a model built only to handle ambient temperature dispersion canbe adequately applied to cold clouds. Consider the possibility that the only heat brought into thecloud is by convected entrainment of air from the atmosphere, and that the molar specific heatsof the air and the contaminant are the same. In this case buoyancy is conserved , and the averagetemperature of the cloud is always determinable from its average concentration. One can seethis by considering heat exchange between a parcel of air and a parcel of cloud. As heat isextracted from the air, and given to the cloud, the air will contract by exactly as much as thecloud expands so that the total volume of air plus cloud is conserved, just as in ambienttemperature mixing. This was shown formally by Webber (1983).

Another consequence of buoyancy conservation is that the cloud will, as it dilutes and warmsapproach air density asymptotically from above, just as it approaches ambient temperature frombelow. It will not therefore become positively buoyant at any stage (though of course passivedispersion must be covered by any model).

One can ask whether the assumptions which lead to these conclusions are valid. As far as know, there is very little strong evidence for an LNG cloud becoming buoyant , either in practiceor in theory (using models with fairly realistic modeling of thermodynamics and heat transferfrom the surroundings), although there are some indications that this might have occurred in twoof the Burro series of field trials (Koopman et al 1982b) but whether this is due to buoyancy isnot certain (Koopman et al 1982a , Morgan et al 1984). The question of buoyancy has beenraised for HF clouds and, more debatably, for NH, where in both cases extra heat is generatedby aerosols forming non-ideal solutions with atmospheric water, but that extra heat is crucial.

Humidity

Atmospheric water vapor might affect an LNG cloud, as air is entrained into the cloud, thewater vapor will condense (and even freeze, though we know of no model which allows for asolid component). In doing so the droplets will tend to increase the density of the cloud, but theheat released by condensation will tend to warm it up and make it less dense. Later as the clouddilutes and warms up, the water will evaporate again, and the two opposing effects on thedensity will each be reversed. Experience suggests that the temperature effects dominate, andthe early effect of water vapor condensation is to warm the cloud up faster than would otherwisebe the case, and make it less dense. Experience also suggests t that, even though entrairunentrates are affected by density, this makes very little difference to the concentration of methane.Thus it was our first conclusion that water vapor in the atmosphere would make very little netdifference in the case of LNG.

However, prompted by questions raised in the early part of this project, we have done a simpletest, and uncovered what must be a caveat for those using models which include the effect ofwater vapor. As noted, the different effects on the density go in different directions from theoutset , and then reverse. A simple analytic argument is therefore difficult. We have thereforeused the Health and Safety Executive (HSE) code DRIFT and carried out two Thomey-Island-like runs using pure methane initially at its boiling point as the dispersing material , and with 0%and 90% humidity respectively. DRIFT considers five components in the cloud: gaseous

In a number of theoretical studies of the behaviour of aerosols in clouds (see e.g. Webber, Jones and Martin 1992Webber, Mercer and Jones I 994 , Kukkonen et al 1994 1995) including HF interactions with water, we observe verydramntic cbanges , induced by atmospheric humidity, in the liquid content of the cloud, and in the cloud temperaturebut very little effect on the actual concentration. This is entirely consistent with the observations in the Goldfishtrials (Blewitt et a11987 , Chan et al 1987)

contaminant, liquid contaminant , water vapor, water liquid, and "dry air" (i.e. the air excludingits water content). For methane the water and liquid methane drops are considered to coexistwithout forming a solution. The whole mixture is considered to be in thermodynamicequilibrium, with very carefully constructed equations of state. For condensation to take placethe cloud must have some liquid initially and so we "seeded" it with a small fraction of themethane initially in the liquid phase. The graphs of concentration against time were verysimilar - the water vapor apparently has essentially no effect. However in the case with watervapor, the cloud warms up faster and becomes less dense. And (as it happens at broadly thetime that LFL was reached) the Richardson number reached (and passed through) zero. Now itmight be tempting to regard the cloud as buoyant at this point and assert that it presents nodanger beyond that point. However we would advise against accepting any such assertion tooreadily, as it probably only means that the cloud has become passive (despite the prediction of aslightly negative Richardson number). Furthermore, as the water vaporizes, extracting its heatof vaporization, the cloud has at least the potential to become dense again. Given that all of thisseems to have such a small effect on the concentration itself, it seems reasonable to expect a lotof supporting evidence if this is to be considered a mitigating factor.

As a consequence we do not believe that even ambient temperature models of LNG dispersionstarting with a cloud of appropriate density, should be ruled out per se.

Geometric considerations: obstacles

Another point of interest is whether dispersion models should include obstacles, or whether it issufficient to ignore them. Some, though perhaps relatively few , integral models considerobstacles. 3D CFD models can usually cope with obstacles.

Far from the source where the cloud is high, obstacles may often make little difference to thedispersion. (An exception occurs at very low wind speed where the cloud remains heavy andlow.) Duijm and Webber (1994) and Jones et al (1992) show how a fence across the flow maybe incorporated into any integral model. The model, which gives a good overall fit to field dataindicated that the fence only makes a difference if the cloud is actually lower than the fencewhen it encounters it, otherwise the turbulent wake of the fence just mixes the already wellmixed cloud. Thus it would appear that individual obstacles only affect cloud dispersion whenthey are bigger than the cloud. Furthermore, a distributed array of obstacles , all much smallerthan the cloud, should be modeled as enhanced ground roughness.

It is possible that concentrations at some locations, close to an obstacle situated near the sourcemay be larger owing to its presence. But far down-stream of the obstacle (where public safety isaffected) the net effect is usually one of extra dilution in the turbulent wake ofthe obstacle.

One does not only have to consider obstacles in the path of a cloud downstream of the source.If a pool of LNG forms just downwind of a large obstacle, then turbulence in the wind field mayenhance the vaporization rate, usually not initially while it is controlled by the heat flux into thepool, but later when the surfaces in contact with the pool have cooled.

In summary: ignoring obstacles will usually result, other things being accurately modeled, in

conservative predictions and we regard this as an entirely acceptable approach when predictinghazard distances.

When modeling obstacles, however, their effects may be over-estimated and non-conservativeresults obtained. Thus any safety case which relies on the existence of obstacles to dilute thecloud, should also present results in the absence of obstacles, and also present a credible casethat the difference has not been overestimated. Simple checks are available: for example Duijmand Webber (1994) indicate that the distance to any given concentration is reduced effectively

by the distance which the cloud would have to travel, in the absence of the obstacle, to increasein height from its height at the obstacle, to the height of the obstacle. If a model produces aneffect greater than that, then further explanation should be sought.

Geometric considerations: terrain

Sloping and complex terrain can also usually be handled by 3D CFD models. Integral modelsof gas dispersion have also considered it, but few have used it in hazard analysis. Interestingly,theoretical analysis (e.g. Webber et a11992) indicates that the process may be better describedby slumping followed by down-slope flow, rather than both simultaneously. For flammablegases, it may not often be important as sites area usually fairly level out to the sort of distancesexpected to (say) half the LFL. However in low wind speeds it can be a factor, and the gas maybe channeled under gravity by obstacles and terrain in quite complicated ways, which might bedifficult to consider in a hazard analysis. Sloping terrain within an impoundment may be veryimportant for liquid flow , possibly by design as a "run-off'

Concentration

Models must be clear on what they mean by concentration. In the literature on heavy gasdispersion , concentrations have been averaged over 0.6 s, over 6 s and assorted other timeintervals, and in the literature on passive dispersion 10 minute averages are common. These arevery different things, although correlations have been produced which attempt to relate them.Models must be clear about what they are predicting. Webber (2002) has reviewed this subjectin some detail , in particular noting various reasons why Y, LFL may be a more valid safetycriterion than LFL (with an appropriate choice of concentration). Further attention is given toaveraging time in Section 2.4.

MODEL EVALUATION

General considerations

Mathematical modeling

Mathematical modeling of a physical process involves schematically the steps summarized asfollows:

1. construct a set of equationswhich one can arguerepresents the process

2. provide a method to solvethose equations (analyticallyor numerically)

3. test the solutions against allknown relevant observationsto check that they dorepresent those observations

4. make new predictions forwhat will be observed inexperiments which have notyet been done

This is a very general summary of what lies at the heart of the scientific method. It is a recipefor gaining an improved understanding of the physical processes in question.

Some things are usually considered so obvious (to academic theorists among others) that theydo not need to be said. These include

(i) Implicit in step I is the idea that the model should be consistent with what oneknows about related phenomena: the model must have a sound scientific basis.Without that the procedure is worthless.

(ii) Implicit in step 2 is the idea that the solution procedure is accurate, and does indeedprovide a solution of the equations

(iii) Implicit in all of it is the concept of a fitness for purpose. There is some purpose forwhich we are studying the process; the model must satisfy that purpose and thesolutions must be accurate enough for that purpose.

Application to consequence analysis for industrial safety

This process underlies consequence modeling in industrial hazard analysis, where anunderstanding of what might happen in an accident is required. Predictions are required for theconsequences of an accident which may never happen, and it may be so severe that anexperimental simulation at full scale is impossible to perform. Therefore the process would

reach as far as step 4 in the flow diagram, with predictions for the accident , but no means testagainst experimental data.

For this reason we require confidence in everything which has been done up to this point. Withthis uppermost in mind , the "Model Evaluation Group" (MEG) was set up by the Conunissionof the European Communities and tasked with providing a very general summary of how thescientific method should be applied to consequence assessments.

They emphasized (MEG, I 994a b) three important aspects of the procedure as:

The scientific basis: the model must be credible and fit for purpose, (see (i) above)

Verification: it must be shown that the solution procedure (usually a computer program)produces solutions of the model equations to satisfactory accuracy, (see (ii) above)

Validation: the model must be shown to agree with relevant experimental observationsto satisfactory accuracy, (see (iii) above)

The concept of "validation" requires some consideration: Comparison with an experiment cannever show that a model is "valid" . The best it can do is fail to show that the model is "invalid"A validated model is therefore one where tests have been performed which could have shown itto be invalid, but which failed to do so.

Therefore, it is often a useful exercise to compare the predictions of different models. If theyare found to disagree significantly then further examination of one or both should be made toidentify the reasons for the discrepancies. Such model comparisons can be done with envisagedaccident scenarios where experimental data may be hard to come by.

Validation is an open ended process (corresponding with the non-closure of the scientificmethod outlined above) but in practice it ends when one has sufficient confidence in the model.

The MEG was , however, aware that the model evaluation procedure would need furtherparticularization and refmement (in different directions) so that it could be applied to differentspecific areas of consequence analysis , and different authors have derived protocols for modelcomparison and validation in different areas. The objective in this report is to derive a protocolappropriate for models used in analyzing the safety of the handling of LNG.

SMEDIS

The ED-sponsored SMEDIS project (Carissimo et al 2001 , Daish et a12000) took as its startingpoint the MEG's reports (MEG, 1994a b) and produced a more detailed protocol for evaluatingheavy gas dispersion models.

In order to develop a specific protocol for LNG dispersion models, the SMEDIS protocol isused as a star1ing point with the following objectives:

to particularize to the physical phenomena expected in LNG dispersion, omittingirrelevant aspects which may be present in other heavy gases;

to give due consideration to the relevant source terms.

The work of Hanna et al

In the USA the prime example of heavy gas dispersion model evaluation was work undertakenby Hanna et al (1991 , 1993). They compared a number of models with a number of data setsand presented overall measures of the average fit of each model to the data, and of thevariability of the fit for continuous passive releases , continuous dense gas releases, andinstantaneous dense gas releases.

It is important to note that Hanna et al. were not involved in developing any of the models, butrather used them as independent assessors. Hanna et al did not review the scientific basis of anyof the chosen models or their verification, and so this should be regarded as a (major) validationexercise rather than a full model evaluation in the sense of the MEG, although they did givesome reasons for certain aspects ofthe results.

In order to compare the models and data on the same basis, Hanna et al created an archive withdata from all the experiments included in a common format. They also wrote pre-processingsoftware to read the data archive and generate the input conditions for each model , and post-processing routines to produce the deemed physical comparison parameters and statisticalperformance measures used to determine ' goodness of fit' . This was a major, four year long,exercise.

Hanna et al also observed that model documentation seldom contained anything concerning themodels ' limitations , but the developers of the models tended to claim a very wide applicabilityfor their models. The MEG make a point of mentioning the desirability of documenting thedesign limitations.

Hanna et al. also observed that allowing models, where possible, to predict the emission rateresulted in much greater discrepancies than would have been the case had they input the bestapproximation to the observed emission rates. This emphasizes the need to evaluate sourcemodels along with dispersion models.

LNG Dispersion Model Validation

Before concluding this section, it is important to note that any model validation exercise focusedon LNG dispersion should not limit itself to data sets involving only LNG. Any appropriatemodel will also cover isothermal gas dispersion and possibly other, more complex situations.Any validation should cover all of these: the validity of a model which fits LNG dispersion databut fails to fit simpler cases, would clearly be in some doubt.

LNG DISPERSION MODEL EVALUATION PROTOCOL

OVERVIEW

This Section describes the LNG Model Evaluation Protocol (MEP). It is based on the SMEDISprotocol for dense gas dispersion models which itself was in a form consistent with the ECModel Evaluation Group (MEG) generic protocol (MEG, 1994a b) and the Heavy GasDispersion Expert Group (HGDEG) protocol (HGDEG , 1996). However, the SMEDIS protocolcontained much content that is not relevant to LNG vapor dispersion and so a specific version ofthe protocol has been developed.

The objectives and guiding principles of the LNG MEP are as follows.

The purpose of the MEP is to provide a comprehensive evaluation methodology fordetermining the suitability of models to accurately simulate the dispersion of vaporsemanating from accidental spills of LNG on land.

The protocol is applicable to a wide range of dispersion models, primarily CFD andintegral models , but also empirical models and shallow layer models. Some of thesemodels will be designed specifically for modeling the dispersion of LNG vaporsothers, particularly the CFD models, will not.

The key steps in the application of the MEP are a scientific assessment, verificationand validation of the model. These three key stages of the MEP are described in detailbelow.

The information required to undertake the scientific assessment is obtained via aquestionnaire that is completed by the ' model developer' or a 'proponent for themodel' . This person doesn t necessarily have to be the original model developer, but itdoes need to be someone who has an intimate knowledge of the model.

The MEP is specific to one and only one version of a model. This will be recordedclearly in the model evaluation report. After the MEP has been applied to a particularmodel it is the responsibility of the user to ensure that the model is the same as thatwhich has undergone the evaluation using the MEP.

The scientific assessment should be carried out by an independent third party who hasthe necessary expertise. However, it is accepted that the validation exercise may becarried out or assisted by the model developer. The validation exercise is not expectedto be a 'blind test' , although any changes made to the model during the validationexercise should be documented.

Parts of the procedure involving the active use of the model must be documented tomake them auditable and the results reproducible. This would apply even if anindependent third party were performing the evaluation.

The MEP should not be biased to anyone model or type of model. Two examplesillustrate the point. The MEP consists of a number of qualitative acceptance criteria(see Section 2. 1) which include a number of physical factors that the model shouldtake into account. These have been designed such that they do not exclude modelswhich can be appropriately used where neglecting these physical parameters leads toconservative results or has negligible effects on the result. Secondly, the validationdatabase consists of two sets of data, one for flat terrain the other for cases with

obstacles / complex terrain, such that even if a model were to perform poorly for thelatter cases , it could still be accepted as suitable for the former.

The performance of the models is quantified using a range of data sets andperformance measures, including quantitative statistical comparison techniques.However, ranking of models according to their performance is not carried out.

10. This MEP is only applicable to dispersion models, although it is recognized thataccompanying models and in particular the associated source model playa veryimportant role in determining the hazard associated with a spill of LNG.

II. Ideally, the information produced by the evaluation should be available to the public.However, it is recognized that for proprietary models this may not be possiblealthough some form of openly-available results would be desirable, and it is hoped theModel Evaluation Report (MER), see below, would be treated in the same way asother documentation on a model.

The development of the MEP includes some uncertainty. It would be appropriate to review andrefllle the MEP after it is has been applied to a number of models, particularly the quantitativeevaluation criteria. Application of the MEP will reveal what a model does and highlight whetherit is obviously poorly designed but in the case of a generally acceptable model the MEP may fallshort of conveying/acquiring a full understanding of the system. Therefore the application of theMEP should be undertaken very much in the spirit of "validation" which never makes a modelvalid: if successful it merely fails to invalidate it.

How to use the MEP

The MEP is split into three main phases: Scientific Assessment, Verification and Validation andfurther details are provided in Sections 2. , 2.3 and 2.4 respectively. The three phases can becarried out in parallel with each other. The validation exercise requires the model to be runagainst a validation database (see Section 2.4.5) and will therefore be the most time consumingof these phases (considerably so for CFD models).

SCIENTIFIC ASSESSMENT

The scientific assessment is carried out by critically reviewing the physical, mathematical andnumerical basis of the model. The information on which this review is based is taken fromliterature made available for this purpose, which may include published material, and acompleted questionnaire that has been specifically designed to extract the necessaryinformation. When this information has been obtained the scientific assessment is carried outand the fmdings are recorded in the MER.

To carry out the scientific assessment the reviewer should have an in-depth understanding of thebehavior of dense gas clouds and the application of dispersion models. To carry out a review ofa CFD model then the reviewer should have additional understanding of this modelingapproach. Clearly the reviewer should also be independent of the model developer and shouldhave no invested interest in the outcome of the model evaluation. However, it is accepted that itmay be appropriate for the model developer to be involved in, or carry out some or all of thevalidation exercise.

Questionnaire

The purpose of this document is to request the information which is needed for the scientificassessment of an LNG dispersion model from the model developer or proponent. The model

developer / proponent should have an intimate knowledge of the model and this may thereforeexclude developers who merely package an existing model. The questionnaire is included as anappendix to this report (see Section 10. 1).

The completed questionnaire should be returned with a set of documentation covering allaspects of the model and , preferably, cross-referenced to the topics in the questionnaire. Theaccompanying documentation should include user manuals, published papers, reports etc.Confidential documents should be clearly indicated. Of particular interest are peer-reviewedapplications of the model in the technical literature, validation exercises and also any othervalidation exercises that may not have been made publicly available (note that confidentialinformation will not be included in the model evaluation report).

The questionnaire includes a set of guidelines following the questions to help the modeldeveloper/proponent to provide the required information. It is essential that these guidelines areused to complete the questionnaire.

The information supplied should refer to a single well-defined version of the model and whichshould be unambiguously identified.

The questionnaire is based on the form derived by the European Commission s SMEDIS projectbut has been particularized (with emphasis on guidelines to support interpretation of thequestions) to models dealing with the dispersion of LNG.

The questionnaire is aimed at all types of model such that they can be compared on an equalbasis, and therefore answers to some questions may be almost trivial in some cases.

The questionnaire is split into the following sections:

(I) General information(2) Information for scientific assessment(3) Information for user-oriented assessment(4) Information on verification(5) Information on validation

(6) Administrative detailsGuidance on completing the questionnaire

Model Evaluation Report (MER)

The MER is the key output following application of the MEP. It contains the full details of thescientific assessment and which in turn is based on information provided in the questionnaireand associated documentation. The model evaluation report is presented in the appendices ofthis report as applied to DEGADIS (Section 10.2), FEM3A (10. 3) and FLUENT (10.4).

The MER provides conclusions on the scientific basis of the model , limitations of the modeluser orientated aspects of the model as well as the evaluation against the MEP qualitativeassessment criteria and validation performed and evaluation against MEP quantitativeassessment criteria.

The MER also makes provision for comments from the model developer which, in practice, willprove to be essential in ensuring that all of the details ofthe model are captured accurately.

The MER is structured as follows:

O. Evaluation informationI. General model description2. Scientific basis of model3. User-orientated basis of model4. Verification performed5. Evaluation against MEP qualitative assessment criteria6. Validation performed and evaluation against MEP quantitative assessment criteria7. Conclusions

Al Actively-generated informationA2 Comments from model supplier / proponent

The main headings above (1-6) are subdivided into the assessment categories which arestructured in a consistent way as follows:

general remarks on the category;

topics of interest describes the subjects relevant to the category. This means that thereshould be information available covering each subject, and the evaluator should checkthat this is the case. If none is available this can be noted together with the reason for itsabsence and recorded later in the MER (this may prove useful in the revision of theprotocol);

assessment and comment gives aspects of the topics of interest that should beconsidered by the evaluator, i.e. this section suggests ways that the information on thetopics of interest can be assessed;

contribution to the evaluation record describes the part to be added to the evaluationrecord, i.e. the MER, for this category. It gives the relevant headings ITom the MERfollowed by the general form of the content under each heading. The contributiontypically involves a combination of some reporting of the information describing thetopics of interest , e.g. summarizing that aspect of the model, followed by assessmentof/comment on the information.

The evaluator must provide a description or an assessment of the model under each category inturn. Note that each category should be addressed to a level of detail that:

concentrates on the most relevant features, and does not reproduce the information to anexcessive degree;

does not demand an umeasonable amount of analysis, e.g. assessing the limits ofapplicability of a model.

VERIFICATION

Introduction

Verification of a model is the process of comparing the implementation of a model with itsmathematical basis. Most commonly this refers to checking that a computer implementation of amodel (computer software) accurately represents its mathematical description.

Verification is essential, and should be demonstrable. A good start is provided if a numericalsolver with a good track record (and published verification) is adopted. But even so, it should bedemonstrated that the solutions presented are indeed solutions of the programmed equations.Models sometimes admit analytic solutions in special cases, and comparison with these is

always useful. In other cases asymptotes can be found analytically and a comparison canprovide a useful test. In yet other cases things are known about the solution which emerge non-trivially ITom the numerical procedure, such as conservation of buoyancy, and this can bechecked. In the specific case of CFD models, the method of manufactured solutions (Roache1998) can also be used as a verification tool.

Verification does require a certain amount of mathematical skill , which is quite different frommany of the engineering skills needed to model hazards.

MEP verification

The verification of a model within the MEP follows the same approach as SMEDIS. This meansthat verification is treated passively as part of the scientific assessment instead of an exercise inits own right. Evidence for verification is therefore sought ITom the model developer and this isassessed and recorded in the MER.

Note that verification of the model is not a qualitative assessment criteria, although it is reported

in the MER. The reason for this is that the absence of information or evidence of verificationwould not be a sufficient reason to reject a model. Also the judgment that needs to be made onwhether a model has been verified is subjective as well as being reliant on claims made by themodel developer/proponent which are impractical to substantiate. For example, two differentreviewers could easily reach different conclusions depending on how rigorous they choose to bein demanding evidence of verification.

VALIDATION

2.4. Introduction

Validation is achieved by a comparison of model predictions against measurements, over themodel's required range of application. A validation database containing these measurementsprovides the means to assess the performance of a model. The end objective of validation is toestablish whether a model replicates reality to an acceptable degree.

Although we use the term ' validation , what we actually mean is ' evaluation . Over a prescribedrange of applications sufficient confidence in a model may be gained by comparison withmeasurements such that the model has been evaluated and found to perform acceptably wellacross this range of applications.

In this Section we set out the basis of the validation procedure based on the approach adoptedand developed during the SMEDIS project (Daish et al 2000, Carissimo et al 2001) and furtherexplained in Duijm & Carissimo (2002).

The validation procedure involves a number of differing aspects, addressed in the followingsteps:

a. Specification of the objective: this being the quantification and assessment of modelperformance for dispersion of LNG vapor from spills on land.

b. Identification of the key physics and variables involved in the dispersion of LNG vaporfrom spills on land.

Identification of target scenarios that cover the key physical processes involved in thedispersion of LNG vapor from spills on land. Ideally these scenarios are sufficiently

wide-ranging that the performance of a model can be tested over the full range of keyphysical variables (source terms , atmospheric conditions , terrain, etc.

d. Identificationof suitable validation data sets.

e. Selection of specific cases from these datasets so as to cover the range of targetscenarios.

Definition of physical comparison parameters (PCP) that are measured or derived frommeasurements and which form the basis of comparisons with model predictions.

g.

Selection of statistical performance measures (SPM) that allow quantitativecomparison of predictions against measurements.

h. Review and defmition of quantitative assessment criteria that defme the acceptablenumerical range of the SPM which result from applying this validation procedure.

*Defme the specific content and design the format of the validation database.

*Populate the validation database.

k. *Provide guidance on model application for the validation cases.

* Apply the validation procedure and refine as necessary in the light of experience.

Steps b) and c) are addressed in Section 2.4.

Steps d) and e) are addressed in Section 2.4.

Steps 1) and g) are covered in Section 2.4.4.

Step h) is covered in Section 2.

*Completion of steps i) to I) is outside the scope of the present project. However, Section 2.4.does provide an illustration of the content of the database for a single, wind tunnel, test case(drawn from the SMEDIS project). This Section also outlines the steps involved in theprocessing offield trials and wind tunnel data, including scaling rules. In addition, Section 2.4.highlights a number of issues which must be addressed when applying models to the validationcases.

Key physics and target scenarios

The key physical processes involved in the dispersion of LNG vapor over land have beendiscussed in Section 1.3.

More comprehensive descriptions of the phenomenology can be found in a series of reports andpapers from the Lawrence Livermore National Laboratory (Koopman et a11982a, Morgan et al1984, Koopman & Ermak 2007). These stem primarily from large-scale, unobstructed , field trialspills of LNG at China Lake, California, in which dispersion occurred over land.Phenomenology in the presence of obstructions comprising a vapor fence and barrier isprovided by Brown et al (1990) and also addressed briefly by Koopman & Ermak (2007).

Additional phenomenology, gained from LNG spills at Maplin Sands in the UK and in whichdispersion occurred over the sea , is provided by Puttock et al (1982) and Colenbrander &Puttock (1983).

Other recent reviews (Luketa-Hanlin 2006 , Cleaver et aI, 2007) also provide further insight intothe key physical processes involved in the dispersion of LNG vapor.

In summary form , the key physical processes involved in the dispersion of LNG vapor over landare as follows:

Formation of a dense cloud due to the low boiling point of LNG.

Gravity-driven spreading.

Advection by the ambient wind field.

Reduction in turbulent mixing due to the resulting stable density stratification.

Dispersion influenced by atmospheric stability.

Other physical processes can also be important, with their significance dependent on theparticular circumstances of a release. These could include:

Enhanced mixing and dilution due to obstacle-generated turbulence.

Influence of terrain on gravity spreading.

Vapor hold-up due to fences or dikes.

Heat addition and removal due to condensation and evaporation of water vapor.

Heat transfer from the ground.

Ideally, scenarios which are used to derIDe the specific test cases in the validation databaseshould encompass the key physical processes. This is most comprehensively achieved byconsideration offield trial spills of LNG.

It is also preferable to test a model over as wide a range of conditions as possible, i.e. for abroad range of scenarios. As outlined in Section 2.4. , these scenarios should ideally be

sufficiently wide-ranging that the performance of a model can be tested over the full range ofkey physical variables (variation in source terms, atmospheric conditions , terrain, etc.

The key physical variables which affect dispersion of LNG vapor over land are as follows:

Source configuration: release rate, duration and pool geometry.

Atmospheric conditions: stability, wind speed , humidity.

Terrain: surface roughness, flat/sloping/complex terrain.

Obstacles: tank, dike, fence, etc.

In principle, a matrix of target scenarios based on the above physical variables, and whichencompass the key physical processes , could be constructed. Test cases would then be derIDedwhich meet entries in this matrix. This was the approach adopted in SMEDIS and which led to amatrix of 45 scenarios for which test cases were found. As will become more apparent inSection 2.4.3 , such an approach is not practicable for dispersion of LNG vapor over landbecause:

data does not exist to allow all target scenarios (i.e. all combinations of factors whichmay be relevant for LNG dispersion in different circumstances) to be met;

even if data were available, the matrix could become impracticably large. (Note that theSMEDIS project extended from 1996 to 1999 and in that time only about one third ofthe entire set of test cases (approximately 300 in total) were computed and analyzed.

This being the case, a modified approach is required such that the main physical processes aretested for a set of scenarios which cover a more focused range of key physical variables. Inpractice the scenarios and test cases are governed by the availability of appropriate data.

Dataset selection

2.4. Data requirements

To be useful for model validation, data must fulfill several requirements. These include:

The quality of the data must be fit for purpose, i.e. model evaluation. Nielsen and Ott(1996) discuss the meaning of data quality in this context and describe methods ofscreening and checking the quality of data for model evaluation.

The test conditions must be known, including source configuration, atmosphericconditions, surface roughness etc. Duijm and Carissimo (2002) stress the importance ofreliable information on the source term and release rate.

The time-averaging applied to the data must be specified. For flammables, as here, datashould be available for short time-averages.

If wind tunnel data are to be used then scaling effects are crucial and must have beenconsidered in the design and reporting of experiments. Scale factors should be withinacceptable ranges. Meroney & Neff (I 982) discusses appropriate scaling rules and scalefactors for wind tunnel simulations of LNG releases.

The data must be available and in suitable formats.

Although the above requirements are quite stringent, they can be met for a range of test caseswhich do address the main physical processes involved in dispersion of LNG vapor over land.

Ideally, multiple realizations of an experiment will also have been undertaken so that ensemble-mean values are available. Note that the vast majority of models produce an output which isessentially an ensemble-mean. Davies (1987) showed, by analysis of wind tunnel trials, that

multiple repeats of an instantaneous release of a dense gas under nominally identical conditionscan produce concentrations at downstream locations which vary by roughly a factor of two.Unfortunately, ensemble-mean data are rarely available, especially for field trials.

2.4. Current status

The most significant and useful datasets resulting from field trial spills of LNG withoutobstructions are ITom Maplin Sands performed by Shell Research in 1980 (Puttock et al1982Colenbrander & Puttock 1983), and the Burro and Coyote trials performed by LawrenceLivermore National Laboratory in 1980 and 1981 (Morgan et aI, 1984). Ermak et al (1988)provide a useful comprehensive overview of these trials.

The only significant field trial spills of LNG in the presence of obstructions are the Falcon trialsundertaken in 1987 , by Lawrence Livermore National Laboratory (Brown et aI , 1990).

These four trials have recently been reviewed by Luketa-Hanlin (2006) and Koopman & Ermak(2007) and so only a brief summary is provided here.

The Maplin Sands trials undertaken by Shell Research in 1980 comprised spills of either LNGor LPG onto water with dispersion occurring over tidal sands (most experiments wereperformed at high tide). Some releases were ignited. Both continuous and instantaneous releasesof LNG were undertaken. Of the thir1een continuous LNG releases, eight were deemed toprovide useful data. These continuous releases were directed vertically downwards onto the seasurface from a range of heights, and in some trials the release impinged on a cone and platedevice designed to restrict initial spreading of the liquid to the horizontal direction. The releaserates ranged from I to 4.5 m /min, in wind speeds of between 2 and 10 mis, all in neutralatmospheric conditions. A large number of sensors were arranged in downstream arcs floatingon 71 pontoons. Data was obtained at a minimum averaging time of 3 s. The trials and theiranalysis are described by Puttock et al (1982) and Colenbrander & Puttock (1983) and a usefulsummary of the test conditions is given by Ermak et al (1988).

The Burro trials were undertaken at China Lake, California, 1980. Dispersion occurred over

land , although the spill was onto a 58 m diameter water pool. The releases were initiallydirected vertically downwards but impinged on a splash plate to limit LNG penetration into thewater. A series of nine tests were undertaken. Spill rates ranged from 11.3 to 18.4 m /min, inwind speeds from 1.8 to 9. 1 mls. With one notable exception all releases were undertaken ineither neutral or slightly unstable atmospheric conditions. The exception was the Burro 8 trialwhich took place in stable conditions (Pasquill-Gifford stability class E) at a relatively low windspeed of 1.8 mls. The Burro 8 test is the only well-instrumented unobstructed field trial releaseof LNG in stable atmospheric conditions. Gas concentration sensors were arranged in four arcsat 57, 140, 400 and 800 m downstream from the release point. Data were obtained at aminimum averaging time of I s. Note that the terrain was not flat, in general tending to slopeupwards downwind from the release, but in a non-uniform manner. The trials are presenteddescribed and analyzed in Koopman et al (1982a , 1982b), Morgan et al (1984) and are alsosummarized in Ermak et al (1988).

The Coyote trials were a follow-up to the Burro trials and were primarily designed to investigateRapid Phase Transitions and the consequences of ignition. Nevertheless, significant dispersiondata were also obtained. Releases took place using the same release configuration as the Burrotrials. Spill rates ranged from 6 to 19 m /min, although for the trials which are useful fordispersion model evaluation the spill rate is in the range 13.5 to 17 m /min. Wind speeds forthese useful dispersion trials ranged from 4.6 to 9.7 mls. Atmospheric stability was eitherneutral or slightly unstable. The gas concentration sensors were clustered in four arcs between140 and 400 m downstream from the point of release. The trials are presented , described andanalyzed in Goldwire et al (1983), Morgan et al (1984) and are also summarized by Ermak et al(1988).

The Falcon trials were undertaken to examine the effectiveness of fences in mitigating theeffects of accidental releases of LNG. The trials were carried out at Frenchman Flat, Nevada in1987. Five trials were undertaken in which LNG was released onto a 40 x 60 m water pond via4 spill pipes. A splash plate was fitted underneath each pipe so that LNG was directed across thesurface of the pond. A fence of height 8.7 m surrounded the water pond. Upwind of the pondbut inside the fence, a 'billboard' structure was located to generate turbulence in a similarmanner to that which could be expected from a storage tank. This billboard was 17.7 m long by13. 3 m high. Spill rates from 8.7 to 30.3 m /min were obtained in these trials. The wind speed

ranged from 1.7 to 5. 2 mls. Significantly, atmospheric stability was either neutral or stableduring these trials. In particular, the Falcon I trial was undertaken in very stable conditions;Pasquill-Gifford stability class G. Gas concentration sensors were clustered along three lines at

, 150 and 250 m from the downwind edge of the fence. The data report (Brown et aI , 1990)provides a comprehensive description of the tests and presents the data in graphical form.

Table 2. 1 summarizes the main features of these four LNG trials.

Trial

Table 2. Summary of field trial spills of LNG

Spill type Release rates Willd speeds/mill) (m/s)5 2-11-18 1.8-13. 17 4.9 - 30 1.7 - 5.

Water poolWater poolWater pool

Water pool

Atmosphericstability

Maplin Sands, 1980Burro, 1980Coyote , 1981Falcon, 1987

Several observations can usefully be made at this point:

Almost all field trial spills of LNG are under neutral conditions and mostly at moderateto high wind speed.

With the exception of the Falcon trials, most field trial spills of LNG are inunobstructed conditions.

The source configuration for field trial spills of LNG is limited to pools, rather than linesources characteristic of spills in a trench.

In addition, although the LNG release rate is essentially steady for a period of typically a fewminutes in these trials - leading to quasi-continuous releases - almost no information exists onthe time-varying dimension of the resulting LNG pool (although where practicable massbalance calculations indicate that the release rate is approximately matched by the overallvaporization rate). This introduces uncertainty in the specification of the vapor source term. Asan illustration of how the pool dimension may vary, HSL has applied a sophisticated liquid spillmodel, GASP (Webber, 1990), to the Burro 8 trial. Figures 2. 1 and 2.2 indicate that the poolradius never reaches a steady-state and is likely to grow to a maximum radius of about 16 m.

16-

14-

12-

IlO

"jj ,

a: -

~ ,

Pool Radius

l ' , , , I ' , , , I ' , , , I ' , , , I ' , , , I ' ,

, , ,

50 100 150 200 250 300

TImets)

Figure 2.1 GASP prediction of LNG pool radius for Burro 8 trial

,-,,'004-

2,.004-

1..004-

8000-

6000-

::0: 4000-

2000-

Mass yaporised

0 -: ' , , , ~O

' , , " ' , , " ' , , ' ' , , ' ' , , '

TIme (s)

Figure 2 GASP prediction of mass of LNG vaporized for Burro 8 trial

Ermak et al (1982) model the Burro trials assuming that the LNG covers the entire surface ofthe 58 m diameter water pond , but the LNG spill model results in Figure 2. 1 indicate that this isunlikely to be the case, Ermak et al recognize this uncertainty in the source term.

This uncertainty in the pool radius , which is common to all four sets of field trials, can howeverbe reduced if an appropriate liquid spill model is used to predict the characteristics of the LNGspill. However, since there is some uncertainty in the LNG vapor source term for these trialsand also because the range of conditions in which experiments have been conducted is mostlylimited to near-neutral conditions , it is also very useful to consider other field trial releases ofdense gas.

Pre-eminent amongst these other field trials are those at Thomey Island, carried out from 1982to 1984 in the UK, in which instantaneous or continuous releases of Freon/nitrogen mixtureswere undertaken over flat terrain. For the continuous cases a release rate of -

/s of gas wasobtained at an initial density ratio of about 2.0. The wind speed ranged from 1.5 to 3. m/s andatmospheric stability from neutral to stable. Gas concentration sensors were located in arectangular grid with distances up to about 800 m from the release point. Data was originallytaken at high frequency but is now only available for a minimum averaging time of 30 s. Thetrials are presented and analyzed in two special editions of the Journal of Hazardous Materials(Volumes II and 16 , 1985 and 1987 , respectively). The continuous releases are presented andanalyzed in McQuaid (1987), Mercer & Nussey (1987) and Mercer & Davies (1987). Tabulateddata for the continuous releases is also presented in Ermak et al (1988).

Wind tunnel data must also be considered for inclusion in the validation database. Indeed anyexclusion of wind tunnel tests would be unduly restrictive to the range of scenarios considered.Wind tunnel tests, if carried out appropriately (Meroney & Neff, 1982), are also widelyrecognized and accepted as a valuable addition to field trials data for model evaluation. Theyallow for more control and repeatability of tests. However, the effects of heat transfer andatmospheric stability are difficult to replicate in a wind tunnel, so most wind tunnel data are fordense, isothermal, releases in neutral stability.

Wind tunnel modeling has recently been completed by the Chemical Hazards Research CenterUniversity of Arkansas, specifically to provide data for the evaluation of LNG dispersionmodels. Three sets of experiments have been undertaken, all with release of CO, as the densegas simulant. Case A is a release without obstacles; Case B is in the presence of a storage tankand 'high' dike; Case B is in the presence of the dike only. A description of the experiments , aswell as the tabulated data, is given in Havens & Spicer (2006). The scaling relations and scalefactor employed (150:1) are discussed in Havens & Spicer (2005 , 2007) together with theimplications of the results. This work repeats earlier work undertaken at the Chemical HazardsResearch Center in the mid 1990' s. However, the earlier work used a smooth floor, whilst thislater work used roughness elements on the floor of the tunnel to create turbulence propertiessimilar to those which might be encountered in full-scale releases. The scaled spill is equivalentto a full-scale LNG release rate of 36 /min.

Extensive wind tunnel modeling of LNG releases has also been carried out at Colorado StateUniversity from about the mid 1970' s to mid 1980' , including examination of the effects oftanks and dikes. This body of work is reported in several publications , for example Meroney &Neff (1979 , 1980 and 1982), Kothari & Meroney (1984). It established the validity and range ofapplicability of wind tunnel experiments for simulating releases of LNG, provides analysis andguidelines on scaling rules and scale factors and reports the phenomenology of LNG releasesincluding the effects of tanks in enhancing mixing and dilution of LNG vapor.

The SMEDIS project also made extensive use of two wind tunnel datasets generated as part of aEuropean Commission-funded project and undertaken by the University of Hamburg, Germany,and TNO, Netherlands, as well as other European organizations. These datasets are commonlyreferred to as BA-Hamburg and BA-TNO. In each case sulfur hexafluoride (SF.) was used asthe dense gas simulant.

A very wide range of configurations was examined in the BA-Hamburg trials including asemicircular fence placed upwind or downwind from a release, a fence completely surroundinga release, crosswind canyons , and sloping terrain. Most configurations were modeled with bothinstantaneous and continuous releases. Multiple repeats of many cases were also undertaken.

The BA- TNO trials consist primarily of continuous releases over flat terrain with or without thepresence of a fence of variable height. The fence is located downwind from the release and iseither perpendicular, or skewed at 150, to the wind direction.

The BA-Hamburg and BA- TNO trials provide data on dense gas dispersion in the presence of awider, more generic, range of obstacle configurations than that of the recent work at theChemical Hazards Research Center, including sloping terrain.

A key requirement is whether field trials and wind tunnel data are available and in a suitableform, ideally in electronic records. HSL have explored this issue in some depth for the fieldtrials and wind tunnel datasets discussed above (with the exception of the Colorado StateUniversity work).

Detailed electronic records are available for the Burro , Coyote, BA-Hamburg and BA-TNOtrials via the REDIPHEM database (Nielsen & Ott, 1996). This database was constructed duringthe REDIPHEM project (1992 - 1995), funded by the European Commission. It comprisesmeasurements from a significant number of field trials and wind tunnel experiments. Thedatabase details the experimental configurations including release and atmospheric conditionssensor positions, and time-series of measured parameters. The time series are typically at Isintervals for the field trials. These time-series can be visualized, processed and exported usingthe REDIPHEM data browser. It is an extremely valuable resource for the evaluation of densegas dispersion models. HSL has been in direct contact with the original developers andcustodians of the REDIPHEM database, at the Riso National Laboratory, Denmark, to obtainthe database - which is freely available.

Detailed data reports for the Burro and Coyote trials are also available (Koopman et a11982bGoldwire et al 1983) and have been obtained.

Electronic records for the key Maplin Sands and Thomey Island field trials are available via theModelers Data Archive (MDA) created by Hanna and co-workers during their extensive densegas model validation exercise (Hanna et a1199l , 1993). The MDA also contains data for theBurro and Coyote trials, as well as other dense gas field trials. The MDA consists of a summaryof the experimental configurations, including release and atmospheric conditions, together withprocessed concentration measurements at arc-wise locations. The MDA provides arc-wisemaximum concentrations and cloud widths. HSL has been in direct contact with the originaldeveloper and current custodians of the MDA , George Mason University, to obtain the databaseand supporting documentation , which is available upon request (Hanna et aI, 1991).

Tabulated data for the key Maplin Sands and Thomey Island trials are also available in Ermak etal (1988).

It is very unlikely that a useful electronic form of the Falcon data can be obtained (Williams2007). However, an extensive data report for these trials is available (Brown et aI , 1990), fromwhich long time-averages of gas concentration can be obtained. This will require digitization ofsensor output currently in graphical form.

The recent wind tunnel work undertaken at the Chemical Hazards Research Center, Universityof Arkansas , is available in tabulated form in Havens & Spicer (2006) and is almost certainlylikely to be available in electronic form upon request.

We have not explored whether electronic records for the Colorado State University wind tunneltrials are still available in a useful form. However, we see the main benefit of this work as beingto prove the validity of wind tunnel modeling of LNG (and other dense gas) releases and toprovide guidance on scaling rules and scale factors.

It should also be noted that the quality of the data in the REDIPHEM and MDA database can beregarded as being the best which is available. In both cases it has undergone in-depth scrutinyby the original developers before acceptance.

2.4. Specific datasets and test cases

Specific test cases for inclusion in the validation database have been selected from the MaplinSands, Burro, Coyote, Falcon and Thomey Island field trials, and the Chemical HazardsResearch Center (CHRC), BA-Hamburg and BA-TNO wind tunnel experiments, and arepresented in Table 2.

Table 2 Specific test cases for the validation database

Trial Field (F) Trial/Case number Atmospheric Data sourceor Wind and/or stabilitytunnel description(WTJ

Maplin 27 dispersion over sea MDA.Sands, 1980 34 dispersion over sea Also Ennak et al

35 dispersion over sea (1988).

Burro, 1980 REDIPHEM. AlsoMDA, Burro datareport, and Ennak et al(1988).

Coyote, 1981 REDIPHEM. AlsoMDA, Coyote datareport, and Ennak et al(1988).

Falcon, 1987 Data report (Brown et, 1990)

Thomey 45 - continuous release MDA.Island 1982- 47 - continuous release Also Ennak et al

(1988).

CHRC, 2006 A - without obstacles Havens & SpicerB - with storage tank & dike (2005 2006 2007) &C - with dike CHRC

BA-Hamburg Unobstructed DAOI20IDAT223 REDIPHEM.Upwind fence 0390511039072 Also seeDownwind fence DA0501IDA0532 Schatzmann et alCircular fence 039094/...095/... 097 (1991), Nielsen & Ot!Slone DAT647/...631/...632/... 637 11996)

BA-TNO TUVO I - unobstructed REDIPHEM.TUV02 - downwind fence Also see Nielsen & Ot!FLS - 3-D mapping (1996)

The selected Maplin Sands test cases are three of the four releases in the MDA of Hanna et al(1991 , 1993). Case 29 is omitted since Ermak et al (1988) state that for this case sub-surfacevaporization was considerable leading to gas jetting as high as 10m in the source area such thatspecification of a vapor source term could prove problematic.

The four selected Burro test cases are those which have been most extensively analyzed andcover the widest wide range of meteorological and spill conditions (Koopman et al 1982aMorgan et al 1984, Koopman & Ermak 2007). They include the Burro 8 case, undertaken instable atmospheric conditions.

The three selected Coyote test cases are again those cases which have been most extensivelyanalyzed (Morgan et a11984) and are regarded as benchmarks for dispersion model validation(Koopman & Ermak, 2007).

The three Falcon test cases are again those which are regarded as benchmarks (Koopman &Ermak, 2007) from the total of five tests carried out in these trials. They include the Falcon Icase, undertaken in very stable atmospheric conditions.

The two Thomey Island test cases are taken from the three continuous release experiments. Oneof these experiments is not included (case 46), since the plume missed many of the gas sensors.

All three ofthe CHRC test cases are selected.

The selected test cases from the very extensive BA-Hamburg trials are those which cover themost pertinent range of obstacle and terrain configurations. For each of the selectedconfigurations there are a number of test cases covering parameter variations, such as a slopeangle. It is proposed that one test case is selected for each configuration, to give a total of fivetest cases from the BA-Hamburg trials. The selection of the specific test case for eachconfiguration requires further analysis of the test conditions and data. This is beyond the scopeof the present project.

The TUVOI and TUVO2 test cases from the BA-TNO trials are similar in outline to some ofthose from the BA-Hamburg trials , but for differing release rates and wind speeds. The FLScase is a very comprehensive 3-D mapping of the concentration field.

Table 2.2 shows that we propose that the validation database consists of a total of 26 test cases.Many of these test cases have of course been used in previous model validation exercises.

All of the selected wind tunnel trials are for continuous releases. The field trial test cases arealso continuous in the case of the Thomey Island data.

For the remaining four field trial spills of LNG, releases were typically carried out over a periodof a few minutes at most. In some cases the release will have ceased before the cloud wouldhave reached sensors at the furthest downstream location. Strictly, these are not continuousreleases (Hanna, Drivas and Chang, 1996), but from the point of view of LNG dispersion (atleast for spills on a water pool) they are probably closer in character to a continuous release thanan instantaneous release.

It is proposed that , where possible, modeling ofthe wind tunnel test cases be carried out at windtunnel scale to avoid any remaining uncertainties introduced as a result of scaling effects. Forcircumstances in which this is not practicable (e.g. due to possible restrictions on the range ofinput data accepted by a model), the wind tunnel data will be scaled to a representative fieldscale using well-established scaling rules (Meroney & Neff, 1982). The validation database willthen contain both data at both model scale and a representative field scale. Section 2.4.contains more information on the scaling rules which should be applied.

2.4. Physical comparison parameters and statistical performancemeasures

2.4.4. Physical comparison parameters

The physical comparison parameters are the physical quantities against which the performanceof a model is evaluated. They can be directly measured or derived from measurements.

Physical comparison parameters can be separated into those which are based on point-wise andarc-wise data. The former involves comparison between model predictions and measurementspaired at specific points. The latter involves comparison between model predictions andmeasurements at specific distances downstream from a release, typically along circular arcs.The advantage of arc-wise comparisons is that uncertainties in wind direction, or thoseintroduced as a result of lateral meandering of a plume, are circumvented. Arc-wisecomparisons are most appropriate for situations in which plume direction is dominated by thewind direction. However, for other situations, for example in which the model performancedepends on the correct prediction of the path of the plume as a consequence of the effects ofobstacles or terrain, then point-wise comparisons should also be made (Duijm & Carissimo2002). SMEDIS included both point-wise and arc-wise based physical comparison parameters(Carissimo et aI, 2001). Hanna et al (1993) based their study on arc-wise comparisons. Furtherdiscussion of the advantages and disadvantages of point-wise and arc-wise comparisons can befound in Duijm et al (1996).

The most commonly-used physical comparison parameter for arc-wise data for the continuousand quasi-continuous releases in Table 2.2 is the maximum concentration across an arc at aspecific distance downwind from a release. This parameter allows evaluation of theperformance of a model in predicting the downwind variation of concentration with distance. Ithas the practical advantage of having been used as a physical comparison parameter in otherdense gas dispersion model evaluation exercises (Hanna et a11993 , Carissimo et a1200l , Chang& Hanna 2004, Hanna et al 2004) which means that there is information available in theliterature on the quantitative values of statistical performance measures for this parameter uponwhich recommended values of quantitative assessment criteria can be based (see Section 2.5.3).Hanna et al (1993), Duijm et al (1996) and Duijm & Carissimo (2002) point out that comparisonof maximum arc-wise concentrations should be combined with comparison of the plume widthat an arc to provide a more comprehensive evaluation of the performance of a model forpredicting concentration in both the downwind and lateral directions. The cloud width istypically calculated using moments of the concentration distribution across the arc (Hanna et al1991 , Carissimo et al 2001). Unfortunately, there is significantly less information available inthe literature on quantitative values of statistical performance measures for cloud width. Itshould also be noted that cloud width appears to be a less discriminating test of a model thanmaximum concentration at downwind distances (Hanna et aI , 1993).

Point-wise time-average concentration at specific locations was also included in SMEDIS as aphysical comparison parameter for continuous releases. This allows credit to be given to modelswhich provide spatial information on the concentration field (e.g. in situations where the cloudis affected by the presence of obstacles and/or terrain) and also provides additional informationon the spatial performance of a model for trials in which an arc contains insufficient sensors toallow determination of cloud width. However, there is unfortunately even less informationavailable in the literature on quantitative values of statistical performance measures for thisparameter than for cloud width. Point-wise comparisons provide a more stringent test of modelperformance then arc-wise comparisons (Carissimo et aI , 2001).

In summary, the recommended physical comparison parameters are as follows: