Beruflich Dokumente
Kultur Dokumente
D a v i d G. Platt
Department of Civil Engineering, Queen's Building, University of Bristol,
Bristol, UK, BS8 1TR
D a v i d G. E l m s & A n d r e w H. B u c h a n a n
Department of Civil Engineering, University of Canterbury,
Canterbury, New Zealand
(Received 20 August 1991; revised version received 17 January 1994;
accepted 19 February 1994)
ABSTRACT
NOTATION
INTRODUCTION
OBJECTIVES
U N D E R L Y I N G C O N C E P T OF T H E M O D E L
O U T L I N E OF T H E F I R E S P R E A D ANALYSIS
Input
barrier
Input
ventilation
Input Input Input ] I Input
thermal compartment spandrel / ~ probability of
resistance data properties & geometry heights / [ doorways
data fuel loads
1 1
1 1 compete compute slze
growth me barrier compartment of name
I modification fire severity projection
factor from window
1 ,j I
l ] compute probability compute probability compute probability
of fire spread via of fire spread via of fire spread via
barrier failure windows open doorway
It form Adjacency
Fire Spread Matrix
I_ I
=- calculate expected
fire spread times
and form
Adjacency Fire
Spread Time Matrix
/l
parameters data
Output
conditional
cost index
form Fire Vectors
It Output
basic cost indices
Global Fire Spread Matrix the result is a set of annual probabilities, one
for each compartment, of fire burning in the compartment either due to
ignition there or due to fire spreading, in a given time t, from ignition in
another compartment elsewhere in the building. The complete set of
these probabilities is known as the Fire Vector.
The following sections of this paper describe in greater detail a
number of the different aspects of the model.
A probabilistic model of fire spread with time effects 375
Flashov~
i
Time
Fully Developed Decay Period L
r -1- q- q
Fig. 4. T i m e - t e m p e r a t u r e curve for c o m p a r t m e n t fire.
Growth period
The growth period is described as the time from the point where fire is
initiated to the stage where flashover occurs. During the growth period
the average temperatures are low and the fire is localized. The growth
period of a fire consists of two distinct stages; the inception period and
the rapid growth period.
The inception period may last a few seconds or several hours, being
primarily a function of the fuel and ignition process. As a result of
uncertainties the inception time is taken as zero, thus assuming the
growth period starts at the point of rapid growth. This appears justified:
in the initial fire c o m p a r t m e n t the analysis starts from the time when
flames appear and the fire is detectable. For the rest of the building, any
new sources of ignition are likely to be accompanied by the 'thermal
m o m e n t u m ' of the already existing fire. Note that with the spread of
fire through a barrier the real concern is the penetration of flames
rather than insulation failure. This point is dealt with in more detail
later in the paper.
The rapid growth stage of the fire is characterized by a progressive
increase in the fire area and a growth in the heat output.
Ramachandran 4'5 developed an exponential model of fire growth that
376 David G. Platt et al.
relates the growth time to the fire area. In moderately sized compart-
ments it may be reasonable to assume that flashover occurs when the
fire area equals the total compartment floor area. However, as yet
statistical data relating to buildings other than warehouses have still to
be analysed.
A more commonly adopted approach is to define the point at which
flashover occurs in terms of a ceiling temperature and burning rate.
Using the criterion that flashover occurs at a certain ceiling temperature
it is possible to determine the rate of heat output necessary to create
this condition. Having obtained a value for the necessary level of heat
output the time taken to achieve this can be obtained by assuming a
parabolic fire growth rate as suggested by Haskestad. 6
The analysis assumes that flashover takes place when the ceiling
temperature reaches 600°C7 To calculate the rate of heat release
necessary to achieve this the model uses the following relationship, ~
which expresses the change in compartment temperature as a function
of the rate of heat release Q.
AT ~f O H~.A, ] (1)
--~o = ] Lg,/2( Cpp,,~T,,A wh ,(2 " g,/2( C,p,, )-----A-wh~2 ]
This may be expressed in the form:
AT
-- c . x ' , ' . (2)
L
where XI and X2 represent two dimensionless groups, the constant C
and exponents m and n being determined from experimental data.
McCaffrey et al. ~' analysed over 100 compartment fires, where gas
temperatures did not exceed 600 °C, and offer the following values for
C, It and m:
AT = 480. X~~-'. X~ '"' (3)
Assuming that the fire growth follows a parabolic relationship then Q
the rate of heat release is given by:
Q = o~(t - t,,) 2 (4)
Values for the coefficient, a, fall in the range, c~ = 0.1, for rapid fire
spread, to a = 0.01 for slowly developing fires. By rearranging eqn (4)
and substituting for Q then, t, the growth period, is given by:
t = t,, + (756 • ~ ~. (H~.A , A . h , ,/2) ~/2) ,.,._ (5)
A discussion concerning the merits of using an approach that is
relatively sensitive to both the ventilation and lining characteristics of a
compartment, when earlier studies have shown these parameters to be
A probabilistic model o f fire spread with time effects 377
The heat flux generated in both the growth and decay periods is small
when compared to the fully developed fire. Therefore, it seems
reasonable to assume that a barrier is only at risk during the fully
developed fire and not during the growth or decay periods. For this
reason the term fire severity shall refer only to the fully developed stage
of the c o m p a r t m e n t fire.
The expected duration of the fully developed fire, S, is given a s "11
S = L .m 1 (6)
The mass loss rate of combustible materials, rh, depends upon whether
the fire is ventilation controlled or fuel bed controlled, this being
determined by the supply of oxygen. If the supply of oxygen is greater
than that required to balance the stoichiometric equation then the fire is
fuel bed controlled. It is usual to assume that after flashover has
occurred the fire is ventilation controlled and the oxygen concentration
is nominally zero. The analysis adopts a simplified approach to this
problem based on a design m e t h o d suggested by Law and O'Brien. 12
Theobald and Heselden 13 found that in domestic furniture fuel bed
controlled fires, the fire duration is about 20 min. Taking this to
represent the m i n i m u m possible duration of a fire, the duration of the
fully developed fire is given by:
~ L . rhf ~ where S > 1200 s
(7)
S = t1200 s where S -< 1200 s
where rhr is the rate at which fuel is consumed in a ventilation
controlled fire. Based on experimental test data T h o m a s and Heselden 14
offer the following empirical expression for rhi:
rhI = 0.18. Awh~v2(W/D) m . (1 - e -°'°36n) (8)
The mass loss rate mr is based on experiment using w o o d e n cribs as the
fuel. Clearly the fuel load in real compartments will be different.
Currently these differences such as the volatility of the fuel and the
different rates of energy output are ignored. Clearly this could have
significant implications when evaluating the severity of a c o m p a r t m e n t
fire using the single parameter time) 5'16
The fuel load L of a c o m p a r t m e n t is the summation of the mass of
378 David G. Platt et al.
all the individual combustible materials. It has been assumed that the
distribution of these values will have a log normal density function. A
comprehensive series of data tables has been published in the CIB 17
report. The report summarizes many European and American surveys
of fire loads, giving average values and standard deviations for various
occupancies. The important parameter is the variability, as indicated by
the coefficient of variation (COW). For a large Swiss study it was found
that for well defined occupancies CofV = 30-50%, and for occupancies
which are rather dissimilar CofV = 50-80%.
As the mass loss rate rhs is derived from empirical data, then strictly
speaking the model should treat it as a probabilistic value. However, at
this stage of development it is assumed that rhy is a deterministic
parameter. Therefore, the duration of the fully developed fire is a
log-normal variate with an expected value and variance given by:
E{S} = rnI IE{L}; Var {S} = rhl 2 Var {L}
The direct use of the duration of the fully developed fire to represent
the severity has certain limitations. For example, it gives no indication
of the temperature, nor the build up of energy within the compartment.
However, it does offer one major advantage, that is, the use of a
common dimension, time, for both the fire severity and the barrier
resistance.
Consider for example, two similar compartments with identical fuel
loads but different ventilation characteristics. The time temperature
curves could appear as shown in Fig. 5. Whilst the total heat released
t~
OC-
TIME (0
Fully Developed Fire -Time
Period
may be the same for both compartments, the duration and temperature
will be quite different. Part of the complication with this stems from the
fact that 40 min at 400 °C is not equivalent to 20 min at 800 °C. This is
because the radiant heat flux is proportional to temperature to the
fourth power. In the example above, radiant heat exposure for 40
minutes at 400 °C is equivalent to 2.5 minutes at 800°C. For an element
of structure exposed to radiant energy, the severity experienced in the
short duration high temperature fire will be greater than that of the
long duration, low temperature fire. The problem is further complicated
if the heat absorbed by a structure, and thus its temperature, is
considered. For example, if a steel structure is clad with an insulation
material then this will retard the rate at which the temperature of the
steel increases. It is possible that the steel structure will reach a higher
temperature in the lower temperature, longer duration fire. Of course,
in a fully developed c o m p a r t m e n t fire thermal energy will not only be
transmitted by radiation but also by conduction and convection. This
further adds to the difficulty of using time as the sole measure of fire
severity.
The assumption in this paper, that the severity of a c o m p a r t m e n t fire
can be measured by its duration, is obviously crude. However, it
represents the first stage of a pragmatic approach to quantitatively
comparing the resistance of a structure to the expected fire load
measured in real time. The next stage of refinement may consider
means of estimating the c o m p a r t m e n t temperature and the quantity of
thermal energy reaching the structure from radiation sources and from
this modifying the expected resistance. In the meantime the above
limitation of the model to realistically analyse the c o m p a r t m e n t fire
must be accepted.
D I F F E R E N T M E A S U R E S OF T I M E
A further complication arises from the use of the ISO standard fire
curve. An important principle in engineering design is the ability to
compare the performance of different structures that have been tested
against a consistent frame of reference. In this respect the ISO standard
fire provides a useful flame against which various structures such as
walls and internal partitions are measured. In design it is usual to
express the duration of the c o m p a r t m e n t fire as an equivalent exposure
time to the ISO standard fire, known as the 'time equivalent' (re),
calculated using expressions such as the one given by the CIB ~7 in eqn
(9).
te = 60. wv. QI (9)
380 David G. Platt et al.
~ ~ R e a l Fire
IOO0 C '
IN.
, NISO Standard
0C --
I TIME ~[t) "
te
t
= 1"32 --~-Tl/ (1 -- e -0"036~) i (10)
te Aw. hw • W
R E S I S T A N C E TO F I R E S P R E A D
As described earlier there are three principal ways by which fire may
spread; through a door, vertically via windows or through a barrier.
Each of these paths has a different resistance to the spread of fire.
A probabilistic model of fire spread with time effects 381
If the windows are aligned vertically and the resulting flame projection
is large enough, fire may spread up the external facade of a building by
'leap frogging' from a window on the fire floor to a window above as
illustrated in Fig. 7. The analysis is based on the following simplified
assumption:
• Flames will emerge from the lower compartment at the time of
flashover.
• Calculations of the probable height and horizontal projection of
the flames from windows are based on an analysis of the
mechanism of fire spread given by Law. TM The mean height of the
flame tip above the window E { z } is given by:
Z = 1 2 " 8 ( v h f / A w ) 2/3 (11)
V///'q/~'//I///~'l///s////////~
XI
']
$
11
~///I/////~,/s//H/////¢////l~
Fig. 7. Fire spread via windows.
382 David G. Platt et al.
Fire resistance tests are performed using the ISO standard time
temperature curve. They are usually carried out on a 'one-off' basis
due, in part, to the high cost, but mainly because sponsors only require
one successful test result to obtain a FRR. As a result there is little
information known about the expected variance for a particular type of
barrier.
In a sur*ey of the fire tests carried out in New Zealand, Platt ~9 found
that the CofV was between 5 and 12%. The larger the approved FRR
the smaller the coefficient of variation. In light construction, such as
timber framed walls lined with plaster-board, the most probable mode
of failure in testing was found to be insulation failure. An integrity
failure occurred on average some 12 min after the insulation failure was
recorded. In most cases insulation failure is unlikely to result in the
spread of fire. It is therefore suggested that the failure time for
modelling fire spread be taken as the integrity failure time. 1~'2° For light
structures this may be taken as the approved FRR plus 12 min.
In the real situation, the nominal FRR of a barrier may be
substantially reduced by the presence of a weakness. These may range
from electrical services and plumbing ducts to doors, etc. They may be
internal or externally mounted and may or may not pass directly
between the two compartments.
The analysis includes these weaknesses by reducing the expected
FRR of the respective barrier. Information concerning the expected
decrease in FRR due to weaknesses is scarce. Quintiere 21 reports that
A probabil&ticmodel of fire spread with time effects 383
THE P R O B A B I L I T Y OF FIRE S P R E A D I N G TO AN
ADJACENT COMPARTMENT
Fire may spread between two adjacent compartments via three possible
path types as shown in Fig. 2.
/Windows
_ Open
Door
~Barrier
unlikely that a floor will have a fire rating less than 30 min even after
adjustment: r e m e m b e r the CofV is usually less than 12%. Therefore, if
both the window and barrier mechanisms exist it is assumed that fire
will spread via a window before the barrier fails. Figure 8 shows a Venn
diagram of the hierarchy.
The arrangement of the three path types in a hierarchy enables them
to be treated as mutually exclusive events. If fire spreads via a doorway,
then the probability of fire spreading via a window is zero, or P[WID]
is zero, likewise P[BID] and P[BIW] are also zero. By expanding
relationship (13) and using the above results the probability of fire
spreading P[FS] can be expressed as:
P[FS] = P[D] + P[W] + P[B] (14)
If the basic events for which probabilities can be calculated are;
D--fire spreads through open doorway;
/)--fire does not spread through open doorway;
W ]/)--fire spreads via windows, given it can do so;
/) U W--fire does not spread through an open doorway or via windows;
B ](/) U 14z)--fire spreads through internal barrier given it can do so;
where a bar above means the negation of the event, then eqn (14) may
be written as:
P[FS] = 1 - (1 - P[D])(1 - P[W I/5])(1 - P[B I(/) U !~,')]) (15)
The probability of fire spreading through a barrier, given fire does not
spread via a doorway or windows, is the probability that the fire
resistance is less than the fire severity, i.e.
P[B 11DU fiT] = P[n < S] (17)
where R is the barrier resistance and S the compartment fire severity.
This expression may be formulated in terms of the ratios Y = R/S. If R
and S are independent log-normal variates then Y is also a log-normal
variate. The case of barrier failure resulting in fire spread would be the
event (Y < 1); therefore, the corresponding probability of fire spread
via a barrier, is:
The probability of the event P[FS] is given by eqn (15), and from the
T h e o r e m of Total Probability the unconditional probability P[Tq = t] is
given by the following expression:
As the last term P[(Tq = t) ]FS]. P[FS] is zero, expression (22) may be
written as:
where fc,~o(t) represents the density function of the variate fire growth
time, given an open doorway, in compartment i.
If the event 'time for fire to spread given it spreads via external
windows' is taken as the fire growth time for the lower compartment i
plus the ignition time Oj for the upper compartment j, then the
probability of the events, P[(T~j = t) I W]. P[W], is given by:
(1 - P [ D ] ) . P[W I/5]
• P[(growth time for comp i + ignition time for comp j = t) I W]
where foj(h) represents the density function of the variate ignition time
of upper compartment j.
The probabilities of events (T~j= t ) [ B and P[B] are considered
separately and expanded as follows. The probability P[(T~i = t) I B] may
be written as:
P[(T~j = t) l B] = P[(Rij = t) fq (S i > t)] (26)
If the barrier resistance Rij and the compartment fire severity Si are
assumed independent then the above expression can be written as:
which is shown as the two areas marked on Fig. l(c). (This is not a
388 David G. Platt et al.
P[B] = P[B I(D U W)]. P[D U W] + P[B I(/) U if')]. P[/5 U !¢~']
= P[B I(/) U I,V)]. P[/) U I~]
But
P[{) U 1~] = 1 - P[D U W]
= 1 - P[D] - P[W] + P[D n W]
= 1 - P [ D ] - ( 1 - P [ D ] ) . P[WI/3] + P [ W I D]. P[D]
So
P[B] = (1 - P [ D ] - (1 - P([D]). P [ W I / 3 ] ) . P[B 1(/3 U if')] (29)
The substitution of eqns (24), (25) and (30) into expression (23) and the
subsequent substitution of the resulting expression into eqn (21) gives
A probabilistic model of fire spread with time effects 389
F I R E S P R E A D TO ANY C O M P A R T M E N T
3
1 4
PATH I ( ~ llnk24 PATH 2
2
~ ~llnk34
(a)
PATH 3 PATH 4
(b)
Fig. 9. (a) Plan of four-compartment building; (b) possible paths by which fire may
spread from (1) to (4).
390 David G. Platt et al.
~o~o~O~O~ .'
FNolFr~ure~ / ~ ~ ~ ( ~ ~nr_.._.7/..
Q~ ~ ~~"2(~ ~
- . - Pa l ®®
USZ3~ @®
i immure
• " " Path3
- - -- Path4
Branchonly
F i g . 10. Modifiedeventtree.
starts in compartment (1) and spreads to (4). Fire may spread from (1)
to (4) by any of the following four paths: (1)---~ (3)--* (4), (1)--0 (2)----~
(4), (1)--0 (2)--0 (3)--0 (4), ( 1 ) ~ ( 3 ) ~ ( 2 ) ~ (4). The simplest approach
to find the probability of fire spreading from (1) to (4) is to use an event
space method• This technique considers exhaustively all possible events•
However, this requires vast computational effort, and is not generally
feasible. Instead a modified event tree, which truncates a branch once
the final outcome is inevitable, is used. A graph of this is shown in Fig.
10. The probability of failure occurring along any one of these branches
is the probability that all the events on a branch occur. Consider for
example the branch marked - - 0 - - 0 - - 0 - - 0 . The probability of this
branch occuring is P[(01) n (12) n (13) N (32) N (24)]• If the events are
assumed i n d e p e n d e n t then the branch probability becomes
P[01]. P[12]. P[13]. P[32]. P(24).
Whilst the modified event tree in Fig. 10 shows seven different
branches by which fire may spread to compartment (4). Figure 9(b)
indicates that there are only four different paths by which the fire may
reach (4). Therefore, the probability of fire spread along a path will be
the summation of the branch failures that constitute that path. For
example, consider path (3), (1)--0 (2)--0 (4):
P[Path(3)] = P[(branch (01), (12), (23), (34), (24))
U (branch (01), (12), (23), (24))]
As each branch is mutually exclusive then the separate branch
probabilities can simply be summed• It is therefore possible to derive an
expression for the probability of fire spread along each path.
A probabilistic model of fire spread with time effects 391
TIME T A K E N F O R F I R E TO T R A V E R S E A G I V E N P A T H
Given that fire does spread from i to j, then let the time taken for fire to
spread along path k be represented by the random variable Zk. Then in
general:
Z, = ~ Tq (32)
all ij on path k
P R O B A B I L I T Y OF F I R E S P R E A D I N G TO ANY
C O M P A R T M E N T IN A G I V E N TIME
Consider again path (3). The probability that fire will enter compart-
ment (4) via path (3) at time t, is:
P[(Path (3))N (Path time to ( 4 ) = t)]
= P[(Path time to (4) -- t) ] Path (3)]. P[Path (3)]
As all the possible paths are mutually exclusive, in general the
probability that fire will enter compartment j via any path from
compartment i at time t, is
(34)
392 David G. Plattet al.
where ET~j I FSgj is the entry time, that is, the time taken for fire to enter
compartment, j, via any path measured from the time, t0, when fire
started in compartment i, conditional on the event fire spreads from i to
]. Substituting for Z~ eqn (34) may be rewritten as:
Thus;
Given that the fire starts in compartment i, the probability that fire will
enter, or will have entered, compartment j, via any path, at or before
time t, where t lies between t and t + dt is given by:
FIRE VECTORS
P[Fireln,] = P[FireStarts,]
fq FireStartsj] (38)
The sprinkler reliability (SPrel) is the probability that the system will
work if a fire occurs, and the sprinkler effectiveness (SPe,) is the
probability that the sprinklers, if they work, will extinguish the fire.
Note that in eqn (39), the probability terms that sprinklers are reliable
and effective modify the overall fire vector rather than individual
elements of the path. This is tantamount to saying that if the sprinkler
system fails to operate for a fire in one compartment, then it will also
fail to operate in any other compartment; secondly if sprinklers fail to
extinguish fire in any compartment then they will fail in all compart-
ments. The justification for the first assumption is that a sprinkler
system usually fails not because an individual head fails to operate but
because of some overall factor such as the water mains being turned off.
The second assumption reflects the belief that if the sprinkler heads
operate in one compartment and fail to control the fire, then the fire
will already have reached such a size that the operation of further
394 David G. Platt et al.
sprinkler heads will also fail to contain it; and furthermore, if a number
of heads are already in operation, the water pressure will be reduced
leading to a decrease in the effectiveness of the remaining heads.
P E R F O R M A N C E INDICES
-- e vT
ev - 1 (41)
APPLICATIONS
t60 i i i i
- - = - 3 0 m t n FRR Ooors
~.~60min FRR ~ ~ . . . . . 50*/* open"
/
120
2 / I" ~ - - 2 5 % open
v v v j .0..~-....-~
3 4 ~ ~" . . . . . . . "
¢1
..a l,O
closed
.~'inkler S
! I f I t I I
0
0 20 zo 60 80 tOO 120
Time ( m i n u t e s /
LIMITATIONS
Probably the most restrictive limitation that prevents the model being
able to accurately predict a real fire situation, is that the fire is assumed
to develop linearly. For example, after a fire starts in a compartment, it
goes through inception and growth phases before becoming fully
developed. If the fully developed fire spreads to a new compartment,
the model assumes that the fire growth period is the same as if the fire
were first ignited in that compartment. This approach ignores the
'thermal m o m e n t u m ' effect of the already developed fire, in particular,
the effect of the radiated heat flux emitted by hot gases. A n important
aspect of fire safety is the evaluation of smoke spread. This has not
been included in the model, although current developments are
extending the model.
CONCLUSION
This paper outlines the development of a probabilistic model to analyse
the spread of fire in multi-compartment buildings following flashover in
one or more compartments. The model is not intended to be a detailed
fire growth model such as developed by others. Fire growth is included
on a relatively crude basis, to allow the time and probability of fire
spread to be estimated. The model allows a single damage index to be
calculated for a given building. This takes into account the probability
of ignition and subsequent spread within the building.
Care has been taken at each stage to maintain a consistent level of
approximation, or crudeness, and to balance this against both the
crudeness of available data and also the use for which the model is
intended. The result is a useful tool for the comparison of different fire
safety strategies.
The approach presented in this paper offers the probability modeller
a means of overcoming the limitations of the more normal fault tree
and event tree analyses; that they cannot allow for the time taken for
events to happen.
ACKNOWLEDGEMENT
The authors thank the Building Research Association of New Zealand
for its financial support of this work.
A probabilistic model of fire spread with time effects 397
REFERENCES