Time- Dependent Event-Tree Method for Fire Risk

Analysis: Tentative Results

TIM0 KORHONEN, JUKKA HIETANIEMI, DJEBAR BAROUDI, and
MATTI KOKKALA
V T T Technical Research Centre of Finland
P.O. Box 1803, FIN-02044 V T T , Finland

ABSTRACT
A fire risk analysis method based on the use of time-dependent event trees is
presented. To concretize the method, a simple example case, a property-loss
risk analysis for a fire in a one-storey industrial hall, is presented. The method
incorporates explicitly the time dependence of the fire and its consequences.
This is achieved so that the fire incident is divided to time intervals and the
events within each time interval are analyzed with a separate event tree. The
event trees form the basic structure of the analysis and it is shown here how
to connect them to a description of the evolution of the fire. This description
can be expressed as a stochastic Markov process and it is shown how to link
the branching probabilities of the event trees to the transfer matrices of the
Markov chain.

KEYWORDS: Fire Safety, Risk Assessment, Time Dependent, Event Tree,

Stochastic Process, Markov Chain, Monte Carlo
INTRODUCTION
The development of fire safety engineering into a well-defined engineering
discipline has enabled administrations around the world to make a change
towards performance-based fire safety regulations. In performance-based design, the most rigorous way to assess the acceptability of a design solution
is t o use quantitative risk analysis, which, in addition to quantifying the
resulting risk level, can also identify the critical factors of the design.
In addition to the use in design of buildings and constructions, fire risk analysis can also be used to examine the fire regulations to establish whether the
requirements are in accordance with their role in reducing of the fire risks.
The method presented in this article is a tool for such a study with focus on
the fire resistance requirements.
The method lays emphasis to the fire resistance requirements of load-bearing
structures. The balance between the level of fire resistance requirements and
their impact on the fire risks are of great economic concern because they often are a significant factor influencing the selection of construction materials
FIRE SAFETY SCIENCE--PROCEEDINGS OF THE SEVENTH INTERNATIONAL SYMPOSIUM, pp. 321-332

Copyright International Association for Fire Safety Science

321

STATE OF THE FIRE AND

EVENTS AT DIFFERENT TIMES

PROCESSES AND ACTIONS

DETECTION: hurnansldetectors

-.

EXTINGUISHING: personn_ebFpbrigad4 ?pr'_"kler?

SELF TERMINATION OF THE FIRE
HEATING OF THE STRUCTURES
.- - - - - - -I
EVACUATION I

--

SMOKE FILLING

==3----

MARKOV CHAIN

SOURCES OF INFORMATION
STATISTICS, RELIABILITY DATA,
EXPERTS' OPINIONS, FAULT TREES
COMPUTATIONAL MODELS,
MONTE CARLO SIMULATIONS

RISKS

Figure 1: Overview of the method. The items encircled by dashed line are
still being developed and thus are not included in the presentation given in
this article.
and construction costs. Presently in Europe, the fire resistance requirements
vary considerably between different countries, or even between different authorities within one country, often for no obvious reason. The regulatory
diversity derives from the fact that the requirements are based more on, e.g.,
historical development rather than science and also the rationale behind the
national requirements may vary considerably, which implies that there is a
definite need to establish a rational risk-motivated basis for the fire resistance
requirements.
There exist several fire risk analysis methods, such as C R ISP [I],FiRECAMTM
[2], and CESARE-RISK [3] to mention some of the recent methods. However,
mainly due t o the strong fire-resistance orientation of the end-use of the
particular application related to the present work, it was decided not to resort
to use of any of the above mentioned methods, but to develop a new method.
The method is not yet completed, but is in a state of continuous development.
For example, a sufficiently versatile and reliable evacuation model is not yet
implemented. Thus, only the potential consequences to property arising from
failure of structures are considered here. The presentation is made concrete
by applying the model to a target building, a single-storey industrial building.
O V E R V I E W OF T H E M E T H O D A N D I T S A P P L I C A T I O N
The time-dependent event tree model and the other methods used to examine
the risks of a chosen fire scenario are shown schematically in Fig. 1. The basic
idea of the model is to analyze the fire incident by using the stochastic design
fire concept developed in this work. Some of the events and processes investigated during a fire are described with computational models. The inherent
variations in some processes are modeled by Monte Carlo simulations. The
fire incident is divided to a relatively small number of discrete time intervals.
At each time interval an event tree is constructed to describe the evolution
of the fire and the potential actions (by humans or automatic systems) to

322

detection manual ext. fire load

1
k7
A No Fire

h
-

*-

A5 No Fire

A3 Fire

Figure 2: The basic event tree used to analyze the fire in the target building at
different times. Shown are also the labels of the time dependent probabilities
associated with the branching points. A bar means the complement of the
probability, e.g., h = 1 - h.
detect and fight the fire. The event trees are accompanied with analysis of
the consequences of the fire at each time interval.
The example event tree structure used in this work to concretize the method
is shown in Fig. 2. The branches of the event trees correspond t o the following three events: detection of the fire (probability h ) , manual extinguishing
(probability IC), and burning out of the fire load (probability f). The system
can be only in five mutually exclusive states at each time interval t:

Ai : No fire (detected, manual extinguishment)

A:: Fire (detected)
A:: Fire (undetected)
At: No fire (detected, fire load has burnt out)
A:: No fire (undetected, fire load has burnt out)
The system is at state A3 when the fire breaks out at time t o = 0. Note, that
in the states A2 and A4 there has been at least one failed extinguishment
attempt, because extinguishment actions are supposed to be started as soon
as the fire is detected.
The event trees that are describing the progress of a fire at a given time interval are subjected to a condition that the fire is not extinguished before the
interval. The determination of the time evolution of the fire can be carried out
either by employing conditional probabilities or by a state transition process
description of the system, which can be described as a Markov process. This
process is shown schematically in Fig. 3. The transition probabilities corresponding to a transition between successive time intervals are also marked in
the figure.
For example, one can end up in the state A4 (fire is detected, manual extinguishment has failed, fire load has ended) by two different ways: 1)from state
A3 by detecting the fire, failing manual extinguishment, and by burning out
the fire load giving a transition probability h i $ ; 2) from state A2 by burning
out the fire load which has a transition probability $. Similarly all the other
transition probabilities between the different states shown in Fig. 3 can be

323

Figure 3: The state-transition process described by the time-dependent event

tree. When the fire starts (time zero) the system is in the state A 3 . States
denoted by the sharp rectangles are final states of the system, i.e. states where
the fire is extinguished. The fire is still burning in states which are denoted
by the rounded rectangles. The time-dependent transition probabilities are
related to the branching probabilities of the event trees as shown.
found. Note, that there is only one extinguishment trial allowed during each
time interval and that this attempt is instantaneous, i.e. if the attempt is
successful then there fire is already regarded extinguished at the beginning
of the next time interval.

Branching probabilities of the event trees

The branching probabilities of the event trees are calculated by using design
fires. A design fire is characterized by the time evolution of its heat release
rate (HRR).In this work, a design fire is considered to be of stochastic
nature, where its growth, maximum size, and duration are varying within
some specific ranges (see Fig. 4a). These input design fires to the system are
assumed to be freely developing fires, i.e., there are no extinguishing attempts
performed (neither by humans nor by automatic devices) [4].
The probability of detection by senses depends on the fire size at a given
time. For a t2-fire growing with normal speed (growth time 300 s) detection
probability is assumed be 0, 0.5, and 0.9 after 5, 10, and 15 minutes from the
ignition, respectively (the times correspond to fire sizes 1 MW, 4 MW, and
9 MW). All fires are assumed to be detected during the first 30 minutes of the
fire. These values are obtained by heuristic reasoning rather than estimation
based on exact data. The values are then interpolated and the probability of
fire detection by senses as a function of time is calculated by a Monte Carlo
average over the stochastic HRR curves.
The probability of automatic detection by heat detectors is modeled by calculating the activation times of the detectors. The gas temperature and flow
velocity in the ceiling jet are modeled using Alperts empirical correlations [5]
and the activation time of the detector is obtained by solving the differential
equation describing its heating. This is done for each simulated HRR curve
and thus, a probability distribution of the automatic detection is obtained.

324

140

1000
900

120

800

100

f
w

700
600

80

s
'L 500

60

400

40

300

200

20

100

20

40

60

80

100

120

Time (minutes)

15

30

45

60

Time (min)

Figure 4: a) Stochastic design fires used in this work, b) Corresponding steel

temperatures calculated for an I-beam (section factor F / V = 150 m-').
The probabilities of the burning out of the fire load are obtained from Monte
Carlo simulations by counting the fraction of fires whose HRR has started to
decrease from its peak value. In this work, the fire load is supposed to be
finished when the HRR of a freely developing design fire has decayed to half
of its peak value.
The probabilities of the manual extinguishing (both by personnel and the fire
brigade) are calculated similarly to the probability of detection by senses,
i.e., values for a few fire sizes are obtained via reasoning based on expert
judgments and then a Monte Carlo average is formed. If the extinguishing
is done by the personnel, the values 0.9 and 0.5 are used for fire sizes 1 MW
and 4 MW, respectively. For the fire brigade the values 0.99, 0.8, and 0.1 are
used for fire sizes 9 MW, 15 MW, and 36 MW, respectively.

St ructura1 failure
Heating of structures is modeled using the localized fire exposure description by Hasemi e t al. [6] revised by an additional parameter introduced by
Myllymaki and Kokkala [7] to describe the proportions of convective and
radiative heat transfer. An example of the calculated steel temperatures
is shown in Fig. 4b for a bended simply supported &beam (section factor
F/V = 150 l / m ) . In this paper, a simple structural failure mode is used.
The steel beams are assumed to fail if their temperature rises above a critical
temperature of 600 "C. The probabilities of structural damage (denoted 21 in
the following) as a function of time is then simply obtained by tabulating
the times when the steel temperature reaches the critical value in each of the
calculated curves.

325

APPLICATION OF THE METHOD TO THE EXAMPLE CASE

Target building and fire scenario
The building examined is a one-storey industrial building with floor area of
5040 m2 and height of 8 m. The load-bearing structures (columns and beams)
are made of unprotected steel and the walls of lightweight concrete. The hall
is practically a single open space. As an example, we considerer only fires
igniting during the afternoons-evenings of working days (2P" - l o p r n ) , when
there are only about 5 occupants in the building.
In this study, the main emphasis has been on the ability of the load-bearing
structures to carry their load. The most severe fire scenario for this purpose
is a large powerful fire which is heating the structures strongly. The selection
of design fires used in this work is based on the study of Paavilainen [8],
where she concludes that the burning of empty wooden pallets stored in piles
represents a severe fire scenario in the target building. Due to the low fire
load and the large size of the building as compared to conceivable sizes of
fires, the probability of a flashover is negligible, and thus localized fires are
chosen as the example fire scenario studied here. Fires which evolve to a
flashover could as easily be studied using the same event trees but then the
probabilities of the branches should be calculated differently as compared to
localized fires.
According to Babrauskas [9], the maximum heat release rate (HRR) of burning wood pallets can be evaluated from the properties of the pallet pile, i.e.,
its floor area, height, wood moisture, and effective heat of combustion. The
variability of the design fire is modeled by treating these parameters as well
as the growth time and decay time constant of the fire as a stochastic quantities. In the calculations moisture was fixed to 10 % while all the other
parameters were selected from suitable distributions. The time evolution of
the fire, which develops freely without external interruptions, is modeled using a t2-design fire with the growth time varying randomly between 300 s
and 600 s and the height of the pallet pile and its floor area varied uniformly
between 1-5 m and 2-20 m2, respectively. As a result we obtain a multitude
of HRR curves (see Fig. 4a), from which, e.g., the probability that the fire
stops to fire load burn out can be derived.
Only the first 120 minutes of a fire are considered in the risk analysis of the
target building. This time period is divided to 6 time intervals, where the
end points of the intervals are tl = 5, t 2 = 10, t 3 = 15, t 4 = 30, t 5 = 60,
and t 6 = 120 minutes. The event tree structure is shown in Fig. 2 and the
branching probabilities of the event trees are summarized in Table 1. Note,
that the fire brigade is assumed to appear at the fire location during the third
time interval (10-15 min). The driving distance is 7.7 km and according to
Tillander and Keski-Rahkonen [lo] the travel time is then just over eight
minutes in the Uusimaa region in Southern Finland.
The probability of detection ( H t ) and the probability that the fire load has
finished (Ft) are tabulated as cumulative probability distributions. (Cumulative distributions are labeled with uppercase letters whereas the branching

326

Table 1: The time dependent probabilities associated with the branching

points of the event trees. Also shown is the probability of damage to the
structures. The symbols used are explained in the text.
5 rnin

10 rnin

15 rnin

30 min

60 min

120 min

Detection, Ht
Manual ext., Ict
No fire load, F

0.08
0.95
0.00

0.79
0.75
0.00

0.95
0.97
0.00

1.00
0.68
0.00

1.00
0.39
0.65

1.00
0.39
1.oo

Damaged, ut

0.00

0.00

0.00

0.35

0.93

0.93

probabilities with lowercase letters.) Cumulative distributions like these can

not be directly used as branching probabilities, because one should use conditional probabilities at the branching points of the time dependent event
trees. For example, in order to detect the fire at time interval 5-10 min, it
should not be detected during the first 5 minutes. To calculate the conditional probability that the fire is detected, u2, during the time interval t l - t 2
Bayes theorem should be used:
u2

P[detected tl - t 2 [not detected t o

t l ] = P[t1,2IG]

Calculation of the time evolution of the probabilities

The probabilities of the states Ai after the break out of a fire are calculated
by using the time dependent branching probabilities of the event trees ht,
ft, and Ict. The following notation is used for the probabilities of the states:
P[A:]is the probability that the system is in the state Ai at time t . At time
t o = 0 the fire ignites and the system is in the state A3 (fire, not detected).
The probability P[Ai]is thus equal to one and the probabilities of all the
other states are zeros.
At the end point of the first time interval, tl = 5 minutes, these probabilities
can be deduced from the event tree shown in Fig. 2

P[A:]= hi k l P[A:] = 0.08

P[A:]= hi E1 . f i P[A:] = 0.08.0.05
P[A:] 7L1 f i . P [ A i ]= 0.92
P [ A f ]= hi . k1 f l . P[Ai]= 0.08 0.05
P[AF]= 7L1 fi PIA:] = 0.92
*

0.95 - 1
1.00 . 1
1.00 . 1
0.00. 1
0.00. 1

= 0.076
= 0.004
= 0.920
= 0.000
= 0.000

(2)
(3)
(4)

(5)
(6)

The probabilities P[Af]and P[A:] are zero because the fire load does not
burn out during the first 30 minutes. Note, that the sum of the probabilities,
E:=,P [ A f ]is equal to one as it should be.

327

Table 2: Probabilities of the states at different times. The numbers are shown
to a greater accuracy than there are physically significant numbers. In this
way, it is demonstrated that the probabilities at each time interval add up
to one exactly as they should. The probability P[Af]is always zero because
the fire is detected by senses before the fire load might end.
5 min

10min

15 min

30min

60 min

120 min

P[A;] 0.076
P[A:] 0.004
P[A:] 0.920
P[A;]
0
P[AS]
0

0.6115
0.1785
0.2100
0
0

0.939845
0.010155
0.050000
0
0

0.980750
0.019250
0
0
0

0.988258
0.004110
0
0.007632
0

0.989861
0
0
0.010139
0

At the beginning of the later time intervals the fire could be in some other
state than in the state A3, so the probabilities of the all five states are calculated for an arbitrary time point ti as (see Fig. 3)

where ti-, is the previous time point (the beginning of the corresponding time
interval). These probabilities at the times ti, i = 1,.. . , 6 , are tabulated in
Table 2.
The probability that there is no fire at time t is the sum of the probabilities
of the states Ai, A:, and A: at this time. The probability that the fire
continues according to the input design fire curve at time t is the sum of
the probabilities of the states A:, A:, A;, and A:. These probabilities are
listed in Table 3 together with the probability that the fire will be manually
extinguished during a time interval.

Probabilities of consequences
It is seen from Table 1 that during the first three time intervals (up to 15
minutes) no structural damage is formed, i.e. the steel structures have not yet
reached the critical temperature. If there still is fire at 30 minutes after the
ignition, then there is a non-zero probability (us = 0.35) that the structure
will have a damage.
The probability 'ui is a conditional probability. It is the probability of a failure
provided that the fire is still design fire at the time ti, i.e. extinguishment has
not been successful,
ui = P[damaged ti I design fire ti] =

P[damaged AND design fire ti]

P[design fire ti]

328

(12)

Table 3: The state of the fire and the building at different times given as
probabilities of the states at the end points of the time intervals. Manual
Ext. is the probability that the fire is extinguished manually during the time
interval corresponding to the given end point.

5 min

10min

15 min

30 min

60 min

120 min

Damaged
No fire
Design fire

0.000
0.076
0.924

0.000
0.612
0.388

0.000
0.940
0.060

0.007
0.981
0.019

0.014
0.996
0.012

0.014
1.000
0.010

Manual ext.

0.076

0.536

0.328

0.041

0.008

0.002

The probability that the structure will have a damage at a given time is thus
@[design fire ti] if there would be no manual extinguishment actions.
The probability that there will be a structural damage at time ti and that
the fire is extinguished manually during the next time interval ti+l is
P[damaged ti AND ext. during &+I]
= P[damaged ti I ext. during ti+l] P[ext. during ti+l]
= P[damaged ti 1 design fire ti] P[ext. during ti+l]
= viP[ext. during ti+l] ,

(13)

because if the fire will be manually extinguished during the time interval
it will be a design fire at time ti. By summing the above values up to
a time ti the probability that the structure is damaged at that time point
is obtained. These values are tabulated on the first row of Table 3. The
value corresponding to the last time point 120 min, which is about 0.014,
is the overall risk that the building will have a structural damage after the
fire incident. By examining the spatial ranges of heat exposures it can be
concluded that most of the damages are likely being restricted only to the
vicinity of the fire.

ti+l,

FORMULATION OF THE METHOD BY USING MARKOV

CHAINS
By examining the equations above it is noticed that the probabilities of the
states at a given time step ti depend only on the probabilities of the previous
time step ti-1 and on the known time-dependent branching probabilities. If
the probabilities of the states are considered to be random variables then the
system is a Markov process. A Markov process does not have memory, i.e., its
future states are fully described by the present values of the random variables.
The process is shown schematically in Fig. 3. The transition probabilities
corresponding to a transition t: ti -+ ti+l between different states are also
marked in the figure.
Define a state vectors Ai = ( A i ,A:, . . . , A:)T and a probability vectors
P[Ai] = (P[Ai],P[A:],. . . ,P[A:] ) T , i = 1,.. . , n. There are p states in the

329

system and n time intervals. Using this notation the above process can be
written as a Markov chain

where M is the transition matrix describing the evolution of the system from
time t i - 1 to time ti.
In our example case is the system is at the state A3 initially, so that the probability vector has a value P[Ao]= (O,O, l , O , O ) T at time zero. Equation (14)
reads as

when the elements of the vectors and the matrix are shown explicitly. Because
the Markov process is dealing with probabilities and the set of states is a
complete one, the columns of the matrix M i should sum up to one, i.e.,
mfl =
mi2 = . . . =
mi5 = 1.

E:=,

~ s = ~

c:=,

Equation (15) represents a system of equations and, for example, the fourth
equation, i.e., the equation for the probability of the state A4, reads as

By comparing this with Eq. 10 it is seen that mfl = 0, mt2 = k i f i , mf3 =

hikifi, mt4 = 1, and mt5 = 0. By inspection of Eqs. 7-11 it is straightforward
to extract all the other matrix elements of the transition matrix, which gives

The transition matrix M can also be form directly by examining the dynamic
process shown in Fig. 3. For example, the probability of the state A3 diminishes by an amount (hlc h % f+ hk f hf) P[A3]at each time step. One is
leftwithl-(hlc+hEf+hEf+hf)=l-(h+hf)=h-hf=h(lf)=hf,
which is just the factor before P[A:-,]in Eq 9.

After the transition matrices M i are formed for each time step it is easy
to calculate the probabilities of states Aj at any time step as follows. One
starts at the initial state probability vector P[Ao]and multiplies this by the
transition matrix M 1 in order to get the probabilities P[A1]at time t l , and

330

the probabilities at all later times are obtained similarly, i.e, by multiplying
previous probability vector with the corresponding transition matrix:

For example, the transition from the initial state A0 to the state A1 at the
first time interval tl goes like (hl = 0.08, IC1 = 0.95, fl = 1.00):

CONCLUSIONS
In this paper, a fire risk analysis method based on the use of time-dependent
event trees is presented with tentative results concerning risks to property
quantified simply as the probability of structural failure. The branching
probabilities describing the momentary state of the fire are obtained through
Monte Carlo simulations, where statistics, computational models, and heuristic reasoning are used to determine the course of events at each simulated
design fires. It is shown how to formulate the problem by using a stochastic
process, which forms a Markov chain, and how to relate the transition probabilities of the stochastic process with the branching probabilities of the event
trees. The method is applied to a simple example case in order to show how
the method works and to clarify the connection between both approaches
used to formulate the problem.
Sprinklers, smoke exhaust systems and other fire safety systems are easy to
implement to the present version of the method. In future, a more realistic
manual extinguishing model will be included, which also models the action
of fire brigade better. In the present method the success of the fire brigade is
calculated using the time passed from the ignition, whereas the arrival time of
the fire brigade depends on the time of detection of the fire. The occurrence of
flashover will be implemented as a separate fire scenario and the safety of life
through implementation of models for life hazard and evacuation calculations.

ACKNOWLEDGMENTS
This work was carried out with a financial grant from the European Coal
and Steel Community (ECSC Steel RTD Programme, Contract No.: 7210PR-251). The funding of Rautaruukki Group and Wood Focus Finland is
gratefully acknowledged.

331

REFERENCES
[1] Fraser-Mitchell, J.N., Object-Oriented Simulation ( CRISP 11) for Fire
Risk Assessment, Fire Safety Science - Proceedings of the Fourth International Symposium, International Association for Fire Safety Science,
Boston, MA, 1994, pp. 793-804.
[2] Yung, D., Hadjisophocleous, G.V., Yager, B., Case Study: The Use of
FiRECAMTM to Identify Cost-Effective Fire Safety Design Options for a
Large 40-Storey Office Building, Proceedings of the Pacific Rim Conference and the Second International Conference on Performance-Based
Codes and Fire Safety Design Methods, International Code Council,
Birmingham, AL, 1998, pp. 441-452.

[3] Beck, V., CESARE - RISK : A Tool for Performance-Based Fire Engineering Design, Proceedings of the Pacific Rim Conference and the Second
International Conference on Performance-Based Codes and Fire Safety
Design Methods, International Code Council, Birmingham, AL, 1998,
pp. 319-330.
[4] Hietaniemi, J., Baroudi, D., Korhonen, T., Bjorkman, J., Kokkala, M.,
and Lappi, E. Influence of Fire Resistance of Structures on Fire Safety
in a Single-Storey Industrial Building: Risk Analysis Using a TimeDependent Event-Tree Model (in Finnish) , VTT Tiedotteita - Research Notes 2123, VTT Building and Transport, Espoo, Finland, 2002.
(http://www .inf.vtt .fi/pdf/ tiedot teet /2002/T2 123.pdf)

[5] Alpert, R.L., Calculation of Response Time of Ceiling-Mounted Fire

Detectors, Fire Technology 8: 181-195 (1972).
[6] Hasemi, Y., Wakamatsu, T., Ptchelintsev, A., Fire Safety of Building Components Exposed to a Localized Fire, Proceedings of the ASIAFLAM 95, Interscience Communications, 1995, pp. 351-361.
[7] Myllymaki, J. and Kokkala, M., Thermal Exposure to a High Welded
I-Beam above a Pool Fire, Proceedings of the First International Workshop on Structures in Fire, University of Liege, Liege, Belgium, 2000,
pp. 211-224.
[8] Paavilainen, A., Verification of the Performance Requirements for Fire
Safety in Industrial Buildings when Using Uninsulated Steel Members
as Load Bearing Structure (in Finnish) , Masters Thesis, Helsinki University of Technology, ESPOO,Finland, 1996.
[9] Babrauskas, V, Burning Rates, The SFPE Handbook of Fire Protection Engineering (,Pded), DiNenno, P.J. (ed), National Fire Protection
Association, Quincy, MA, 1995, p. 3-1.
[lo] Tillander, K. and Keski-Rahkonen, O., The Influence of Fire Department Intervention to the Safety of a Building Assessed Using Fire
Risk Analysis, Proceedings of the Third International Conference on
Performance-Based Codes and Fire Safety Design Methods, Society of
Fire Protection Engineers, Bethesda, MD, 2000, pp. 247-256.

332