You are on page 1of 14

Engineering Structures 58 (2014) 152–165

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Fire load: Survey data, recent standards, and probabilistic models


for office buildings
Negar Elhami Khorasani a, Maria Garlock a,⇑, Paolo Gardoni b
a
Dept. of Civil and Environmental Engineering, Princeton University, Princeton, NJ, United States
b
Dept. of Civil and Environmental Engineering, University of Illinois at Urbana–Champaign, IL, United States

a r t i c l e i n f o a b s t r a c t

Article history: To enable a probabilistic performance-based approach to fire design, probabilistic models to represent
Available online 31 August 2013 the fire load are needed. Such probabilistic models are presented in this paper for office buildings. First,
a literature review of recent fire load density surveys is presented. These surveys indicate a large range of
Keywords: fire load density values, and strong correlation between fire load density, compartment area, and use.
Fire load However, current codes and standards (such as Eurocode and a recent publication of NFPA 557) that
Fire load density are used to estimate fire load density do not account for these variables and specify constant values.
Fire temperature
Based on survey data, a Bayesian probability approach is used to develop probabilistic models to predict
Compartment fire
Office
the fire load density in office buildings (one for light-weight use and one for heavy-weight use). The mod-
Survey els consider the size of the compartment and the office room use (general office, library, storage, etc.). The
Probabilistic model proposed models correlate well to the data and have a better fit than that obtained, using the Eurocode
Bayesian and NFPA 557. The proposed models for fire load density are then used to develop probabilistic models
Performance-based design for the maximum fire temperature in a given compartment. Several scenarios with different floor areas
and openings are defined and the fire load models developed in this paper are used to investigate the
range of possible maximum fire temperatures and their corresponding probabilities. It is found that
the proposed maximum temperature model results in a range of temperatures that correlates well with
the test data and the Refined Tanaka Method proposed by a recent SFPE standard. It is shown that both
the fire load density and the maximum temperature probabilistic models are well suited for application
in a probabilistic performance-based approach to fire design.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction that falls into the specified occupancy category (for example all of-
fice rooms would have the same fire load density value). Statistics
The first step in the design of structures under fire is to define of data collected from fire load density surveys are the basis for the
the fire load density and develop the temperature–time evolution design values in codes and standards. These surveys indicate a
of fire. During a fire event, compartment temperatures depend large range of fire load density values, therefore codes and stan-
mainly on three factors: (1) geometric characteristics of the dards recommend a conservative value (80% fractile or higher) that
compartment, (2) thermal properties of the boundary of enclosure has a small probability of reaching. The surveys also show strong
and (3) fire load density. All three groups of variables have correlation between fire load density, compartment floor area,
uncertainties related to them; however, Hurley [1] indicates that and use; but current codes and standards do not consider these
thermal properties of the enclosure linings have the least amount correlations. Therefore, the use of probabilistic models that con-
of uncertainty. On the other hand, the fire load density strongly sider these correlations and produce a range of results with related
influences the temperature achieved during a fire event, due to probabilities allows the designer to make evaluations based on
the great deal of uncertainty in its value. more realistic fire load values.
Determination of fire load density, i.e., the available material as This paper consists of the following original contributions to the
fuel, is not an easy task. The design fire load density for a specified literature on fire design:
room, that is evaluated based on codes and standards, is generally
a deterministic value that would be the same for any other room (1) A literature review of recent fire load density surveys [2–6]
is given. The data are compared to older data [7–15], design
values suggested by the Eurocode [16], and a recent publica-
⇑ Corresponding author. Tel.: +1 609 258 2728.
tion of NFPA [17].
E-mail addresses: nelhami@princeton.edu (N. Elhami Khorasani), mgarlock@
princeton.edu (M. Garlock), gardoni@illinois.edu (P. Gardoni).

0141-0296/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.engstruct.2013.07.042
N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165 153

(2) Based on survey data, a Bayesian probability approach [18] Results of the questionnaire method were not included in [2]
is used to develop two novel probabilistic models to predict because the sample size was small and the reported statistics
the fire load density in office buildings (one for light-weight would not be representative. Comparison of fire load density for
use and one for heavy-weight use). The models consider the the three other methods under consideration is presented in
size of the compartment and office room use (general office, Table 1.
library, storage, etc.), both of which have a strong correlation The results show that in most cases the inventory method pre-
to fire load density in the surveys. dicted larger fire load densities followed by the combination
(3) The proposed models for fire load density are then used to method, while the weighing method predicted the lowest fire
develop probabilistic models for the maximum temperature load density. The report concludes that the combination method
in a given compartment. Several scenarios with different is the best survey technique since it is a combination of best prac-
floor areas and openings are defined and the fire load models tices from both methods and therefore minimizes the degree of
developed in this paper are used to investigate the range of uncertainty and error. Table 1 shows that there is human error
possible maximum fire temperatures and their correspond- in the sampling, given the variability in the measurements using
ing probabilities. different methods. The study also concludes that cubicle offices
(4) The proposed probabilistic models for fire load density are have greater fire load densities (MJ/m2) than enclosed offices,
compared to deterministic values given by the Eurocode while combustible content (MJ) of cubicle offices are less than
[16] and NFPA [17]. Further, the proposed probabilistic those of enclosed offices. Finally, fire load density would decrease
model for maximum temperature is compared to a recent with increasing floor area, but the area is not reported so there is
publication by SFPE [19]. no data to quantify a correlation between area and fire load
density.
This work is significant for the development of a probabilistic
performance-based approach to fire design, which is a ‘second- 2.2. Survey performed in Europe in 2005
generation’ performance-based engineering approach. This ap-
proach uses a probability distribution for the variables with uncer- Fire load data of 95 industrial and commercial enterprises in
tainties and ‘‘improve[s] . . . risk decision-making through Switzerland were surveyed in 2005 by ETH (Eidgenössische Tech-
assessment and design methods that have a strong scientific basis nische Hoschule Zürich) and VKF (Swiss Fire Authority, Bern) [3].
and that express options in terms that enable stakeholders to make The report does not specify details of the survey method but states
informed decisions’’ [20]. A probabilistic performance-based ap- that the data are based on ‘‘instantaneous snapshots’’ that can be
proach to fire design thus has the potential to provide higher levels subject to relatively large fluctuations. Therefore, in many cases,
of fire safety and also lead to more efficient and cost effective the minimum and maximum surveyed values are both presented.
designs. Given the observed range of data and based on experience, the
study proposes a design value for fire load density as shown in
Table 2.
2. Fire load density survey data in office buildings
2.3. Survey performed in India in 1993
Fire load density is the amount of available combustible mate-
rials per unit floor area of a compartment (MJ/m2). The combusti-
In Kanpur, India, 388 rooms in eight government office build-
ble materials in a compartment include wall and ceiling linings,
ings were surveyed for fire load density data from July 1992 to July
and furniture in the room. To investigate a realistic range for the
1993 [4]. In the buildings surveyed, no office space was configured
value of fire load density for office buildings, actual collected data
as an open plan environment (cubicle). The survey was completed
are studied. It should be noted that many of the fire load surveys
using the inventory method. Results of the survey are presented in
for offices were conducted in the 1970s and 1980s [5,7,21,22]
Table 3. The floor areas presented in the table are the total covered
while the number of new fire load surveys conducted in recent
floor areas that were surveyed.
years is limited. This section presents results of the available few
recent surveys and compares them to the CIB W14 collection of
2.4. Survey performed in the US in 1975
surveys [7] and the most comprehensive survey performed by Cul-
ver in 1975 [5].
Based on a 1975 survey, Culver [5] reported statistics of fire load
density for 23 typical US office buildings located in various regions
2.1. Survey performed in Canada in 2011 of the United States. The data collection procedure was based on
the inventory technique. At the time, the concept of inventory
A recent research project, supported by the National Fire Protec- method had been used to some extent, and Culver fully utilized
tion Association (NFPA), was carried out in Canada to perform a the concept [5]. Fig. 1 shows the frequency distribution of room fire
sensitivity analysis of the key survey methods for evaluating fire load density in clerical and general offices in government and pri-
load density. As part of the sensitivity analysis, results of the fire vate buildings and Table 4 shows the results of the survey. Culver’s
load density surveys were also published and are therefore studied report, although published in 1976, still has the most complete
in this paper [2]. Four main survey methods were considered: (1) information and data on fire load density for office buildings. The
weighing method, (2) inventory method, (3) combination of report studied the influence of several factors on the value of fire
weighing and inventory and (4) questionnaire. The objective of load density. Results show that the type of office building (govern-
the Canadian study was to establish an approach to validate fire ment or private offices), geographic location, building height, and
load survey methods and to enhance guidance on means to cali- building age have little or no influence on the value of fire load
brate the results of fire load data surveys. A total of 103 offices in density. However, the room use affects the fire load density and
five office buildings were surveyed, including 27 enclosed offices the report concludes that influence of the room area on fire load
and 76 open offices (cubicles). The office buildings were located density requires further study.
in the provinces of Ontario and Quebec and included private and Fire loads in Culver’s report are in lb/ft2 of floor area but val-
federal government buildings. The survey was performed from ues in this paper are converted to MJ/m2. In order to convert
November 2010 to April 2011. from lb/ft2 to MJ/m2, the caloric value of fuel (combustible
154 N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165

Table 1 (paper, cardboard, cotton, etc.) [16]. Also, Buchanan [23] makes
Comparison of fire load density [2]. the argument that the caloric value depends on moisture content
Sample Inventory Weighing Combination of wood and can range from 16.7 MJ/kg for wood with 12% mois-
size method method method ture content to 19 MJ/kg for a dry wood. Buchanan also suggests
Mean (MJ/m2) 103 852 530 557 that, based on other studies and documents [24,25], wood caloric
Standard deviation 103 484 257 286 value can range from 13–20 MJ/kg while NFPA fire protection
(MJ/m2) handbook gives a typical range of 18–21 MJ/kg [26]. Based on
the above discussion, a net caloric value of 18 MJ/kg is assumed
in this paper.
Table 2 A total of 2433 offices are surveyed, with 8.5% of them (207 offi-
Results of fire load density survey for offices [3]. ces) discarded due to error. Out of the remaining 2226 surveyed
Office use Surveyed fire load Proposed fire load rooms, 1354 of them are randomly selected. The report presents
(MJ/m2) (MJ/m2) the fire load density for the 1354 randomly selected rooms,
Offices – general 300–900 700
grouped into different room categories: General, Clerical, Lobby,
Offices – meeting rooms 200–400 300 Conference, File, Storage and Library. Also, office spaces are split
into four categories based on floor area Af: less than 4.65 m2
(50 ft2), between 4.65 to 9.30 m2 (50 to 100 ft2), 9.30 to 27.90 m2
(100 to 300 ft2) and larger than 27.90 m2 (300 ft2). The detailed sta-
materials in the compartment) needs to be assumed. Most of the tistics for each category (floor area range, mean, standard deviation
fuel in building fires comes from cellulosic material; therefore and maximum values for observed fire loads (MJ/m2)) in 834
caloric value of wood (MJ/kg) is typically used to convert from private office buildings are provided in Table 5. Similar detailed
units of mass to energy. EC1 assumes a net caloric value of results of the survey for 520 government office buildings can be
17.5 MJ/kg for wood and 20 MJ/kg for other cellulosic materials found in Culver’s report [5].

Table 3
Influence of room use on fire load density [4].

Room use Number of rooms Floor area (m2) Fire load density (MJ/m2)
Maximum load Mean load Standard deviation
Reception 19 246.7 1540 537 349
Storage and file 40 1217.9 1860 601 372
Technical 21 623.8 1240 434 302
Clerical 84 3617.8 1760 432 200
General 108 2802.6 1080 300 191
Corridor 36 1922.2 1150 153 140
Conference 8 575.5 317 189 39
Lavatory 62 461.9 762 146 143
Miscellaneous 10 251.4 862 278 270
All rooms 388 11719.8 1860 348 262

(a) (b)

Fig. 1. Frequency distribution of room fire load data for (a) government, and (b) private office buildings (adapted from [6]).
N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165 155

Table 4 uncertainty in predicting fire load density in an office room. This


Results of fire load density survey for government and private offices in MJ/m2 [6]. observation is consistent with the results of CIB W14 report. An-
Building type Mean St. dev. Min Max other important observation made from the collected surveyed
Government buildings 641 641 53 9365 (storage) data in Table 7, is the influence of the room use and size on fire
Private buildings 720 562 132 4551 (storage) load density. Most of the above studies realize that meeting rooms
or relatively large offices have a smaller fire load density than stor-
2.5. CIB W14 collection of surveys in 1986 age rooms, file areas, or smaller offices in an office building.
Except for the Indian survey [4], recent surveys do not provide
A complete study on the value of fire load density was per- detailed statistics of surveyed floor areas or their relation to the fire
formed in early 1980s and the results were published in the CIB load density. The Indian survey provides statistics on the variation
W14 workshop report [7]. Based on what was available at the time, of fire load density with floor area; however, the data has the low-
the report collected data of surveys for different countries, differ- est mean with a relatively small standard deviation compared to
ent types of occupancy, and different room use [5,8–15]. Table 6 other surveys as shown in Table 7. The maximum surveyed fire
shows a summary of some of the relevant data for the purpose of load in the Indian Survey is 1860 MJ/m2 (Table 3) whereas Culver
this study. The table confirms that there is a wide range of recorded reported values as high as 9365 MJ/m2 (Table 4). At the time, the
fire load data. It can also be concluded that, aside from the Swiss Indian survey [4] confirmed that their fire loads are smaller than
data where the standard deviation is not available, the USA surveys Culver’s and that their data is closer to Baldwin’s survey performed
estimated larger fire load densities compared to the European and in UK in 1970 [21] than Culver’s survey. They also explained that
Swedish surveys. The CIB W14 workshop reported data from Cul- smaller fire loads were partly due to an increase use of lighter fur-
ver’s study [5] plus other sources from the USA; but Table 6 only niture in offices. However, the European and Canadian surveys,
presents surveyed fire load density from sources other than Cul- which were performed after the Indian survey, report larger fire
ver’s. Also, Table 6 does not include results of the survey in [13] loads.
as the reference reports a single value for permanent fire loads The USA survey performed by Culver, although older than other
and not a distribution of data. surveys in Table 7, interestingly enough, reports a mean value
within a reasonable range when compared to recent surveys. Cul-
ver also provides the most comprehensive data collection when
2.6. Comparison of survey data compared to recent surveys. One may have concern that with the
use of computer accessories, the nature of office furniture might
Similar to CIB W14 workshop report [7], Table 7 compares the have changed. This change could affect the caloric value of furni-
results of fire load density surveys collected in this study (Sections ture. Having the same concern about using Culver’s report, NIST
2.1–2.4). It can be seen that the Canadian survey with the Inven- completed a survey nearly 20 years later in 1996 [27]. The report
tory Method estimates the highest mean value (852 MJ/m2) while investigated the changes in office furniture, including computers
the Indian survey reports the smallest mean value (348 MJ/m2). and furnishings to create workstation (partitions). The study
Available surveys show a wide range in the recorded mean fire load included buildings at the University of Maryland College Park
density values, which implies a considerable amount of and at the General Services Administration (GSA) Headquarters

Table 5
Influence of room use and size on fire load density for private office buildings [5].

Room use Floor area Af (m2) No. of rooms Total fire load (MJ/m2)
Mean St. dev. Max.
General offices Af 6 4.65 3 1529 1538 3269
4.65 < Af 6 9.30 107 800 413 1977
9.30 < Af 6 27.90 331 650 334 2513
Af P 27.90 38 554 378 2100
Clerical Af 6 4.65 3 923 325 1195
4.65 < Af 6 9.30 21 580 519 2118
9.30 < Af 6 27.90 78 580 281 1239
Af P 27.90 44 466 281 1318
Lobby Af 6 4.65 1 – – 228
4.65 < Af 6 9.30 4 360 255 685
9.30 < Af 6 27.90 29 536 404 2074
Af P 27.90 11 237 211 659
Conference Af 6 4.65 0 – – –
4.65 < Af 6 9.30 3 641 255 826
9.30 < Af 6 27.90 45 554 431 2470
Af P 27.90 9 290 88 387
File Af 6 4.65 2 3427 782 3981
4.65 < Af 6 9.30 10 1345 1142 4262
9.30 < Af 6 27.90 5 817 44 1933
Af P 27.90 3 431 44 457
Storage Af 6 4.65 27 1213 1107 4552
4.65 < Af 6 9.30 18 1169 1090 3085
9.30 < Af 6 27.90 26 1046 993 4095
Af P 27.90 6 694 800 2672
Library Af 6 4.65 0 – – –
4.65 < Af 6 9.30 2 1520 0 1520
9.30 < Af 6 27.90 7 2153 1089 3506
Af P 27.90 1 – – 2601
156 N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165

Table 6 qf,k for different occupancies (such as office, dwelling, and hospital)
Summary of fire load density data in office buildings from the CIB W14 workshop but it does not impose a particular fractile of qf,k to use in design. It
report [7].
does, however, imply that the value should be equal to or greater
Source Mean (MJ/m2) St. dev. (MJ/m2) than the 80% fractile. Given the source data of EC1 [7] and another
Swedish data [8,9] 411 334 study by [28], the 95% fractile is a reasonable design value for fire
European data load density for an office compartment. Therefore, according to
Source 1 [10] 420 370 source data of EC1 for an office building, which has a mean fire load
Source 2 [11] 410 330 density of 420 MJ/m2 and an 80% fractile of 511 MJ/m2, the 95%
Source 3 [12] 330 400 fractile is estimated to be 760 MJ/m2 [28]. From this point on in
Swiss data [14] this paper, qfk,d will be designated as the design characteristic fire
Technical offices 580 – load density based on EC1 95% fractile with a value of 760 MJ/m2.
Administrative offices 750 –
To compare the EC1 design value with Culver’s data, a set of
USA data (government buildings) [8,15] 555 625 data points are randomly generated. In Culver’s report, data are ar-
ranged in tables (similar to Table 5) such that mean, standard devi-
ation and maximum values of fire load density for different ranges
Table 7 of floor area are provided. Therefore, with the provided data and
Comparison of survey results for fire load density in office buildings.
assuming a minimum value for the fire load density, Beta distribu-
Reference Mean (MJ/m2) St. dev. (MJ/m2) tion is assumed for modeling the load. This distribution is selected
US Survey-1975 [5,6] because it can take a wide range of shapes for the probability dis-
Government: light-weighta 584 Not reported tribution function. Also, similar to Table 5, three ranges for floor
Government: heavy-weighta 1213 Not reported area are considered (4.65 < Af < 9.30 m2, 9.30 < Af < 27.90 m2 and
Government: total 641 641
27.90 < Af < 46.45 m2). In this paper, floor areas less than 4.65 m2
Private: light-weighta 633 Not reported
Private: heavy-weighta 1294 Not reported (50 ft2) are not considered due to insufficient data reported for this
Private: total 720 562 range. The maximum possible Af is limited to 46.45 m2 (500 ft2) to
Indian Survey-1993 [4] 348 262 impose a reasonable upper bound on the compartment size. For
every range, it is assumed that Af follows a uniform distribution
European Survey-2005 [3]
General 300–900 – since the compartment sizes in each range are equally probable.
Meeting rooms 200–400 – Based on the surveyed fire load density and floor areas by Cul-
Canadian Survey-2011 [2] ver, and the assumed distributions, for each range under consider-
Inventory methods 852 484 ation, five random floor areas and five random fire loads are
Weighing method 530 257 generated. As explained in Section 2.4, Culver surveyed both pri-
Combination method 557 286 vate and government office buildings, with 834 observations for
a
Light-weight category includes general offices, clerical, lobby and conference private offices and 520 observations for government buildings.
room, and heavy-weight category includes file, storage and library. The randomly generated data in this study are based on the data
from both office types.
Building in Washington, DC. The results show that papers and In order to relate the corresponding load to an area, it is as-
books contributed 54% of total load, furniture contributed 25.1%, sumed that as area increases, fire load density decreases. Therefore,
partitions contributed 11.4%, computer equipment contributed the largest area would correspond to the smallest fire load density.
6.2% and miscellaneous items contributed 3.3%. Partitions are This assumption is made because Culver’s analysis on all data (gov-
made of fabric panels that are fitted into metal frames, not having ernment plus private offices), indicates that the mean load de-
a caloric value significantly larger than other furniture. It is only creases as the area increases. The trend can also be observed for
the 6.2% computer equipment that has a larger caloric value. every room category in Table 5, except for library. Since there are
On a similar note, a new standard recently published by NFPA only twelve observations available for the category of library out
[17] lists the 1993 Indian survey [4] as one of the sources for infor- of 1354 randomly selected rooms, no data for this study is gener-
mation on fire load density; yet again it uses Culver’s data from ated from that category.
1975 to determine the statistics published in the Standard. NFPA Generated data points are organized into two main categories:
used a value of 15 MJ/kg for the net caloric value when using Cul- (1) light-weight compartments (general, clerical, lobby and confer-
ver’s data. Parker and Tran [25] showed that the net caloric value ence) and (2) heavy-weight compartments (file and storage). Based
or effective heat of combustion of wood is in the range of 13– on this classification, a total of 60 data points are generated for
15 MJ/kg. Based on the discussion in Section 2.4 on the caloric va- light-weight category, and 30 data points for heavy-weight
lue of fuel, and the above explanation on the changes in office fur- category.
niture, the net caloric value of 18 MJ/kg selected in this study is a Fig. 2 compares the EC1 design characteristic fire load density
conservative value that can compensate for the larger caloric value qfk,d with randomly generated data from Culver’s report qCulver.
of computer equipment. Therefore, Culver’s survey [5] was The EC1 data fall on a straight line since qfk,d is a constant value
adopted as the primary basis for this study. of 760 MJ/m2 for all office types. Fig. 2 also shows qualitative Prob-
ability Distribution Functions (PDFs) of qCulver with the obtained
3. Current codes and standards compared to survey data mean values for light and heavy weight categories (455 MJ/m2
for light-weight and 1305 MJ/m2 for heavy-weight categories).
3.1. Eurocode 1 and fire load density The comparison between qfk,d and qCulver shows that, based on
EC1, qfk,d is mostly overestimated for light-weight categories and
Eurocode 1 (EC1) [16] defines qf,k as the characteristic fire load underestimated for heavy-weight categories.
density per unit floor area and this is the quantity that is best re-
lated to the surveyed data. In Annex E of EC1, it is stated that qf,k 3.2. NFPA standard and fire load density
may be determined ‘‘from a national fire load classification occu-
pancies and/or specific for an individual project by performing a The National Fire Protection Association (NFPA) recently
fire load survey.’’ EC1 provides mean and 80% fractile values of published ‘‘NFPA 557: Standard for Determination of Fire Loads
N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165 157

for Use in Structural Fire Protection Design’’ [17]. The standard 557 reported values are based on the combined data from govern-
uses a risk framework in determination of fire load densities. It ment and private buildings while Culver’s report provides the
considers two types of fire loads: localized fire loads and distrib- mean and standard deviation values for government and private
uted fire loads. Localized fire loads are those with concentration buildings separately and the combined values are not known to
of combustible materials and the potential of severe thermal expo- the authors of this paper. Also, Culver published data in units of
sure. Distributed fire loads reflect the total fire load throughout a mass per unit area (psf), while the values in this paper and NFPA
compartment. 557 are in units of energy per unit area (MJ/m2). To convert from
The procedure to calculate the design fire load density for a dis- mass to energy, in this paper, the effective heat of combustion is
tributed fire is based on statistical distribution of fire loads in assumed to be 18 MJ/kg (see Section 2.4 for discussion), whereas
buildings, fire initiation frequency and effectiveness and reliability NFPA 557 assumes 15 MJ/kg. The NFPA 557 assumption is based
of the fire protection features that contribute to fire control in the on the reported values in SFPE handbook of fire protection engi-
early stages of the fire. Also, the standard divides the combustible neering [36].
material in a compartment into two types [17]: (1) content fire NFPA 557 also provides an annex on the procedure to perform a
load: which consists of ‘‘all moveable or secured contents and fur- fire load survey. The document explains the permitted simplifying
nishing and all occupant possessions within a compartment, assumptions, recommendations on survey process and creating a
including items that can be placed into a compartment or taken survey form and finally the relevant guidelines on analysis of the
out of it without causing structural damage’’, and (2) fixed fire data.
load: which includes all combustible materials used as structural Having specified the mean and standard deviation of the fire
elements or as interior finish. load density, the NFPA 557 standard proceeds to calculate the de-
For office buildings, the standard specifies an average content sign fire load density based on a risk objective and the Gumbel dis-
fire load density of 600 MJ/m2 with standard deviation of 500 MJ/ tribution. The risk objective is related to a specified risk
m2. Also, it specifies an average fixed fire load density of 130 MJ/ performance criterion and frequency of structurally significant
m2 with standard deviation of 40 MJ/m2 for buildings of noncom- fires. It should be noted that the frequency of structurally signifi-
bustible construction materials. Combined together, an average va- cant fires is a function of floor area, where a larger floor area im-
lue of 730 MJ/m2 (sum of the averages for content and fixed fire plies higher risk of significant fires.
load densities) with standard deviation of 502 MJ/m2 (square root
of sum of the squares of standard deviations) is obtained for a typ-
ical office building. These values are similar to Culver’s data [6] (Ta- 4. Proposed probabilistic fire load density models
ble 7, USA Survey) for reasons that will be explained shortly.
NFPA 557 provides an annex on fire load survey data. The annex In this section, novel probabilistic fire load density models are
explains that the fire load density values in the document are developed using a Bayesian approach to correct current codes
based on a study of Refs. [4,5,21,29–35]. Their study shows that and standards accounting for the information from the survey data.
with the exceptions of Culver [5] and Ingberg [29], none of the
other documents provides information on space usage in offices. 4.1. Bayesian framework
Ingberg’s survey [29] was conducted from 1928 to 1940; therefore,
given the date, it is not likely that the data represents current fire Based on the work of Gardoni et al. [18], a Bayesian framework
loads and is not used in developing the fire load density in NFPA is applied to develop probabilistic models for predicting quantities
557. Therefore, Culver’s data [5] is used to develop the specified of interest given deterministic values and observed experimental
fire load density in the document. This explains why the specified data. In this procedure, a model is a mathematical expression that
values by NFPA 557 (mean of 730 MJ/m2 and standard deviation of relates a quantity of interest to a set of measurable variables. A
502 MJ/m2) are close to Culver’s data (mean of 641 MJ/m2 and model has the general form C = C(x, H) where H is a set of param-
standard deviation of 641 MJ/m2 for government buildings, and eters introduced to fit the model to the observed data,
mean of 720 MJ/m2 and standard deviation of 562 MJ/m2 for pri- x = (x1, x2, . . .) is a set of measurable variables and C is the quantity
vate buildings). The values do not match exactly because NFPA of interest. The general model form can be written as:

(b)
(a)

Conservative
Mean=455MJ/m2 Non-conservative
qfk,d (MJ/m2)
qfk,d (MJ/m2)

2
Mean=1305 MJ/m

53 data points out of


60 points: conservative 16 data points out of 30
points: non-conservative

Fig. 2. Comparison of Culver’s observations (randomly generated) with EC1 [16] design characteristic fire load density (a) light-weight, and (b) heavy-weight categories.
158 N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165

Cðx; HÞ ¼ ^cðxÞ þ cðx; hÞ þ re ð1Þ Eq. (1) and the formulation explained in Section 4.1 are used to
develop the proposed models. The proposed fire load density (qcor-
where H = (h, r), h = (h1, h2, . . .) denotes the set of unknown model
rected in Eqs. (6) and (7)) is calculated based on the natural loga-
parameters, ^cðxÞ is the selected deterministic model that is
rithm of suggested code value ðq ^Þ that is corrected by Eq. (2). The
expressed as a function of the variables x, c(x, h) is the correction
natural logarithm is a variance stabilizing transformation, required
term for the bias inherent in the deterministic model, re is the
to fulfill homoskedasticity assumption of Gardoni’s formulation.
model error that captures the remaining scatter in the residuals,
The proposed equations include the model error re. The three un-
in which e is the random variable with zero mean and unit variance
known parameters in the proposed formulation (h1 and h2 in Eq. (2)
and r represents the standard deviation of the model error. In writ-
and r) are calculated using the maximum likelihood estimation.
ing Eq. (1), three assumptions are made: (1) e follows a standard
Eqs. (6) and (7) are the final form of the proposed model:
normal distribution (normality assumption), (2) r does not depend
Light-weight compartment:
on x (homoskedasticity assumption), and (3) additivity of c(x, h)
(i.e., the correction can be added to ^cðxÞ instead of being, for exam- ^Þ þ h1;LW  0:0047ðAf  10:76Þ þ 0:5712e
qcorrected ¼ exp½lnðq ð6Þ
ple, a multiplicative term.) These assumptions can typically be sat-
isfied by considering variance stabilizing transformations of the Heavy-weight compartment:
original quantity of interest [38], and verified using appropriate ^Þ þ h1;HW  0:0081ðAf  10:76Þ þ 0:5508e
qcorrected ¼ exp½lnðq ð7Þ
diagnostic plots [39].
For the purpose of this study, C(x, H) is the natural logarithm of where q ^ is the deterministic value of fire load density from current
the corrected (i.e. proposed) characteristic fire load density, de- codes and standards in units of MJ/m2, and Af is in units of m2. The
noted by qcorrected. Based on the available measured data, variable value of h1,LW and h1,HW depends on the selected value of q ^ and must
x is the floor area (Af) of the compartment, ^cðxÞ is the natural log- be calculated such that ðlnðq ^Þ þ h1LW Þ ¼ 6:951 for light-weight and
arithm of a deterministic value of the fire load density from current ^Þ þ h1;HW Þ ¼ 8:252 for heavy-weight. For example:
(lnðq
codes and standards (henceforth denoted q ^), and c(x, h) is defined  Using the mean EC1 value for q ^ ¼ 420 MJ=m2 , h1,LW = 0.910,
as: h1,HW = 2.212.
 Using NFPA 557 value for q ^ ¼ 730 MJ=m2 , h1,LW = 0.358,
cðx; hÞ ¼ h1 þ h2 Af ð2Þ
h1,HW = 1.659.
where h1 captures a potential constant bias in ^cðxÞ and h2Af brings in
the dependency of the bias on the floor area of the compartment. Regardless of the value of q^ chosen, qcorrected generates the same
A Bayesian framework is considered to estimate the parameters distribution as long as h1,LW and h1,HW are determined correctly. A
H. In this approach the distribution of the parameters are assessed more generic way to write Eqs. (6) and (7) is therefore:
using the following updating rule: Light-weight compartment:
f ðHÞ ¼ jLðHÞpðHÞ ð3Þ qcorrected ¼ exp½6:951  0:0047ðAf  10:76Þ þ 0:5712e ð8Þ
where f ðHÞ is the posterior distribution of H that reflects the objec- Heavy-weight compartment:
tive information in the likelihood function LðHÞ and the subjective, a
priori, information in the prior distribution pðHÞ, and j is a normal- qcorrected ¼ exp½8:252  0:0081ðAf  10:76Þ þ 0:5508e ð9Þ
izing factor.
Based on Eqs. (8) and (9), Eqs. (10) and (11) show the second-order
Applying Eq. (1) to the measured data point Ci with the given xi
approximation of the expected value of qcorrected for light and heavy
results in C i ¼ ^cðxi Þ þ cðxi ; hÞ + rei or rei = ri (h), where
weight categories given the probabilistic models in Eqs. (6)–(9). Eqs.
ri ðhÞ ¼ C i  ^cðxi Þ  cðxi ; hÞ and ei denotes the outcome of the model
(10) and (11) provide relatively accurate predictions for the mean of
error term at the ith measured data point. With the assumption of
qcorrected. Approximate standard deviation equations are not pro-
exact measurements, and under the assumption of statistically
vided here because the approximation results in an error in the or-
independent observations, the likelihood function has the form of:
Y der of 20%. The general procedure to obtain moments of nonlinear
LðHÞ / P½rei ¼ r i ðhÞ ð4Þ functions and the relevant formulas can be found in [40].
all data Light-weight compartment:
and since e follows the standard normal distribution, Eq. (4) can be 1215
rewritten as: Mean ðqcorrected Þ  ð10Þ
expð0:0506Af Þ
Y  1 r i ðhÞ
LðHÞ / / ð5Þ Heavy-weight compartment:
all data
r r
4418
where /() denotes the standard normal probability density Mean ðqcorrected Þ  ð11Þ
expð0:0872Af Þ
function.
In the above formulation, it is assumed that there is no error in Fig. 3 shows a comparison of fire load densities from Culver’s survey
the measured data. If measurement errors are present, they can be results qCulver with the median ‘‘corrected’’ values as designated by
accounted for by using the general formulation in [18]. In this pa- qcorrected in Eqs. (8) and (9). The 1:1 lines drawn on the plots repre-
per, it is also assumed that no prior information is available about sent a perfect match between the data and model, and the dashed
the parameters H before collecting the surveyed data. lines in the figures delimit the regions within one standard devia-
tion of the model. The corrected models for both light and heavy
4.2. Model formulation weight categories demonstrate an improved prediction of fire load
density values when compared to EC1 (see Fig. 2). Fig. 3 shows that
Fig. 2 showed that the EC1 model used for fire load density can the correction improves the accuracy of the prediction in particular
be significantly improved. This section proposes new probabilistic for qCulver < 1500 MJ/m2 in light-weight categories (Fig. 3a) and for
models where current codes and standards (e.g., EC1 and NFPA qCulver < 4000 MJ/m2 in heavy-weight categories (Fig. 3b). The cor-
557) are modified based on the observed data, while including rection does not work as well for higher values of qCulver; but such
the influence of floor area and room use. values are highly improbable for the given category. Further, for
N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165 159

heavy-weight categories, even in these large ranges, the proposed The above two methods were selected based on another study
model fits better to the data than the EC1 predictions. in which SFPE compared the maximum fire temperature in a
compartment from test experiments with those of 14 available
5. Maximum gas temperature predictive methods in the literature [37]. Results from 702 test
experiments were compiled from which 321 tests were conducted
In design, the fire load densities must be converted to a gas tem- under the auspices of CIB. The study concludes that ‘‘if the predic-
perature in relation to time for a given compartment. The method- tion of the peak temperature were the only metric concern, [. . .]
ology to do the conversion and calculating the gas temperature is the most logical selection would be either Tanaka Refined [meth-
discussed based on the provided guidelines in two documents: od] or a 1200 °C constant exposure’’ [37].
Eurocode and The Society of Fire Protection Engineers (SFPE) stan- Fig. 4a and b are from the SFPE document [37] and compare the
dard. In this section, first a brief overview of the SFPE standard is test results with predictions from the constant temperature and
provided, and then the procedure by Eurocode is used in seven dif- Tanaka refined methods. It can be seen that, in both figures, almost
ferent scenarios to perform a parametric study on the maximum all the data points are on the conservative side; therefore both
gas temperature. The objective is to demonstrate an application methods suggested by SFPE predict a conservative envelope for
of the proposed probabilistic fire load models (Eqs. (8) and (9)) the maximum fire temperature in the compartment.
in a probabilistic study of compartment fires.
5.2. Parametric study and seven scenarios
5.1. SFPE standard and the maximum temperature
Eurocode 1 Annex A [16] provides a methodology to calculate
The Society of Fire Protection Engineers (SFPE) recently pub- parametric temperature time curves. The formulation is valid for
lished ‘‘SFPE Engineering Standard on Calculating Fire Exposures compartments up to 500 m2 of floor area, without roof openings
to Structures’’ [19]. To predict temperature–time fire curves, the and for a maximum compartment height of 4 m. The gas tempera-
SFPE standard considers two categories of fire exposure: Enclosed ture is a function of fire load density, the floor area and surface area
exposure vs. local exposure. The thermal response should be eval- of the compartment, window sizes, the fire activation risk, fire-
uated based on an enclosed fire exposure if it is determined that fighting measures, and other parameters. In addition, Harmathy
flashover will occur. But, if a post-flashover exposure is not possi- provides a methodology for determining the transition between a
ble, then the local fire exposure methodology should be employed. ventilation and a fuel controlled fire [41]. This is important to
Most critical fires are ventilation-controlled with post-flashover investigate since post flashover fires in typical rooms are ventila-
behavior; therefore two enclosed exposure methodology will be tion controlled. For convenience, a summary of these methodolo-
discussed here. gies is provided in Appendices A and B.
According to [19], two methods are proposed for predicting a The developed probabilistic fire load models (Eqs. (8) and (9))
temperature–time fire curve for enclosed exposure: and the Eurocode parametric time–temperature formulation
(Appendix A, Eq. (A.1)) are used to study the range of maximum
(1) Constant compartment temperature method: ‘‘in this method temperatures and the probability of reaching them in seven differ-
the compartment temperature shall be 1200 °C for all times ent scenarios (shown in Table 8). Also, Harmathy’s formulation
after ignition but before the burnout time. The temperature (Appendix B, Eq. (B.4)) is used to indicate the transition between
shall decrease to ambient condition at a constant rate of fuel and ventilation controlled fires.
7 °C/minute after the burnout time.’’ [19] The scenarios represent realistic floor areas and opening factors
(2) Refined Tanaka Method: this is a method to calculate ventila- in an office building and include small and large office areas and
tion-controlled post-flashover transient fire temperatures. opening factors. Opening factor ‘‘O’’ is calculated according to
EC1 guidelines expressed as:
The constant compartment temperature method is suggested pffiffiffiffiffiffiffi
for design; however, the Refined Tanaka Method is an option for Av heq
O¼ ð12Þ
more detailed analysis. At

(a) (b)
qCorrected (MJ/m2)

qCorrected (MJ/m2)

Fig. 3. Comparison of Culver’s observations (randomly generated) with corrected fire load qcorrected (a) light-weight, and (b) heavy-weight categories.
160 N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165

where Av is the total area of vertical openings (m2), heq is the Table 8
weighted average of window heights on walls (m) and At is the total Description of seven scenarios.

surface area of enclosure (m2). Scenario Af (m2) At (m2) O (m0.5) Af/O (m1.5)
All scenarios are analyzed for both light-weight and heavy- S-1a 24 108 0.0786 305
weight cases, represented by adding letters ‘‘L’’ for light or ‘‘H’’ S-1b 24 108 0.0262 916
for heavy weight categories to designate each scenario shown in S-2a 9 54 0.068 132
Tables 9 and 10 (e.g. S-1a-L or S-1a-H). S-2b 9 54 0.051 176
S-2c 9 54 0.034 265
To perform a probabilistic analysis, fire load density from cur-
S-3a 15 78 0.0725 207
rent codes and standards is corrected using Eqs. (8) and (9) while S-3b 15 78 0.0353 425
the model error re provides the variability in qcorrected. EC1 formu-
lation is then applied to calculate qf,d and qt,d (Appendix A, Eqs.
(A.2) and (A.3)), which are now based on qcorrected, which replaces
where unburned gaseous fuels mix with outside air [23]. In con-
qf,k in the EC1 formulation. It is assumed that the office space has
trast, a fuel controlled fire is similar to burning in the open air
boundary of enclosure with density of 2300 kg/m3, specific heat
which is controlled by surface area of the fuel. Fuel controlled fire
of 880 J/kg K and thermal conductivity of 1.4 W/mk. Also, the office
is common for well-ventilated rooms and fuel items with limited
space has automatic water extinguishing system, automatic fire
area of combustible surface [23]. The probability of reaching a ven-
detection and alarm (by smoke), safe access routes and firefighting
tilation-controlled fire is also shown in Tables 9 and 10 based on
devices.
the fractile of the limiting qcorrected, given the mean and standard
It should be noted that the EC1 requires that qt,d be greater than
deviation of qcorrected. Results in Table 10 show that for heavy-
or equal to 50 MJ/m2, meaning that smaller values of qt,d should not
weight categories, the probability of reaching ventilation-con-
be used. Also, the EC1 limits the time to reach maximum temper-
trolled fire in the compartment for 5 of the 7 scenarios is above
ature (tmax) to 15 min for a relatively fast fire growth rate. These
94%.
two limits imposed by the EC1 are conservatively established for
design purposes but these limits are not considered here since this
study is concerned with evaluating the most realistic fire load den- 6. Proposed probabilistic maximum fire temperature models
sities and gas temperatures.
For design purposes, it is desirable to develop an estimate of the
maximum fire temperature in the compartment based on a proba-
5.3. Results of parametric study bilistic basis. This section provides closed-form solutions that re-
late parameters of the probabilistic distribution of the maximum
For each scenario, Tables 9 and 10 show the results of Monte fire temperature to the compartment geometry and openings. This
Carlo Simulations (MCS) for mean and standard deviation of qcor- makes it possible to estimate the probabilistic distribution of the
rected (Eqs. (8) and (9)) and the maximum temperature in the com-
fire temperature without performing MCS, saving time and compu-
partment Tmax. Results of each scenario are based on 10,000 tational effort that otherwise would be required.
random realizations of e in Eqs. (8) and (9). The average fire load The probabilistic fire load density models presented in Section 4
density values obtained from MCS range from 364 to 782 MJ/m2 do not depend on any code or standard; the models are solely
for light-weight categories, and 550 to 2051 MJ/m2 for heavy- based on survey data. However, sufficient data is not available to
weight categories. This shows how the probabilistic analysis pro- develop a probabilistic maximum fire temperature model without
vides a range of probable fire loads when compared to the design dependency on a code or standard. Therefore the EC1 equations
value of 760 MJ/m2 (qfk,d) from EC1 and 730 MJ/m2 from NFPA 557. that formulate a maximum temperature based on our fire load
The tables also include the limiting values that separate fuel density model (as discussed in Section 5) is used as a basis.
and ventilation controlled fires (Appendix B, Eq. (B.4)). In addition
to room temperatures, it is useful to know the fire burning rate for 6.1. Light-weight categories
structural design under fire [23]. Ventilation vs. fuel controlled
fires would determine the rate of burning. In a ventilation-con- Fig. 5 shows the range of maximum fire temperatures and their
trolled fire, there is insufficient air for the combustible gases to corresponding Cumulative Distribution Function (CDF) for all se-
burn inside the room, therefore it is typically assumed that all glass ven scenarios in the case of light-weight categories. Normal distri-
windows are broken and flames extend outside of the windows bution fits well for all the cases except Scenario 1b where the

Fig. 4. Comparison of test temperatures and constant temperature method: (a) constant 1200 °C, and (b) Refined Tanaka Method (figures from [37]).
N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165 161

Table 9
Results of MCS for qcorrected and Tmax for light-weight categories.

Scenario qcorrected (MJ/m2) Tmax (°C) Limit (ventilation vs. fuel controlled)a Prob. of ventilation control (%)
2
Mean St. Dev. Mean St. Dev. qcorrected (MJ/m ) Tmax (°C)
S-1a-L 364 229 812 79 606 970 12
S-1b-L 365 224 643 108 202 777 77.5
S-2a-L 775 481 857 81 700 953 45.5
S-2b-L 782 477 819 79 525 926 66
S-2c-L 779 487 762 82 350 867 87
S-3a-L 580 363 845 80 646 960 31
S-3b-L 577 364 744 85 315 876 78
a
Values larger than the limiting value indicate a ventilation-controlled fire. Lower values are fuel-controlled.

Table 10
Results of MCS for qcorrected and Tmax for heavy-weight categories.

Scenario qcorrected (MJ/m2) Tmax (°C) Limit (ventilation vs. fuel controlled)a Prob. of ventilation control (%)
2
Mean St. Dev. Mean St. Dev. qcorrected (MJ/m ) Tmax (°C)
S-1a-H 550 327 871 78 606 970 33
S-1b-H 555 331 717 87 202 777 94
S-2a-H 2049 1203 1001 82 700 953 95
S-2b-H 2051 1252 958 82 525 926 98.5
S-2c-H 2021 1182 897 79 350 867 99.8
S-3a-H 1222 733 954 82 646 960 80.5
S-3b-H 1200 717 848 76 315 876 98.5
a
Values larger than the limiting value indicate a ventilation-controlled fire. Lower values are fuel-controlled.

distribution of the data is slightly skewed. However, the deviation all the cases except Scenario 1b. Again, it is assumed that the
from Normal distribution in Scenario 1b is not significant and maximum fire temperature in a compartment follows normal
therefore, it is assumed that the maximum fire temperature in a distribution.
compartment follows normal distribution. Table 12 shows the 95% fractiles for the maximum temperature
Given the mean and standard deviation of the maximum tem- in the compartment given the mean and standard deviation ob-
perature in the compartment (Table 9) and assuming a normal dis- tained from MCS (Table 10) and assuming a normal distribution.
tribution, the 95% fractile for all scenarios are calculated and The mean for heavy-weight categories ranges from 717 to 1001 °C.
presented in Table 11. The results show that the mean of the max- Fig. 8 shows the linear relation between mean and standard
imum temperature in a light-weight compartment ranges from deviation of the maximum temperature with a fitted Normal Dis-
643 to 857 °C. tribution (Table 12) and the factor Af/O. Eqs. (15) and (16) are the
It is desirable to be able to predict parameters of the normal dis- proposed closed-form solutions that can be used to predict normal
tribution without performing MCS. Fig. 6 shows that the parame- distribution parameters (lH and rH) of the maximum fire temper-
ters of the Normal distribution in Table 11 change linearly with ature in a given compartment assuming it belongs to a heavy-
the ratio of floor area to opening factor (Af/O). This way, the param- weight category.
eters are related to the compartment geometry and its openings.
Eqs. (13) and (14) are the proposed closed-form solutions that
can be used to predict normal distribution parameters (lL and
rL) of the maximum fire temperature in a given compartment
assuming it belongs to a light-weight category.

Af
lL ¼ 0:260 þ 873:4 ð13Þ
O

Af
rL ¼ 0:0379 þ 71:6 ð14Þ
O
Table 11 summarizes the calculated lL and rL from the pro-
posed equations and the 95% fractile based on the obtained lL
and rL. Comparison of the results shows that the predictions from
the proposed equations (Eqs. (13) and (14)) are within 6% of the re-
sults obtained from MCS for light-weight categories. These equa-
tions can therefore be used to perform probabilistic analyses
without performing MCS.

6.2. Heavy-weight categories

Fig. 7 shows the range of maximum fire temperatures and their


corresponding CDF of all seven scenarios for heavy-weight catego-
ries. Similar to Section 6.1, Normal distribution has the best fit for Fig. 5. Cumulative distribution function for Tmax (°C) for light-weight categories.
162 N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165

Table 11
Parameters of normal distribution for Tmax in case of light-weight categories.

Scenario Tmax (°C) – MCS Tmax (°C) – Eqs. (13) and (14)
Mean St. dev. 95% fractile Mean St. dev. 95% fractile
lL rL lL rL
S-1a-L 812 79 942 794 83 931
S-1b-L 643 108 821 635 106 809
S-2a-L 857 81 990 839 77 966
S-2b-L 819 79 949 828 78 956
S-2c-L 762 82 897 805 82 940
S-3a-L 845 80 977 820 79 950
S-3b-L 744 85 884 763 88 908

Af
lH ¼ 0:337 þ 1009:1 ð15Þ
O

Af
rH ¼ 0:0064 þ 78:7 ð16Þ
O
Finally, Table 12 compares the calculated lH and rH from the
Fig. 7. Cumulative distribution function for Tmax (°C) for heavy-weight categories.
proposed equations and the 95% fractile based on the obtained
lH and rH. The comparison shows that, on average, the predicted
values from Eqs. (15) and (16) are within 4% of the MCS results. Table 12
Parameters of normal distribution for Tmax in case of heavy-weight categories.

Scenario Tmax (°C) – MCS Tmax (°C) – Eqs. (15) and (16)
7. Current codes and standards compared to proposed models
Mean St. dev. 95% fractile Mean St. dev. 95% fractile
lH rH lH rH
7.1. Eurocode 1, fire load density and gas temperature
S-1a-H 871 78 999 906 80.7 1038
S-1b-H 717 87 860 700 84.6 839
Fig. 3 in Section 4.2 shows that the proposed models for fire S-2a-H 1001 82 1136 965 79.5 1096
load density (Eqs. (8) and (9)) cover a wide range of qcorrected values S-2b-H 958 82 1093 950 79.8 1081
up to approximately 1000 MJ/m2 for light-weight categories and S-2c-H 897 79 1027 920 80.4 1052
2500 MJ/m2 for heavy-weight categories. On the other hand, the S-3a-H 954 82 1089 939 80.0 1071
S-3b-H 848 76 973 866 81.4 1000
95% fractile characteristic fire load density in EC1 is calculated to
be 760 MJ/m2. The comparison confirms that EC1 may be conserva-
tive for light-weight categories and underestimate fire load density
for heavy-weight categories and that better predictions can be developed in this paper provides a more realistic range of values
made. for each scenario given the room use.
Table 13 summarizes the 95% fractiles of the maximum fire
temperatures based on the proposed equations in Section 6 (Eqs.
(13)–(16)) and compares them to the EC1 design temperature 7.2. NFPA 557 and fire load density
(Appendix A, Eq. (A.1)). The EC1 values are calculated assuming a
medium fire growth rate (tlim = 20 min). The comparison shows Table 14 compares the mean and standard deviation of fire load
that the 95% fractiles range from 809 to 966 °C for light-weight cat- density from NFPA 557, Culver’s data and the 7 scenario probabilis-
egories, 839 to 1096 °C for heavy-weight categories, and 785 to tic analysis in Section 5. As discussed in Section 3.2, NFPA 557
1019 °C based on EC1 formulation. This range of temperatures specified values are based on Culver’s data and therefore very sim-
indicates that the proposed equations are offset higher compared ilar to each other. In comparison, the generated fire load densities
to those of the EC1 design values. For four scenarios, the EC1 values for light and heavy weigh categories for seven scenarios cover a
are smaller than that predicted by the proposed equations. In wide range with smaller mean values for light-weight and larger
addition, the probabilistic approach with the fire load equations mean values for heavy-weight categories, as expected. NFPA and

900 120
-Normal Distribution

-Normal Distribution

800 L = 0.0379Af /O + 71.6


100
700
600 µL = -0.260Af /O + 873.4
80
500
60
400
300 40
200
20
100 (a) (b)
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Af /O (m1.5) Af /O (m1.5)

Fig. 6. Normal distribution parameters vs. Af/O ratio (m1.5) for light-weight categories.
N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165 163

1200 100

σ - Normal Distribution
μ - Normal Distribution
1000 80
σH = 0.0064Af /O + 78.7
800
μH = -0.337Af /O + 1009.1 60
600
40
400

200 20

0
(a) (b)
0
0 200 400 600 800 1000 0 200 400 600 800 1000
Af /O (m1.5) Af /O (m1.5)

Fig. 8. Normal distribution parameter vs. Af/O ratio (m1.5) for heavy-weight categories.

Table 13 load density, and strong correlation was observed with compart-
Comparison of maximum gas temperatures (°C) from the proposed models and EC1.
ment area, and compartment use. Surveys indicate that the larger
Scenario Tmax (°C) – 95% fractile EC1 design value the compartment area the smaller the fire load density. Further,
Light-weight Heavy-weight ‘‘heavy-weight’’ compartment use (such as those categorized as
S-1a 931 1038 1019
file, storage, and library) had a mean fire load density that was
S-1b 809 839 785 on the order of two times larger than ‘‘light-weight’’ compartment
S-2a 966 1096 976 use (such as office, clerical, lobby, and conference). However, cur-
S-2b 956 1081 888 rent codes and standards do not account for these variables (area
S-2c 940 1052 782
and use) and specify constant values for fire load density.
S-3a 950 1071 995
S-3b 908 1000 804 This survey data for fire load density was compared to older
data [5], to design values suggested by the Eurocode [16], and to
a recent NFPA 557 Standard [17]. Examining the mean and stan-
dard deviation, the older data documented by Culver in 1975 [5]
Table 14
Comparison of fire load densities (MJ/m2) from the NFPA 557, Culver and seven
compared reasonably well to data of the last 10 years (taken in
scenarios. Canada and Switzerland). The values provided by the Eurocode
[16] and NFPA 557 [17] were shown to be conservative compared
Source Mean St. dev.
to data for ‘‘light-weight’’ use, and non-conservative compared to
NFPA 557 [17] 730 502 data for ‘‘heavy-weight’’ use.
Culver’s data [6] Since Culver [5] presented the most comprehensive survey data,
Government Bldgs 641 641 and the mean and standard deviation was within range of more re-
Private Bldgs 720 562
cent surveys, Culver data was used to develop probabilistic models
7 Scenarios to predict the fire load density in office buildings (one for light
Light-weight 364–782 224–487
Heavy-weight 550–2051 327–1252
weight and one for heavy weight use). The proposed models con-
sider (1) the size of the compartment and (2) room use by distin-
guishing between light-weight and heavy-weight as defined
above. By including these two parameters good correlation was ob-
Culvers’ statistics, both fall within the range of predictions from served between the survey data and the models. The models fit the
the probabilistic analysis performed in Section 5. data better compared to Eurocode and NFPA 557 (neither of which
considers area and use for calculating fire load density).
7.3. SFPE and gas temperature These models for fire load density were then used to develop
the gas temperature–time evolution (based on Eurocode) in a given
The conditions and compartment properties from which the re- compartment. Several scenarios with different floor areas and
ported SFPE test temperatures were obtained (Fig. 4) are not openings were defined and the fire load density models were used
known to the authors so a direct comparison of test data to the in Monte Carlo Simulations (MCS) to investigate the range of pos-
proposed equations cannot be made. However, some general sible fire temperatures and their corresponding probabilities. Using
observations indicate that the proposed equations predict reason- the results of these MCS, normal distribution parameters were de-
able maximum temperatures. For example, Fig. 4, in Section 5.1, rived to predict a normal distribution for the maximum gas tem-
shows that the maximum average test temperature ranges from perature based on the floor area and opening factor. Excellent
as low as 240 °C up to slightly higher than 1200 °C. On the other correlation was observed between this closed-form normal distri-
hand, the probabilistic study generated temperatures in the range bution model and the MCS. Further, the model was compared to
of approximately 350 °C up to 1200 °C considering all seven sce- a recent SFPE publication [19] where two models are proposed
narios (Figs. 5 and 7). This comparison indicates that the range of for predicting the maximum temperature for enclosed exposure:
values obtained from the proposed fire load models resembles Refined Tanaka Method and the constant compartment tempera-
the test data. In addition, the comparison shows that the range of ture method. It was found that the proposed maximum tempera-
obtained temperatures is similar to that predicted by the Refined ture model results in a range of temperatures that correlates well
Tanaka Method (Fig. 4b). with the test data and the Refined Tanaka Method.
Both the fire load density and maximum temperature probabi-
8. Summary and conclusions listic models provide realistic and reasonable predictions for fire
load density and maximum temperature. These models are well
This paper provided a literature review of recent fire load den- suited for application in a probabilistic performance-based
sity surveys. The surveys indicated a large range of values for fire approach to fire design, which can lead to improved fire safety,
164 N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165

efficiency, and economy through a more realistic assessment of the ventilation factor and thermal insulation in a different way. The
expected fire load in an office building. resulting design decay rate is given by:
  sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
dT dT O=0:04
Appendix A. Eurocode 1 parametric fire curves ¼ ðA:4Þ
dt dt ref b=1900
Eurocode 1 Annex 1 [16] defines the fire temperature–time where b is the square root of thermal inertia for boundary of
curves in a compartment as: enclosure.
  Based on the ISO 834 testing standard, the Eurocode uses a ref-
hg ¼ 20 þ 1325ð1  0:324e0:2t  0:204e1:7t
erence decay rate (dT/dt) equal to 625 °C per hour for fires with a

 0:472e19t Þ ðA:1Þ burning period less than half an hour, decreasing to 250 °C per
hour for fires with a burning period greater than 2 h [23].
where
hg = gas temperature in the fire compartment (°C),
Appendix B. Fuel-controlled vs. ventilation-controlled fires
t* = t  C (h),
t = time (h),
In general, the rate of burning depends on the opening size and
C = is a function of opening factor and properties of boundary of
surface area of the fuel. These two factors determine the transition
enclosure.
between ventilation and fuel-controlled fires. In the case of a fully
developed fire, if the air supply from openings is not sufficient, the
The maximum hg in the heating phase occurs when t  ¼ tmax
rate of burning will depend on the opening size (ventilation-con-
where:
trolled). If the air supply is sufficient, the rate of burning will de-
t max ¼ t max C ðhÞ pend on the surface area and burning characteristics of the fuel.
Based on Ma and Makelainen study [41] Harmathy developed a
with t max ¼ ð0:2  103  qt;d =O; t lim Þ
semi-empirical equation to determine the transition region of
where compartment fires:
qt,d = the design value for the fire load density related to total pffiffiffi
qa g 1=2 Aw h
surface area At of the enclosure, whereby qt,d = qf,d  Af/At (MJ/ ¼ 0:263 ðB:1Þ
Afuel
m2), with limits as 50 6 qt,d 6 1000 (MJ/m2) (Eq. (A.2)),
qf,d = the design value of the fire load density related to the floor qa = density of air (kg/m3),
area Af of the floor, taken from Annex E of Eurocode 1 (MJ/m2). g = acceleration due to gravity (m/s2),
(Eq. (A.3)), Aw = area of windows (m2),
tlim = 25 (min) for a slow fire growth rate, 20 (min) for a med- h = height of windows (m),
ium fire growth rate, and 15 (min) for a fast fire growth rate, Afuel = surface area of fuel bed (m2).
in (hours).
The surface area of fuel bed is hard to estimate; therefore, Afuel is
Also Annex E defines: expressed as:

qt;d ¼ qf ;d  Af =AT ðMJ=m2 Þ ðA:2Þ Afuel ¼ uG0 ðB:2Þ

where G0 is the total fire load of equivalent woods (kg), and u is the
qf ;d ¼ qf ;k  m  dq1  dq2  dn ðMJ=m2 Þ ðA:3Þ surface area ratio of fuel (m2/kg). Based on experience, the value of
u is between 0.1 and 0.4 m2/kg with most common values of 0.12
Af = the floor area of the compartment (m2),
and 0.18 for general furniture.
m = the combustion factor,
Taking the values of qa = 1.205 kg/m3 as density of air and
qf,k = the characteristic fire load density per unit floor area (MJ/
g = 9.81 m/s2 as the acceleration due to gravity, the resulting equa-
m2),
tion will be:
qf,d = the design value of the fire load density based on floor area
(MJ/m2), O=q ¼ 0:07ju ðB:3Þ
qt,d = the design value of the fire load density based on the total
O = opening factor,
compartment area (MJ/m2),
q = fire load density per unit floor area (kg of wood/m2),
dq1 = factor taking into account the fire activation risk due to the
j = ratio of floor area to total surface area.
size of compartment
Finally, based on Eq. (B.3):
dq2 = factor taking into account the fire activation risk due to the
( )
type of occupancy if q > 0:07Oju ventilation-controlled
dn ¼ P10 ðB:4Þ
i¼1 dni is a factor taking into account the different active if q < 0:07Oju fuel-controlled
firefighting measures (i.e. sprinkler, detection, automatic alarm
transmission, firefighter access and location, etc.) Eq. (B.4) can be used to estimate the fire load q (kg of wood/m2)
at which a transition between fuel-controlled and ventilation con-
In Eurocode, as shown in Eq. (A.1), the reference temperature is trolled fire will occur. The fire load q in kg of wood/m2 can be con-
related to a fictitious time. Feasey and Buchanan [42] indicate that verted to MJ/m2 assuming a caloric value of 18 MJ/kg.
the fictitious time that is derived to be used during the burning
period has not been justified for use in the decay period and the References
outcome of this model gives unsatisfactory results with extremely
fast decay rates for large openings in well-insulated compartments [1] Hurley M. Design fire scenarios. In: Proceedings of NIST-SFPE workshop for
and extremely slow decay rates for small openings in poorly development of a national R&D roadmap for fire safety design and retrofit of
structures. NISTIR 7133:107-117; 2004.
insulated compartments. Feasey and Buchanan [42] demonstrate [2] National Fire Protection Association (NFPA). Validation of methodologies to
that it is more accurate to modify the reference decay rate for determine fire load for use in structural fire protection – final report. The fire
N. Elhami Khorasani et al. / Engineering Structures 58 (2014) 152–165 165

protection research foundation – research in support of the NFPA mission. [21] Baldwin R, Law M, Allen G, Griffiths LG. Survey of fire-loads in modern office
Carleton University; May 2011. buildings – some preliminary results. Fire Research Note #808. Borehamwood,
[3] VKF and AEAI. Note Explicative De Protection Incendie – Evaluation en vue de England: Fire Research Station; 1970.
la determination de la grandeur des compartiments coupe-feu. Vereinigung [22] Choi ECC. Data structure and data processing procedures for live loads and fire
Kantonaler Feuerversicherungen (VKF) and Association des établissements loads in office buildings. Technical record 524. National Building Technology
cantonaux d’assurance incendie (AEAI); 2007. Centre; 1988.
[4] Kumar S, Kameswara Rao CVS. Fire loads in office buildings. J Struct Eng [23] Buchanan AH. Structural design for fire safety. Chichester: John Wiley & Sons
1997;123(3). Inc.; 2001.
[5] Culver C. Survey results for fire loads and live loads in office [24] Babrauskas V. Burning rates. SFPE handbook of fire protection
buildings. Washington, DC: National Bureau of Standards; 1976. engineering. USA: Society of Fire Protection Engineers; 1995 [Chapter 3-1].
[6] Culver C. Characteristics of fire loads in office buildings. J Fire Technol [25] Parker WJ, Tran HC. Wood materials. In: Babrauskas V, Grayson SJ, editors.
1978;14(1):51–60. Heat release in fires. Elsevier Applied Science; 1992 [Chapter 11].
[7] CIB W14 Workshop Report. Design guide – structural fire safety. Fire Saf J [26] NFPA, Fire protection handbook, 18th ed. Quincy, MA, USA: National Fire
1986;10(2):77–137. Protection Association; 1997.
[8] Pettersson O, Magnusson SE, Thor J. Fire engineering design of steel [27] Caro TC, Milke JA. A survey of fuel loads in contemporary office building. NIST-
structures. Stockholm: Swedish Institute of Steel Construction; 1976 GCR-96-697; September 1996.
(Swedish Edition 1974). [28] Bukowski RW. Determining design fires for design-level and extreme events.
[9] European Recommendations for the fire Safety of Steel Structures. Fire safety In: SFPE 6th international conference on performance-based codes and fire
of steel structures. Fire exposure, Technical Committee 3, European safety design methods, Tokyo; 2006.
Convention for Constructional Steelwork, Brussels; July 1981 [chapter 2]. [29] Ingberg SH, Dunham JW, Thompson JP. Combustible contents in
[10] Bryl S. Brandbelastung in Hochbau. Schweizerische Bauzeitung, special reprint buildings. Washington, DC: National Bureau of Standards, U.S. Dept. of
from: 93, Jahrgang, Heft 17; 1975. Commerce; 1957.
[11] Bryl S. Brandbelastung im Stahlbau, Teil III, Brandbelastung in Burogebauden, [30] Caro T, Milke J. A survey of furl loads in contemporary office buildings. NIST
ECCS-III-74-2-D. Rotterdam: European Convention for Constructional report GCR-96-697. Gaithersburg, MD: NIST; 1996.
Steelwork; 1974. [31] Green M. A survey of fire loads in hackney hospital. J Fire Technol February
[12] Bonetti M, Kree P, Kruppa J. Estimation des Charges Mobilieres d’Incendie dans 1977;13(1):42–52.
les Immeubles a Usage de Bureaus, Construction Metallique, No. 3, Centre [32] Barnett CR. Pilot fire load survey. Project #3580 CRB, McDonald Barnett
Technical Industriel de la Construction Metallique (C.T.I.C.M.); September Partners, Auckland, NZ; 1984.
1975. [33] Korpela K, Keski-Rahkonen O. Fire loads in office buildings. In: Proceedings –
[13] Combessis JC, Fauconnier R, Cluzel D. Enquetes et Charges d’Incendie; 3rd international conference on performance-based codes and fire safety
Etablissements Recevant du Public, Charges Indendie Coubes, Temperature/ design methods, society of fire protection engineers, Bethesda, MD; 2000.
Temps Correspondantes, Comande D.S.C. No. 005110. Institute Technique du [34] Culver C, Kushner J. A program for survey or fire loads and live loads in office
Batiment et des Travaux Publics; September 1983. buildings. NBS technical note 858. Gaithersburg, MD: National Bureau of
[14] Beilage 2 der SIA-Dokumentation 81. Brandrisikobewertung/ Standards; 1975.
Berechnungsvertahren SIA. Schweizerischer Ingenieur –und [35] Thauvoye C, Zhao B, Klein J, Fontana M. Fire load survey and statistical
Architektenverein, Zurich; 1984. analysis. Fire Saf Sci 2008;9:991–1002.
[15] Campbell JA. Confinement of fire in buildings. Fire protection handbook. [36] Tewarson A. Generation of heat and gaseous, liquid, and solid products in fires.
National Fire Protection Association; 1981. The SFPE handbook of fire protection engineering, 4th ed. Quincy, MA; 2008.
[16] European Committee for Standardization (CEN). Actions on structures – Part [37] Society of Fire Protection Engineers (SFPE). Evaluation of enclosure
1–2: General actions – actions on structures exposed to fire, Eurocode 1, temperature empirical models, SFPE S.01; 2011.
Brussels, Belgium; 2002. [38] Box GEP, d Cox DR. An analysis of transformations (with discussion). J R Stat
[17] National Fire Protection Association (NFPA). NFPA 557 standard for Soc Ser B 1964;26:211–52.
determination of fire loads for use in structural fire protection design, MA; [39] Rao CR, Toutenburg H. Linear models, least squares and alternatives. New
Ed. 2012. York: Springer; 1997.
[18] Gardoni P, Der Kiureghian A, Mosalam KM. Probabilistic capacity models and [40] Ang AH-S, Tang WH. Probabilistic concepts in engineering – emphasis on
fragility estimates for reinforced concrete columns based on experimental applications to civil and environmental engineering. John Wiley & Sons Inc.;
observations. J Eng Mech 2002;128(10). 2007.
[19] Society of Fire Protection Engineers (SFPE). SFPE engineering standard on [41] Ma Z, Makelainen P. Parametric temperature–time curves of medium
calculating fire exposures to structures. SFPE technical, report; November compartment fires for structural design. Fire Saf J 2000;34:361–75.
2010. [42] Feasey R, Buchanan A. Post-flashover fires for structural design. Fire Saf J
[20] ATC. Engineering design parameters for structural framing systems. ATC-58 2002;37(1).
Project task report, Phase 2, Task 2.2; 2004.

You might also like