Beruflich Dokumente
Kultur Dokumente
0, v < v
crit
g(v), v
crit
v < v
tot
1, v
tot
v.
(1)
The qualitative trend of G for a single building is plotted in Fig. 1.
The total monetary loss is obtained by multiplying G with the total
value or the reconstruction costs W of the building.
Summation of all single monetary losses results in the total loss
of a number of N buildings.
Total loss = loss
1
+loss
2
+ +loss
N
= G
1
W
1
+G
2
W
2
+ +G
n
W
n
=
N
i=1
G
i
(v
i
)W
i
. (2)
For an exact solution of this equation one would have to know
the function G with its variables v
crit
, v
tot
and g(v), the maximum
wind speed v as well as the total value W for every building(!).
It is theoretically possible with a certain effort to determine
W and v for every building. The damage functions G would
have to be determined for every single building by deterministic
approaches [28,20,21,6] which is practically impossible due to
the large amount of buildings and the great diversity of building
structures.
However, known values are the number of buildings N within a
postal-code zone or municipality, the total value W within this
area and wind speeds v during storm events on a 1 km 1 km
raster by numerical wind field simulation.
P. Heneka, B. Ruck / Engineering Structures 30 (2008) 36033609 3605
Fig. 2. Built-up areas with the borders of a zip code zone or a municipality.
2.2. Model derivation
Due to the lack of data, assumptions have to be made to
construct a fully applicable damage model. As the distribution of
buildings and damages within a spatial unit is unknown, the model
is used to calculate the total number and amount of damage in a
spatial unit. The mainassumptions for the derivationof the damage
model are listed below:
1. All buildings have an equal value W = W/N. Consequently,
the weight of the damage ratios for the summation is equal.
2. The calculated wind speeds are averaged over the built-upareas
of the spatial unit and are therefore applied for all buildings
(Fig. 2).
3. Critical and total wind speeds of all buildings within a stock
(postal-code zone, municipality) can be described with suitable
probability distribution functions f (v
crit
) and f (v
tot
). Generally,
every function can be used.
4. The function for damage increase g(v) is valid for all buildings
within an area and represents therefore an averaged damage
propagation.
The distribution function for critical wind speeds f (v
crit
)
simulates the inhomogeneities of the buildings resistance to
wind loads. Damage will generally not start at a common wind
speed. Moreover, some buildings will suffer at lower wind speeds
than others and some will resist even very high wind speeds.
For the same purpose, Sill and Kozlowski [25] are using a
triangular function over the square of the wind speeds. In Fig. 3,
a distribution function for the critical wind speeds is plotted for
example. Integration of the function f (v
crit
) with the upper limit
v (maximum wind speed) results directly in the ratio of affected
buildings CR (claim ratio) for which the critical wind speed was
exceeded and damage occurred.
CR(v) =
v
0
f (v
crit
) dv
crit
. (3)
The integral is equal to the cumulative density function of f (v
crit
)
and plotted as the grey area in Fig. 3.
The damage ratio of buildings in an area is calculated as follows
and illustrated in a graphical way in Fig. 4. A small proportion
of buildings f
1
with high vulnerability suffers high damage ratios
G
1
up to the wind speed v. A larger proportion of buildings f
n
is damaged near the maximum wind speed v and each of those
therefore suffers a smaller damage ratio G
n
. Summation of these
proportions yields
DR(v) = DR
1
+DR
2
+ +DR
n
= f
1
G
1
+f
2
G
2
+ +f
n
G
n
=
n
i=1
f
i
G
i
. (4)
Fig. 3. Probability density function of critical wind speeds.
Fig. 4. Graphical presentation of damage calculation.
For n , one receives the integral for the calculation of the
damage ratio DR for a number of buildings:
DR(v) =
v
0
f (v
crit
) G(v) dv
crit
. (5)
Consequently, Eqs. (3) and (5) are used to calculate damage and
claim ratio of a building stock when hit by a wind speed v during a
stormevent. Upto this point, the derivationof the damage function
was kept as general as possible to ensure a maximum of choices
for the probability distribution functions f (v
crit
) and f (v
tot
) as well
as for the damage increase functions g(v). Within the presented
mathematical framework, all types are possible and the following
considerations will help to choose suitable functions.
Distribution function for critical wind speeds
The critical wind speed of a building is determined by
the weakest construction detail and, therefore, depends on the
construction design of the structure [20]. Additional factors like
construction quality, age and maintenance also play an important
but unknown role. We will have to face a large spreading of
critical wind speeds in a building stock which shall be simulated
by means of a suitable distribution function.
As there is no further evidence about the shape of the
distribution function, we use the Normal distribution function
with the shape parameters
crit
and
crit
to describe this variation.
Introduction of the Normal function in Eq. (3) and multiplication
with the total number of buildings N results in the number of
damaged buildings at a maximum wind speed v
Buildings(v) = N
crit
2
exp
(v
crit
crit
)
2
2
2
crit
dv
crit
.(6)
Hence, the number of damaged buildings is solely described by the
shape parameters
crit
and
crit
.
Damage increase function for buildings
It is often discussed whether the increase of damage with
higher wind speeds can be based on physical assumptions instead
of pure empirical evidence [15,14]. From investigation of the
3606 P. Heneka, B. Ruck / Engineering Structures 30 (2008) 36033609
European 1990 stormseries [18] a power 3 relationship was found
(damage v
3
) which was traced back to the fact that the
available kinetic energy of a flow is a cubic function of wind
speed. This may be true for observations of single buildings but
not for a building stock where the increased monetary damage
is mainly a result of the increased amount of damaged buildings
and not of a significantly higher average damage. In the proposed
model, these 2 mechanisms are treated separately; the damage
increase denotes this of a single average building while the above
mentioned distribution of critical wind speeds is valid for an
amount of buildings.
The damage increase function is assumed to be proportional
to the power of a wind speed and writes with the boundary
conditions g(v
crit
) = 0 and g(v
tot
) = 1 as:
g(v) =
v v
crit
v
tot
v
crit
. (7)
Here, for = 2 the average damage of a single building would be
proportional to the wind force, for = 3 to the kinetic energy of
the wind flow.
Distribution function for total wind speeds
Analogous to the critical wind speeds, the total wind speeds
are also described by a distribution function. Generally, these two
functions will not be independent fromeach other (correlated with
correlation coefficient ) and can be written as a 2-dimensional
probability distribution function f (v
crit
, v
tot
, ). For reasons of
simplicity and no further evidence, we assume that the total wind
speeds also follow a Normal distribution and are fully correlated
( = 1) with the critical wind speeds. Consequently, v
tot
for every
single building can be written as
v
tot
= v
crit
+v. (8)
As a result, the difference between total and critical wind speed is
constant for all buildings within an area. Substitution of Eqs. (7)
and (8) in (5) and multiplication with the total building value W
results in the damage at a given wind speed v
Damage(v)
= W
v v
crit
v
crit
2
exp
(v
crit
crit
)
2
2
2
crit
dv
crit
. (9)
Besides the shape parameters
crit
and
crit
, damage is additionally
dependent on v and .
Again, the damage increase function and distribution functions
chosenfor v
crit
andv
tot
are relatively simple approaches and, hence,
subject to inaccuracies. However, within the proposed model
equations (3) and (5), any function can be used.
2.3. Implementation of further parameters
To describe damage in more detail, it would be desirable to
have different damage functions for different types of buildings or
different exposures. The influence of these additional parameters
has to be quantified. For example, Sill and Kozlowski [25] and
Khanduri andMorrow[13] proposedadditional curves for different
types of buildings, Schraft et al. [24] for storm duration.
In our model,
crit
,
crit
, and v will have to be evaluated to
best fit the damage data available which is classified by additional
parameters.
2.4. Uncertainties of damage assessment
The analysis of past storm events show a large variation of
damages when plotted against wind speed which is not entirely
Fig. 5. Illustration of uncertainty of damage functions.
explainable with the available information [4,11]. These variations
have to be treated as random variation. It is therefore necessary
to at least quantify these variations in order to obtain an estimate
of the uncertainty of the damage assessment. The introduction of
artificial uncertainty is a way of dealing with this problem in a
practical sense. This is realised by a simple add-on to the model
which consists of a random variation of the shape parameter
crit
following a distribution f (
crit
). Here, also a Normal distribution
is used with the mean
crit
=
crit
and standard deviation
crit
.
With Monte Carlo simulations for each wind speed a variation for
the damage is obtained (Fig. 5). The determination of
crit
was
performed by meeting the percentiles of the modelled damage and
the observed damage, e.g. the 16th and 84th percentile.
2.5. Remarks to the model parameters
The parameters
crit
,
crit
, andv are neededinorder to apply
the model. Generally, there are two ways of obtaining these values:
(1) determination of the parameters in an analytical way and (2)
calibration to available wind speed and storm damage data of past
storm events. In our case damage and wind speed data of 4 storm
events are available. We therefore use both approaches: After the
determination of reasonable starting values for the 4 parameters,
the calibration procedure was performed as follows in order to
minimize the difference between calculated and observed total
building damage:
Calculationof number of buildings anddamage for every postal-
code zone and storm.
Summation of results for all postal-code zones for every storm.
Calculation of difference to observed overall damage.
The mean of the normal distribution
crit
is equal to the
wind speed where exactly 50% of the buildings in an area are
damaged. Hence, given a suitable observation database of past
building damages, this value can be directly determined. Sill
and Kozlowski [25] propose 48 m/s for this wind speed; from
publications of the Munich Re [19] a value of 4550 m/s is read.
The standard deviation
crit
can roughly be estimated based on the
following thoughts: First significant damage to buildings occurs at
wind speeds approximately 2025 m/s. Taking into account that
for a normal distribution of critical wind speeds, approx. 0.1% of
the values lay below
crit
3
crit
, we deduce values of 710 m/s as
a starting estimate for the standard deviation.
The mean of the total wind speeds
tot
represents the wind
speed where 50% of the buildings suffer total damage (DR = 1).
Wind speeds of this magnitude are not expected for winter storms
in Germany and are even rarely observed in tornadoes. However,
an approach to obtaining this wind speed is a comparison with
F5 tornado damage where values of 120130 m/s are observed
P. Heneka, B. Ruck / Engineering Structures 30 (2008) 36033609 3607
Fig. 6. Damage model 1 (left) and 2 (right) and observed storm damage.
Table 2
Suggested parameters for residential buildings in Germany
Parameter Description Model 1 Model 2
Absolute values Relative values
crit
Wind speed where 50% of the buildings are damaged 50.5 m/s ( = 2.5) 1.31 ( = 0.04)
crit
Standard deviation of critical wind speed distribution 7.8 m/s 0.20
Increase of damage with wind speed 2 2
v Difference in wind speed between start of damage and total damage 70 m/s 1.85
for these damage states [5]. Taking these values a first rough
estimate for
tot
, the difference v between v
tot
and v
crit
calculates
7090 m/s. There is also no evidence about the shape parameter
for the damage increase function but, for reasons of model
consistency, the above mentioned values indicate that = 2. A
higher would have the consequence that the mean value for the
total damage distribution
tot
is expected already at wind speeds
of 70 m/s which is not a realistic value for the German building
stock.
Summarizing, with knowledge of building damage and corre-
sponding wind speeds, starting values of the model parameters can
be determined in a rational way.
3. Application to winter storms in Germany
The model is used to calculate the number and monetary
amount of damages caused by winter storms in Germany. Damage
data on a postal code base are available for 4 storm events with
different severities for the German state of BadenWrttemberg
situated in the South-West of Germany. The maximum gust fields
of these past events have been calculated in a raster with a
horizontal resolution of 1 km 1 km by the IMK, University of
Karlsruhe [10].
In a previous publication [11] it was found that the number
and monetary amount of damage is rather a function of a relative
wind than of the absolute wind speeds which occur during storm
events. The relative wind speed is hereby defined as the ratio of
the maximum gust during a storm event and the 50-year wind
gust. The 50-year wind speed is denoted as wind climate and is
exceeded with a probability of 2% in a time period of one year.
This consideration of wind climate has the effect that wind damage
occurs solely if the local wind climate is exceeded. In regions
with high wind climate the building structures are used to also
withstand higher gusts, thus damage is not as high as in other
regions with lower wind climate.
In order to show the consequences for the damage calculation,
we use both approaches to run the model. Model 1 is run with
absolute wind speeds, model 2 with relative wind speeds. For the
available damage and wind data, the parameters suggested for the
damage model for residential buildings in Germany are listed in
Table 2. The values fit very well to the theoretical considerations of
Section 2.5.
Both the damage data as well as the damage model are plotted
in Fig. 6, for the two approaches, respectively. The plots show the
damage ratio (above) and the claim ratio (below) in respect to
the absolute gust speed (left) and relative gust speed (right). The
dots show wind speeds and damage for single postal-code zones.
Bothmodels represent very well the meanincrease of damage with
wind speed.
As the data points scatter very much, anuncertainty assessment
is given. In each case, 16% of the data points lay below and above
the slash-dotted curves and consequently more than 2/3 of the
data points lay between these borders. The size of the area is
therefore a measure for the uncertainty of the model; the smaller
the better the model is capable of explaining the data. At this point
it becomes clear that the damage model is solely representing the
average damage trend without taking into account further local
differences or uncertainties.
The overall damage is calculated by the summation of the
number and amount of damage for all postal-code zones in the
state andcomparedwiththe observeddamage (Table 3). The errors
range from 0% to 40% for the number of affected buildings and
3608 P. Heneka, B. Ruck / Engineering Structures 30 (2008) 36033609
Table 3
Observation and model calculations for four past winter storm events
Storm Storm 1986 Wiebke Lore Lothar
Date 20.10.1986 01.03.1990 28.01.1994 26.12.1999
Observed buildings 3 34 17 197
(Thousand)
Simulated buildings 3 40 22 175
Model 1 (Thousand) (21%) (+18%) (+26%) (10%)
Simulated buildings 2 34 19 181
Model 2 (Thousand) (41%) (+0%) (+12%) (7%)
Observed damage 5 51 24 304
(Mil. e)
Simulated damage 2 61 32 278
Model 1 (Mil. e) (54%) (+21%) (+37%) (10%)
Simulated damage 1 45 26 258
Model 2 (Mil. e) (71%) (11%) (+10%) (15%)
Table 4
Correlation coefficients for storm damage and number of damaged buildings
Name Sturm 1986 Wiebke Lore Lothar
Date 20.10.1986 28.2.-1.3.1990 28.1.1994 26.12.1999
Damage model 1
Damage 0.19 0.28 0.13 0.61
Buildings 0.21 0.38 0.23 0.75
Damage model 2
Damage 0.25 0.57 0.34 0.87
Buildings 0.28 0.64 0.47 0.86
from 10% to 70% for the monetary damage. Ignoring the weakest
event of 1986 which generally caused little damage, model 2
is better capable to calculate storm damage of the 3 strongest
events. Especially the number of affected buildings is sufficiently
reproduced with a maximum error of 7% for the 1999 event. This
holds also for the monetary damage where the maximum error is
15% compared to 40% for model 1.
The reasons for the insufficient calculation of the 1986 event
is of a general nature: At lower wind speeds where only very
few damage per spatial unit is observed, an inaccurate calculation
results easily in large errors.
A comparison of the spatial distributions of storm damage for
winter storm Lothar is shown in Fig. 7. The amount of monetary
damage per postal-code zone is represented by black bars and
the average wind speeds are marked in colour. By eye, a better
correspondence of model 2 to the observed damage patterns is
visible, especially in the western, southern and central parts of the
state. This impression is proven by the calculation of the spatial
correlation coefficients where we get 0.87 for model 2 and 0.61
for model 1. For the number of affected buildings and all the other
winter storms, the correlation coefficients are listed in Table 4. It
is obvious that model 2 reproduces the spatial distribution for all
events much better than model 1.
4. Conclusions
The most important conclusions of this work are listed below:
A damage model was developed which is embedded in a
mathematical framework and which offers space for arbitrary
improvements. It is nowpossible to run this model with a mini-
mumof available data. The advantage of the model compared to
empirical damage functions is that the extrapolation for higher
wind speeds is based on a set of logical assumptions instead of
best-fit functions.
The model was calibrated with damages of past storm events
in Germany and was able to reproduce the damage numbers. It
is shown that damage functions based on relative gust speeds
Fig. 7. Residential building damage per postal-code zone in Mil. e for winter storm
Lothar in 1999.
(model 2) are better capable to calculate damage than those
based on absolute gust speeds (model 1) as the precision
P. Heneka, B. Ruck / Engineering Structures 30 (2008) 36033609 3609
and especially the spatial correlation coefficients are higher
throughout.
The developed storm damage model represents the average
winddamage relationship of residential buildings within
postal-code zones or equivalent spatial areas. However, the
discrepancy between modelled and observed damage for single
zones may be huge as at this stage no further influences like
building structure, age or storm duration are considered.
The derived set of model parameters is used to recalculate
the damage of all storm events that have been available.
Changes of the model parameters result in different overall
damage predictions while the spatial correlation coefficients
are mainly insensitive to changes. However, compared to the
general uncertainty of the model, the sensitivity of the results
to changes of the model parameters is low. This indicates that
there should be no larger problems to determine the model
parameters to further building data.
Although calibrated to data of a specific region, model 2 with
its model parameters is applicable also elsewhere as it uses the
wind climate as a proxy for local vulnerability of buildings (see
also [11]). This holds only for regions where the construction
of buildings is similar to the calibration region which is at least
true for Germany andWesternEurope. However, due to missing
damage data of other regions, this applicability could not yet be
finally proven.
Acknowledgements
The authors would like to thank the SV Gebudeversicherung
Stuttgart, Germany, for the provision of stormdamage data and the
Institute for Meteorology and Climate Research of the University of
Karlsruhe for the provisionof meteorological data. This work is part
of the project Risk map Germany of the Center for Disaster Man-
agement and Risk Reduction Technology (http://www.cedim.de), a
joint venture of the GeoForschungsZentrumPotsdam(GFZ) andthe
Technical University of Karlsruhe (TH). We thank the GFZ Potsdam
and the University of Karlsruhe for financial support.
References
[1] Blackmore P, Tsokri E. Windstorm damage to buildings and structures in the
UK during 2002. Weather 2004;59(12):3369.
[2] Blong R. A new damage index. Nat Hazards 2003;30:123.
[3] Buller P. Wind damage to buildings in the United Kingdom 19701976.
Technical report CP42/78. Garston: Building Research Establishment; 1978.
[4] Dorland C, Tol R, Palutikof J. Vulnerability of the Netherlands and Northwest
Europe to storm damage under climate change. Climatic Change 1999;43:
51335.
[5] Dotzek N. Tornadoes in Germany. Atmos Res 2001;56:23351.
[6] Ellingwood B, Rosowsky D, Li Y, Kim J. Fragility assessment of light-frame
wood construction subjected to wind and earthquake hazards. J Struct Eng
2004;30:192130.
[7] Fujita TT. Tornadoes and downbursts in the context of generalized planetary
scales. J Atmos Sci 1981;38:151134.
[8] Hart G. Natural hazards: Tornado, hurricane, severe wind loss models.
Technical report NTIS no. PB294594/AS. Redondo Beach(CA): National Science
Foundation; 1976.
[9] Heneka P, Ruck B. Development of a storm damage risk map of Germany
A review of storm damage functions. In: Proceedings of the international
conference for disasters and society. 2004.
[10] Heneka P, Hofherr T, Ruck B, Kottmeier C. Winter storm risk of residential
structuresmodel development andapplicationto the Germanstate of Baden-
Wrttemberg. Nat Hazard Earth Sys 2006;6:72133.
[11] Heneka P. Schden durch Winterstrme - das Schadensrisiko von Wohnge-
buden in Baden-Wrttemberg. Ph.D. thesis. University of Karlsruhe; 2007.
[12] Katz R. Stochastic modeling of hurricane damage. J Appl Meteorol 2002;41(7):
75462.
[13] Khanduri, Morrow. Vulnerability of buildings to windstorms and insurance
loss estimation. J Wind Eng Ind Aerod 2003;91(4):45567.
[14] Klawa M, Ulbrich U. A model for the estimation of storm losses and the
identification of severe winter storms in Germany. Nat Hazard Earth Sys 2003;
3:72532.
[15] Lamb H. Historic storms of the North sea. In: British Isles and Northwest
Europe. Cambridge: Cambridge University Press; 1991.
[16] Leicester R, Reardon G. A statistical analyses of the structural damage by
cyclone Tracy. Civil Eng Trans 1976;18(2):504.
[17] Meaden GT. Tornadoes in Britain: Their intensities and distribution in space
and time. J Meteorol 1976;1:24251.
[18] Munich Re . Winter storms in Europe (I): Analysis of 1990 losses and future
loss potential. Mnchen: Mnchener Rckversicherungs-Gesellschaft; 1993.
[19] Munich Re . Winter storms in Europe (II): Analysis of 1999 losses and loss
potentials. Mnchen: Mnchener Rckversicherungs-Gesellschaft; 2001.
[20] Nateghi-A F. Assessment of wind speeds that damage buildings. Nat Hazards
1996;14:7384.
[21] Pinelli J, Simiu E, Gurley K, Subramanian C, Zhang L, Cope A. Hurricane
damage predictionmodel for residential structures. J Struct Eng 2004;130(11):
168591.
[22] Rootzn H, Tajvidi N. Extreme value statistics and wind storm losses: A case
study. Scand Act J 1997;1:7094.
[23] Sacr C. Extreme windspeedinFrance: The 99 storms andtheir consequences.
J Wind Eng Ind Aerod 2002;90:176371.
[24] Schraft A, Durand E, Hausmann P. Storms over Europe: Losses and scenarios.
Zurich: Swiss Reinsurance Company; 1993. p. 28.
[25] Sill B, Kozlowski R. Analysis of storm-damage factors for low-rise structures. J
Perf Construc Fac 1997;11(4):16877.
[26] Sparks P, Schiff S, Reinhold T. Wind damage to envelopes of houses and
consequent insurance losses. J Wind Eng Ind Aerod 1994;53:14555.
[27] Spence R, Fawcett W, Brown A. Windstorm vulnerability of the UK building
stock. In: Proceedings of the wind engineering society conference. 1998. p.
2338.
[28] Unanwa C, McDonald J, Mehta K, Smith D. The development of wind damage
bands for buildings. J Wind Eng Ind Aerod 2000;84:11949.
[29] Watson C, Johnson M. Hurricane loss estimation models, opportunities for
improving the state of the art. Bull Am Meteor Soc 2004;85:171326.