Predicting Design Wind Speeds from Anemometer Records: Some Interesting

Findings
Melissa D. Burton1, Andrew C. Allsop2
1
Wind Engineer, Arup AT&R, London, UK, [email protected]
2
Director, Wind Engineering, Arup AT&R, London, UK, [email protected]

ABSTRACT
Designers are more frequently relying upon wind climate analyses of wind records from local
meteorological sites for the design of modern structures. As a result, it is increasingly important
to understand the effects that various analysis and recording techniques have on these
predictions. Highlighted herein are the findings of recent studies that have raised questions on
the accessibility, validity, and identity of wind records, and the problems associated with not
knowing the anemometer type and exposure characteristics, incorrectly assuming the averaging
time of the wind speeds, and carrying out an assessment using only one meteorological site. The
findings of various case studies are presented and the implications (which in some circumstances
are alarming) that the extreme value predictions from the case studies have on the design of
modern structures is discussed.

INTRODUCTION
A vital part of the design of modern structures lies in the accurate assessment of design wind
speeds with probability of exceeedance of either 0.02 or 0.01. Designers typically rely upon
either codified values or results of wind climate analyses for the wind speeds to be used in
design; both of which depend upon accurate recordings of long-term wind records from
meteorological sites. Almost always the wind speeds needed for design are of a return-period
which is significantly longer than the wind record itself. Uncertainty in the predictions of the
design wind speeds can account for variations in design loading forecasts that are greater than
the value of the load factor.
Previous investigations by Arup have involved the analysis of wind records of various
length, anemometer type and recording methodology for much of the world. Various analysis
procedures have been employed, including those of Gumbel and Harris to derive extreme value
distributions of maximum wind speeds for which estimates of design wind speeds have been
made. Some advantages and disadvantages of these assessment methodologies will be discussed
in this paper.
In addition, this paper aims to highlight some concerns with various wind records from
different parts of the world. The discussion will range from the length of the averaging time
associated with the wind speeds, the anemometer type utilized in the recording, and the standard
methodology in maintaining wind records through manual and/or automatic recordings. The
implications on the design wind speed predictions will be brought to light.
GENERAL PROCEDURE FOR ASSESSING DESIGN WIND SPEEDS
IDENTIFYING SUITABLE WIND RECORDS
In the studies listed herein long-term climatological wind records for periods of recording of
longer than ten years are analyzed using statistical theory of extreme values. Before any
assessment is begun, pictorial documents and historical auditing reports are accessed (if
available) to confirm any dramatic changes. These changes would include: surroundings (i.e.
development construction, tree growth), anemometer location (i.e. movement from a sheltered
location to a less sheltered location), anemometer recording methodology (i.e. manual recordings
to automatic recordings), and/or anemometer type (i.e. pressure tube-anemometer to rotating-cup
anemometer).
The general procedure adopted for analyzing the wind records is; first - ensure the quality
of measurements and then separate mixed climate records, second - apply exposure correction by
transposing the wind records from the anemometer site (considering elevation, terrain, and
topographical correction) to be consistent with open country (Engineering Sciences Data Unit
(ESDU [1]) z0 = 0.03m) or open desert terrain, third - assess the frequency of occurrence of the
regular winds on a seasonal basis to gain further understanding of the parent distribution, and
fourth - carry out an extreme value statistical analysis using discrete mean and/or gust wind
speeds.
A standard procedure for separating mixed climate data is adopted. This involves
filtering mean wind records for speeds which are greater than 15 m/s and three times greater than
both the previous and post hourly recording. Wind speeds fitting this criterion are divided and
analyzed separately. “Mean” wind speed records which increase sharply in this manner are
attributed to thunderstorm downbursts or other convective behavior which is better suited to a
gust analysis.1
In recent years, there has been an increasing tendency to use mean speeds to predict
extremes. They have become favoured over gust speeds because of the theoretical statistical
stability of measurement and the known problems with interpreting and comparing gust speed
records from different meteorological sites and recording systems. On the other hand, gust
speeds are much less sensitive to changes or estimates of terrain roughness and site exposure and
with longer data records and consistent automatic measurements the extreme value gust
predictions also become reliable.
The choice between gust or mean wind speeds for the prediction of the design speed is
also related to the wind climate specific to the region in question. For example, the strength of
longer duration wind events (such as monsoons) is better predicted using mean wind speeds
whereas shorter events (such as thunderstorms) require an analysis of gust speeds. Therefore in
well-known thunderstorms climates, such as the Middle East or the eastern coat of Australia,
gust records are preferred (but aren’t always available!).
Unfortunately, from the majority of meteorological sites around the world the mean
speeds tend to be the most accessible wind record. Often the gust speeds are only recorded in
significant storms (and are sporadic at best) and therefore, in most cases, cannot be depended
upon to cross-check design wind speed predictions derived using mean speeds.

1
Convective activity is best recorded through comparison of both mean and gust wind speeds; although gust data is
not always available. For taller buildings, convective activity can often be disregarded since the strong horizontal
winds associated with these events are localized towards the ground [2] and have relatively little impact on the upper
portions of tall buildings where loading is most important for structural design.
INSTRUMENTATION
Since the early 1900’s gust and mean wind speeds around the world have been monitored using
either pressure-tube anemometers (most commonly used in British Commonwealth countries)
and rotating cup anemometers. In more recent years many meteorological sites have been
replacing pressure-tube anemometers with rotating-cup types, as the latter is thought to provide a
more accurate estimate of the gust wind speed. Previous research has shown that although the
two recording instruments exhibit little variation in the recordings of mean wind speeds, the
pressure-tube anemometer tends to record gust speeds 6-7% higher than the rotating-cup
anemometer [3]. The ‘Papillion’ anemometer utilized throughout France in the 1970s is also well
known for over recording wind speeds [4], but the mean wind speeds in this case.
In the move to upgrade legacy wind sensors in meteorological sites around the world
many places are moving to automated recording systems. A prime example of this is the
Automated Surface Observing System (ASOS) program which, in the 1990’s, moved to replace
all weather monitoring stations across the United States with an automated recording and
statistical processing system. This is in contrast to other areas around the world, such as parts of
China and most of the Middle East, which still continue to depend on manual recording methods.
In addition to the difficulty of understanding the anemometer types and recording
methodology, it can also be difficult to ascertain the averaging times associated with the wind
records. Even for mean wind speeds there are (at least) four averaging times in use: an hourly
mean, a 10-min continuous mean, or a 10-min mean measured over a 10-min period near the
hour or before every three-hour period. These are clearly not the same thing, yet they are
commonly treated as being the same during processes of analysis when the exact acquisition
process is unknown.

ENVIRONMENTAL COEFFICIENTS
Meteorological sites are seldom homogenous in their environment and rarely similar in their
surroundings. To account for the environmental variations it is necessary to correct wind records
to obtain a reference wind speed which is independent of local surrounding particularities, such
as nearby obstacles and terrain characteristics. These particularities are dependent upon wind
direction. Meteorologically homogeneous sites would be recorded over terrain with the same
roughness characteristics (over the duration of the record), at a standard height above ground and
averaged over the same time interval.
The generation of a standard reference speed allows for the like-for-like comparison of
wind speeds from meteorological sites with differing exposure and provides a basis for
comparing wind speeds recommended for design (i.e. hourly mean wind speed, in open-desert
(zo = 0.005m), at 10 m height, 50-year return period). Often the height of the anemometer is not
explicitly stated and although the standard height is 10 m, anemometers are sometimes mounted
higher in an attempt to compensate for the shelter of surroundings, although the accuracy of this
adjustment is obviously uncertain.
To review inconsistencies in wind records, regional analyses are preferred. This requires
the acquisition of wind records from a minimum of three meteorological sites, and ideally a
significant number more. In these cases the wind records from each of the regional sites are
considered separately to ensure that local topographical shelter is not influencing the wind
speeds recorded. Depending on the quality of the recordings, the proximity of the anemometers
and the suitability of the terrain correction factors it is possible to combine the various
anemometers into a ‘super-station’, where mutually independent wind records from these nearby
meteorological sites are consolidated into a single wind record.2 The ‘super-station’ approach
enables the reduction of sampling errors. In this type of assessment it is imperative that
independence of extreme events is ensured.
Terrain Roughness Adjustments
The widely accepted Deaves and Harris wind model of the atmospheric boundary layer, as
defined in ESDU Item 01008 [1], is commonly used to assess the upwind effects of ground
roughness.
Detailed surveys of terrain roughness are carried out for each meteorological site using
satellite imagery to assess roughness length. The analysis is completed for twelve sectors of
wind direction and up to a distance of 100 km from the meteorological site. Surface roughness
length, zo, is chosen to be consistent with the model used in ESDU. Some researchers in the past
have opted to not correct for the effect of variation in upwind terrain, however, it can be
observed in numerous wind records that a trend in time exists that correlates well with
development of upwind building density.
Due to the balance between gain/loss of turbulence and loss/gain of kinetic energy over a
rough/smooth surface, correction factors for the gust speeds are smaller than those for the mean
speeds.
Topography (Orography) and Altitude Adjustments
A rough upwind terrain can act to slow wind speeds down and a large mountain or a steep cliff
may act to steer or accelerate the winds. Several methods exist for assessing mean speed-up or
slow-down flow over hills, including the methods of Jackson and Hunt (eg. FLOWSTAR [7]),
wind tunnel testing and code methods that have been derived from these. It is however in
practice impossibly to parameterize all the factors of hill-shape that can be important, especially
when the hill slopes become steeper than about 30%, where wind flow may detach from the
surface. There are even greater difficulties with assessing the turbulence, especially in regions of
intermittent flow detachment. Depending on the size of the hills, mountains, valleys, cliffs, etc.,
topographic effects can be very difficult to correct and data from such regions is probably best
applied only locally.
In the UK and Eire, wind records are corrected for the local effects of altitude. The
altitude factor (SA) of the UK codes, is based on the elevation above sea level of the anemometer
as recommended in Cook [6]. The altitude factor follows the form:
S A  1 0.001* A (1)
Where A is the elevation above sea level in meters. This can be a useful correction for residual
effects of topography not covered by other code methods. This approach is believed to work
reasonably well over relatively short distances within the UK, but is known to be unreliable for
larger scale changes. More research into this and better methods of topography corrections in
general could help to clear up a number of interesting inconsistencies.

2
This process of creating a ‘super-station’ should be done with care. In a recent study carried out in the Emirates a
super-station was generated from 20-year gust records from four meteorological sites. The records were corrected
for terrain and merged into an equivalent super-station to obtain an enlarged gust database. The analysis resulted in
a 50-year open desert design gust speed equivalent to 37.8 m/s; this corresponds well with previous published
literature [5], however, upon further inspection it was revealed that 91% of the gust speeds used in the independent
storms analyses were recorded at one of the contributing four stations. Therefore, the design gust speed is more
representative of one area of the region and not necessarily the entire region.
 Obstacles such as warehouses or aircraft hangers in the immediate vicinity of
meteorological sites can also have a significant effect on wind speed measurements. Minor
effects of these obstacles can be specifically accounted for, however, for directions in which
major shelter is obvious, wind speeds and directions need to be adjusted conservatively.

EXTREME VALUE PREDICTIONS

Until the last twenty-thirty years, the estimate of extreme winds around the world was mostly
based on classical extreme value theory [8], and this is still used. However, statistical extreme
value analysis (EVA) has evolved in the last thirty years to include analysis techniques which
potentially offer more accurate predictions of the tails of the EVA distribution. Two EVA
techniques (modified Gumbel [10], and independent storms (proposed initially by Cook [11] and
improved by Harris [12]) are used herein to determine the probability of exceeding a given wind
speed in a given return period.
Design wind speeds depend on both the mean and the variability of the time series, the
design wind speed being positively correlated with both of these. The accuracy of the estimated
design speed generally improves with the length of the time series but this also depends on the
consistency and accuracy of the records.
In the majority of synoptic wind climates the parent wind distribution, irrespective of
wind direction, is reasonably well represented by a Weibull distribution (specific characteristics
of the distribution function were analyzed by Holmes and Moriarty [13]).

PV  1  exp  V / c 
k
 (2)
The extreme value x is equated to the wind velocity (mean or gust) squared, the
distribution of which is closer to exponential, as recommended in Cook [10] (to ensure a rapid
convergence to the extreme value distribution and to reduce the methodological bias).3 The
work of Cook and Harris shows that the fastest convergence to a standard Gumbel Fisher-Tippett
Type-I (FT-I) extreme value occurs if the underlying function is exponential and this would be
achieved by fitting wind speed to the power of k. Subsequent work by Cook [14] has shown that
fitting velocity squared (dynamic pressure) is likely to be at least as valid.
If attempts are made to fit velocity to a FT-I curve, then the resultant data fits will tend to
curve downwards in the tail. Clearly for a given number of data points it is possible to vary the
form of the Gumbel curve (using the second parameter) and get an equally good statistical fit.
[Occam's Razor would say it is best to use the simplest model that fits the data.]
Theoretically, for a set of data conforming to a Weibull distribution the cumulative
distribution function of the extremes will converge towards a FT-I distribution of the form:

Px  exp  exp y   (3)
Where the reduced variate y, is given by:

3
The methodology of fitting the Gumbel distribution to the dynamic pressure has been championed by a
number of researchers ([11], [12], [15]) as offering statistically the best fit (a Rayleigh distribution) to extreme wind
speed distributions, further researchers ([13],[17]) have shown the reverse Weibull distribution to be a more
appropriate probabilistic model of extreme wind speeds.
Although the two methodological assessments have been pitted against each other for some time, Simiu has
clearly shown in a very informative research study ([18]) that the probability plot coefficients (PPCC) goodness of
fit tests, which were used to measure the suitability of the reverse Weibull and the Rayleigh distribution of fitting
extreme wind speeds, were similar for the two distributions, varying only slightly in most cases on the third-
significant digit. This difference has limited significance when it is well known that there is a scatter-band of
reliability on the predictions at least an order of magnitude higher!
y   ln[ ln( Px )]
(4)
And Px is the probability that an extreme value will be less than a value x in any one year.
The modified Gumbel analysis makes use of the annual maximum wind records, whereas
the independent storms analysis uses the maximum ‘2.4 x years of data’ wind records [19] from
the same wind mechanism. The major advantage of the independent storms method over the
modified Gumbel method is that more wind records are available for the extreme value
assessment, including those which might otherwise be hidden by a more extreme storm which
happens to occur in the same year. This leads to a greater confidence in the estimation of design
wind speeds.
The Harris independent storms methodology applies a probability weighted least squares
to each dynamic pressure to associate lower confidence to wind speeds occurring less frequently.
Statistically independent observations exceeding a certain threshold wind speed are
considered in the analysis. Pairs of wind records separated by three days or less are treated as one
event to ensure that only independent storm maxima are retained. A reasonably high threshold is
adopted to ensure that only events arising from extreme storms are included in the analysis.
Gross et al. [18] has shown that extremes cannot be inferred from weaker winds, these winds can
often be considered to be noise which obscures the extreme processes of interest; ‘let the tails
speak for themselves’.

DESIGN WIND SPEED PREDICTIONS, INTERESTING FINDINGS

EFFECT OF ANEMOMETER TYPE
In one recent study wind time histories (of length up to 50 years) of hourly mean and
daily gust wind speeds for a dozen near-by meteorological sites were utilized to predict 50-year
design wind speeds for the region. All of the meteorological sites were initially equipped with a
Dines pressure tube anemometer; these were replaced with either a Vaisala or a Vector rotating
cup anemometer at a mid-way point in the record. Concurrently to the anemometer change the
weather stations became automated and all of the wind statistics (hourly means and daily gusts)
were calculated using automated software. About half of the sites converted to the Vaisala-cup
and the other half to the Vector-cup. There was no other significant change to any of the
meteorological stations at the time of conversion.
The time histories of the wind records were separated by the type of recording instrument
and extreme value analysis was carried out on each wind record independently. This was
possible as the length of the shortest record was seven years; the shortest length of record needed
to permit a reasonable estimate of design winds [20]. The results of the analysis revealed
significant variations in the 50-year design wind speed predictions from the three anemometer
types.
The sites which moved to the Vector-cup showed considerable reduction in the estimated
50-year design wind speed. Similar reductions were not observed at the sites that moved to the
Vaisala-cup.
The average reduction in the hourly-mean design wind speed predictions from the records
of the Vector-cup anemometer was approximately 15% less than the predictions from the Dines.
A similar percent reduction is observed for both the daily gust and hourly mean speeds (as shown
in the left plot of Figure 1). This reduction in design wind speeds can not be explained by
changes in terrain or simply by the frequency and strength of storms.
 From the right plot shown in Figure 1 it can be observed that the sites that were changed
to the Vaisala-cup anemometers show markedly different results. The hourly mean and daily
gust design wind speed predictions were similar with only small changes from certain directions
that can be attributed to changes in the upwind terrain.
It is hypothesized that the Vector-cup anemometer is predicting relatively low hourly
mean and daily gust wind speeds. Reasons for this are not yet clear but might include a
calibration or instrument matching issue or possibly some unintended influence of wind
turbulence.4
Clearly such discrepancies could have a serious impact on the design of structures if only
the most recent wind records are used to quantify the design wind speeds or if a mixed record
from the two anemometers was used without initially separating the records by anemometer type.

Figure 1 - 50-year directional design wind speed predictions, comparing Dines pressure tube anemometers
with Vector and Vaisala rotating cup anemometers.

EFFECT OF DATA SOURCE (NCDC VS. LOCAL AUTHORITIES)

It is a standard procedure of wind tunnel laboratories and consultancy groups to purchase wind
records from meteorological sites around the world through the National Climatic Data Centre’s
(NCDC) website. These wind records are then often used to predict wind speeds to be used in
the design of substantial structures.
Over the past couple of months, after acquiring wind records from local meteorological
sites directly, some alarming discoveries have been made. Attempting to directly correlate wind
statistics between the NCDC and the local meteorological sites has proved impossible.
Shown below in Figure 2 is a 19-year time history (1986 – 2004) of yearly maximum
hourly mean wind speeds acquired from the NCDC and the Australian Meteorological Bureau
(AMB). Both records are for Melbourne Airport and uncorrected for terrain. The NCDC record
is provided as hourly mean speeds in miles per hour whereas the AMB records are given in
kilometers per hour (both are shown in meters per second below for comparison).

4
From conversations the authors have had over the years, this is not a unique example, see also [3].
 The expectation is that the yearly maximum values are identical from the two data
sources, however in actuality only 5 of the 19 (~20%) records are synonymous. In some cases,
even the dates when the yearly maximum storm occurred are dissimilar (shown on the top and
middle of the Figure for the NCDC and AMB respectively; format MM,DD).
It has been noted anecdotally that the NCDC is given, through international agreements,
not the time history of the wind records from the meteorological sites around the world but an
hourly updated log file. This log file may often be composed of manually transcribed wind
records.5

Figure 2 - Maximum yearly mean wind speeds for Melbourne airport from two sources, time history showing
storm dates.
As a further example, twenty-three years of yearly maximum hourly mean wind records
(1983 – 2005) were acquired from the Don Muang Airport in the north of Bangkok from both the
Thai Meteorological Department (TMD) and the NCDC. Unfortunately in this case the dates of
occurrence of the strongest yearly winds were not available from the TMD and therefore a date
of occurrence correlation to the records from NCDC was not carried out.
Gumbel extreme value analysis was used to determine the 50-year open-country mean
hourly design wind, as shown in Figure 3 (yearly wind speeds are shown corrected for terrain).
The results of the analysis show that using the NCDC wind record, design wind speed

5
In at least one case in recent history a wind laboratory has utilized a 20-year wind record acquired from NCDC
from solely one meteorological site to predict design wind speeds for a tower approximately 400 m tall. This can be
significantly unconservative in light of the fact that there is a level of uncertainty in what the NCDC is providing.
Further problems with the NCDC wind records, to do with unit conversion, have been reported in Gatey and Miller
[21].
predictions of the order of 36 m/s are derived; this is compared to less than 32 m/s derived from
the TMD wind record.6
In this case the NCDC wind record has erred on the conservative side, however a similar
analysis was carried out for a meteorological site in northern China and the design wind speeds
given by NCDC were approximately 20% lower than those quoted by the Chinese
Meteorological Administration. A 20% difference in wind speed results at least a 44%
difference in loading and possibly a 70% difference in accelerations! Under-predicting wind
speeds to this order very quickly removes the risk coverage given by the load factor.

Figure 3 - Maximum yearly mean wind speeds for Bangkok airport from two sources, basic Gumbel plot with
50-year return period prediction.
Another issue with wind records is that they are often “binned” or rounded, resulting in
significant numbers of records of the same value. For such records it is necessary either to
randomize the values inside the box (for which some knowledge of the expected distribution is
required) or, perhaps better, to fit at the ends of each box. Unfortunately current Harris/Cook
fitting methods are not suitable for the latter, and the Lieblein BLUE methodology [22, 23] may
give better results.

6
Interestingly, these values are 44% and 28%, respectively, in excess of the 25 m/s 50-yr open-country hourly mean
design wind speed specified in the Thai Wind Loading Code (1311-50). It is believed that both the NCDC and the
TMD yearly maximums shown in Figure 3 are mean speeds confounded with short duration gust speeds (as opposed
to strictly consisting of hourly mean speeds). This was found to be the case after running the NCDC wind record
through a mixed climate analysis which identified almost 30% of the yearly maximums included in Figure 3 as
suspicious mixed record outliers.
Figure 4 - Independent storms extreme value analysis of highly grouped data.

IMPLICATIONS OF TRUSTING THE PREDICTIONS

It can misleading to rely solely on the point estimate from the independent storms EVA of the
longer return period events. There is a significant chance that the true longer return period wind
speed is larger than the point estimate. This is at least a normal result of variability within the
wind records.
In his 2000 paper [25], Harris suggests a way of generating artificial storms, assuming an
underlying exponential distribution of dynamic pressure and an arbitrary number of storms per
year (e.g. 100). The simulation of thousands of wind storms has become a popular area of
research in recent years as a greater attempt to demonstrate both the goodness of fit of the
various extreme value techniques as the reliability of design wind speed predictions has become
more important in structural design. By comparing the fit in an arbitrary period of records (say
20 years) over a large number (say 10,000) of sets of 20 year records this demonstration is
possible.
An independent storm simulation (as suggested by Harris [24]) was carried out to assess
the reliability of the point estimate of the 50-year design wind speeds that wind tunnel
laboratories and consultancy groups often depend upon. Ten thousand sets of 2000 storms (20
years * 100 storms per year) were generated. An independent storms EVA was carried out using
the maximum forty-eight (2.4 * 20 years) storms for each 20 year simulation.
The range of the 50-year predictions (99% confidence limits) varied by approximately
50% on the dynamic pressure; the 99th percentile design wind speed was approximately 20%
greater than the 1st percentile design wind speed, and 10% greater than the 50th percentile design
wind speed! Variations as great as these can result in an under-(or over-)prediction of the along-
wind loads of 20% and the across-wind loads up to 50%. Under-predictions of loads by
magnitudes as great as these have obvious implications on the design of modern day structures,
remembering that some of this is built into typical design safety factors.
 Two examples of the independent storms analysis of the simulated wind records
(resulting in upper and lower design wind speed predictions) are shown in Figures 5 ad 6; the
storm maxima are shown as black squares. The results of the EVA of those storm maxima is
shown in a black dotted line, the 50-year wind speed prediction is 25 m/s [Figure 5] and 20.9 m/s
[Figure 6] (the actual value of the design wind speed is 22.6 m/s (shown as a thick solid black
line); the basis of the storm simulation).7 Also shown in the figures are the 99% confidence
bounds on the simulated wind records and the 99% range of the design wind speed predictions
(black triangles).

Figure 5 - Wind storm simulation, an example of an independent storms assessment of one 20 year data set
(upper range of predictions).

7
In the wild it is found that recorded wind maxima are outside reasonable confidence limits. There are three
common reasons accounting for these obscure wind records. Firstly, the value may originate from an event with a
different mechanism (maybe not even a wind event) and a study which correlated the date of the occurrence of the
extreme event with meteorological records would need to be employed. Secondly, the event may have a smaller
probability of occurrence than the analysis would indicate, due to the finite record length not containing other
intermediate storm events which would occur eventually (i.e. a single storm of ~1000 year return). Thirdly, the
value could be associated with a manual recording/rounding error. In the case of this simulation only the second
explanation applies.
Figure 6 - Wind storm simulation, an example of an independent storms assessment of one 20 year data set
(lower range of predictions).
From these independent wind storm simulations it has been observed that:
a) The statistical expected value in a 20 year period is of 36 year return and similarly
for other exposure periods. (This can also be predicted from extreme value
theory, without the Monte-Carlo simulation by factoring by the exponential of
0.5772).

b) In reality, the slope (mode/dispersion) of the fit of the dynamic pressure should be
equal to the natural log of the number of storms per year (in the UK the slope is
~5 = 148 storms/annum).

c) The main element of scatter in the predictions from the various 20 year sets of
data is associated with the prediction of the slope (see Figures 5 and 6) and not
with the prediction of the annual mode. It is proposed, that in microclimates with
known slope (e.g.. slope = 5 in the UK); the slope is used to establish the 50-year
design wind speed from the annual mode, thereby enhancing the reliability of the
prediction. This has the potential to reduce the scatter of the design wind speed
predictions from 0.044 to 0.021.8

8
To take account of this it is necessary to be confident of the slope through a regional, rather than a local, climatic
study.
FINAL COMMENTS
The above is a guide to the practical use of real wind data. In many cases a subjective judgment
about whether particular data should be used or not has to be made. In general, the following is
recommended:
 Check sources of wind records and wind record consistency thoroughly.

 Separate storm types as far as is practical.

 Never rely on only one meteorological site.

 Check the slopes of the EV predictions against regional expectations.

With designers more frequently relying upon wind climate analyses (which utilize these
wind records from local meteorological sites) for the design of modern structures it is
increasingly important to understand the effects that various analysis and recording techniques
have on design wind speed predictions. Even a relatively small difference in design wind speed
can result in a dramatic difference in design wind loads and calculated building motions. This
can have significant financial and/or safety implications.

ACKNOWLEDGEMENTS
The authors would like to acknowledge the assistance of Alex To for liaising with the relevant
authorities to secure wind records directly from the various meteorological bureaus and for
carrying out some of the extreme value analyses discussed herein.

REFERENCES
[1] Engineering Sciences Data Unit (ESDU), World-wide extreme wind speeds. Part 1: origins and
methods of analysis. Item 87034b, ESDU International.

[2] Holmes, J.D., Wind loading of structures. Spon Press, London, New York 2001.

[3] Logue, J. J., The estimation of extreme wind speeds over standard terrain in Ireland. Technical Note
No. 51, Meteorological Service, Dublin, 1989.

[4] Sacre, C., Moisselin, J.M., Sabre, M., Flori, J.P., Dubuisson, B., A new statistical approach to
extreme wind speeds in France. Journal of Wind Engineering and Industrial Aerodynamics, Vol. 95, pp.
1415-1423, 2007.

[5] Irwin, P.A., Baker, W.F., Weismantle, P.A., The Burj Dubai tower, wind tunnel testing of cladding
and pedestrian level. Structure Magazine, November, pg. 47-51, 2006.

[6] Cook, N.J., The design guide to wind loading of building structures. Part 2: static structures.
Butterworths, London, 1985.

[7] FLOWSTAR, Model for airflow in complex terrain. Cambridge Environmental Research Consultants
(CERC).

[8] Gumbel, E., Statistics of extremes, Columbia University Press, 1958.

[10] Harris, R.I., Gumbel re-visited – a new look at extreme value statistics applied to wind speeds.
Journal of Wind Engineering and Industrial Aerodynamics, Vol. 59, pg. 1-22, 1996.

[11] Cook, N.J., Towards better estimation of extreme winds. Journal of Wind Engineering and
Industrial Aerodynamics, Vol. 9, pp. 292-323, 1982.

[12] Harris, R.I., Improvements to the method of independent storms. Journal of Wind Engineering and
Industrial Aerodynamics, Vol. 80, pp. 1-30, 1999.

[13]Holmes, J.D., Moriarty, W.W., Application of the generalized Pareto distribution to extreme value
analysis in wind engineering. Journal of Wind Engineering and Industrial Aerodynamics, Vol. 83, pp. 1-
10, 1999.

[14] Cook, N.J., Further note of directional and seasonal assessment of extreme winds for design.
Journal of Wind Engineering and Industrial Aerodynamics, Vol. 71, Issue 3, pp. 201-208,1999.

[15] Naess, A., Estimation of long return period design values for wind speeds. Journal of Engineering
Mechanics, Vol. 124, pg. 252-259, 1998.

[16] Simiu, E., Heckert, N.A., Extremewind distribution tails: a “peaks over threshold” approach,
Journal of Structural Engineering, Vol. 122, pg. 539-547, 1996.

[17.]Simiu, E., Heckert, N.A., Filliben, J.J., Johnson, S.K., Extreme wind load estimates based on the
Gumbel distribution of dynamic pressures: an assessment, Journal of Structural Safety, Vol. 23, pg. 221-
229, 2001.

[18] Gross, J.L., Heckert, N.A., Lechner, J.A., Simiu, E., Extreme wind estimates by the conditional mean
exceeedence procedure. National Institute of Standards and Technology, NISTIR 5531, 1995.

[19] Harris, R.I., Control curves for extreme value methods. Journal of Wind Engineering and Industrial
Aerodynamics, Vol. 88, pp. 119-131, 2000.

[20] Cook, N.J., The design guide to wind loading of building structures. Part 1: background, damage
survey, wind data and structural classification. Butterworths, London, 1985.

[21] Gatey, D. A., Miller, C.A., An investigation into 50-year return period wind speed differences for
Europe. Journal of Wind Engineering and Industrial Aerodynamics, Vol. 95, pg. 1040-1052, 2007.

[22] Lieblein, J., Efficient methods of extreme-value methodology. Report NBSIR 74 602, National
Bureau of Standards, Washington, 1974.

[23] Engineering Sciences Data Unit (ESDU), World-wide extreme wind speeds. Part 1: origins and
methods of analysis. Item 87034b, ESDU International.

[24] Harris, R.I., The accuracy of design values from extreme values analysis. Journal of Wind
Engineering and Industrial Aerodynamics, Vol. 89, pp. 153-164, 2001.