Tests of the Generalized Pareto Distribution for Predicting Extreme Wind Speeds

SEPTEMBER 2000 BRABSON AND PALUTIKOF 1627 Tests of the Generalized Pareto Distribution for Predicting Extreme Wind Speeds B. B. BRABSON Department of Physics, Indiana University, Bloomington, Indiana J. P. PALUTIKOF Climatic Research Unit, University of East Anglia, Norwich, United Kingdom (Manuscript received 27 June 1998, in final form 21 October 1999) ABSTRACT Extreme wind speed predictions are often based on statistical analysis of site measurements of annual maxima, using one of the Generalized Extreme Value (GEV) distributions. An alternative method applies one of the Generalized Pareto Distributions (GPD) to all measurements over a chosen threshold (peaks over threshold). This method increases the number of measurements included in the analysis, and correspondingly reduces the statistical uncertainty of quantile variances, but raises other important questions about, for example, event independence and the choice of threshold. Here an empirical study of the influence of event independence and threshold choice is carried out by performing a GPD analysis of gust speed maxima from five island sites in the north of Scotland. The expected invariance of the GPD shape parameter with choice of threshold is utilized to look for changes of characteristic wind speed behavior with threshold. The impact of decadal variability in wind on GEV and GPD extreme wind speed predictions is also examined, and these predictions are compared with those from the simpler Gumbel and exponential forms. 1. Introduction a. The need Safe design of structures at exposed locations (bridges, radio masts, wind turbines, etc.) requires knowledge of maximum expected wind gust speeds. The design criteria for such structures are often stated in terms of quantile values. For example, a structure only able to withstand the 50-yr quantile x 50 would have an unacceptably high predicted failure rate of 2% yr 1. A time series of gust speed measurements can be used to extract both the parameters of the statistical distribution of the extreme gust speeds and the corresponding quantiles. For example, the Generalized Extreme Value (GEV) distribution has been used to quantify the behavior of measured annual maximum (AM) gust speeds (Davison 1984; Smith 1989; Stedinger et al. 1993). As the standard errors for both distribution parameters and quantiles are proportional to 1/ n, where n is the number of measurements used, a longer time series gives smaller standard errors, the desired result. Unfortunately, long gust speed time series are not always available at the site of interest. Ten years of data are regarded as Corresponding author address: B. B. Brabson, Department of Physics, Indiana University, Bloomington, IN 47405. E-mail: brabson@indiana.edu the absolute minimum for a GEV analysis based on annual maxima (Cook 1985). To overcome the problem of a time series of inadequate length, several alternatives to a GEV analysis of annual maxima have been developed. Among these are analyses of the parent distribution directly (Gomes and Vickery 1977; Milford 1987), the use of statistical simulation techniques (Kirchoff et al. 1989; Kaminsky et al. 1991; Dukes and Palutikof 1995), r-largest events analysis (Weissman 1978; Smith 1986), the Method of Independent Storms (Cook 1982; Harris, 1999), and the Peaks over Threshold (POT) technique (Hosking and Wallis 1987; Abild et al. 1992; Simiu and Heckert 1996) where all measurements above a chosen threshold,, are included. The Generalized Pareto Distribution (GPD), introduced by Pickands (1975) and explored by several authors including Hosking and Wallis (1987), is often the distribution of choice because of its special relationship to GEV, as explained in part b of this section. Using a POT analysis with GPD, a number of authors have carried out extreme value analyses based on as few as 10 yr of data (e.g., Coles and Walshaw 1994; Abild et al. 1992). In this paper using 13 yr of hourly gust speed data from Sumburgh, Shetland, and some 40 yr of daily maximum gust speed data from each of four additional island sites in northern Scotland, we examine the following questions. 2000 American Meteorological Society

1628 JOURNAL OF APPLIED METEOROLOGY VOLUME 39 R With the greater number of gust speed measurements above threshold in a POT selection, more time-correlated events (events from a single storm, for example) are potentially included in the analysis. In section 2 we describe an empirical method, based on the Poisson distribution of independent events above a sufficient threshold, to examine and quantify this clustering. Using this method, we measure the number of such correlated events as a function of choice of threshold and event separation, and examine the influence of these events on the extreme speed quantile estimates x T. An additional measure of this event clustering is provided by an autocorrelation analysis. R When performing a GPD analysis on POT selected data, the choice of threshold is important. Only at sufficiently high thresholds does the GPD asymptotic behavior approach GEV. Using the long time series data with clustering removed, we are able to explore the GPD threshold dependence across a range from values well below GPD to GEV convergence to gust speeds of 40 m s 1. This long range allows us to examine the convergence process itself. R A characteristic of the GPD distribution that makes it suitable for extreme value analysis is termed the threshold stability property (Hosking and Wallis 1987; Dargahi-Nougari 1989). If the random variable X is GPD distributed, then the conditional distribution of (X ), where X, is also a GPD with the same shape parameter. That is, for a single underlying mechanism following a GPD distribution, should not change with choice of threshold. However, a mixture of underlying climatic components with significantly different gust speed distribution shapes and intensities could result in a change in shape parameter with threshold. In section 3 we investigate the evidence for the presence of multiple storm mechanisms by examining the shape parameter as a function of. R For short time series the preferred use of GPD over GEV is based on the assumption that the quantile estimates will be improved by increasing the sample size. In section 4 we investigate the validity of this assumption. The study of event independence and the GEV/GPD comparisons were done using an hourly time series of maximum 3-s gust speeds from 1972 84, and annual maxima for the overlapping period from 1972 96, both from the Sumburgh Airport in the Shetland Islands. These data are likely to be simpler than at other onshore sites. The site has relatively few obstacles to wind, and is surrounded by sea on three sides. The wind speed diurnal cycle is weak. We ignore the effects of obstacles, roughness, and terrain. To give a flavor for the Sumburgh data set, Fig. 1 shows a histogram of the hourly gust speed maxima, and Fig. 2 the time series during two high wind speed episodes. An investigation of GPD threshold dependence and the corresponding evidence for a mixed climate requires FIG. 1. The logarithm of the histogram of 1972 84 Sumburgh gusts. longer time series with greater numbers of events at large gust speeds. Thus, the analysis was performed using gust speed maxima from Sumburgh and four additional 37- to 39-yr daily maximum time series from the island sites of Benbecula and Stornoway Airport (Western Islands), Kirkwall (Orkney), and Lerwick (Shetland). b. GPD extreme gust speed calculations The Sumburgh distribution of 3-s maximum hourly gust speeds, shown in Fig. 1, illustrates the Weibull-like distribution typical of wind speeds. A GPD extreme gust analysis of these data focuses on the high-end tail, by selecting only those independent events above a threshold value. A GPD distribution of Type I has an exponential tail and a shape parameter of zero. For positive values of, Type III, the tail falls faster than exponential and unlike the exponential is finite in extent. Conversely, a negative shape parameter gives the Type II GPD a faster falling slope near threshold and an infinitely long, thicker tail. As explained by Hosking and Wallis (1987), for Z max(0, X 1,...,X N ) where X i are iid according to GPD, Z is essentially GEV distributed with corresponding types (I, Gumbel, 0; II, Frechet, 0; III, reversed Weibull, 0). We investigate this correspondence further in section 4 of this paper. Traditionally, the distribution of extreme gust and wind speed data for temperate latitudes is assumed to be Type I (e.g., Ross 1987; Gusella 1991; Abild et al. 1992), although Walshaw (1994) disputes this assumption. Lechner et al. (1992) found that of 100 wind records in the United States, 61 showed a Type III limiting form, 36 were of Type I, and 3 were of Type II. The Type II distribution is seldom found, and might be indicative of a mixed wind series (Gomes and Vickery 1978). No assumptions about distribution type are made here.

SEPTEMBER 2000 BRABSON AND PALUTIKOF 1629 FIG. 2. The gust speed (m s 1 ) time line during two high-wind episodes at Sumburgh. The GPD has a cumulative distribution function (cdf) F(x;, ) 1 [1 (/)(x )] 1/, (1) where x / for 0, and x for 0. For the special case when 0, F(x; ) 1 exp[(x )/], (2) where x. The derivative of the cdf in Eq. (1) gives the probability density function (pdf), [ ] (1)/ 1 (x ) f (x;, ) 1. (3) To the extent that gust speed exceedances follow a GPD distribution, we can estimate the GPD quantiles. With the definition of crossing rate,, as the expected number of peaks above threshold per year, the expression for the number of exceedances in t yr is x t(1 F). (4) Taking (1 F) from Eq. (1), setting x to unity, and t to T, that is, one exceedance in T yr, we can solve for the quantile, x T, GPD x /[1 (T) T ]. (5) The problem of estimating quantiles then devolves to estimating the GPD parameters and for a given threshold,. In this study we use two techniques for parameter estimation, the Probability Weighted Moments (PWM) method, (Hosking and Wallis 1987; Abild et al. 1992; Stedinger et al. 1993), and the Maximum Likelihood (ML) method (Davison 1984; Smith 1984; Hosking and Wallis 1987; Davison and Smith 1990). The PWM estimators have been shown to perform well when 0, particularly when 0.2, while ML estimators have higher efficiency when is near 0 and outperform PWM for 0 (Hosking and Wallis 1987). Therefore, both are presented in this analysis. As will be shown, the two methods can produce quite different results.

1630 JOURNAL OF APPLIED METEOROLOGY VOLUME 39 2. The question of statistical independence It is a condition of extreme value analysis that the selected extremes must be independent (Poisson distributed). Related to this condition, it is important to understand the quantitative impact of nonindependent extremes on the calculations of extreme quantiles. Annual maximum wind speeds chosen from each wind year (July June) are virtually always statistically independent. However, when several data points are taken from each year, as in a POT analysis, there may well be several clustered hourly maximum gust speeds above threshold from a single storm (see Fig. 2). Such events are unlikely to be statistically independent. Here, we pose the question of the effect of using nonindependent, correlated extremes in a POT analysis. Various strategies are invoked to remove dependent events before proceeding with a statistical analysis. A simple strategy is to require a minimum time separation or deadtime between selected events. Working with European wind climates, Cook (1985) and Gusella (1991) use a deadtime of 48 h, whereas Coles and Walshaw (1994) extend this to 60 h. In a second strategy, Walshaw (1994) uses a modification of the mean residual life (MRL) plot to establish that a dataset of extremes consists of truly independent events. The MRL plot is simply x, the sample mean of the events above threshold minus the threshold, plotted against the threshold (Davison and Smith 1990). His modification of MRL he calls the reclustered excess plot and he uses the linear behavior of this plot to identify acceptable values of the threshold where extreme events can be deemed independent. Alternatively, Smith and Weissman (1994) discuss the concept of an extremal index, a measure of the degree of clustering of extreme events in a time series. The cluster size is taken to be the number of observations of extreme events that occur within a chosen storm length or block size. The extremal index is the reciprocal of the mean cluster size, lies between 0 and 1, and can be thought of as the fraction of independent events in a selected sample. They give two methods for determining this extremal index, and describe how it relates the parent distribution to the asymptotic GEV distribution. In this section we take an empirical and complementary approach of examining the different timescales of clustering of extremes in our sample data, deciding which to regard as dependent, and measuring the fraction of such dependent events in the sample. We then calculate the desired extreme quantile wind speeds x T for that sample, and adjust both extreme event threshold and separation in time until stable x T are achieved. In this manner, an acceptable level of dependent event contamination is determined. We begin our analysis by choosing as our parent distribution the hourly maximum gust speeds (of 3sduration) recorded at Sumburgh, Shetland, from 1972 84. FIG. 3. Number of events above threshold for the Sumburgh maximum gust speed data, 1972 84. The event separation is then varied in nine steps from short (3-h) to longer (96-h) intervals. A sample of extreme wind speed events is chosen by selecting an observation above a threshold only when this observation is not followed by a higher wind speed in a time period of duration d. Figure 3 shows the number of events above threshold as a function of these two parameters. The number of events included in the data sample rises rapidly as the deadtime between events is allowed to decrease below about 18 h. a. The distribution of time intervals between storms Associated with the rapid rise in number of exceedances for low thresholds and short deadtimes, we expect an increase in the number of statistically dependent (clustered) events. A simple consistency check for statistical independence can be made. Because such events should be Poisson distributed in time, the frequency distribution of the time intervals t between events should demonstrate exponential behavior f(t) e t, (6) where is the crossing rate defined as the average number of selected events per unit time. A time interval plot provides a simple way of identifying event correlations at different timescales. For example, Fig. 4 shows the t plot for all events above a threshold of 24 m s 1 and event separation of 3 h. By examining the t plot at different scales, bin widths of 2 h in Fig. 4a and of 10 h in Fig. 4b, we discover several interesting time structures in the Sumburgh wind data. This plot is well represented by exponentials with different slopes, each corresponding to a different Poisson event crossing rate, and each characteristic of an underlying correlating mechanism (i.e., average storm duration). The two exponential lines in Fig. 4a correspond to crossing rates of one event every

SEPTEMBER 2000 BRABSON AND PALUTIKOF 1631 4 h (1/ 4 h) and one event every 27 h (1/ 27 h), and represent the short time interval part of the spectrum quite well. When these same events are shown on a longer timescale in Fig. 4b, an additional exponential is needed, corresponding to a crossing rate of an event every 160 h. At this point we wish to develop an index giving a measure of event independence of a set of POT data and to use this index in identifying appropriate deadtimes and thresholds in our subsequent GPD analysis. This amounts to counting the fraction of independent events in a given dataset. Thus, we must decide which components of our data to identify as independent. A conservative choice to assure event independence at the individual storm level is to exclude both of the short time interval 4-h and 27-h components when counting independent events. This choice, we will find is consistent with the autocorrelation calculations in part d of this section and also with the literature on storm lengths in the United Kingdom, a summary of which is given in Palutikof et al. (1999). Thus, the area under the 160-h exponential in Fig. 4b represents independent events, while the events in the excess at the shorter t above the 160-h exponential are not considered to be independent. This corresponds to our understanding of the behavior of the atmospheric circulation over the eastern Atlantic. Families of depressions track from west to east across the Atlantic, arriving over the United Kingdom at intervals of around 6 days (Cook 1985). We define the independent event index as the ratio of the area under the 160-h exponential (events without short-term correlation) to the total area. independent events. (7) total events b. Event independence, thresholds, and dead times We are now in a position to quantitatively identify proper combinations of dead time d and threshold, where the selected events have near 1.0, free of shortterm correlations. The quantity and its errors are tabulated in Table 1. At relatively low thresholds of 24 to 26ms 1, exceeds 0.85, for example, only for d 48 h or longer. At higher thresholds of 34 36 m s 1, on the other hand, a dead time of 12 h is sufficient to ensure 0.9. Dead times less than 12 h do not work. c. Quantile variation with dead time We first illustrate the nature of the 100-yr quantile dependence on d and on. Figure 5 shows the 100-yr FIG. 4. Semilogarithmic plots of the time interval frequency histogram for a 3-h dead time and 24 m s 1 threshold, using (a) 2-h bins, and (b) 10-h bins. The superimposed exponential lines are explained in the text.

1632 JOURNAL OF APPLIED METEOROLOGY VOLUME 39 TABLE 1. Independent event index at Sumburgh for different thresholds (m s 1 ) and dead times in hours. Standard errors are given in parentheses. Threshold dead time 24 m s 1 26ms 1 28ms 1 30ms 1 32ms 1 34ms 1 36ms 1 38ms 1 3h 6h 12 h 18 h 24 h 36 h 48 h 72 h 96 h 0.31 (0.02) 0.45 (0.03) 0.61 (0.03) 0.68 (0.04) 0.73 (0.04) 0.79 (0.04) 0.88 (0.04) 1.05 (0.05) 1.10 (0.04) 0.37 (0.02) 0.54 (0.03) 0.73 (0.04) 0.69 (0.05) 0.83 (0.04) 0.81 (0.04) 0.86 (0.04) 0.98 (0.05) 1.07 (0.04) 0.39 (0.03) 0.55 (0.04) 0.68 (0.04) 0.71 (0.04) 0.77 (0.04) 0.81 (0.04) 0.83 (0.04) 0.94 (0.04) 1.01 (0.04) 0.49 (0.04) 0.71 (0.04) 0.88 (0.04) 0.92 (0.04) 0.95 (0.04) 0.90 (0.04) 0.90 (0.04) 1.01 (0.04) 1.03 (0.04) 0.49 (0.04) 0.71 (0.05) 0.86 (0.05) 0.90 (0.05) 0.91 (0.04) 0.89 (0.05) 0.85 (0.05) 0.98 (0.04) 0.98 (0.04) 0.54 (0.05) 0.72 (0.06) 0.91 (0.05) 0.88 (0.05) 0.92 (0.05) 0.91 (0.05) 0.96 (0.05) 0.82 (0.14) 1.08 (0.04) 0.68 (0.08) 0.76 (0.07) 0.91 (0.06) 0.85 (0.06) 0.90 (0.06) 0.88 (0.06) 0.91 (0.06) 0.75 (0.16) 1.02 (0.03) 0.59 (0.09) 0.81 (0.09) 0.95 (0.05) 0.95 (0.05) 0.95 (0.05) 0.66 (0.20) 0.66 (0.20) 0.68 (0.21) 0.95 (0.05) quantile x 100 as a function of d for a threshold of 24 m s 1 on the left and the same data as a function of on the right. Note that quantiles from the ML method have less variability with dead time than the PWM method. As the dead time increases, the 100-yr quantile x 100 first drops and then stabilizes for dead times above 36 h. As a function of, the quantiles stabilize above of 0.8. Thus, is a useful parameter for identifying quantile stability and indicates that a clustered event contamination of as much as 20% has relatively little impact on x T. Table 1 also indicates that at thresholds as high as 30 m s 1, a dead time of 24 h is sufficient to remove clustered events. This interval is short in comparison with the usual choice of 48 60 h. To understand the higher x T at small, we compared the probability density distributions for event samples with high and low. The clustered events occur predominantly at relatively low gust speeds, just above threshold. When included in a GPD analysis, they force a less positive shape parameter, corresponding to higher gust speed quantiles, thus explaining the artificial rise in x 100 observed in Fig. 5 at low dead time. A second negative consequence of an event sample with low is an underestimation of the quantile standard errors. These errors are proportional to 1/ n independent (Hosking and Wallis 1987; Abild et al. 1992), and will be underestimated if n independent is inflated by the addition of clustered events. In summary, we have identified two reasons to avoid heavy contamination from nonindependent events in this kind of statistical analysis. First, these events distort the shape of the gust speed distribution producing anomalously high gust speed quantiles, and second, the quantile standard error, proportional to 1/ n, will be underestimated by an inflated n. However, provided a sufficiently high threshold is selected, shorter dead times than are normally found in the literature can be utilized. FIG. 5. The 100-yr return quantiles x 100 for gust speeds above 24 ms 1 as (a) a function of dead time and (b) independent event index. Example PWM and ML quantile errors are included. d. Event independence, autocorrelation, and longer time series Up to this point the exponential behavior of the time interval frequency and the corresponding stabilization of the gust speed quantile x T with increasing event separation d have been used to identify a d sufficient for Poisson event independence. As an additional check, we show in Fig. 6 the autoregressive parameter r k (Wilks 1995) for the 24 m s 1 threshold data of Fig. 5. The value of r 1 is a measure of correlation of neighboring

SEPTEMBER 2000 BRABSON AND PALUTIKOF 1633 FIG. 6. Autoregressive parameter as a function of lag for 24 m s 1 sample data. wind events and is seen to move to values less than 0.05 for d of 12 h or greater. The hourly gust speed maxima from Sumburgh provided the detail necessary for this event independence study. To confirm the usefulness of the technique, we also calculated for the four 37- to 39-yr time series from the neighboring island sites. For each site and for thresholds 24 m s 1, separating events by at least one intervening observation is sufficient to bring into the range of 0.9, enough to remove the undesirable effects of event correlation discussed for Sumburgh above. Therefore, in discussion of quantile variation with threshold, and the related discussions of mixed climate in section 3, we utilize these longer time series with a minimum of one intervening observation and shall refer to this minimum separation as a deadtime of one day. FIG. 7. (a) A histogram of Benbecula gust speed data, 16 m s 1, d 1 day, (b) (), (c) (), (d) x 100 (). 3. Threshold behavior GPD and the question of mixed climate Simiu and Heckert (1996) have examined the question of whether the finite-tailed (Reverse Weibull, 0) GPD distribution is an appropriate extreme wind speed model in a POT analysis. Analyzing daily data from 44 U.S. stations, they answer the question affirmatively by examining the threshold dependence of dehaan s tail-length parameter c, the negative of our shape parameter. We extend their study by examining GPD () for the four northern Scottish island sites where prevailing westerlies are expected to provide the dominant component of this high wind speed climate. The high average wind speeds at these sites provide an extended threshold range over which to examine the question of GPD to GEV convergence. Upon convergence, the threshold stability property of GPD would predict a constant for a simple wind climate described by a single GPD distribution. At a higher level of complexity with a mixed climate of two GPD distributions with different values of, () should shift from one value of to another. As we will discuss, maximum wind speeds at these four island sites exhibit a more complex behavior, where changes continuously with threshold. We also examine a related variable, the mean residual life (MRL), another quantity sensitive to a mixed climate (Davison and Smith 1990) when plotted against threshold. a. The threshold behavior of GPD When GPD fits are done as a function of threshold, they provide additional information about a time series not available to a single GEV fit. The longer time series from Benbecula, Kirkwall, Lerwick, and Stornoway Airport give us a greater gust speed threshold range than does the 13-yr time series from Sumburgh alone. Because these four time series have similar characteristics, we begin by looking at one (Benbecula) in some detail. Figure 7a is simply a histogram of maximum daily gust speeds separated by a deadtime of one day. It shows clearly the roll-over below 20 m s 1 and the approximately exponential high gust speed tail. Superimposed on this histogram are the properly normalized pdfs from GPD maximum likelihood fits for thresholds of 20 m s 1,32ms 1, and 40 m s 1. The 20 m s 1 threshold hasafit of 0.18 and is seen to have a faster-falling tail than the more exponential 32 m s 1 fit with of 0.03. For still lower thresholds, 16 20 m s 1, the GPD fit is forced by the turn-over at the left-hand (low gust speed) end to still more positive values of, and, it turns out, to poorer fits. The 40 m s 1 fit is spurious, illustrating the behavior of a GPD when fit to a very small dataset. At these very high thresholds there is little constraint from low gust speeds and GPD creates a highly curved () fit consistent with essentially no events beyond those actually observed.

1634 JOURNAL OF APPLIED METEOROLOGY VOLUME 39 Figure 7d shows the 100-yr return quantile x 100 as a function of. In the region from 20 to 36 m s 1, x 100 is roughly constant even as the individual GPD parameters are rapidly varying. In this same threshold interval the maximum quantile speed for positive, limit /, is relatively constant because it depends only on the ratio, /. More variability is seen in the probability weighted moments estimators than in the maximum likelihood estimators, but both show the same trend. Effectively, we have a number of different estimates of the extreme gust speed quantiles. Though these estimates are not independent, they do derive from different portions of the original dataset and tell us something about those portions. A conservative estimate of the quantile x 100 is given by the maximum observed quantile over the selected threshold range. For Benbecula, such an estimate of the 100-yr quantile speed would come from the 36 m s 1 threshold point and give a quantile estimate of x 100 52.7 m s 1 4.6 m s 1. Thus, a GPD parameterization as a function of threshold provides an estimation procedure for extreme quantiles that includes the possibility of a number of contributing mechanisms. If the gust speed extremes are well described by a single GPD distribution, then () should either be constant and equal to GEV, the GEV shape parameter, or it should approach this value asymptotically. Alternatively, the five sites in northern Scotland might show evidence for a second component of wind climate. We pursue this mixed climate hypothesis in sections 3b and 3c. b. Search for mixed climate using GPD The question of uncovering underlying dynamical mechanisms by observing the character of maximum gust speed distributions has been addressed by several authors (e.g., Gomes and Vickery 1978; Coles and Walshaw 1994; Hannah et al. 1995). For example, Gomes and Vickery merged statistical models from different climates (hurricanes, prevailing westerlies, and convective storms) to reproduce observed gust speed maximum distributions, and corresponding quantiles. In this analysis, we choose to use the properties of the GPD distribution itself to search for underlying mechanisms in these Scottish gust speed data. Although should be constant with threshold for simple one-gpd samples, two-gpd wind climate data might well show shifts in (). We now examine the actual GPD maximum likelihood fits for the four long time series gust data. Figure 8 shows () and () and for each of these datasets, both and demonstrate a decrease from 16 m s 1 to 36ms 1, with an approximately linear decrease above 20ms 1. These data are neither constant nor does the shape parameter demonstrate an asymptotic approach to the GEV shape parameter GEV. FIG. 8. GPD maximum likelihood parameters, (a) () and (b) () for POT data with one-day separation from Benbecula, Kirkwall, Lerwick, and Stornoway Airport. The sites are displaced vertically by even units to facilitate their comparison. For these four datasets the observed extreme POT gust speed distributions do not seem to be consistent with either a simple GPD or combinations of two GPD distributions. The POT tail is more complex. For each of the four datasets the highly positive shape parameters at low threshold evolve to near zero (or even negative) at large thresholds. Although extreme value theory tells us that the distribution of annual maxima taken from a simple GPD POT distribution converges to a GEV distribution with the same, it does not dictate that real wind distributions have GPD tails. We have found examples where they do not. Summarizing this section, we find evidence to support the view that POT data taken from areas with high average wind speeds in northern Scotland are not well represented by GPD with the threshold stability property expected at sufficiently high thresholds. Nevertheless, we find that a GPD POT analysis provides useful ex-

SEPTEMBER 2000 BRABSON AND PALUTIKOF 1635 FIG. 9. MRL plots for the four 37- to 39-yr series with one-day event separation. Here the sites are displaced vertically by single units to facilitate their comparison. treme quantile predictions x T as a function of threshold, information not available from an annual maximum GEV analysis. Conservatively, the extreme quantiles can be taken as the highest of the GPD determinations with threshold. c. Search for mixed climate using mean residual life In an MRL plot, x, the sample mean of the events above threshold minus the threshold, is plotted against the threshold. For an exponential distribution, for example, the MRL has zero slope and intercept. More generally, Davison and Smith (1990) point out that for a pure GPD distribution, the MRL will have a slope of /(1 ), and an intercept of /(1 ). MRL plots for Benbecula, Kirkwall, Lerwick, and Stornoway are shown in Fig. 9 for an event separation of 1 day. At each site the MRL slope is negative at a threshold of 20 m s 1 and approaches zero near 36 m s 1. Though the variability increases as the number of data points above threshold decreases, there is evidence near 36 m s 1 for a slope change to near zero. Above approximately 43 m s 1 the MRL drops rapidly toward zero as we reach the last events in the distribution tail. From this information alone, we would predict a positive and decreasing in the region 16 36 m s 1 followed by a short interval of exponential tail ( 0) above 36 ms 1. Thus, we find additional evidence for the long gradual decrease in seen in Fig. 8 above. We do not find an MRL with constant slope, expected if were constant with threshold. 4. The question of quantile estimation in a variable climate Extreme value analyses of wind data usually assume the stationarity of the parent distribution, though genuine long-term linear trends have been identified in such analyses (Tawn 1988; Palutikof et al. 1999). Palutikof et al. (1987), in a study of temporal and spatial variability of wind, found evidence for decadal variability of average wind speeds in the United Kingdom. This immediately raises the question of the impact of such variation on extreme quantile predictions. First, we begin with a comparison of GEV with GPD for the 13 yr (1972 84) of hourly 3-s gust speed maxima originally available at Sumburgh. Second, we include 25 yr (1972 96) of annual maximum hourly 3-s gust speeds from Sumburgh in an attempt to validate the 13- yr results from GEV and GPD. Both the generalized distributions (GEV and GPD) and the simpler distributions (Gumbel and exponential) are included in these comparisons. Last, we carry out a similar comparison for the four longer time series sites of Benbecula, Kirkwall, Lerwick, and Stornoway Airport, here again investigating the influence of decadal variation. a. The GEV distribution Beginning with a parent distribution whose cumulative distribution function (cdf) is F(x), we sample this distribution n times, and take the maximum value of the n samples. This maximum value has a cumulative distribution function of its own of simply F n (x). This relationship leads to the simplicity in extreme value theory, pointed out by Fisher and Tippett (1928), that for sufficiently long sequences of independent and identically distributed (iid) random variables, the maxima of these sequences can be fit to one of three limiting distributions. The cdfs of these three limiting distributions are summarized in the GEV distribution, [ ] 1/ (x ) F(x) exp 1, (8) where x / for 0, / x for 0, and x for 0. The three classes of GEV are designated by the value of the shape parameter,. Gumbel or Fisher Tippett Type I has of 0, while Frechet or Fisher Tippett Type II has 0, and Reverse Weibull or Fisher Tippett Type III has 0. Similar to the GPD distribution, the scale parameter determines the width of the distribution. The GEV distribution mode is designated here by. For Type I, Gumbel, the above equation takes a simpler form: [ ] x F(x) exp exp. (9) As an example we consider a sample from the Sumburgh set of hourly gust speed maxima. The annual maxima of the hourly gust speeds will almost certainly be independent events and should then be GEV distributed.

1636 JOURNAL OF APPLIED METEOROLOGY VOLUME 39 FIG. 10. Sumburgh 1972 84 gust speed quantiles as a function of return period for GEV, Gumbel, and BLUE-corrected Gumbel. b. The application of GEV to Sumburgh Like the GPD distributions discussed in sections 2 and 3, the GEV shape parameter strongly influences the nature of the upper tail of the distribution. The tail is exponential for Gumbel ( 0), thicker than exponential for Frechet ( 0), and finite in extent for Reverse Weibull ( 0). Thus, the GEV extreme gust speed quantiles depend sensitively on (Simiu and Heckert 1996). As we did for GPD, we calculate unbiased estimates of the GEV parameters,,, and both from the probability weighted moments ( 0, 1, and 2 in Stedinger et al. 1993) and separately using maximum likelihood. Here, these strategies agree quite well. We calculate the quantiles of the GEV distribution from xt {1 [ln(f)] }, (10) where F is the cumulative probability, given in terms of the return period by F (T 1)/T. Figure 10 shows the PWM GEV quantile gust speed, x T, as a function of the return period, T (dashed curve). On the same plot we place the 13 annual gust speed maxima for 1972 84 (plus signs). As sample estimates of the cumulative probability for GPD given by Eq. (1), the plotting positions are taken as F i 1 (i 0.44)/ (n 0.12), where x 1 x n are the ordered annual maxima (Hosking et al. 1985; Linacre 1992; Hosking and Wallis 1995). The PWM GEV quantile plot does indeed follow the annual maxima from which it is derived. However, the shape parameter has the peculiar value of 0.61 corresponding to a strongly truncated tail, with a predicted maximum gust speed of 45.8 m s 1, an unlikely physical result, given that wind speeds as high as 44.0 m s 1 are already observed during the 13-yr recording period. For purposes of comparison we superimpose the FIG. 11. Sumburgh 1972 84 gust speed quantiles ( 32ms 1 and d 24 h) as a function of return period for GPD and exponential. Dotted lines reproduce Fig. 10. PWM Gumbel quantile plot for these same data. The Gumbel distribution (Stedinger et al. 1993) with its shape parameter 0, has an exponential high-end tail and is shown as the solid curve in Fig. 10. The Gumbel quantile function is simply x T ln[ln(f)]. (11) The Gumbel distribution follows these data less well. Part of the difficulty, as explained in Cook (1985, p. 111), lies in the asymmetry of the Gumbel distribution and the consequent unequal confidence among the estimates of the ranked annual maxima. The largest of 13 annual maxima is also Gumbel distributed, and on average, will have a value higher than the mode of this (skewed) distribution. A method devised by Lieblein (1974) defines a corrected set of weights called best linear unbiased estimators (BLUE) to remove this bias. As Cook explains, the higher ranked data points are weighted less. We plot the BLUE corrected Gumbel plot (dash-dotted curve in Fig. 10), as well. Though this BLUE Gumbel fit gives higher quantile values for large return periods, as expected, it differs relatively little from the uncorrected Gumbel. By eye, neither fits the 13 yr of annual maxima well. c. A comparison of GEV with GPD Figure 11 shows the quantiles of Fig. 10 (dotted curves) with the added quantile predictions from the peaks-over-threshold GPD ML fit (solid line), described in section 1b. The GPD parameters for this fit are 0.262, and 4.69 m s 1. With these parameters, we calculate the GPD quantiles as a function of return period using Eq. (5). For this comparison we select the GPD POT analysis with a threshold of 32 m s 1 and dead time of 24 h, comfortably located in the region of independent event index near one, where the quantile

SEPTEMBER 2000 BRABSON AND PALUTIKOF 1637 TABLE 2. Sumburgh: a comparison of 100-yr quantiles from GEV, Gumbel, GPD, and exponential analyses. GPD values in the table are done with a threshold of 32 m s 1 and dead time of 24 h. Sumburgh analysis type and period Parameters x 100 (m s 1 ) SE (m s 1 ) GEV, PWM Param, 13 yrs of AM Gumbel, PWM Param, 13 yr of AM GEV, PWM Param, 25 yr of AM GEV, ML Param, 25 yr of AM Gumbel, PWM Param, 25 yr of AM GPD, ML Param, 13 yr of POT EXP, PWM Param, 13 yr of POT GPD, PWM Param, 13 yr of POT 0.61 0.005 0.008 0.36 0.26 4.06 2.96 4.00 3.99 3.97 5.39 3.35 4.92 39.1 38.1 38.7 38.7 38.6 32.0 32.0 32.0 45.3 51.7 56.8 56.4 56.9 45.4 53.8 47.3 2.0 3.3 5.6 7.6 3.2 2.4 3.3 3.1 speeds are stable. These choices yield 86 events above threshold. We plot these 86 data points using the relationship T 1/(1 F) from Eq. (4), and the plotting position F i 1 (i 0.44)/(n 0.12) as above. Note that GPD GEV x100 (solid line) lies close to x100 (lower dotted line). To complete these comparisons from the 13-yr dataset, we add the exponential GPD quantile prediction (dashed line in Fig. 11). This lies well above both the GEV and GPD predictions, in the region of the Gumbel predictions (upper dotted lines). Critical to the usefulness of these maximum gust speed predictions are their associated standard errors [see Hosking et al. (1985) and Abild et al. (1992) for calculation procedures]. Table 2 contains both the GEV and GPD 100-yr quantiles and standard errors, based on the 13 yr from 1972 to 1984. We first draw attention GEV to the 1.8 standard deviation separation between x 100 45.3 2.0 m s 1 Gumbel and x 51.7 3.3 m s 1 100. The GEV analysis fits the annual maxima significantly better GPD than Gumbel. The ML x 45.4 2.4ms 1 100 agrees GEV well with the x 100 but is approximately 1.5 standard Gumbel deviations away from the x 100 and 2.1 standard deexp. viations away from 53.8 3.3 m s 1. Based on x 100 FIG. 12. Sumburgh annual maximum 3-s gust speeds for 1972 96. the 13-yr dataset alone, one would conclude that the generalized distributions (GEV and GPD) give a better fit, and disagree with the more highly constrained (Gumbel and exponential) distributions with shape parameter, 0. d. A longer dataset at Sumburgh In an attempt to validate this analysis of the 1972 84 Sumburgh data, we now examine the annual maxima beyond 1984, taken from the Monthly Weather Report produced by the U.K. Meteorological Office, and shown in Fig. 12. Using 25 yr of annual maxima from 1972 to 1996, we look again at both GEV and Gumbel predictions of quantiles for Sumburgh. Figure 13 shows these quantile functions, superimposed on the 25 annual maxima. During this extended period four very large annual maxima occurred in the period, 1991 94 (see Sumburgh in Fig. 12). These dramatically affect the GEV GEV quantile predictions, raising x 100 from 45.3 2.0 ms 1 to 56.8 5.6 m s 1, a change well outside the standard errors calculated on the basis of the 13-yr sample. However, the Gumbel predictions are less affected by Gumbel these new data. The x 51.7 3.3 m s 1 100 from the 13-yr sample changes to 56.9 3.2ms 1. The Gumbel BLUE analysis de-weights the extreme data points and changes the 100-yr quantile even less, from BLUE 52.7 m s 1 BLUE x100 for 13 yr to x100 53.0 m s 1 for 25 yr. As expected, we find that the Gumbel and Gumbel BLUE quantile analyses are less sensitive than the more flexible GEV quantile analysis. A key comparison can now be made between a POT analysis of 13 yr of data (with 86 points) and an AM analysis of 25 yr (Fig. 14). Though more data points are used in the GPD analysis, with concomitant smaller standard errors, the 25-yr GEV and Gumbel analyses and the 13-yr exponential analysis are in better agreement with the 1972 96 annual maxima. We conclude that the additional data points from a POT analysis are not necessarily a replacement for a longer time series. The 25-yr dataset shows that both of the generalized models, GEV and GPD, when based on 13 yr of data, though very good at following the detailed shape of the

1638 JOURNAL OF APPLIED METEOROLOGY VOLUME 39 FIG. 14. Annual maxima for the five northern Scotland island sites. FIG. 13. (a) Quantile functions from the 25-yr 1972 96 period. Note that Gumbel and GEV are superimposed. (b) A comparison of the 13-yr GPD and exponential quantiles with the 25-yr GEV and Gumbel quantiles from (a) (dotted lines). gust speed tails, fail to predict maximum gust speeds accurately. The 0 versions of these models have less flexibility, and in this case, make a more accurate prediction of high speed winds, even when based on only 13 yr of data. This, of course, does not detract from the utilization of the generalized models as a useful parameterization for mixed climates analysis as described in section 3 of this paper (Abild et al. 1992; Gomes and Vickery 1978), but does caution against their use in quantile prediction with datasets as short as 13 yr. e. Decadal variability and longer time series In Fig. 14 we plot the annual maxima for the longer daily records for Benbecula, Kirkwall, Lerwick, and Stornoway Airport, provided by the British Atmospheric Data Centre (BADC). At the Sumburgh site we found that the average annual maximum gust speed during the 13-yr period from 1972 to 1984 was exceeded by each of the years in the period from 1989 to 1992. This apparent variability on a decadal timescale is also observed at the other four sites in this study. For example, the annual maxima at Lerwick, 25 km farther north on Shetland have enhanced maximum gust speeds in the 1989 92 period and an additional period of high speed annual maxima in the period from 1962 to 1967. This decadal variability in wind in the United Kingdom has been studied by Palutikof et al. (1987), where they state, In recent years we find that the years 1962 67 were dominated by westerly circulation patterns and display generally higher windspeeds over Britain, whereas from 1968 onward an increased frequency of anticyclonic days has been linked to lower windspeeds. A number of authors have suggested that this low-frequency variability at the decadal rate may be linked to the North Atlantic oscillation (Hurrell 1995). Using the 13- and 25-yr time series from Sumburgh we have demonstrated the difficulty in making quantile exceedance predictions using either a GEV or a GPD analysis in a wind climate with decadal variations. With an even longer time series at the neighboring site of Lerwick, we can now look further into the question of length of time series needed in GEV and GPD analyses to provide valid quantile predictions in a wind climate with this level of variability. Lerwick has an average annual maximum gust speed close to that of Sumburgh (39.8 m s 1 at Lerwick compared to 40.9 m s 1 at Sumburgh). We have carried out GEV and GPD at 13, 25, and 39 yr in an attempt to identify the point of quantile stability. The Lerwick 100- yr quantiles with errors are given in Table 3 and show that the large difference between 13 and 25 yr is not repeated when the analysis is extended to 39 yr, either for GEV or GPD. For example, the x 100 quantiles from a GPD analysis of 13, 25, and 39 yr are 47.3 3.4,

SEPTEMBER 2000 BRABSON AND PALUTIKOF 1639 TABLE 3. Lerwick: a comparison of 100-yr quantiles from GEV, Gumbel, and GPD analyses. The GPD analyses were done with a minimum separation of 1 day and threshold of 30 m s 1. Lerwick analysis type and period (m s 1 ) (m s 1 ) x 100 (m s 1 ) SE (m s 1 ) GEV, PWM with 13 yr of AM GEV, PWM with 25 yr of AM GEV, PWM with 39 yr of AM GPD, ML with 13 yr of POT GPD, ML with 25 yr of POT GPD, ML with 39 yr of POT Gumbel, PWM with 13 yr of AM Gumbel, PWM with 25 yr of AM Gumbel, PWM with 39 yr of AM 0.305 0.119 0.058 0.083 0.016 0.057 4.16 4.39 3.93 3.77 3.79 3.98 3.36 3.97 3.73 38.0 37.7 37.5 30.4 30.4 30.4 37.4 37.5 37.4 48.2 53.2 53.4 47.3 52.8 52.5 52.9 55.7 54.6 2.3 4.0 3.1 3.4 4.1 3.3 3.8 3.2 2.4 52.8 4.1, and 52.5 3.3 yr, respectively. A similar pattern is seen in Table 3 for the corresponding three GEV analyses, with similar errors. In both cases we find evidence that a 25-yr analysis yields stable quantile predictions at return periods of 100 yr. Summarizing this section, we find that although 13 yr of data are not sufficient to get stable quantile estimates, 25 yr or more do give stable quantiles within errors for both GEV and GPD analyses. 5. Conclusions This study examines extreme gust speeds at five island sites in northern Scotland. It focuses on hourly data from Sumburgh, in Shetland, but uses daily data from four similarly exposed Scottish sites to substantiate its conclusions. It has centered on three issues in extreme value analysis: the importance of event independence, the existence of mixed climate indicators, and the use of generalized extreme value distributions. As part of this study we have examined three methods for the determination of GPD parameters and have examined the sensitivities of each to choice of threshold, dead time, and length of the time series. The main conclusions can be summarized as follows. R Event independence: Using a simple method to extract the fraction of independent events in a time series, we find that correlated or nonindependent events in a gust speed time series have two specific effects on extreme gust speed quantile estimates. First, these events distort the shape of the gust speed distribution producing anomalously high gust speed quantiles, and second, the quantile standard errors, proportional to 1/ n, are underestimated when n, the total number of events, is inflated by the addition of nonindependent events. Also, for high thresholds, a 24-h dead time between events is sufficient to remove the effects of correlated events for these data. R Mixed climate: The analysis of the four relatively long (37 to 39 yr) daily time series at high average wind speed sites provides evidence that a single GPD model or simple combination of GPD components is not able to describe the actual wind climate at these sites. Categorizing extreme wind data from this region by GEV/ GPD type is therefore not easily done because the result depends on the threshold chosen. This result is seen in both a GPD analysis and an MRL analysis at these four sites. R Generalized distribution: The generalized models, GEV and GPD, have been brought to bear on 13 yr of Sumburgh gust speed maxima. Though very good at following the detailed shape of the 13-yr gust speed tail, when extrapolated, both fail to predict the actual maximum gust speeds observed over a longer 25-yr period. We attribute this failure largely to the apparent nonstationarity in the wind climate in this region. This does not detract from the utilization of the generalized models as a useful parameterization for mixed climate analysis as described in section 3 of this paper but does caution against their use as long-term quantile estimators with datasets as short as 13 yr. Using the longer time series from neighboring Lerwick, Shetland, we find that in this wind regime with apparent decadal variation, extreme quantiles have already stabilized for both GEV and GPD analyses of 25 yr. Acknowledgments. The daily wind speed data for Benbecula, Kirkwall, Lerwick, and Stornoway Airport were provided by the BADC. Also, B. B. Brabson wishes to acknowledge the generous hospitality and stimulating scientific environment of both the Climatic Research Unit and the School of Environmental Science at the University of East Anglia. REFERENCES Abild, J., E. Y. Andersen, and D. Rosbjerg, 1992: The climate of extreme winds at the Great Belt of Denmark. J. Wind Eng. Ind. Aerodyn., 41 44, 521 532. Coles, S. G., and D. Walshaw, 1994: Directional modelling of extreme wind speeds. Appl. Stat., 43, 139 157. Cook, N. J., 1982: Towards better estimation of extreme winds. J. Wind Eng. Ind. Aerodyn., 9, 295 323., 1985: The Designer s Guide to Wind Loading of Building Structures. Butterworths, 371 pp. Dargahi-Nougari, G. R., 1989: New methods for predicting extreme wind speeds. J. Eng. Mech., 115, 859 866. Davison, A. C., 1984: Modelling excesses over high thresholds, with an application. Statistical Extremes and Applications, J. Tiago de Oliviera, Ed., Reidel, 461 482.

1640 JOURNAL OF APPLIED METEOROLOGY VOLUME 39, and R. L. Smith, 1990: Models for exceedances over high thresholds. J. Roy. Stat. Soc., B52, 393 442. Dukes, M. D. G., and J. P. Palutikof, 1995: Estimation of extreme wind speeds with very long return periods. J. Appl. Meteor., 34, 1950 1961. Fisher, R. A., and L. H. C. Tippett, 1928: Limiting forms of the frequency distributions of the largest or smallest members of a sample. Proc. Cambridge Philos. Soc., 24, 180 190. Gomes, L., and B. J. Vickery, 1977: On the prediction of extreme wind speeds from the parent distribution. J. Ind. Aerodyn., 2, 21 36., and, 1978: Extreme wind speeds in mixed wind climates. J. Ind. Aerodyn., 2, 331 344. Gumbel, E. J., 1958: Statistics of Extremes. Columbia University Press, 375 pp. Gusella, V., 1991: Estimation of extreme winds from short-term records. J. Struct. Eng., 117, 375 390. Hannah, P., P. Rainbird, J. Palutikof, and K. Shein, 1995: Prediction of extreme wind speeds at wind energy sites. ETSU Rep. W/11/ 00427/REP, 250 pp. Harris, R. I., 1999: Improvements to the Method of Independent Storms. J. Wind Eng. Ind. Aerodyn., 80, 1 30. Hosking, J. R., and J. R. Wallis, 1987: Parameter and quantile estimation for the generalized Pareto distribution. Technometrics, 29, 339 349., and, 1995: A comparison of unbiased and plotting-position estimators of L moments. Water Resour. Res., 31, 2019 2025.,, and E. F. Wood, 1985: Estimation of the generalized extreme-value distribution by the method of probability weighted moments. Technometrics, 27, 251 261. Hurrell, J., 1995; Decadal trends in the North Atlantic oscillation: Regional temperatures and precipitation. Science, 269, 676 679. Kaminsky, F. C., R. H. Kirchoff, C. Y. Syu, and J. F. Manwell, 1991: A comparison of alternative approaches for the synthetic generation of wind speed time series. J. Sol. Energy Eng., 113, 280 289. Kirchoff, R. H., F. C. Kaminsky, and G. Y. Syu, 1989: Synthesis of wind speed using a Markov process. Proc. Eighth ASME Wind Energy Symp., Houston, TX, American Society of Mechanical Engineers, 17 22. Lechner, J. A., S. D. Leigh, and E. Simiu, 1992: Recent approaches to extreme value estimation with application to wind speeds. Part I: The Pickands method. J. Wind Eng. Ind. Aerodyn., 41 44, 509 519. Lieblein, J., 1974: Efficient methods of extreme-value methodology. National Bureau of Standards Report, NBSIR Rep. 74 602. Linacre, E., 1992: Climate Data and Resources. Routledge Press, 366 pp. Milford, R. V., 1987: Annual maximum wind speeds from parent distribution functions. J. Wind Eng. Ind. Aerodyn., 25, 163 178. Palutikof, J. P., P. M. Kelly, T. D. Davies, and J. A. Halliday, 1987: Impacts of spatial and temporal windspeed variability on wind energy output. J. Climate Appl. Meteor., 26, 1124 1133., B. B. Brabson, D. H. Lister, and S. J Adcock, 1999: A review of methods to calculate extreme wind speeds. Meteor. Appl., 6, 119 132. Pickands, J., 1975: Statistical inference using extreme order statistics Ann. Stat., 3, 119 131. Ross, W. H., 1987: A peaks-over-threshold analysis of extreme wind speeds. Can. J. Stat., 15, 328 335. Simiu, E., and N. A. Heckert, 1996: Extreme wind distribution tails: A peaks-over-threshold approach. J. Struct. Eng., 122, 539 547. Smith, R. L., 1984: Threshold methods for sample extremes. Statistical Extremes and Applications, J. Tiago de Oliviera, Ed., Reidel, 621 638., 1986: Extreme value theory based on the r largest annual events. J. Hydrol., 86, 27 43., 1989: Extreme value theory. Handbook of Applicable Mathematics, W. Ledermann, Ed., John Wiley and Sons, 437 472., and I. Weissman, 1994: Estimating the extremal index. J. Roy. Stat. Soc., B56, 515 528. Stedinger, J. R., R. M. Vogel, and E. Foufoula-Georgiou, 1993: Frequency analysis of extreme events. Handbook of Hydrology, D. R. Maidment, Ed., McGraw Hill, 18.1 18.66. Walshaw, D., 1994: Getting the most from your extreme wind data: A step by step guide. J. Res. Nat. Inst. Stand. Technol., 99, 399 411. Weissman, I., 1978: Estimation of parameters and large quantities based on the k largest observations. J. Amer. Stat. Assoc., 73, 812 815. Wilks, D. S., 1995: Statistical Methods in the Atmospheric Sciences: An Introduction. Academic Press, 467 pp.