#.“.”

AIAA-2001-0044

Extreme Load Estimation for Wind Turbines:Issues and Opportunities for Improved Practice

Paul S VeersWind Energy Technology Department

Sandia National LaboratoriesAlbuquerque, NM 87111-0708

psveers@,sandia.~ov505.844.5522

Sandy ButterfieldNational Wind Technology Center

National Renewable Energy LaboratoryGolden, CO 80401-3393

sandy butterfield@,nrel.~ov303.384.6902

ABSTRACT

Current design load estimation procedures for windturbines often do not accurately treat the statisticalnature of loads. Current practice for wind turbine loadanalysis is reviewed. The authors’ opirdons on theshortcomings of these practices are discussed.Experience gained from recent research on statisticalload extrapolation methods is reviewed. Statisticalmodeling techniques are presented. Open questions oncurrent tecludques are summarized and critical issuesthat need to be resolved for art accurate statistical loadextrapolation method are discussed.

INTRODUCTION AND BACKGROUND

Stochastic Environment

Wind turbines must be designed to operate in a verystochastic enyiromnent for at least 20 years accordingto International Electrotechnical Commission (IEC)standards, In addition to the cyclic nature of loadsinduced by their own inertial effects, loads result fromspatial and temporal changes in wind spee~ direction,shear and vorticity. ‘Ilk has challenged designerk formany years. Inhially designers felt this level of detailin wind modeling was impossible and unnecessary.Very simple techniques were used in the late 70s andearly 80s. These tecludques worked when the designswere very simple and conservatively designed or thewind conditions were benign. As wind turbines becamelarger it was too expensive to use large safety margins.They were also being installed in very turbulent sites.

*‘Hispaperis declareda workof the U.S.Governmentandis notsubjectto copyrightprotectionin the UnitedStates.Sandiais amultiprograrnlaboratoryoperatedbySandiaCorporation,a LockheedMartincompany,forthe U.S.Departmentof EnergyundercontractDE-AC04-94AL85000.

Failures caused by inaccurate estimation of designloads mandated more accurate prediction techniqueswhich did account for more detail in the inflow.

Detailed structural dynamic models were developed andbecame the workhorses for the wind industry by themid 1990s. Included in these computer codes wereturbulence models, which simulated stochastic inflowfields, aerodynamic models, which predictedaerodynamic loads from the turbulent inflow, andcontrol algorithms, which commanded pitch, yaw, andbraking actions. The aerodynamic loads were appliedto the structural dynamic model which was then run in atime marching fashion. Figure 1 shows a flow chart ofthe general analysis models. With this approachdesigners finally had tools which could simulate all theimportant operational features of the entire windturbine, even the control system.

One can axgue about the accuracy of the various modelsthat make this overall system of models, but even intheir current state they provide a f%rmore accurate andpower tool than ever before. Armed with these newtools the designer must decide how to use them. He/sheis now able to simulate ahnost any wind andoperational condition. The designer is still fhced withestimating the fatigue life and peak loads over 20 years.This implies running 20 years of computer simulationswhich, at the present time, only run near real time.Obviously tlds is not practical. How should a subset ofsimulations be used to extrapolate to a representative 20year spectrum of loads?

..... .... . . . —..-..,- ,--y, .W / ~. -. ,.,4 ,., . . . . . ,., . . . . . . . . , ..,- --

DISCLAIMER

This report was prepared as an account of work sponsoredby an agency of the United States Government. Neitherthe United States Government nor any agency thereof, norany of their employees, make any warranty, express orimplied, or assumes any legal liability or responsibility forthe accuracy, completeness, or usefulness of anyinformation, apparatus, product, or process disclosed, orrepresents that its use would not infringe privately ownedrights. Reference herein to any specific commercialproduct, process, or service by trade name, trademark,manufacturer, or otherwise does not necessarily constituteor imply its endorsement, recommendation, or favoring bythe United States Government or any agency thereof. Theviews and opinions of authors expressed herein do notnecessarily state or reflect those of the United StatesGovernment or any agency thereof.

“ a- .,--7- %5.,7-YES... . .. . ., .T-—r--=.=. . . .. . . . . . . . . . . . . w,... . . ..U-J’. .m+a?.,, ,-&,,.., ,Lm , ;. , y .,, —- * .=. ,- .;. - :~ .,,--- ?---

DISCLAIMER

Portions of this document may be illegiblein electronic image products. Images areproduced from the best available originaldocument.

LoadProcessing IaCOnlrOls

Figure 1. Typical Turbine Model Flow Chart

Tv~ical Load Prediction Process

To address the problem of extrapolation to 20 years offatigue loads and extreme loads designers typically usethe models to simulate Ioadmg conditions over a rangeof operational and extreme wind speeds. Fault statesare also simulated. Generally only one ten-minutesimulation is run for each condition and state. In theevent of extreme conditions a short simulation is runusing a discrete wind gust model.

Wmd speed probability density iimctions are used toestimate the duration of time that each wind speedcondition will occur over the life of the turbine (lowfrequency variations in the wind). Turbulence modelsare used to represent the short-term variations (highfrequency end of the spectrum). A ten-minute loadsimulation is assumed to represent the dynamicresponse to the spatial and temporal character of thewind for a mean wind speed bin. Ten-minutesimulations are run for a range of mean winds. If itwere practical simulations would be run for 20 years ofturbine life using an accurate probability densityfunction to represent the low frequency wind variabilityand turbulence models with a characteristic distributionof turbulence intensity to represent the high frequencyvariability. In practice the results of one ten minutesimulation for each wind speed bin are cycle countedand the number of cycles are multiplied by the 20 yearduration. These discretid cycle matrices are summedover the fidl range of wind speeds for a completefatigue load distribution.

Extreme loads are estimated by simulating extremeinflow conditions. Both a 50-year extreme ten minuteaverage wind speed with turbulence and an extremediscrete wind speed model are used. The highest loadtlom among all conditions is used for design andcertification purposes. When the highest load comesfrom a turbulent wind simulation, it is sometimes usedwithout further statistical extrapolation.

This method is flawed in at least two ways. Firstj itdoes not account for rare events, which will begenerated as longer turbulence simulations are run,essentially filling in the tails of the stochastic inflowdistributions for each wind speed bin. SeconL becausethe simulations do not capture accurate extreme

statistics they do not estimate the peak loa~ which willoccur over the operating life of the machine. There arenot enough data to make reliable estimates.

The fatigue simulations capture the essentialaerodynamics, structural dynamics and control featuresbut lack the statistical depth needed to establishstatistical significance for the lifetime load predictions.Accurate estimation of uncertainties, statisticalsignificance and confidence in extreme load estimationis needed to build reliable machmes.

Even if statistical confidence could be established thedesigner has to choose a set of conditions that willcover the range of sites where the machme is likely tobe installed. Standards have been developed to assureconsistency of design-load conditions across theindustry.

Desipn Requirements (lEC Standard~

The International Electrotechnical Commission (IEC)standards formaliid design requirements. While theydo not specifically prescribe the design methodsdescribed above they are based on the assumption thatdesigners will use these methods. Their goal is toprovide a consistent set of models and parameters fordesigning machines to an implicit level of robustness toachieve hQjh reliability at many economically attractivesites. If the design methods are incorrect turbines canbe over designed or under designed with respect to siterequirements.

Turbulence Simulations

IEC 61400-11 specifies design classes with associatedannual average wind speeds, turbulence intensities andextreme wind speeds (Table 1). Turbulence models arealso specified. With these parameters, models, andother environmental conditions, analysts are presumedto have the essential “external conditions” needed toestablish their design conditions.

The IEC standards do specify target levels ofprobability and confidence for load predictions but donot specifj how many turbulence simulations areneeded to establish these levels. They do not provideany incentive to increase confidence in the loadpredictions. The methods for achieving specified levelsof confidence are not clear. Minimum safety factors arecodified to account for uncertainty in the load-prediction process and material properties but they areunproven aud they are not intended to account forstatistical uncertainty implied by fhite duration datasamnles.

#

WTGS I II HI Iv sClass

V,ef (m/s) 50 42,5 37,5 30

v ave (m/s) 10 8,5 7,5 6 Values

to be

A 115 (-) 0,18 0,18 0,18 0,18 specified

a(-) 2 2 2 2 by the

B 115 (-) 0,16 0,16 0,16 0,16 designer

a(-) 3 3 3 3*wnere:

. the values apply at hub height,

. A designates the category for higherturbulence characteristics,

● B designates the category for lowerturbulence characteristics,

● 115is the characteristic value of theturbulence intensity at 15 m/s,

● a is the turbulence slope parameter

Table 1- Basic parameters for WTGS classes

Discrete Models

IEC offers discrete models in an attempt toacknowledge that the turbulence simulations do notcapture the extreme load conditions accurately. Thesemodels include a 50 year extreme gust profile, shear,direction change, and several other wind conditions.Time dependent magnitudes are specified so that thetransient response of the wind turbine can be simulated.These models are design class dependent.

It is not clear that these models accurately represent thetails of the distribution. Confidence in the models isderived from experience; their use has resulted inmachines that demonstrate high inhial reliability. Asthey are currently defied they are not related to theimplied extremes of stochastic turbulence dktributionand therefore could represent an inconsistency betweenthe discrete, event-based and turbulence-based models.

Need for improved load estimation methods

The design community has grown to mistrust thediscrete models and rely heavily on turbulencesimulations for establishing design loads. Improvedmodeling power has given designers the basic toolsneeded to more accurately represent the stochasticnature of their loads. But there is little guidance onhow to use these tools to achieve the full range of lifetime fatigue loads. Given the current state ofknowledge and modeling power, the design loads are

likely to be best determined from exhaustivesimulations of the turbulent inflow and subsequentstatistical analysis of the turbine response. In thatevent the design community will need to base loadestimates on a broad range of turbulent inflowconditions. It is also doubffil that there can be a singlewind condition (combination of mean and turbulencelevel) that will be a worst case for all Iypes of turbines.

Statistically robust methods of load estimation andfatigue life extrapolation would allow designers tointelligently match their designs to the stochastic designenvironment. They would enable them to quanti& theconfidence of load predictions and increase themthrough more extensive simulations. They would allowextreme load extrapolations.

As the wind industry becomes more sophisticated, theywill be able to apply more accurate load estimationtechniques to reduce unnecessary conservatism. Windenergy may soon be competitive with fossil fiels if itcan reduce its cost by 20%-30V0. These reductionsmay be possible simply through more accurate designmethods and consequent reductions in safety factors.

The goal is to develop a set of design methods thatincorporates the stochastic nature of the designenvironment over the intended life of the machine.These methods must be compatible with existingstructural dynamic simulation tools. They must bepractical, verifiable and accurate. They must beinternally consistent from fatigue load estimation toextreme load estimation.

Other industries have already embraced statisticaldesign methods that are compatible with the stochasticnature of their design environment. Offshore structuresare subjected to stochastic loads. Earthquake loads arestochastic. Designers in these industries use statisticaldesign methods to achieve cost effective designs. Thewind energy industry needs similar tools.

RECENT EXPERIENCE AND DIRECTION

There has been some recent progress in examining theway the standard design loads are derived and whatapproaches might be better. Madsen* has begun the re-examination of the extreme loads with a look at how theIEC loads compare with statistical extrapolations ofparked and operating loads. It is clear that prescribedextreme events do not always result in the highestloads. Ronold has published studies of both extremeload3 and fatigye4’5 reliability, proposing new loadmodeling approaches that lead to the ability to extractreliabilities (probabilities of failure) from the design

data. Methods for statistical modeling of the loads havealso been explored by WintersteinG’7 Ad hoc cases ofstatistical models populated the European Wind EnergyConference in recent years!-12 All of this activityhighlights the need to apply statistical methods to windturbine design loads in ways that improve the designprocess. The level of uncertainty in each applicationshould drive the design loads.

Summarv of what we think we have learned

Deterministic load cases such as those specified in theDesign Load Cases (DLCS) in the IEC Safety Standard*will always provide useful checks on a design. Table 2lists the various load cases. Load cases are acombination of external conditions and operatingconditions. Many of the critical DLCS include discreteexternal conditions (e.g. the 50 year extreme operating

gust EOG50) rather than normal turbulence models(NTM). Deterndnktic evaluations have utility in theirown right. They need to be used in addition to full-fieldturbulence models. They are much more importantthan being a mere convenience. They are also likely tocontinue for a long time as a primary tool for designevaluation because there is a difference between designevaluation and the process of design itself.Deterministic calculations will remain important inproviding spot checks, which a design evaluator canmake independently. They may, however, assume alesser importance in driving future turbine designs asstatistical extrapolation methods become more widelyused.

It is simply not possible to construct a set of load casesthat will be equally applicable to all wind turbinearchitectures and control implementations. Thus, whilethe deterministic load cases may produce an onerouscondition for one turbine, they may be relatively benignfor another. The ability to evaluate the turbine loadresponse with respect to realistic wind inflowconditions is already well established. It is also notclear that a set of deterministic load cases can beconstructed to adequately represent every possibleinflow event for all potential turbine sites. Therefore, itis highly likely that simulated response to turbulentinput winds will be used to calculate response andstatistical methods will be used to translate the short-term response time series into long-term design loads.

There are two distinctly different aspects of designanalysis that need to be treated separately, either ofwhich can provide the controlling design loads for aparticular component within a particular system. Theparked extreme wind loads must be treated differentlyfrom the operating loads

Parked loads are the simplest to analyze because thehighest loads have usually been assumed to occur in thehighest winds. Madsen, et al: have thrown some doubton that assumption by showing the large variability inextreme loads at a given wind condition. Parked loadsalso seem to be the easiest to fit to statistical models,being no different than the response of any stationarystructure to turbulent winds. The extreme wind Ioadmgcase (e.g., 50 year maximum 10 minute average windspeed) can be specified. Once specifie~ the meanextreme load (average of all possible realizations of this50 year maximum condition) can be estimated frommultiple simulations. It is of coume assumed that thisload case will never be derived fkom measurementsbecause the 50 year extreme will never be measuredwithin the duration of a ~ical prototype evaluationprogram. The parked case is therefore a matter ofspecifjhg the controlling parameters of the worstinflow condition and estimating the mean turbineresponse to that stochastic inflow. It is therefore ahnostexactly analogous to the case of offshore structuresloaded by storms, an application for which the designcriteria are already quite advanced.13

Loads during operation raise an interesting difficulty.As wind speed decreases, most system loads will alsodecrease from a maximum that might occur at either thehighest operating wind speed or around rated windspeed. As wind speed decreases the amount of time atthe wind speed increases. The more time spent at aparticular operating condhion, the greater theopportunity to experience higher response levels,farther out in the tails of the distribution. Therefore,one can not assume that the highest loads will begenerated by the highest winds. Instea4 it is necessaryto combine the responses at all wind speeds into asingle long-term distribution of load extremes. Onlytlom this combined distribution can the highest loadover the entire design lifetime be extracted.

The extreme operating load therefore begins toresemble the fatigue load case where the long-termdistribution of fatigue load cycles is required tocalculate the total damage over the design lifetime.Although the load quantities are different (peakresponse for extreme load estimation and rainflowranges for fati~e) the statistical modeling problemlooks similar. In both cases, it is necessary todetermine a short-term response distribution conditionalon the input wind conditions (currently speed andturbulence intensity). Then the short-term response isintegmted over all wind speeds to generate the long-term distribution using the “total probability theorem.”

4

-~-,-.,T~--

, -....- ,.,.,.7 ..m, .--------

Design situetion DLC Wind condition) Other conditions Type of Partialanalysis safety

factors

1) Power production 1.1 NTM Vh.b = Vr I I u I Nor VOut

1.2 NTM Vln< Vhub I I F I●<Vout

1.3 ECD ~huh = V, I IUIN

1.4 NWP Vh”b= v, I External electrical fauitor Voti

u I N

I

1.5 EOGI Vh.b = Vr I Loss of electricalor Vout connection

u I N

I

u I N1.6 E0G50 Vh.b = Vr

or Vout I1.7 EWS Vh.b = v,

or Vout

EDC50 Vhub = Vr

or VOut

ECG Vh.b = Vr

1

u I N1.8

u IN1.9

2) Power productionplus occurrence of fault

2.1 NWP Vh”b = V, I Controi system fauit I u I Nor Vout

2.2 NWP Vh”b = v, Protection system oror Vout preceding internai

electrical fault

u A

2.3 NTM Vin< Vh.b I Controi or protection I F I●<V.”t svstem fauit

3) Startup 3.1 NWP Vin< Vhub F ●

< Vout

EOGI Vhub = Vin, u NVr or Vout

EDCI Vhub = Vln, u NVr or Vout

NWP Vin < Vhub F ●

< Vo”t

EOGI Vh.b = Vr u Nor Vout

3.2

3.3

4.1

4.2

5.1

6.1

6.2

7.1

4) Normai shut down

5) Emergency shutdown

NWP Vhub = Vr I I u I Nor VOut

6) Parked (standing stiilor idling)

EWM Vh”b = Possible loss of u NVeso electrical power

network

NTM Vh”b<0.7 F ●

Vref

EWM Vhub = Vel u A7) Parked and faultconditions

1) if no cut-out wind speed V..t is defined, the Value of V~f should be used.

Table 2 IEC Design Load Cases (Ref. 1)

5

The practice of defining fatigue loads empirically, withhistograms of the number of cycles in each loadamplitude bin, has been used for this purpose.14”15Theresulting long-term distribution is derived by asummation over all the (usually discrete) wind speedbins.

N(L,),ong.,em =xl’U~,)sho.-tmlYlQV/

N&),~Ofl.h is the number of cycles per

(1)

fixed timepefi6d of load amplitude L, at av&age wmd speed U,P(fi) is the probability that the winds are in the z windspeed bin and N(Lj)I..&ta is the resulting long termnumber of cycles at load amplitude Lfi P(Vj) is foundby integrating the probability density fhnction of windspeed, j(V), over the width of the VIbin.

This empirical approach has some definite weaknesses.The first is difficulty in determining the high-amplitudetail of the load distribution. The fatigue damage inmaterials typically used in wind turbine blades have ahigh fatigue exponent that leads to fatigue damagegoverned by the high-amplitude tail. The empiricalapproach requires an enornious amount of data todefine the tail of each short-term distribution. Thesecond difficulty is in translating the data from one siteinto a load distribution at another site with differentinflow characteristics. Finally, there is no systematicway to make use of increased information, i.e., betterdescriptions of the loads based on a larger data base.Ideally, more data should lead to higher confidence inthe loads resulting in reduced design margins.

Statistical modeling can improve design load estimatesfor both fatigue and ultimate load applications. First astatistical model can be used to extrapolate from theexisting data to more rare, higher-amplitude events (notperfectly, but with calculable uncertainty). Second, themodels can be used to parametrize the response withrespect to wind conditions so loads in other windconditions can be estimated. And finally, the statisticaluncertainty in the models can be used as a basis ofimproved design Ioad standards by accounting for theuncertainty based on the amount of loads data used toestimate the extremes and fatigue load spectra.

Statistical Modeling

Statistical modeling uses probability distributions todescribe the loads data at each set of short-term windinput conditions. The distribution fictions are definedin terms of parameters which in turn are derived fromcertain statistics of the dam usually mean, variance,

Pand perhaps other hi er statistical moments (i.e.,

3’5This approach can be calledskewness and kurtosis).

parametric based on its need to define the parameters ofdistribution fimctions.

Calculating long-term distributions is similar to Eq. (l),except that there is no need to discretize the calculation.The resulting continuous long-term probabilitydistribution fimction, F(L)long.tm is derived from thecontinuous short term distributions, whose parametersare defined as fimctions of wind spee~ as follows.

W),ong-,enn = jWI %..+J(WVv

(2)

Equation (2) in various forms appears throughout thepapers in the Design Loads Estimation special sessionof the 2001 ASME Wind Energy Symposium. It is thebasis for load estimates in the offshore oil business,13for extreme loads on wind turbines16”7 and for fatigueloads.18 In a tower clearance study, Laino19 highlightsthe value of statistical analysis while illustrating thedifficulty of doing so without the use of Eq. 2. It isfoundational to a systematic way of describing loadsthat depend on a distribution of environmentalcondhions such as wind speeds or wave heights. Forextremes, the distributions are of extreme values whilefor fatigue, the distributions are of rainflow countedamplitudes. Equation 2 applies to both. Of coume, Eq.2 does not explicitly account for the frequency ofoccurrence of the loads in question. The number of loadevents is derived from the time of exposure and thefrequency and will determine the probability level atwhich the design load is evaluated.

Wind speed alone is not capable of describing theloading response. Turbulence is also responsible fordetermining the response intensity of a wind turbinerotor. Since the two are relate~ it is difficult to sayexactly which is more importan~ but is safe to say thatboth are crucial. Existing approaches incorporateturbulence by prescribing a functional relationshipbetween wind speed and turbulence standard deviation,which permits the continued use of Eq. 2. Byspeci@ng V, the related turbulence level, T, isautomatically defined. This often results in the need todefine the turbine response at turbulence levels thatmay never have been measured.zo

Alternatively, the response parameters can be mappedto both wind speed and turbulence level through somesort of regression. Equation (2) is also transformed intoa double integral over both environmental inputs.

I Blade out-of-plane bending moment (kNm) I

I Simulationno. ILoad extremes (10 mln tuna)

I C.35T. ( I

I Simulation no. I

Rotor pltchlng moment (kNm)

Figure 2Variability of Extreme Loads from 100Independent Ten Minute Simulations.

Maximum valuas of out-ef-plane blade bending moment-oparetion at 14mls

100 Wmintimseries

1,00m0,s0 -*0,60

~ 0,40 R

0,20 ,,:”,:...? ...

..’ .-0,00

3,25 3,50 3,75 4,02 4,25 4,50 4,75 5,WBendingmemantInkNm

Figure3 Extreme Value Model Fit to MaximumLoad Simulations

While it is known that statistical modeling, parametricloads definitions and explicit inclusion of turbulence inthe loads definitions offer significant advantages, thereare still several unknowns that need to be resolved.

Thinm we do not know

There are a number of details of statistical modelingand analysis that need to be resolved to demonstratethat statistical models can be used constructively toproduce load estimates at prescribed confidence levels.Some of the papers in this session go a long way toimproving the state-of-the-art in this area.

Figure 2 illustrates how the maximum value of each tenminute simulation will change by only changing theseed for the turbulence model. Obviously a singlesimulation does not form a good basis for extrapolationto a 20 year life. Madsen et al.2 showed that a linearstructure (a non-operating turbine) obeys an extremevalue type 1 model, and shows that it works pretty welleven for the non-linear system of an operating turbine(Figure 3).

On the other hand recent attempts to apply full responsemodels to extreme value estimation for operating loadshave not worked well.z6’21 These models attempt toincrease the statistical certain~ by using more data (allthe time series instead of just the local extremes) toextrapolate to the long-term extremes. It turns out thatthe cyclic Ioadmg due to gravity and wind shear makesit difficult to apply a statistical model. It is better to usethe extremes of the loads to predict ultimate loads.Exactly which extremes have not yet been resolved.Candidates include the maximum load in ten minutes,maximum for some other specified time perio~maxima between mean-level crossings, maximabetween some higher-level crossings, and maximum ineach revolution. In the case of fatigue loads, rainflowcounted cycle amplitudes are uniformly used. However,this choice too may be questioned if it bringsunnecessmy difficulties into design loads estimationwithout demonstrable added value.

We do not yet know what exact choice of statisticalparameters (moments, parameters of probabilitydistribution fimctions, etc.) should be used to define theturbine response as a function of input conditions.Although there are slight advantages to some choicesover others, the chosen parameters are not likely to beof central importance. However, good choices shouldresult in miniial bias and reduced uncertainty. It islikely that the best parameters will be independent ofprobability distribution type and statisticallyindependent. Central moments are front runners at thktime.5’6 However, any response quantities that can beused to derive extreme loads or define fatigue-loaddistributions are candidates for statistical modeling.

Regression over wind speed and turbulence level willmost likely be required to create the map between inputconditions and turbine response. The exact nature of

the regression will likely have an influence on theresults if particularly poor choices are made. However,it is likely that any relatively good regression model,that is any model that fits the data well, should produceequivalent results.

Lastly, the turbulent inflow description is liiely to bevery important in specifjhg the loads both for fatigueand extremes. There is growing consensus that thecurrent approach in the standards of using average windspeed and a single (safe) value of turbulence intensity isinadequate. We need to at the very least account for thevariability in turbulence levels or we are likely to missthe extreme events.

There seems to be fair agreement that the use of the raw10-minute turbulence intensity to describe theturbulence is also inadequate, since it can not possiblycapture the nature of the inflow important to windturbine response (including frequency conten~ spatialvariability, lateral and vertical wind components, etc.).The MOUNTURB programz has identified the along-wind turbulence standard deviation as the mostimportant factor in predicting turbine response, butthere are a host of other factors that also play a role.These factors include the cross-wind and vertical windspeed standard deviations, length scale, coherencedecay factor, and average shear. Recent work byKelle# has shown the importance of the meanshearing stress and a@ospheric stability in turbineresponse. Filtered turbulence standard deviation thatfocuses on high tlequency turbulence content has alsoshown some promise.% Additional experiments andassociated research need to be done to determine howto describe the inflow to best correlate to both extremeand fatigue loads on with turbine structures.

CRITICAL ISSUES FOR FUTURE WORK

Perhaps the most crucial issue is how the characteristicvalues and partial safety factors will be defined toadequately cover the uncertainties while allowing thediligent designer to lower margins as fm as possible byapplying better and better load estimates. (Similarly,resistance estimates based on improved or additionalinformation should also be rewarded with smallerdesign margins.) The design standards need to have astructure capable of using the statistical analysis of theloads data to generate associated design margins.

There is currently no value incorporated in thestandards for expending the effort to get additionalsimulations or field measurements. In some cases it is

19 perhaps the characteristic load couldjust the reverse.be tied to a specified confidence level, and thisconfidence level determined from the statistical analysisof the loads. Then it should be possible to reduce the

design load with additional data (assuming that theadditional data does not reveal a load that had beenunderestimated due to the smaller data se~ which for a95’%.confidence level should only occur once every 20tries). The safety factors could then cover the non-statistical uncertainty relating to the host of factorsinfluencing design loads, ranging from numericalmodel (or measurement) error to various environmentalunknowns.

Because current standards are based on past experienceand industry consensus rather than objective, risk-basedanalysis, it may be dangerous to remove conservatismfrom one area without also checking elsewhere. Oneconservatism may be covering for an unknown lack ofconservatism elsewhere in the design process. Ingeneral, the current standards give a load calculation“recipe” that results in some specific reliability level. Ifthese current reliability levels are deemed adequate onaverage (over various cases), one cannot reduceconservatism in turbulence specification withoutadjusting the recipe to compensate elsewhere; e.g.,through use of a higher load factor. Note that a morestatistically based alternative procedure may result inmore uniform reliability across a range of machinetypes and site characteristics. An important unknown iswhat probabili~ of failure is currently achieved throughthe experience-based design rules. Calibration studies(e.g., Ronol& et al~-s) should be done to estimate theimplicit safety level produced by current standards.

A very important issue for &her research is the waythe standards describe the turbulence of particular sites.It is quite clear that the raw 10-minute turbulenceintensity has no hope of being an adequate descriptor ofthe site inflow characteristics. The required researchwill need to include detailed measurements of both theinflow field and turbine response. The data must beanalyzed to determine the inflow statistics that governthe load response statistics. Site characterization workwill then have to include summaries of the criticalstatistics. It may be an additional burden, but couldhave a significant payoff in reduced cost of energy fromparticular sites.

Site-specific design becomes more attractive as thesoph~tication of the industry increases to the pointwhere individual components can be reinforced toaccount for particularly intense site-driven loadings.Alternatively, less expensive components could besubstituted in particularly benign locales. There areltiltations to site specific design. There are two,sometimes conflicting, purposes of the standard thefirst is to provide minimum standards of practice whichspeci~ the necessary elements to consider in designand the second is to provide a set of external condition

8

-Vxr7T T.-. —-.—-.. ..

.,.

parameters that lead to turbines of consistent robustnessincomparable applications. Thesecond purpose is notsewed by having only site-specific designs. The IECstandard is more of a product standard than a code ofpractice although there are elements of both in it. As aproduct standar~ it is entirely appropriate to have a setof arbitrary design conditions that must be used. Thebroad site classes used in current standards may providesignificant value as a product standard. Thus they aidin reducing cost through increased manufacturingvolume that comes from standard products. However,it is worth further investigation to determine the effectof standard class definitions on wind turbine cost.

SUMMARY

The process of estimating design loads can besubstantially improved through the use of statisticalmodels. Standard definitions of characteristic loads andassociated partial safety factors should reflect the levelsof uncertainty calculated using the statistical models.The details of the models may be less important thanthe construction of the safety standard. Thecharacteristic loads should be estimated by 1) finding agood probability distribution model for the short-termresponse at given input conditions, 2) detining therelationship between controlling parameters of the shortterm distributions and the input conditions, 3)integrating over the input distributions to get long-termdistributions, 4) quantifying the uncertainty in theestimate of the long-term d~tribution, and 5) applyingthe uncertainty to determine the loads at a specifiedconfidence level. Perhaps the details of each step maybe Iefi to the individual application. Although thisprovides a good first step, the bigger issues are likely tolie in the less quantifiable area of improved partialsafety factors. As odd as it may soun~ it is not possibleto improve the safety factors without basing them on afirm foundation of statistical uncertainty analysis.Otherwise the safety factors themselves will need toprovide the largest margins required to account for theworst possible job of loads estimation. We can dobetter than that.

REFERENCES

1. IEC/TC88, 61400-1 Wind Turbine GeneratorSystems - Part 1: Sajety Requirements,International Electrotechnical Commission (IEC),Genev% Switzerland 1998.

2. Madsen, P. H., Pierce, K. and Buhl, M.,“Predicting Ultimate Loads for Wind TurbineDesign? A collection of the 1999 ASME WindEnergy Symp., at the 444 Aerospace SciencesMtg., Reno, Neva~ AL04-99-0069, January1999, Pp. 355-364.

3. Ronol& K. O. and Larse% G. C., , “ReliabilitY-based design of wind-turbine rotor blades againstftilure in ultimate loading: Elsevier, Engineering.Structures 22,2000, pp. 565-574.

4. RonolL K. O., Wedel-Heinen, J., and Christensen,C. J., “Calibration of Partial Safety Factors forDesign of Wind-Turbine Rotors Blades AgainstFatigue in Flapwise Bending? 1996 EuropeanUnion Wind Energy Con$, Goteborg, Sweden, 20-24 May 1996.

5. Ronol& K. O., Wedel-Heinen, J. and Christensen,C. J., “Reliability-based fatigue design of wind-turbine rotor blades: Elsevier, Engineering.Structures 21, 1999, pp. 1101-1114.

6. Wintenstein, S. R and Kashefi T., “Moment-BasedLoad and Response Models with WindEngineering Applications: A collection of the1999 ASME Wind Energv Symp., at the AMAAerospace Sciences Mtg., Reno, Neva& AIAA-99-0068, Jauuary 1999, pp. 346-354.

7. Veers, P. S. and Winterste@ S. R, “Application ofMeasured Loads to Wind Turbine Fatigue andReliability Analysis: Journal of Solar EnergvEngineering Trans. of the ASME, Vol. 120, No. 4,November 1998.

8. Cheng, P. W., Probabilistic Approach of ExtremeLoading for Offshore Wmd Energy Converters:1999 European Wind Energy Con$, Nice, France,1-5 March 1999, pp. 228-231.

9. Petersen, S. M., Larsem G. C., Antoniow I., Lin&s. o., and Courtney, M, “ExperimentalInvestigation of Ultimate Loads: 1999 EuropeanWind Energy Con$, Nice, France, 1-5 March 1999,pp. 199-202.

10. Andre% A. Argyriadis, K., and Folhichs, U., NewIEC 61400-1 and Site Conditions in Reality: 1999European Wind Ener~ Con$, Nice, France, 1-5March 1999, pp. 593-596.

11. Enevoklsen, P., Stiesdal, H., and Vinther, S. “HowDo We Measure the Real Loads on Windturbines?: 1999 European Wind Ener~ Con$,Nice, France, 1-5 March 1999, pp. 620-623.

12. Rono14 K. O. and Larsen, G. C., “Variability ofExtreme Flap Loads During Turbine Operation:1999 European Wind Energy Conj, Nice, France,1-5 March 1999, pp. 224-227.

13. Haver, S., “Application of Stochastic Methods inStructural Desigm The Offshore Experience; Acollection of the 2001 ASME Wind Energy Symp.,at the 4L4 Aerospace Sciences Mtg., Reno,Nevad& AIAA-2001-O043, January 2001.

14. McCoy, T. J., Malcolm, D. J., and Griffin, D. A.,“An Approach to the Development of TurbineLoads in Accordance with IEC 1400-1 and 1S02394/ A collection of the 1999 ASIUE WindEnergy Symp., at the 444 Aerospace Sciences

9

..... .: ~:,{, ., ;, -— —---—-- -- .-

##.

15.

16.

17.

18.

19.

20,

21.

22.

23.

24.

Mtg., Reno, Neva& AIAA-99-0020, January1999, pp. 1-9.Antonio% I. and Peterse% S. M., “Wind TurbineTesti Structural Loads ELKRAFT lMW (StallRegulated Operation): RisO-I-865(EN), RisONational Laboratory, Roskilde, DenmarkDecember, 1995.Cheng, P. W. and Bierbooms, W., “Extreme GustLoading for Wmd Turbines During Operatio%” Acollection of the 2001 ASME Wind Energy Symp.,at the AMA Aerospace Sciences Mtg., Reno,Nevati AIAA-2001-O045, January 2001.Fitzwater, L. M. and Winterstein, S. R, “PredictingDesign Wind Turbine Loads from Limited DatiComparing Random Process and Random PeakModels: A collection of the 2001 ASME WindEnergy Symp., at the AMA Aerospace SciencesMtg., Reno, Neva@ AIAA-2001-O046, January2001.Manuel, L., Veers, P. S., and Winterstein, S. R,“Parametric Models for Estimating Wind TurbineFatigue Loads for Designfl A collection of the 2001ASME Wind Energy Symp., at the AL4A AerospaceSciences Mtg., Reno, Neva@ AIAA-2001-O047,January 2001.Laino, D., “Statistical Analysis of Wind TurbineRotor Tower Clearance:’ A collection of the 2002ASME Wind Energy @rep., at the ALU AerospaceSciences Mtg., Reno, Nevad% AIAA-2001-O046,January 2001.IECITC88, Drajl IEC 61400-13 l% Ea! 1: Windturbine generator systems – Part 13: Measurementof mechanical Ioaak, 88/120/CDV, InternationalElectrotechnical Commission (IEC), Genev%Switzerland, 1999.Laino, D. J. and Pierce, K. G., “EvaluatingStatistical Loads Extrapolation Methods: Acollection of the 2000 ASA4E Wind Energy Symp.,at the ALU Aerospace Sciences Mtg., Reno,Neva@ AIAA-2000-O064, January 2000, pp. 424-432.MOUNTURB, Load and Power MeasurementProgram on Wind Turbines Operating in ComplexMountainous Regions, Volumes. I - III, Editor P.Chaviaropoulos, Coordinator A. N. Fragoulis,CRES, RISO, ECN, NTUA-FS, published byCRES, Pikermi, Greece, November 1996.Kelley, N. D., “A Case for Including AtmosphericThermodynamic Variables in Wind TurbineFatigue Loading Parameter Identification.”NRILXP-500-26829, Second Symposium on WindConditions for Wind Turbine Design, IEA AnnexXI, Roskilde, Denmark 12-13 April 1999.Kashefi T. and Winterstein, S. R, “RelatingTurbulence to Wmd Turbine Blade Loads:Parametric Study with Multiple Regression

Analysis: A collection of the 1998 ASME WindEnergy @rep., at the AMA Aerospace SciencesMtg., Reno, Nevadaj AIAA-2001-O046, January1998, pp. 273-281.

10

I