You are on page 1of 26

EXTREMESTATISTICS

ExperiencesEs+ma+ng100YearDesignCondi+ons

SteveWinterstein
stevewinterstein@alum.mit.edu

WhatIsthe100YearWaveHeight?
NorthSea:Wavesmeasured
frominstrumentsmountedon
jacketstructures
Measure:waveeleva+on(t)
Report:HS=4 every3hours
(N=2920data/yr)
Seek:H100=100yearHS,from
18yearsofdata
Reference:Winterstein,S.R.,Kleiven,G.,andHagen,O.,ComparingExtremeWaveEs+mates
fromHourlyandAnnualData,Proceedings,ISOPE2001,StavangerNorway,June1722,2001
URL:sites/h^p://sites.google.com/site/stevewinterstein/publica+ons1/archivalpublica+ons

EsAmaAngH100:TwoApproaches
Approach1:UseallHSdata(onevalueper3hrs)
N=2920data/yr
FitWeibullmodelofHS
P[HS>H100]=(100*2920)1=3.4x106[per3hrseastate]

EsAmaAngH100:TwoApproaches
Approach1:UseallHSdata(onevalueper3hrs)
N=2920data/yr
FitWeibullmodelofHS
P[HS>H100]=(100*2920)1=3.4x106[per3hrseastate]
Approach2:UseonlyHANN=annualmaximadata(onevalue
peryear)
FitGumbelmodelofHANN
P[HANN>H100]=(100)1=1x102[peryear]

EsAmaAngH100:TwoApproaches
Approach1:UseallHSdata(onevalueper3hrs)
N=2920data/yr
FitWeibullmodelofHS
P[HS>H100]=(100*2920)1=3.4x106[per3hrseastate]
Approach2:UseonlyHANN=annualmaximadata(onevalue
peryear)
FitGumbelmodelofHANN
P[HANN>H100]=(100)1=1x102[peryear]

WHICHIS
BETTER?

Approach1:FiJngtoAll3HourData
Fitto3
Moments
ofHS
ResulAng
EsAmate:
H100=14.5m

Approach2:FiJngtoAnnualMaxima
Fitto2
Moments
ofHANN
ResulAng
EsAmate:
H100=13.2m

Approach2:FiJngtoAnnualMaxima
Fitto2
Moments
ofHANN
ResulAng
EsAmate:
H100=13.2m
10%
lowerthan
Weibull
esAmate
H100=14.5m

H100=13.2mvs14.5WhyDierent?
Hypothesis1:SeasonalEects.
Fit4dierentWeibulldistribu+onsineachseason
Result:noeect

H100=13.2mvs14.5WhyDierent?
Hypothesis1:SeasonalEects.
Fit4dierentWeibulldistribu+onsineachseason
Result:noeect
Hypothesis2:Clustering.
H100=14.5assumesall3hourdataindependent
Actual3hrdataclustershouldreduceH100

H100=13.2mvs14.5WhyDierent?
Hypothesis1:SeasonalEects.
Fit4dierentWeibulldistribu+onsineachseason
Result:noeect
Hypothesis2:Clustering.
H100=14.5assumesall3hourdataindependent
Actual3hrdataclustershouldreduceH100
Hypothesis3:StaAsAcalUncertainty.
H100=13.2usesonly18annualmaxima
Considerableuncertaintyshouldfavor
H100=14.5m

ClusteringModel:Markovian

ClusteringModel:Markovian
ResulAng
EsAmate:
H100=14.2m
vs
H100=14.5m
ifindep.
Clustering
hasliZle
eecton
extremes

StaAsAcalUncertaintyfromMoments:
CONSIDERMOMENTFITGUMBELMODELOFHANN:
UNCERTAINTYINMOMENTSGIVEN18DATA:

COMBININGTHESERESULTS:

DIFFERENCE14.513.2=1.3m=1.4UNLIKELY(Normal:p=8%)

Focus:UpperTailofAllData

TailofF_3hrfallsosharplyaboveh=10m;obscuredbyglobalttoalldata
Favorsmodelsofannualmaxs(orstorms=peakHsoverthreshold8m)

OMAEPaper200979691,Presented4June2009,Hawaii

TurkstraModelsofCurrentProles
SteveWinterstein
StanfordCon,nuingStudies

SverreHaverand
EinarNygaard
StatoilHydro

NORTHSEA
CURRENTS@
OrmenLange

NORTHSEA
CURRENTS@
OrmenLange
ISSUES:
Acous+csensors:
currentdataover
depth>1km,
T=12yrs
Concernforrisers:
dragloads,VIV

CurrentSpeedsintheNorthSea
16monthlypeaksat6depths

16MonthlyMaximaat200m:GumbelScalePlot

Whathappensifweremovelowest4points(quietmonths912)?

ExtrapolaAngto100YearLevelsatd=200m

ExtrapolaAngto100YearLevelsatd=200m

ALL16MonthlyData>X100=1.54*Xmax,obs
ExcludeMonths912>X100=1.32*Xmax,obs
1.54/1.32=1.17=17%dierenceincurrentspeed
1.172=37%dierenceindragload

Seasonal
Eects
Among
Monthly
Maximain
Current
Datasets

CONCLUSIONS
WAVEHEIGHTEXAMPLE:
Modellingall3hourdataobscurestrendsintails
Be^ermodelsfocusdirectlyonuppertails:
annualmaxima
peaksoverthreshold

CONCLUSIONS
WAVEHEIGHTEXAMPLE:
Modellingall3hourdataobscurestrendsintails
Be^ermodelsfocusdirectlyonuppertails:
annualmaxima
peaksoverthreshold
CURRENTSPEEDEXAMPLE:
Globalmodel(ofallmonthlymaxs)obscuresseasonaleects
Be^ermodelsfocusonuppertails:
excludequietseasons
peaksoverthreshold
Shorterdatasetsofcurrentspeedsmoresta+s+caluncertainty

CONCLUSIONS
WAVEHEIGHTEXAMPLE:
Modellingall3hourdataobscurestrendsintails
Be^ermodelsfocusdirectlyonuppertails:
annualmaxima
peaksoverthreshold
CURRENTSPEEDEXAMPLE:
Globalmodel(ofallmonthlymaxs)obscuresseasonaleects
Be^ermodelsfocusonuppertails:
excludequietseasons
peaksoverthreshold
Shorterdatasetsofcurrentspeedsmoresta+s+caluncertainty
GENERAL:
Bewareofmodelsthatincludealldatavsallrelevantdata

You might also like