# EXTREME STATISTICS

Experiences Es+ma+ng 100‐Year Design Condi+ons

Steve Winterstein
stevewinterstein@alum.mit.edu

What Is the 100‐Year Wave Height?
•  North Sea: Waves measured
from instruments mounted on
jacket structures
•  Measure: wave eleva+on η(t)
•  Report: HS=4ση every 3 hours
(N=2920 data/yr)
•  Seek: H100 = 100‐year HS, from
18 years of data
Reference: Winterstein, S.R., Kleiven, G., and Hagen, O., “Comparing Extreme Wave Es+mates
from Hourly and Annual Data, Proceedings, ISOPE 2001, Stavanger Norway, June 17‐22, 2001

4 x 10‐6  [per 3‐hr seastate]  .EsAmaAng H100: Two Approaches…  •  Approach 1: Use all HS data (one value per 3 hrs)  N=2920 data/yr  Fit Weibull model of HS  P[HS  > H100] = (100*2920)‐1  = 3.

4 x 10‐6  [per 3‐hr seastate]  •  Approach 2: Use only HANN  = annual maxima data (one value  per year)  Fit Gumbel model of HANN  P[HANN  > H100] = (100)‐1  = 1 x 10‐2  [per year]  .EsAmaAng H100: Two Approaches…  •  Approach 1: Use all HS data (one value per 3 hrs)  N=2920 data/yr  Fit Weibull model of HS  P[HS  > H100] = (100*2920)‐1  = 3.

4 x 10‐6  [per 3‐hr seastate]  •  Approach 2: Use only HANN  = annual maxima data (one value  per year)  Fit Gumbel model of HANN  P[HANN  > H100] = (100)‐1  = 1 x 10‐2  [per year]  WHICH IS  BETTER?  .EsAmaAng H100: Two Approaches…  •  Approach 1: Use all HS data (one value per 3 hrs)  N=2920 data/yr  Fit Weibull model of HS  P[HS  > H100] = (100*2920)‐1  = 3.

5m   .Approach 1: FiJng to All 3‐Hour Data  Fit to 3  Moments  of HS  ResulAng  EsAmate:  H100=14.

2m   .Approach 2: FiJng to Annual Maxima   Fit to 2  Moments  of HANN  ResulAng  EsAmate:  H100=13.

Approach 2: FiJng to Annual Maxima   Fit to 2  Moments  of HANN  ResulAng  EsAmate:  H100=13.5m   .2m    10%  lower than  Weibull  esAmate  H100=14.

Fit 4 diﬀerent Weibull distribu+ons in each season  Result: no eﬀect  .H100=13.2m vs 14.5  Why Diﬀerent?  •  Hypothesis 1: Seasonal Eﬀects.

Fit 4 diﬀerent Weibull distribu+ons in each season  Result: no eﬀect  •  Hypothesis 2: Clustering.5  Why Diﬀerent?  •  Hypothesis 1: Seasonal Eﬀects.2m vs 14.H100=13.  H100 = 14.5 assumes all 3‐hour data independent  Actual 3‐hr data cluster  should reduce H100  .

H100=13.  Fit 4 diﬀerent Weibull distribu+ons in each season  Result: no eﬀect  •  Hypothesis 2: Clustering.2m vs 14.5 assumes all 3‐hour data independent  Actual 3‐hr data cluster  should reduce H100  •  Hypothesis 3: StaAsAcal Uncertainty.  H100 = 14.  H100 = 13.5  Why Diﬀerent?  •  Hypothesis 1: Seasonal Eﬀects.2 uses only 18 annual maxima  Considerable uncertainty  should favor  H100=14.5m  .

Clustering Model: Markovian  .

 Clustering  has liZle  eﬀect on  extremes  . Clustering Model: Markovian  ResulAng  EsAmate:  H100=14.5m  if indep.2m  … vs  H100=14.

StaAsAcal Uncertainty from Moments:  CONSIDER MOMENT‐FIT GUMBEL MODEL OF HANN:  UNCERTAINTY IN MOMENTS GIVEN 18 DATA:   COMBINING THESE RESULTS:    DIFFERENCE 14.5‐13.2=1.4σ UNLIKELY (Normal: p=8%)  .3m=1.

obscured by global ﬁt to all data  Favors models of annual maxs (or storms = peak Hs over threshold 8m)   .Focus: Upper Tail of All Data  Tail of F_3‐hr  falls oﬀ sharply above h=10m.

OMAE Paper 2009‐79691.nuing Studies  Sverre Haver and  Einar Nygaard  StatoilHydro  . Presented  4 June 2009. Hawaii  Turkstra Models of Current Proﬁles  Steve Winterstein  Stanford Con.

NORTH SEA  CURRENTS @  Ormen Lange  .

T=1‐2 yrs  •  Concern for risers:  drag loads.NORTH SEA  CURRENTS @  Ormen Lange  ISSUES:  •  Acous+c sensors:  current data over  depth > 1km. VIV  .

Current Speeds in the North Sea…  16 monthly peaks at 6 depths      .

16 Monthly Maxima at 200m: Gumbel Scale Plot  What happens if we remove lowest 4 points (“quiet” months 9‐12)?  .

ExtrapolaAng to 100‐Year Levels at d=200m…  .

17 = 17% diﬀerence in current speed  • 1.obs   • Exclude Months 9‐‐12 ‐‐‐>     X100 = 1.54/1.obs  • 1.54 * Xmax.172                          = 37% diﬀerence in drag load  .ExtrapolaAng to 100‐Year Levels at d=200m…  • ALL 16 Monthly Data ‐‐‐‐>     X100 = 1.32 * Xmax.32=1.

Seasonal  Eﬀects  Among  Monthly  Maxima in  Current  Datasets  .

CONCLUSIONS  WAVE HEIGHT EXAMPLE:  Modelling all 3‐hour data obscures trends in tails  Be^er models focus directly on upper tails:      annual maxima      peaks over threshold  .

CONCLUSIONS  WAVE HEIGHT EXAMPLE:  Modelling all 3‐hour data obscures trends in tails  Be^er models focus directly on upper tails:      annual maxima      peaks over threshold  CURRENT SPEED EXAMPLE:   Global model (of all monthly maxs) obscures seasonal eﬀects  Be^er models focus on upper tails:      exclude “quiet” seasons     peaks over threshold  Shorter data sets of current speeds  more sta+s+cal uncertainty  .

CONCLUSIONS  WAVE HEIGHT EXAMPLE:  Modelling all 3‐hour data obscures trends in tails  Be^er models focus directly on upper tails:      annual maxima      peaks over threshold  CURRENT SPEED EXAMPLE:   Global model (of all monthly maxs) obscures seasonal eﬀects  Be^er models focus on upper tails:      exclude “quiet” seasons     peaks over threshold  Shorter data sets of current speeds  more sta+s+cal uncertainty  GENERAL:  Beware of models that include “all” data vs “all relevant” data  .