23.12.2014 Views

Petra Friederichs - Chair of Statistics STAT

Petra Friederichs - Chair of Statistics STAT

Petra Friederichs - Chair of Statistics STAT

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Predicting High-Impact Weather<br />

<strong>Petra</strong> <strong>Friederichs</strong> and Sabrina Bentzien<br />

Meteorological Institute, University <strong>of</strong> Bonn<br />

in cooperation with<br />

Susanne Theis and Co-workers<br />

German Meteorological National Service (DWD), Offenbach<br />

Marco Oesting and Martin Schlather<br />

Institute <strong>of</strong> Mathematical Stochastics, Goettingen<br />

Ascona, 11 – 15 July 2011<br />

P.<strong>Friederichs</strong> Extreme weather 1 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Why high-impact weather<br />

◮ Mathematically:<br />

Defined as block maxima or excedances <strong>of</strong> large thresholds.<br />

Events that lie in the tails <strong>of</strong> a distribution<br />

◮ Perseption:<br />

Rare, exceptional, ”large” and high impact<br />

◮ Problems:<br />

◮<br />

◮<br />

95% quantile <strong>of</strong> daily precipitation: ≈ 10 − 15mm/d<br />

≈ 2-5 years <strong>of</strong> data – only few extremes events for verification<br />

P.<strong>Friederichs</strong> Extreme weather 2 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Outline<br />

Predicting High-Impact Weather<br />

◮ Atmospheric scales and mesoscale atmospheric dynamics<br />

◮ Numerical weather prediction<br />

◮ Forecast uncertainty and ensemble prediction systems<br />

Ensemble postprocessing<br />

◮ Postprocessing: Extract and calibrate information<br />

Application<br />

◮ Ensemble postprocessing – precipitation<br />

◮ Spatial modelling – wind gusts<br />

P.<strong>Friederichs</strong> Extreme weather 3 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Mesoscale Extremes<br />

Predicting high-impact weather<br />

Ensemble Postprocessing<br />

Mesoscale Weather Prediction<br />

◮ Strong and disastrous impact <strong>of</strong> many weather extremes calls<br />

for reliable forecasts<br />

◮ ”Although forecasters have traditionally viewed weather<br />

prediction as deterministic, a cultural change towards<br />

probabilistic forecasting is in progress.” (T N Palmer, 2002)<br />

◮ Weather extremes do not come ”Out <strong>of</strong> the Blue”<br />

◮ Numerical weather forecast models provide reliable forecasts<br />

<strong>of</strong> the atmospheric circulation prone to generate extremes<br />

◮ Combination <strong>of</strong> dynamical and statistical analysis methods<br />

P.<strong>Friederichs</strong> Extreme weather 4 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Mesoscale Extremes<br />

Predicting high-impact weather<br />

Ensemble Postprocessing<br />

Atmospheric scales and mesoscale dynamics<br />

◮ Different scales exhibit different<br />

dominant force balances,<br />

different wave dynamics<br />

◮ Mesoscale on horizontal scales<br />

2km – 2000km<br />

◮ Complex force balances<br />

Steinhorst, Promet 35, 2010<br />

P.<strong>Friederichs</strong> Extreme weather 5 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Mesoscale Extremes<br />

Predicting high-impact weather<br />

Ensemble Postprocessing<br />

Mesoscale extremes<br />

Squall-line 14 July 2010<br />

1000<br />

998<br />

A. Kelbch (2010)<br />

35<br />

30<br />

996<br />

25<br />

994<br />

20<br />

Gartenstation MIUB<br />

6 9 12 15 18 21 24<br />

July 2010<br />

P.<strong>Friederichs</strong> Extreme weather 6 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Mesoscale Extremes<br />

Predicting high-impact weather<br />

Ensemble Postprocessing<br />

High-impact weather<br />

◮ European mid-latitudes:<br />

Heavy rain and/or violent wind, lightning, hail, . . .<br />

◮ Deep midlatitude convection – Warm season heavy<br />

precipitation clusters<br />

◮ Processes with time scale <strong>of</strong> ∼1h - 1d<br />

◮ Convective cells – organized and embedded within larger<br />

weather systems – initiated on much smaller scales<br />

www.dwd.de<br />

◮ THORPEX - European Section<br />

P.<strong>Friederichs</strong> Extreme weather 7 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Mesoscale Extremes<br />

Predicting high-impact weather<br />

Ensemble Postprocessing<br />

Predictability<br />

◮ Inherent limit <strong>of</strong> predictability<br />

◮ Fastes error growth at smallest<br />

scales<br />

◮ Predictability strongly depends<br />

on flow regime<br />

◮ Moist convection is primary<br />

source <strong>of</strong> forecast-error growth<br />

◮ Mesoscale forecasts are issued<br />

for ≤18h-24h<br />

E N Lorenz, Tellus 21, 19 (1969)<br />

P.<strong>Friederichs</strong> Extreme weather 8 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Mesoscale Extremes<br />

Predicting high-impact weather<br />

Ensemble Postprocessing<br />

COSMO-DE Ensemble Prediction System (EPS)<br />

◮ 2.8 km grid spacing – convection<br />

permitting NWP model<br />

◮ Operational forecasts 0-21 hours (DWD)<br />

◮ EPS with 20 (40) members since 12/10<br />

◮ Uncertainty due to<br />

Initial conditions<br />

Boundary conditions<br />

Model parameterization<br />

◮ High-impact weather<br />

Postprocessing is an integral part <strong>of</strong> an EPS<br />

GME<br />

COSMO-DE<br />

COSMO-EU<br />

Initial State Boundary Model<br />

Theis, DWD (2010)<br />

P.<strong>Friederichs</strong> Extreme weather 9 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Mesoscale Extremes<br />

Predicting high-impact weather<br />

Ensemble Postprocessing<br />

E (κ) [m 3 s 2 ]<br />

1e−02 1e+01 1e+04 1e+07<br />

λ [km]<br />

628 314 125 62 31 12 6<br />

(−5 3)<br />

κ<br />

1e−05 5e−05 2e−04 5e−04<br />

κ [rad/m]<br />

Bierdel, Bentzien, <strong>Friederichs</strong> (2011)<br />

κ −1.55 21.08.2009<br />

COSMO-DE-EPS<br />

(Experiment 2009, 20 members)<br />

◮ Forecast hour 10<br />

◮ Mesoscale k −5/3 spectrum<br />

◮ No spin-up effect<br />

◮ Effective resolution:<br />

4 to 5 · ∆x<br />

P.<strong>Friederichs</strong> Extreme weather 10 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Mesoscale Extremes<br />

Predicting high-impact weather<br />

Ensemble Postprocessing<br />

Ensemble Postprocessing<br />

◮ COSMO-DE forecasts 1 July 2008 – 30 April 2010<br />

◮ lagged average ensemble<br />

12 h accumulated precipitation between 12 and 00 UTC<br />

forecast (LAF)<br />

• Umgebung<br />

• 4x9 member<br />

• 4x25 member<br />

• ...<br />

0 3 6 9 12 15 18 21 lead time<br />

Bentzien, <strong>Friederichs</strong> (2011)<br />

6<br />

First guess: probability, mean and 0.9-quantile<br />

from 5 × 5 neighborhood and 4 COSMO-DE forecasts<br />

P.<strong>Friederichs</strong> Extreme weather 11 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Mesoscale Extremes<br />

Predicting high-impact weather<br />

Ensemble Postprocessing<br />

from Hamill (2006)<br />

Probabilistic forecasting: Maximize sharpness <strong>of</strong><br />

the predictive distribution subject to calibration<br />

Calibration:<br />

Raw ensemble data need adjustmend: biased and underdispersive<br />

Gneiting et al. (2005)<br />

P.<strong>Friederichs</strong> Extreme weather 12 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Semi-Parametric Approaches<br />

Parametric Approaches<br />

Forecast Verification: Proper Scoring Rules<br />

Statistical postprocessing<br />

◮ Marginal distributions:<br />

Conditional on weather forecast<br />

Standard ensemble postprocessing<br />

◮ Ensemble MOS (e.g., Wilks, 2006; Gneiting et al., 2004)<br />

◮<br />

◮<br />

Ensemble dressing – Ensemle kernel dressing (e.g., Wang and<br />

Bishop, 2004; Broecker and Smith, 200)<br />

Bayesian model averaging (e.g., Hoeting et al, 1999; Raftery et al,<br />

2005)<br />

◮ Dependence structure:<br />

Transformed Gaussian random fields (e.g., Berrocal et al., 2008)<br />

Max-stable random fields (e.g., Kabluchko et al., 20098)<br />

P.<strong>Friederichs</strong> Extreme weather 13 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Semi-Parametric Approaches<br />

Parametric Approaches<br />

Forecast Verification: Proper Scoring Rules<br />

Semi-parametric approach<br />

probability<br />

1.0<br />

0.8<br />

0.6<br />

0.4<br />

0.2<br />

F(1)<br />

F(10)<br />

density<br />

1.0<br />

0.8<br />

0.6<br />

0.4<br />

0.2<br />

quantile [mm/12h]<br />

50<br />

10<br />

3<br />

1<br />

0.3<br />

0.5 quantile<br />

0.75 quantile<br />

0.95 quantile<br />

0.0<br />

0.0<br />

0.1<br />

0.1 0.3 1 3 10 50<br />

rain [mm/12h]<br />

0.1 0.3 1 3 10 50<br />

rain [mm/12h]<br />

0.0 0.2 0.4 0.6 0.8 1.0<br />

probability<br />

◮ Logistic regression: π = 1 − F Y (y | X) = logit −1 (β T X)<br />

◮ Censored quantile regression: q = F −1<br />

Y (τ | X) = max(0, βT X)<br />

P.<strong>Friederichs</strong> Extreme weather 14 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Semi-Parametric Approaches<br />

Parametric Approaches<br />

Forecast Verification: Proper Scoring Rules<br />

Parametric approach – Precipitation<br />

probability<br />

1.0<br />

0.8<br />

0.6<br />

0.4<br />

0.2<br />

F(1)<br />

F(10)<br />

density<br />

1.0<br />

0.8<br />

0.6<br />

0.4<br />

0.2<br />

quantile [mm/12h]<br />

50<br />

10<br />

3<br />

1<br />

0.3<br />

0.5 quantile<br />

0.75 quantile<br />

0.95 quantile<br />

0.0<br />

0.0<br />

0.1<br />

0.1 0.3 1 3 10 50<br />

rain [mm/12h]<br />

0.1 0.3 1 3 10 50<br />

rain [mm/12h]<br />

0.0 0.2 0.4 0.6 0.8 1.0<br />

probability<br />

◮ Mixture model:<br />

Y ∼ F y (y; Θ 1 , Θ 2 ) with Θ 1/2 = Θ 1/2 (X)<br />

F y (y; Θ) = π + F Γ (y|y > 0, y ≤ u; Θ 1 ) + F GPD (y|y > u; Θ 2 )<br />

P.<strong>Friederichs</strong> Extreme weather 15 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Semi-Parametric Approaches<br />

Parametric Approaches<br />

Forecast Verification: Proper Scoring Rules<br />

Parametric approach – Wind gusts<br />

probability<br />

1.0<br />

0.8<br />

0.6<br />

0.4<br />

0.2<br />

F(1)<br />

F(10)<br />

density<br />

1.0<br />

0.8<br />

0.6<br />

0.4<br />

0.2<br />

quantile [mm/12h]<br />

50<br />

10<br />

3<br />

1<br />

0.3<br />

0.5 quantile<br />

0.75 quantile<br />

0.95 quantile<br />

0.0<br />

0.0<br />

0.1<br />

0.1 0.3 1 3 10 50<br />

rain [mm/12h]<br />

0.1 0.3 1 3 10 50<br />

rain [mm/12h]<br />

0.0 0.2 0.4 0.6 0.8 1.0<br />

probability<br />

◮ Mixture model: Y ∼ F y (y; Θ) with Θ = Θ(X)<br />

F y (y; Θ) = F GEV (y|Θ)<br />

P.<strong>Friederichs</strong> Extreme weather 16 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Semi-Parametric Approaches<br />

Parametric Approaches<br />

Forecast Verification: Proper Scoring Rules<br />

Quantile verification score (QVS)<br />

QVS(F −1<br />

Y<br />

(τ), y i) = ∑ i<br />

ρ τ (y i − F −1<br />

Y<br />

(τ | x i))<br />

Brier Score (BS)<br />

BS(F Y (u), y i ) = ∑ i<br />

(F Y (u | x i )) − I u−yi ) 2<br />

Continuous rank probability score (CRPS)<br />

CRPS(F Y , y i ) = ∑ ∫<br />

(F Y (u | x i ) − H(u − y i )du<br />

i<br />

Skill scores<br />

SS =<br />

S(F , y)<br />

S ref (F ref , y)<br />

P.<strong>Friederichs</strong> Extreme weather 17 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Ensemble Postprocessing<br />

◮ COSMO-DE forecasts 1 July 2008 – 30 April 2010<br />

◮ 12 h accumulated precipitation between 12 and 00 UTC<br />

n.train<br />

20,30,..,90 Tage<br />

30<br />

T.<br />

!<br />

Jul.08 Okt.08 Jan.09 Apr.09 Jul.09 Okt.09 Jan.10 Apr.10<br />

P.<strong>Friederichs</strong> Extreme weather 18 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Radar vs. rain gauge<br />

QVSS (0.9 quantile), training: stations<br />

verification:<br />

radar field<br />

60<br />

52.5<br />

0.6<br />

quantile verification skill score (%)<br />

50<br />

40<br />

30<br />

●<br />

●<br />

latitude<br />

52.0<br />

51.5<br />

51.0<br />

0.5<br />

0.4<br />

0.3<br />

0.2<br />

20<br />

●<br />

verification:<br />

verification:<br />

●<br />

83 stations<br />

83 radar points<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

RD ST RD ST RD ●<br />

● ST<br />

●<br />

●<br />

data for statistical model training<br />

●<br />

●<br />

●<br />

●<br />

Bentzien, <strong>Friederichs</strong> (2011)<br />

50.5<br />

6 7 8 9<br />

longitude<br />

0.1<br />

0.0<br />

Merging rain radar and station data PhD <strong>of</strong> K. Krebsbach within TR32<br />

P.<strong>Friederichs</strong> Extreme weather 19 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Reliability<br />

conditional observed frequency<br />

1.0<br />

0.8<br />

0.6<br />

0.4<br />

0.2<br />

0.0<br />

MIX<br />

fgp<br />

(b) 1mm threshold<br />

0.9<br />

0.6<br />

0.3<br />

0.0<br />

0.9<br />

0.6<br />

0.3<br />

0.0<br />

BS=0.094 REL=0.001 RES=0.086<br />

N=47202 (MIX)<br />

0 0.2 0.4 0.6 0.8 1<br />

forecast distribution<br />

BS=0.102 REL=0.008 RES=0.085<br />

N=47202 (fgp)<br />

0.0 0.2 0.4 0.6 0.8 1.0<br />

0 0.2 0.4 0.6 0.8 1<br />

forecast probability<br />

forecast distribution<br />

P.<strong>Friederichs</strong> Extreme weather 20 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Reliability<br />

conditional observed frequency<br />

1.0<br />

0.8<br />

0.6<br />

0.4<br />

0.2<br />

0.0<br />

MIX<br />

fgp<br />

(d) 10mm threshold<br />

0.9<br />

0.6<br />

0.3<br />

0.0<br />

0.9<br />

0.6<br />

0.3<br />

0.0<br />

BS=0.019 REL=0.002 RES=0.005<br />

N=47202 (MIX)<br />

0 0.2 0.4 0.6 0.8 1<br />

forecast distribution<br />

BS=0.021 REL=0.004 RES=0.005<br />

N=47202 (fgp)<br />

0.0 0.2 0.4 0.6 0.8 1.0<br />

0 0.2 0.4 0.6 0.8 1<br />

forecast probability<br />

forecast distribution<br />

P.<strong>Friederichs</strong> Extreme weather 21 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Quantile forecasts - Skill Score<br />

QVSS for q 0.90 QVSS for q 0.99<br />

52.5<br />

0.6<br />

52.5<br />

0.8<br />

52.0<br />

0.5<br />

52.0<br />

0.6<br />

0.4<br />

latitude<br />

51.5<br />

0.3<br />

latitude<br />

51.5<br />

0.4<br />

51.0<br />

0.2<br />

51.0<br />

0.2<br />

50.5<br />

0.1<br />

50.5<br />

0.0<br />

0.0<br />

6 7 8 9<br />

longitude<br />

6 7 8 9<br />

longitude<br />

P.<strong>Friederichs</strong> Extreme weather 22 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Quantile forecasts - Skill Score<br />

QVSS<br />

0.7<br />

0.6<br />

0.5<br />

0.4<br />

0.3<br />

0.2<br />

0.1<br />

0.0<br />

●<br />

0.99 quantile<br />

●<br />

+<br />

●●● ●●●●●●●● +<br />

+<br />

+<br />

+<br />

+ +<br />

+<br />

+<br />

● + +<br />

+ + + ++ ●●●●●●●●●●●●● ++ +<br />

+<br />

+<br />

++ + ++ + + + +<br />

+ +<br />

+ + ++<br />

+<br />

++ + ●●●● ●●●●●●●●●●●●●●●●●●<br />

+++ + ++<br />

++<br />

+<br />

+ +<br />

+<br />

+++<br />

+<br />

+ + +<br />

+<br />

+ +<br />

+<br />

++ ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● +++ +<br />

+<br />

+<br />

+ + +<br />

+ + +<br />

+ + +<br />

+ + ++ + +<br />

+ +<br />

+ +<br />

++ +<br />

+<br />

+<br />

+<br />

+<br />

++ + ++ + +<br />

+ ++<br />

+ +<br />

++<br />

+<br />

+ + + +<br />

+<br />

●<br />

+<br />

+ +<br />

●<br />

+<br />

+ +<br />

+<br />

●<br />

+<br />

+<br />

+<br />

+ + +<br />

●<br />

+<br />

+ +<br />

QR<br />

MIX<br />

MIX.t<br />

GPD<br />

QR MIX MIX.t GPD<br />

0 10 20 30 40 50 60 70 80<br />

P.<strong>Friederichs</strong> Extreme weather 23 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Quantile forecasts - Skill Score<br />

QVSS<br />

0.8<br />

0.7<br />

0.6<br />

0.5<br />

0.4<br />

0.3<br />

0.2<br />

0.1<br />

0.0<br />

−0.1<br />

−0.2<br />

−0.3<br />

0.99 quantile<br />

● +<br />

+<br />

+<br />

+ + + +<br />

●●●●●●●●● ●● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● +<br />

++<br />

+ +<br />

+<br />

+ +<br />

+<br />

+ + +<br />

+ ++ + + + +<br />

+ + +<br />

+<br />

+ + + +<br />

+<br />

++ +<br />

++<br />

+<br />

+++<br />

+ +<br />

+ ++<br />

+++ + ++<br />

+ + + + +<br />

+ + +<br />

+ +++ ++<br />

+<br />

+<br />

+ + + + +<br />

+ + +<br />

+<br />

+ +<br />

+<br />

+ +<br />

+<br />

● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● +<br />

+<br />

+ + +<br />

+ + +<br />

++ ++<br />

+ +<br />

+ +<br />

+ +<br />

+ + +<br />

+<br />

+ +<br />

+<br />

+ +<br />

+ + ++ + +<br />

+<br />

++<br />

+ + + + ++ +<br />

+<br />

+ + +<br />

●<br />

+<br />

+ +<br />

++ ++ + QR<br />

MIX<br />

MIX.t<br />

GPD<br />

QR MIX MIX.t GPD<br />

0 10 20 30 40 50 60 70 80<br />

P.<strong>Friederichs</strong> Extreme weather 24 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Quantile forecasts, station and radar measurements<br />

rain [mm/12h]<br />

50<br />

40<br />

30<br />

20<br />

10<br />

0<br />

●<br />

Kleve<br />

●<br />

● ● ● ● ● ● ● ●<br />

●<br />

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●<br />

●<br />

●<br />

●<br />

rain gauge<br />

radar<br />

q0.99<br />

q0.95<br />

q0.90<br />

q0.75<br />

Bentzien, <strong>Friederichs</strong> (2011)<br />

Jul 05 Jul 15 Jul 25 Aug 04<br />

P.<strong>Friederichs</strong> Extreme weather 25 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

03.07.2009, init 12 UTC<br />

rain [mm/12h]<br />

03.07.2009, init 09 UTC<br />

rain [mm/12h]<br />

03.07.2009, radar<br />

rain [mm/12h]<br />

52.5<br />

52.5<br />

52.5<br />

80<br />

80<br />

80<br />

52.0<br />

52.0<br />

52.0<br />

60<br />

60<br />

60<br />

51.5<br />

latitude<br />

51.5<br />

latitude<br />

51.5<br />

latitude<br />

40<br />

40<br />

40<br />

51.0<br />

51.0<br />

51.0<br />

20<br />

20<br />

20<br />

50.5<br />

50.5<br />

50.5<br />

longitude<br />

longitude<br />

longitude<br />

6 7 8 9<br />

03.07.2009, init 06 UTC<br />

rain [mm/12h]<br />

6 7 8 9<br />

03.07.2009, init 03 UTC<br />

rain [mm/12h]<br />

6 7 8 9<br />

03.07.2009, 0.99 quantile (mixture model)<br />

rain [mm/12h]<br />

52.5<br />

52.5<br />

52.5<br />

140<br />

80<br />

80<br />

120<br />

52.0<br />

52.0<br />

52.0<br />

60<br />

60<br />

100<br />

51.5<br />

latitude<br />

51.5<br />

latitude<br />

51.5<br />

latitude<br />

80<br />

40<br />

40<br />

60<br />

51.0<br />

51.0<br />

51.0<br />

40<br />

20<br />

20<br />

50.5<br />

50.5<br />

50.5<br />

20<br />

longitude<br />

6 7 8 9<br />

Bentzien, <strong>Friederichs</strong> (2011)<br />

longitude<br />

6 7 8 9<br />

longitude<br />

6 7 8 9<br />

P.<strong>Friederichs</strong> Extreme weather 26 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Forecasting + Verification – German Weather Service<br />

type <strong>of</strong> gust threshold m/s km/h knots<br />

near gale > 14 50-60 28-33<br />

gale 18 − 24 65-85 34-47<br />

storm 25 − 28 90-100 48-55<br />

violent storm 29 − 32 105-125 56-63<br />

hurricane force > 33 > 120 > 64<br />

P.<strong>Friederichs</strong> Extreme weather 27 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Spatial modeling <strong>of</strong> peak wind speed observations<br />

◮ Marginal distributions:<br />

Spatial interpolation <strong>of</strong> marginal distribution<br />

- conditional on large scale weather situation<br />

- conditional on local conditions (surface roughness,<br />

altitude,. . .)<br />

→ Transform marginals to standard Fréchet<br />

◮ Dependence structure:<br />

Gaussian random fields to construct max-stable processes<br />

P.<strong>Friederichs</strong> Extreme weather 28 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Conditional marginal distribution<br />

Non-stationary process is characterised by atmospheric circulation<br />

◮ COSMO-DE forecasts: wind velocity 10m above ground<br />

◮ Diurnal and annual cycle<br />

◮ Topography and surface roughness<br />

◮ Spatially and temporally constant hyperparameters (MLE)<br />

µ t = Z(X t ) T γ ; σ t = Z(X t ) T ρ ; ξ t = ξ<br />

→ Diploma thesis <strong>of</strong> S. Memmel: BHM to spatially model parameter surfaces<br />

P.<strong>Friederichs</strong> Extreme weather 29 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

∼400 (139) weather stations<br />

◮ 6-hourly wind maxima in 1m/s<br />

◮ from April 2003 to May 2011<br />

COSMO-DE forecasts<br />

◮ 10m wind velocity, 6 hour mean<br />

◮ surface maximum wind velocity<br />

◮ January 2010 to April 2010<br />

(available until now)<br />

Forecasts and observations 28 Feb 2010<br />

P.<strong>Friederichs</strong> Extreme weather 30 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Max-stabile Prozesse<br />

A stochastic process {Y (r), r ∈ Z} is called max-stable, if its<br />

scaled maxima follow the same distrinution.<br />

P.<strong>Friederichs</strong> Extreme weather 31 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Construction <strong>of</strong> max-stable process – Smith (1990)<br />

( ∫<br />

)<br />

Prob(Y i ≤ y i , ∀i ∈ N) = exp − {y −1 f (r, R i )}dr<br />

R d max<br />

i∈N<br />

where f (r, R i ) = f 0 (r − R i ), f 0 is a multivariate normal distribution<br />

i<br />

P.<strong>Friederichs</strong> Extreme weather 32 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Construction <strong>of</strong> max-stable process – Schlather (2002)<br />

W i : stationary, standard Gaussian random process in R d<br />

Poisson point process (U i ) on (0, ∞) with intensity u −2 du<br />

Y (r) = √ 2π max{U i max{0, W i }}, r ∈ R d<br />

i∈N<br />

is stationary and max-stable with standard Fréchet marginals<br />

P.<strong>Friederichs</strong> Extreme weather 33 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Construction <strong>of</strong> max-stable process – Kabluchko (2009)<br />

W i : zero-mean Gaussian random process R d with variance σ 2 (x)<br />

Poisson point process (U i i) on (0, ∞)<br />

with intensity u −2 du<br />

Z(r) = e −σ2 (r)/2 max{U i e W i (r) }, r ∈ R d<br />

i∈N<br />

is stationary and max-stable with standard Fréchet marginals<br />

P.<strong>Friederichs</strong> Extreme weather 34 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Extremal coefficient<br />

Extremal coefficient Θ N defined as<br />

with Θ N = V (1, . . . , 1)<br />

Prob(Y (r i ) ≤ y, ∀i ∈ N) = exp<br />

(<br />

− Θ )<br />

N<br />

,<br />

y<br />

Measure for degrees <strong>of</strong> freedom<br />

P.<strong>Friederichs</strong> Extreme weather 35 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Madogram<br />

F -Madogram: Cooley (2007)<br />

ν F (r i , r j ) = 1 2 E [|F (Y (r i)) − F (Y (r j ))|] ,<br />

F -Madogram determines extremal coefficient<br />

Θ 2 (r i , r j ) = 1 + 2ν F (r i − r j )<br />

1 − 2ν F (r i − r j ) .<br />

P.<strong>Friederichs</strong> Extreme weather 36 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Extremal coefficient<br />

1.0 1.2 1.4 1.6 1.8 2.0<br />

DWDnorth<br />

F−Madogram<br />

Smith<br />

Schlather<br />

Whittle−Matern<br />

Powered Exponential<br />

Cauchy<br />

Brown−Resnick<br />

0 50 100 150 200 250 300<br />

Distance in km<br />

P.<strong>Friederichs</strong> Extreme weather 37 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Extremal coefficient<br />

1.0 1.2 1.4 1.6 1.8 2.0<br />

DWDnorth<br />

F−Madogram<br />

Smith<br />

Schlather<br />

Whittle−Matern<br />

Powered Exponential<br />

Cauchy<br />

Brown−Resnick<br />

0 50 100 150 200 250 300<br />

Distance in km<br />

P.<strong>Friederichs</strong> Extreme weather 38 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Extremal coefficient<br />

1.0 1.2 1.4 1.6 1.8 2.0<br />

DWDnorth<br />

F−Madogram<br />

Smith<br />

Schlather<br />

Whittle−Matern<br />

Powered Exponential<br />

Cauchy<br />

Brown−Resnick<br />

0 50 100 150 200 250 300<br />

Distance in km<br />

P.<strong>Friederichs</strong> Extreme weather 39 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Extremal coefficient<br />

1.0 1.2 1.4 1.6 1.8 2.0<br />

DWDeast<br />

F−Madogram<br />

Smith<br />

Schlather<br />

Whittle−Matern<br />

Powered Exponential<br />

Cauchy<br />

Brown−Resnick<br />

0 50 100 150 200 250 300<br />

Distance in km<br />

P.<strong>Friederichs</strong> Extreme weather 40 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Ensemble Postprocessing – Precipitation<br />

Case study – 3rd July 2009<br />

Spatial Modelling – Wind Gusts<br />

Thanks to M. Oesting and M. Schlather for the simulations <strong>of</strong> the BR<br />

P.<strong>Friederichs</strong> Extreme weather 41 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Conclusions<br />

◮ Weather forcasts provide information that conditions occurence <strong>of</strong><br />

extremes<br />

◮ Linear (non-linear) statistical modeling extracts information<br />

◮ Extreme value theory provides distributions tailored for extremes<br />

◮ Parametric method less uncertain than non-parametric method and<br />

non-linear dependecy (shape parameter) is not parsimony<br />

◮ High-impact weather: insufficient data available for training and for<br />

validation<br />

P.<strong>Friederichs</strong> Extreme weather 42 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Challenges<br />

◮ Improve physical understanding <strong>of</strong> generation processes <strong>of</strong> extremes<br />

◮ Application to multi-variable and spatio-temporal predictionswhich<br />

◮<br />

◮<br />

◮<br />

Combine spatial statistics with model postprocessing (Berrocal<br />

et al., 2007)<br />

Develop methods for multivariate postprocessing<br />

Develop ensemble methods tailored to extremes<br />

◮ Verification tailored to extremes<br />

◮ Verification for probabilistic multivariate and spatial forecasts<br />

P.<strong>Friederichs</strong> Extreme weather 43 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Towards a next-generation mesoscale prediction system<br />

for extreme weather<br />

VolkswagenStiftung funded project:<br />

◮ Physical understanding <strong>of</strong> mesoscale weather extremes<br />

◮ Theory and methodology for spatial and multivariate statistical<br />

modeling, uncertainty assessment and verification <strong>of</strong> extremes<br />

◮ Operational ensemble prediction system<br />

◮ Postprocessing for multivariate and spatial weather variables<br />

PhD position on “Conditional statistical modeling <strong>of</strong> spatial extremes<br />

using non-homogeneous marked point processes” (M. Schlather)<br />

go to www.wex-mop.uni-bonn.de<br />

P.<strong>Friederichs</strong> Extreme weather 44 / 45


Mesoscale Weather Prediction<br />

Statistical Postprocessing<br />

Applications<br />

Conclusions and Challenges<br />

Thank you for your attention!<br />

P.<strong>Friederichs</strong> Extreme weather 45 / 45

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!