23.11.2014 Views

Seamless prediction

Seamless prediction

Seamless prediction

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.

<strong>Seamless</strong><br />

<strong>prediction</strong><br />

Prediction from days to<br />

decades<br />

Wilco Hazeleger


<strong>Seamless</strong> <strong>prediction</strong>


<strong>Seamless</strong> <strong>prediction</strong>


<strong>Seamless</strong> <strong>prediction</strong><br />

The science case:<br />

Same physical principles for weather and climate (but different<br />

processes acting on different scales)<br />

Prediction problem with different role initial and boundary conditions<br />

Seasonal to decadal <strong>prediction</strong> as focus<br />

The modelling case:<br />

Convergence of NWP-Climate model development<br />

<strong>Seamless</strong> earth system <strong>prediction</strong><br />

The application case:<br />

Calibration (weights) of climate <strong>prediction</strong>s & projections<br />

<strong>Seamless</strong> connection to impact studies


3 pillars of 21st century scientific method<br />

Theory (since antiquity)<br />

Combined with experiment<br />

(since Galilei & Newton)<br />

….and with simulation<br />

(since 1940s, Teller, von Neumann, Fermi,..)


Sources and limits of<br />

predictability


Natural variability, chaotic atmosphere<br />

Lorentz limit ~15 days<br />

Selten et al, Sterl et al 2008


Natural and `forced variability


Predicting natural variability: more than noise?<br />

• 0-hypothesis: ocean integrates white noise (weather)<br />

(Hasselmann 1976)<br />

dX ( t)<br />

dt<br />

= −α X ( t)<br />

+ζ ( t)<br />

• With α damping coefficent and ζ(t) random variable (AR1 process à red noise)<br />

• When variability stands out of red noise, e.g. oscillations due<br />

to internal dynamics, dynamical <strong>prediction</strong>s may be possible<br />

• But… even red noise is a source of predictability (e.g. damped<br />

persistence)


Spectra of global mean temperature: peaks?


Monthly: Soil moisture, Sea Surface temperatures,<br />

Atmospheric composition<br />

l Dry soils in spring can lead to high summer temperatures


Seasonal to interannual: El Niño<br />

Sea surface temperature anomaly during El Niño


Seasonal to interannual: Coupled ocean-atmosphere<br />

variability


El Nino – Southern Oscillation (ENSO)<br />

" Predictable 3-9 month ahead (depending on the season)<br />

" World-wide impact<br />

Precipitation SON<br />

Temperature DJF


Decadal: Atlantic Multidecadal Oscillation<br />

Knight et al. 2005


Decadel: Pacific decadal variability


Multi-decadal: anthropogenically forced change


Modelling climate changes: uncertainties<br />

Natural<br />

fluctuations<br />

GHG emission<br />

uncertainty<br />

Model uncertainty<br />

Hawkins and Sutton, 2009


Predictability beyond the Lorentz-limiet<br />

l<br />

The atmospheric circulation is chaotic and unpredictable beyond<br />

~15 days, the averaged weather is in some areas and in some<br />

seasons predictable beyond the Lorentz limit<br />

l<br />

l<br />

(damped) persistence beyond 15 days<br />

Showly fluctuating patterns of variability (often involving slow<br />

ocean responses): ENSO, Indian Ocean Zonal Mode, Pacific<br />

Decadal Oscillation, Atlantic Multidecadal Oscillation,…,…<br />

l<br />

Initial conditions of the earth system provide memory: ocean,<br />

ice, soil, atmospheric composition<br />

l<br />

Anthropogenic forcing (trends) predictable


The ideal case: potential<br />

predictability


What to expect from <strong>prediction</strong>s?<br />

• Diagnostic potential predictability: ratio of variances in a long<br />

control simulation<br />

DPP<br />

=<br />

σ<br />

σ<br />

m<br />

2<br />

σ<br />

2 − 1<br />

v<br />

2<br />

• With σ v<br />

2<br />

variance of m-year means<br />

(e.g. Boer et al, Pohlmann et al)


Diagnostic potential predictability in EC-Earth<br />

10 yr


What to expect from <strong>prediction</strong>s?<br />

• Prognostic potential predictability: ensemble spread in relation<br />

to total variance<br />

PPP = 1−<br />

1<br />

N(<br />

M<br />

−1)<br />

i=<br />

1<br />

2<br />

• With X ij is the i th member of j th ensemble, N is number<br />

ensembles, M number of ensemble members<br />

N<br />

∑<br />

j=<br />

1<br />

M<br />

∑<br />

σ<br />

[ X<br />

ij<br />

( t)<br />

− X<br />

j<br />

( t)]<br />

2<br />

Griffies and Bryan 1997, Collins et al, Pohlmann et al


Atlantic MOC (30N)<br />

MOC (Sv)<br />

Collins et al 2006


Prognostic Potential predictability in EC-Earth (T2m, yr<br />

1-10)<br />

T. Koenigk, SMHI, pers. comm.


Prognostic potential predictability in EC-Earth (T2m, yr<br />

1-10; without trend)<br />

T. Koenigk, SMHI, pers. comm.


Making <strong>prediction</strong>s


Prediction systems<br />

l<br />

Statistical<br />

l<br />

l<br />

l<br />

l<br />

analyse 100 years of data for relationships<br />

Easy to fool yourself (and others)<br />

Much used in tropics: Monsoons India, Indonesia, West Africa,<br />

hurricanes, seasonal <strong>prediction</strong> US<br />

Often misused: summer and winter <strong>prediction</strong>s for Europe


Statistical: Precipipation in Curaçao<br />

l<br />

Rain related to ENSO<br />

4 months in advance


Prediction systems<br />

l Dynamical<br />

Use numerical weather <strong>prediction</strong> model, run it longer. Use climate<br />

model, run it shorter<br />

" Essential: good initial state of ocean, snow, soil, aerosols,<br />

greenhouse gasses<br />

" ECMWF, EC-Earth, Met Office, Météo France, NCEP, ...


Prediction experiment<br />

Initialize atmosphere-ocean-sea ice-land models<br />

from observed/analyzed ocean and sea ice state<br />

Perturb initialized models to generate ensembles<br />

Verify the results against own analyses and<br />

independent observations


Complex Global Circulation Model<br />

Based on laws of physics (F= m x a<br />

and thermodynamics)<br />

Forcing specified (Q= e.g. radiation)<br />

Small scales not resolved<br />

Big science: 2 million lines of code<br />

and 10 variables on 10^9 gridpoints


Initialization, in particular the ocean:<br />

Limited subsurface ocean observations<br />

1960 1980<br />

2007


Initialisation ocean: ocean analyses<br />

A. Koehl, pers. comm.


Perturbing the ensemble<br />

" Perturbations which grow most rapidly in slow component (e.g. in<br />

ocean, for instance Kleeman et al. for ENSO, Hawkins and Sutton<br />

for 3D ocean, bred vectors B. Kirtman etc.)<br />

" Consistent with the observational uncertainties<br />

" Can be useful for identifying regions where additional observations<br />

would be most valuable to improve <strong>prediction</strong>s


Perturbing ocean<br />

• E.g. linear Inverse Modelling (Penland & Sardeshmukh 1995, Hawkins &<br />

Sutton 2009)<br />

dx<br />

dt<br />

x(<br />

t<br />

P<br />

T<br />

= Bx + ζ x represents<br />

+ τ ) =<br />

Px<br />

0<br />

= λx<br />

P x(<br />

t)<br />

τ<br />

0<br />

leading EOFs<br />

Eigenvectors are<br />

optimal growing<br />

perturbations<br />

In practice, pragmatic approaches (perturbing<br />

atmosphere, different ocean states, perturbing ocean<br />

diffusivity)<br />

Hawkins & Sutton 2009


Measures of skill


Measures of skill


Some simple deterministic measures of skill<br />

N<br />

• Mean error (additive<br />

1<br />

bias)<br />

ME = ∑(<br />

Fi<br />

− O<br />

N<br />

i=<br />

1<br />

i<br />

)<br />

• Root mean square error<br />

RMSE<br />

=<br />

1<br />

N<br />

N<br />

∑<br />

i=<br />

1<br />

(<br />

−<br />

F i Oi<br />

)<br />

2<br />

• Anomaly correlation<br />

• (F=forecast,<br />

=observation,<br />

C=climatology)<br />

AC<br />

( F −C)(<br />

O −C)<br />

= ∑<br />

2<br />

( F −C)<br />

( O −C)<br />

2


Skill weather <strong>prediction</strong>: 500hPa height


Skill dynamical seasonal <strong>prediction</strong>s: temperature<br />

Dec-Feb temperature, predicted from 1st of November<br />

Note: bias correction always needed


Skill dynamical seasonal <strong>prediction</strong>s: precipitation


Probabilistic scores<br />

Reliability diagram: How well do the predicted probabilities of an<br />

event correspond to their observed frequencies?<br />

The range of forecast probabilities is divided into K bins (for example, 0-5%, 5-15%, 15-25%, etc.).<br />

The sample size in each bin is often included as a histogram or values beside the data points.


Skill single and multi-model ensemble<br />

From EU Demeter project (Doblas Reyes)


Probabilistic score<br />

Relative Operation Characteristic: What is the ability of the forecast<br />

to discriminate between events and non-events?<br />

Plot hit rate (POD) vs false alarm rate (POFD), using a set of increasing probability thresholds<br />

(for example, 0.05, 0.15, 0.25, etc.) to make the yes/no decision.<br />

The area under the ROC curve is frequently used as a score.


Temperature<br />

Precipitation


ROC score, 3 month lead time<br />

Temperature<br />

Sea level pressure


Probabilistic scores<br />

Rank histogram: How well does the ensemble spread of the forecast<br />

represent the true variability (uncertainty) of the observations?<br />

1. At every observation (or analysis) point rank the N ensemble members from lowest to highest.<br />

2. Identify which bin the observation falls into at each point<br />

3. Tally over many observations to create a histogram of rank.<br />

Flat - ensemble spread about right to represent forecast uncertainty<br />

U-shaped - ensemble spread too small, many observations falling outside the extremes of the<br />

ensemble<br />

Dome-shaped - ensemble spread too large, most observations falling near the center of the ensemble<br />

Asymmetric - ensemble contains bias


Always use a 0-hypothesis<br />

• Verify against<br />

simplest statistical<br />

X ( t + τ )<br />

model (AR1,<br />

damped persistence, A(<br />

τ ) = 0<br />

or climatology) A(<br />

τ ) = 1<br />

=<br />

A(<br />

τ ) X ( t)<br />

(climatology)<br />

(persistence)


Skill dynamische verwachtingen<br />

l<br />

l<br />

l<br />

l<br />

Temperatuur: trend + persistentie + ENSO<br />

Neerslag: ENSO + beetje meer, beter dan puur statistische<br />

modellen<br />

Kan regionaal verbeterd worden met downscaling technieken (bv<br />

SVD tussen waargenomen en verwachte velden om spatiële biases<br />

te corrigeren)<br />

Zal beter worden naarmate meer relevante processen goed<br />

gecalibreerd maagenomen worden.


The model and bias


Based on ECMWF weather<br />

and seasonal <strong>prediction</strong><br />

system<br />

Regular merging with<br />

operational ECMWF<br />

NWP system<br />

Resolution standard runs:<br />

Atmosphere: T159 L62<br />

(AMIP runs up to T799)<br />

Hazeleger et al BAMS, October, 2010<br />

Ocean: 1 degree L42 (with<br />

equatorial refinement<br />

and tripolar grid)


Low (T159) and high resolution (T799)


EC-Earth: temperature and bias


EC-Earth: precipitation and bias


EC-Earth: El Nino


EC-Earth: the North Atlantic Oscillation<br />

Observations<br />

EC-Earth


EC-Earth: the Atlantic Ocean overturning


EC-Earth: global mean temperature<br />

EC-Earth<br />

Observations


‘Weather’ <strong>prediction</strong>s


Seasonal <strong>prediction</strong>s


Seasonal <strong>prediction</strong>s: mean bias<br />

May<br />

Nov<br />

Bias of first month near-surface air temperature re-forecasts wrt ERA40/Int over 1976-2005.


Seasonal <strong>prediction</strong>s: ENSO (2-4 month lead time)<br />

EC-Earth<br />

Ratio sd: 1.34<br />

Corr: 0.82<br />

RPSSd: 0.48<br />

ECMWF System 3<br />

Ratio sd: 0.84<br />

Corr: 0.86<br />

RPSSd: 0.68<br />

Niño3.4 time series for ERA40/Int (red dots), ensemble range (green box-and-whisker)<br />

and ensemble mean (blue dots) 2-4 month (JJA) re-forecasts over 1981-2005.<br />

Doblas-Reyes, IC3-group


Correlation skill<br />

1-month<br />

lead time<br />

JJA<br />

DJF<br />

7-month<br />

lead time<br />

Ensemble-mean correlation of EC-Earth near-surface air temperature re-forecasts<br />

wrt ERA40/Int over 1976-2005. Dots for values statistically significant with 95% conf.


Seasonal forecast ENSO


Seasonal forecast July-Sept 2011 temperature<br />

Left: <strong>prediction</strong>; anomalies probability above mean<br />

Right: ROC skill (hit rate vs false alarm rate)


Seasonal <strong>prediction</strong>s: tropical cyclones


Decadal <strong>prediction</strong>s


Decadal <strong>prediction</strong>s in EC-Earth: global mean SST<br />

Wouters et al, in prep


Decadal <strong>prediction</strong>s in EC-Earth, drift corrected<br />

Wouters et al, in prep


Verification multi-model EU-ENSEMBLES decadal<br />

<strong>prediction</strong>s<br />

van Oldenborgh, Doblas-Reyes, Wouters and Hazeleger, subm


Skill multiannual <strong>prediction</strong>s<br />

anomaly correlations ensemble mean – observations<br />

lead time 2-5 years and 6-9 years averaged


Verification multi-model decadal <strong>prediction</strong>s<br />

van Oldenborgh, Doblas-Reyes, Wouters and Hazeleger, subm


Skill initialized <strong>prediction</strong>s at multiannual lead times<br />

• Trend, as defined by regression on global mean CO 2 concentrations,<br />

removed<br />

• Seasonal predictability comparable to ECMWF S3


Skill initialized <strong>prediction</strong>s at multiannual lead times<br />

R=0.77 R=0.80 R=0.19


Skill initialized <strong>prediction</strong>s at multiannual lead times<br />

• Trend, as defined by regression on global mean CO 2 concentrations,<br />

removed<br />

• Seasonal predictability comparable to ECMWF S3


Predicting Atlantic Multidecadal Oscillation<br />

R=0.67 R=0.84 R=0.57


Skill of very simple statistical model


Final remarks<br />

Predictability beyond the Lorenz limit: (damped) persistence &<br />

(damped) oscillations<br />

Memory provided by slow components of the earth system: soil<br />

moisture, snow, ocean temperatures<br />

à Challenge in initialization of dynamical models (atmosphere as in<br />

weather <strong>prediction</strong>, but memory quickly lost. Ocean, snow, ice, land<br />

surface relevant for longer time scales)<br />

Statistical models perform well in tropics, dynamical models can<br />

outperform: predictability at multiannual time scales found (AMO,<br />

PDO)<br />

Always verify! If possible probabilistically


http://www.cawcr.gov.au/projects/verification

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

Saved successfully!

Ooh no, something went wrong!