X General COSMO meeting Cracow, 15-19 September 2008

A.Montani; COSMO-LEPS status, verification, …X COSMO meeting - Cracow - 15-19 September 2008

COSMO LEPS system:

status, verification, applications

Andrea Montani,

C. Marsigli, T. Paccagnella

ARPA-SIM, Bologna, Italy

F. Fundel, A. Walser, M. A. Liniger, C. Appenzeller

Meteoswiss

X General COSMO meetingCracow, 15-19 September 2008


Outline

• Introduction• Present status and recent upgrades• Verification results (SYNOP on the GTS):

– performance from 2002 onwards

• Application of COSMO-LEPS: the reforecast exercise

• Main results


COSMO-LEPS (developed at ARPA-SIM)

• What is it?It is a Limited-area Ensemble Prediction System (LEPS),

based on COSMO-model and implemented within COSMO (COnsortium for Small-scale MOdelling, which includes Germany, Greece, Italy, Poland, Romania, Switzerland).

• Why?It was developed to combine the advantages of global-

model ensembles with the high-resolution details gained by the LAMs, so as to identify the possible occurrence of severe and localised weather events (heavy rainfall, strong winds, temperature anomalies, snowfall, …)

generation of COSMO-LEPS to improve the Late-Short (48hr) to Early-Medium (132hr) range forecast of

severe weather events.


COSMO-LEPS suite @ ECMWF before and during D-PHASE Operations Period (from 1 June to 30 November 2007)

d-1d-1 dd d+5d+5d+1d+1 d+2d+2 d+4d+4d+3d+3

older EPSolder EPS

younger EPSyounger EPS

clustering clustering periodperiod

0000

1212

Cluster Analysis and RM identificationCluster Analysis and RM identification

4 variables4 variables

Z U V QZ U V Q

3 levels3 levels

500 700 850 hPa500 700 850 hPa

2 2 time time stepssteps


European European areaarea

Complete Complete LinkageLinkage

COSMO-LEPS

Integration Domain

1616 Representative Representative Members driving the Members driving the 1616

COSMO-model COSMO-model integrations (weighted integrations (weighted according to the cluster according to the cluster

populations)populations)

employing either employing either Tiedtke or Kain-Fristch Tiedtke or Kain-Fristch

convection scheme convection scheme (randomly choosen) (randomly choosen)

COSMO-LEPS

clustering area

• suite runs as a ”time-critical application” managed by ARPA-SIM;

• Δx ~ 10 km; 40 ML;• COSM0 v3.20;• fc length: 132h;• Computer time (4.7 million

BU for 2007) provided by the COSMO partners which are ECMWF member states.


Recent updatesChanges to the operational suite on 1 December 2007:Changes to the operational suite on 1 December 2007:• new model version (COSMO-LM 4.0);• new numerics (Runge-Kutta scheme); • use of multi-layer soil model;• new perturbations introduced relative to the maximal turbulent length scale (tur_len) and

to the length scale of thermal surface patterns (pat_len) to increase spread and maintain skill.

Old system and new system ran in parallel for 52 days (October and November 2007);

Larger T2m spread in new system for all forecast ranges;

Comparable skill (in terms of T2m root-mean-square error) of the two systems.


COSMO-LEPS suite @ ECMWFafter D-PHASE Operations Period (since December 2007)

d-1d-1 dd d+5d+5d+1d+1 d+2d+2 d+4d+4d+3d+3

older EPSolder EPS

younger EPSyounger EPS

clustering clustering periodperiod

0000

1212


4 variables4 variables

Z U V QZ U V Q

3 levels3 levels

500 700 850 hPa500 700 850 hPa

2 2 time time stepssteps


European European areaarea

Complete Complete LinkageLinkage

COSMO-LEPS

Integration Domain

1616 Representative Representative Members driving the Members driving the 1616

COSMO-model COSMO-model integrations (weighted integrations (weighted according to the cluster according to the cluster

populations)populations)

employing either employing either Tiedtke or Kain-Fristch Tiedtke or Kain-Fristch

convection scheme convection scheme (randomly choosen) (randomly choosen)

++perturbations in perturbations in

turbulence schemeturbulence scheme

COSMO-LEPS

clustering area

• suite runs as a ”time-critical application” managed by ARPA-SIM;

• Δx ~ 10 km; 40 ML; fc+ 132h;• COSM0 v4.0 since Dec07

(with RK + multi-layer);• computer time (5.8 million

BU for 2008) provided by the COSMO partners which are ECMWF member states.


– SYNOP on the GTS

Objective verification of COSMO-LEPS

Main features:

variable: 12h cumulated precip (18-06, 06-18

UTC);

period : from Dec 2002 to Aug 2008;

region: 43-50N, 2-18E (MAP D-PHASEPHASE

area);

method: nearest grid point; no-weighted fcst;

obs: synop reports (about 470

stations/day);

fcst ranges: 6-18h, 18-30h, …, 102-114h, 114-126h;

thresholds: 1, 5, 10, 15, 25, 50 mm/12h;

system: COSMO-LEPS;

scores: ROC area, RPSS, Outliers

both (69) monthly and (23) seasonal scores were computed

verification over the full domain (about 1400 stations/day) is underway


Time series of ROC area Area under the curve in the HIT rate vs FAR diagram. Valuable forecast systems have ROC area values > 0.6.

ROC The positive

impact of increasing the ensemble size in 2004 is evident for all thresholds and for different forecast ranges.

Poor performance of the system in Spring and Summer 2006 (both particularly dry), despite system upgrades.

Good performance during DOP 2007.

Small positive impact of 2007 system upgrades.


Seasonal scores of ROC area Performance of the system assessed for the last 5 summers (JJA) and the last 5 autumns

(SON). Consider the “event” 10 mm/12h; valuable forecast systems have ROC area values > 0.6.

ROC Need to take

into account the different statistics for each season (JJA 2004 and SON 2006 less rainy than the others).

Good scores obtained throughout the full DOP in 2007.

Best performance for JJA 2008, more evident for longer ranges.


Outliers: time series + seasonal scores How many times the analysis is out of the forecast interval spanned by the ensemble

members. … the lower the better … Performance of the system assessed as time series and for the last 5 summers (JJA).

OUTL Evident seasonal

cycle (more outliers in winter), but overall reduction of outliers in the years up to 2007.

Reduction of outliers from one summer to the other, related to the increase of ensemble size (evident for the 10m to 16m increase).

Need to take into account the different statistics for each season.

Best results for the last 2 summers.


COSMO-LEPS reforecasts (v 4.0)Motivations:• COSMO-LEPS is often not reliable; • probabilities might be wrong! need

for calibration

Setup• Period of 30 years (1971-2000)• 1 deterministic run of COSMO-LEPS• fc+90h, 12:00 initial time (every 3rd

day)• ERA40 as initial/boundary conditions.• Calculated and archived at ECMWF• Convect. scheme = Tiedtke/Kain-Fritsch• Random physics (turlen & patlen)

Limitations• New climatology needed with each

model version change• Needs time and is costly (~ 2000

KBU/yr)


• 24h Tot_Prec (0600-0600 UTC)• Domain: Switzerland• interp. on CLEPS Grid (C.Frei)• Apr06 – Aug07 (verification)• 1971-2000 (calibration)

Observation Data

Verification of calibrated vs raw forecasts

model topography [m

]


calibratedraw

BSSD 24h precipitation BSSD is written as 1-BS/(BSref +D). Useful forecast systems if BSSD > 0. BS measures the mean squared difference between fcst and obs in

probability space. Equivalent to MSE for deterministic forecast.

winter summer


Main results (1)

• Since November 2002, COSMO-LEPS system has been running on a daily basis.

• Since June 2008, COSMO-LEPS data are disseminated to JRC for the use in the European Flood Alert System (EFAS) as input to the flood forecasting system over the major European basins ==> success for COSMO both from a scientific and a visibility point of view.

• COSMO-LEPS products routinely used in met-ops rooms across COSMO community as well as in research Projects (e.g. COPS, D-PHASE, PREVIEW: see the poster).

increase in ROC area scores and reduction in outliers percentages; positive impact of increasing the population from 5 to 10 members (June

2004); some deficiency in the skill of the system can be identified after the system

upgrades occurred on February 2006 (from 10 to 16 members; from 32 to 40 model levels), BUT

scores are encouraging throughout DPHASE Operations Period; system upgrades of Dec07 (RK + multi-layer + new perturbations)

brought small but positive impact.

Time-series verification scores indicate the following trends:


Main results (2)

• Calibrating with reforecasts improves the forecast skill significantly (especially in winter).

• Calibrated warnings without using observations are possible.

• Calibrating frequent precipitation events does not require a large calibration period, while calibrating extreme events does (not shown).


Thank you !


Present systemx = 10 kmz = 40 MLt = 90 sngp = 306x258x40 =

3.157.920fcst range = 132hcost = 640 BU x runelapsed time = 45 min

Feasibility study: COSMO-LEPS at 7 km (the answer to forecasters’ dream?)

New systemx = 7 kmz = 40 MLt = 66-72 sngp = 510x405x40 =

8.262.000fcst range = 132hcost = 1925 BU x runelapsed time = 138 min

… cosmoleps_7 is about 3 times more expensive than present configuration

new computer at ECMWF being installed

Computer resources for each ECMWF member state will increase by a factor of 5 (five) and ….


… the dream is possible

COSMO-LEPS 10 kmCOSMO-LEPS 7 km

the grid of cosmoleps_7 would be almost identical to that of COSMO-EU, this making easier and cleaner the use of initial fields provided by DWD (e.g. soil moisture analysis).


• test the use of the Soil Moisture Analysis fields provided by DWD;

• run cleps_7 for ~ 40 days in autumn 2008 and assess the impact;

• within TIGGE-LAM, develop coding of COSMO-LEPS output files in GRIB2 format;

Future plans (2008 and 2009)

• migration to the new machine at ECMWF;

• use a better snow analysis (possibly provided by DWD or Meteoswiss);

• extend the cluster analysis so as to consider not only ECMWF EPS, but also UKMO MOGREPS as global ensemble providing ic’s and bc’s (first tests);

• implement cosmoleps_7;

• gaining from COSMO-SREPS experience, introduce more model perturbations;

• test COSMO-LEPS nested on the under-development ECMWF EDA over MAP D‑PHASE period;

• optimise use of reforecasts + calibration of wind gust;

• support CONSENS.


Feasibility study of COSMOLEPS at 7 km

Motivations: Provide a more detailed description of

mesoscale processes by incresing the horizontal resolution.

Do not lose a “reasonable advantage” against ECMWF EPS, which will go to x=25 km during 2009.

“Keep the pace” with deterministic model (x~ 2-3 km): if the gap in resolutions between deterministic and probabilistic systems is too large, the two systems do not support each other and go for different solutions (that is, they forecast different weather!).from 10 to 7 km (plus small domain extensions) does not

seem a lot


but …


Present methodology


Operational set-up

Core products: 16 perturbed COSMO-model runs (ICs and 3-

hourly BCs from 16 EPS members) to generate, “via weights”, probabilistic output (start at 12UTC; t = 132h);

Additional products: 1 deterministic run (ICs and 3-hourly BCs from the

high-resolution deterministic ECMWF forecast) to “join” deterministic and probabilistic approaches (start at 12UTC; t = 132h);

1 “hindcast” run (ICs and 3-hourly BCs from ECMWF analyses) to “downscale” ECMWF information (start at 00UTC; t = 36h).


Recent updatesChanges to the operational suite on 1 December 2007:Changes to the operational suite on 1 December 2007:• new model version (COSMO-LM 4.0);• new numerics (Runge-Kutta scheme); • use of multi-layer soil model;• new perturbations introduced relative to the maximal turbulent length scale (tur_len) and

to the length scale of thermal surface patterns (pat_len) to increase spread and maintain skill.

Old system and new system ran in parallel for 52 days (October and November 2007);

Larger T2m spread in new system for all forecast ranges;

Comparable skill (in terms of T2m root-mean-square error) of the two systems.

back to plans


Time series of ROC area for COSMO-LEPS (2)

Area under the curve in the FAR vs HIT diagram. Valuable forecast systems have ROC area values > 0.6

ROC

the positive impact of increasing the ensemble size in 2004 is evident for all thresholds and for different fcst ranges.

poor performance of the system in Spring and Summer 2006, despite system upgrades.

Jun04: 5m 10m

Feb06: 10m 16m; 32ML 40 ML

fc step: 78-90h


Time series of Brier Skill Score BSS is written as 1-BS/BSref. Sample climate is the reference system. Useful forecast systems

if BSS > 0. BS measures the mean squared difference between forecast and observation in probability

space. Equivalent to MSE for deterministic forecast.BSS

improvement of performance detectable for all thresholds along the years;

still problems with high thresholds, but good trend in 2007.

fc step: 30-42h

Jun04: 5m 10m

Feb06: 10m16m; 32ML 40 ML


Semi-diurnal cycle in COSMO-LEPS scores

BSS score … the higher the better … Performance of the system assessed for 5 different Summers (JJA).

BSS Evident 12-hour cycle in

BSS scores (the same holds for RPSS, while less evident for ROC area scores).

Better performance of the system for “night-time” precipitation, that is for rainfall predicted between 18Z and 6Z (ranges 30-42h, 54-66h, …).

The amplitude of the cycle is somewhat reduced throughout the years and with increasing forecast range.

The bad performance in Summer 2006 is confirmed.


Score dependence on the domain size (1)

Verification of COSMO-LEPS against synop reports over the MAP D-PHASE area (~ 470 stations; MAPDOM) and the full domain (~ 1500 stations; fulldom):

different statistics of the verification samples; up to now, performance of the system over the 2 domains assessed only for the

last 6 months (March-August 2007). difficult to draw general conclusions


Score dependence on the domain size (2)

RPSS

RPSS score… the higher the better… (and positive).

ROC area… the higher the better… (and above 0.6). Smoother transitions from month to month in “fulldom” scores.

Slightly better performance of COSMO-LEPS over the MAPDOM, but the signal varies from month to month.

Higher predictability with orographic forcing?

Need to check individual regions and/or to stratify for type of stations.

OUTL

ROC

Outliers percentage … the lower the better.


COSMO-LEPS

16-MEMBER EPS

51-MEMBER EPS

tp > 1mm/24h

tp > 5mm/24h

Average values (boxes 0.5 x 0.5)MAM06

As regards AVERAGE precipitation above these two threshols, the 3 systems have similar performance.


Main results

• COSMO-LEPS system runs on a daily basis since November 2002 and it has become a “member-state time-critical application” at ECMWF ( partial involvement of ECMWF operators in the suite management).

• Time series scores identify positive trends in the performance of the system:• increase in ROC area scores and reduction in outliers percentages;• positive impact of increasing the population from 5 to 10 members (June

2004);• some deficiency in the skill of the system identified after the system

upgrades occurred on February 2006 (from 10 to 16 members; from 32 to 40 model levels + EPS upgrade!!!) need of further investigation.

• High-resolution verification shows better scores of COSMO-LEPS with respect to EPS in forecasting precipitation maxima within boxes.• No clear impact of the weighting procedure as regards precipitation, BUT COSMO-LEPS ensemble mean for 2m temperature has better skill (lower standard deviation to observations) with weighting the ensemble members according to cluster populations compared to the mean with unweighted members.


Time series of ROC area Area under the curve in the HIT rate vs FAR diagram. Valuable forecast systems have ROC area values > 0.6.

ROC the positive

impact of increasing the ensemble size in 2004 is evident for all thresholds and for different forecast ranges.

poor performance of the system in Spring and Summer 2006 (both particularly dry), despite system upgrades.

good performance during DOP 2007.

Jun04: 5m 10m

Feb06: 10m 16m; 32ML 40 ML

fc step: 30-42h

DOP

fc step: 78-90h

Jun04: 5m 10m

Feb06: 10m 16m; 32ML 40 ML

DOP


Seasonal scores of ROC area Performance of the system assessed for 5 different summers (JJA) and 5 different autumns

(SON). Consider the “event” 10 mm/12h; valuable forecast systems have ROC area values > 0.6.

ROC Need to take

into account the different statistics for each season (JJA 2003 and SON 2006 less rainy than the others).

Best scores obtained during DOP 2007 for either seasons.


Time series of Ranked Probability Skill Score

A sort of BSS “cumulated” over all thresholds. BSS is written as 1-BS/BSref. Sample climate is the reference system. BS measures the mean squared difference between forecast and observation in probability space.

Useful forecast system, if RPSS > 0.

RPSS the improvement of

the system performance is detectable for all forecast ranges along the years;

the poor performance of the system in summer 2006 is confirmed;

RPSS always positive throughout 2007 and especially high during DOP.

Jun04: 5m 10m

Feb06: 10m16m; 32ML 40 ML

DOP


Outliers: time series + seasonal scores How many times the analysis is out of the forecast interval spanned by the ensemble

members. … the lower the better … Performance of the system assessed as time series and for 5 different summers (JJA).

OUTL Evident seasonal

cycle (more outliers in winter), but overall reduction of outliers in the years.

Reduction of outliers from one summer to the other, related to the increase of ensemble size (more evident for the 5 to 10 increase).

Need to take into account the different statistics for each season (JJA 2003 less rainy than the others).

Best results for D-PHASE summer!

Jun04: 5m 10m Feb06: 10m 16m; 32ML 40 ML

DOP


• Verification vs “DPHASE observations”:

– assess performance over different domains (North and South of the Alps),

– study individual case studies,

– consider basin-by-basin performance.

• “Think about” increasing horizontal resolution to 7 km.

• Calibrate COSMO-LEPS fcsts using reforecasts (F. Fundel ,

Meteoswiss).

Future plans

• Implement “seamless COSMO ensemble system” merging

COSMO-LEPS and COSMO-SREPS.

• Develop “hybrid” clustering technique (take boundary conditions

from a “grand-global” ensemble provided by mixing ECMWF-EPS and

UKMO-MOGREPS).


Dissemination during DOP

probabilistic products (www.smr.arpa.emr.it/prodotti and www.d-phase.info )

“deterministic” products (individual 16 COSMO-LEPS runs) derived probability products (EM, ES) images and alerts


Time series of Ranked Probability Skill Score

A sort of BSS “cumulated” over all thresholds. BSS is written as 1-BS/BSref. Sample climate is the reference system. BS measures the mean squared difference between forecast and observation in probability space.

Useful forecast system, if RPSS > 0.

RPSS the improvement of

the system performance is detectable for all forecast ranges along the years;

the poor performance of the system in summer 2006 is confirmed;

RPSS always positive throughout 2007 and especially high during DOP.

Lesser skill in 2008 (so far).


Dim 2

Initial conditions Dim 1 Dim 2

Possible evolution scenarios

Dim 1 Initial conditions

ensemble size reduction

Cluster members chosen as representative members (RMs)

LAM integrations driven byRMs

LAM scenario

LAM scenario

LAM scenario

COSMO-LEPS methodologyCOSMO-LEPS methodology


Why calibrate?

OBS-CLEPS/(OBS+CLEPS)/2 1971-2000 LT: 42h

Jan

Jul

Q0.8 Q0.95

COSMO-LEPS is not reliable,Probabilities might be wrong!

Need for calibration


Reliability Diagram:

Verification


• raw forecast overconfident, very limited skill

• strong improvement in reliability

• long lead-time forecasts more reliable

calibratedraw

Winter 06/07 Q0.8 24h precipitation

+ (18-42h) + (66-90h)


• raw forecast overconfident

• modest improvement of reliability

• long lead-time forecasts more reliable

calibratedraw

Summer 06 & 07 Q0.8 24h precipitation

+ (18-42h) + (66-90h)


Special case BSS:

M: Ensemble size

Weigel et al. 2006,Mon. Wea. Rev.

The BSS debiased

DBS

BSBSS

REFD

1

)1(1

ooM

D

X General COSMO meeting Cracow, 15-19 September 2008

Documents

cosmo leps system

cosmoleps suite

cosmo partners

generation of cosmo

cosmo consortium

onwardsapplication of

cosmomodel integrations

new model version cosmolm