Priority project « Advanced interpretation and verification of very high resolution models »

Federal Department of Home Affairs FDHAFederal Office of Meteorology and Climatology MeteoSwiss

Priority project « Advanced interpretation and verification

of very high resolution models »

2 COSMO General meeting ¦ Cracow, September 2008Pierre.Eckert[at]meteoswiss.ch

Topics

1. Advanced postprocessing of weather parameters

2. Verification of very high resolution models, incl. fuzzy verification methods

3. Hydrological applications


3. Hydrological applications

Hydrology (precipitation adaptation):

Presentation by A. Mazur

Snow parametrisation:

Presentation by E. Machulskaya


1. Recognition of weather elements

• Done last year: recognition of thunderstorms with the boosting algorithm:

• Choice of predictors

Perler, Kohli, Walser


1. Kalman filtering of COSMO LEPS

V. Stauch, poster outside


2. Verification of very high resolution models

Goals

• 1-3 km scale (VHR)• Focus on precipitation• Is VHR (~2km) better than HR (~7km)?• Model intercomparison• Generate products related to the verification• Way to define the scores could depend on the

application (value)• Use synop, (high resolution rainguage network),

radar, evt. composition of all (gridded observations)


Which rain forecast would you rather use?

Mesoscale model (5 km) 21 Mar 2004

Sydney

Global model (100 km) 21 Mar 2004

Sydney

Motivation

Observed 24h rain

RMS=13.0 RMS=4.6

B. Ebert


Motivation: precipitation pattern

7km

2km


Fuzzy Verification F. Ament

Verification on coarser scales than model scale: “Do not require a point wise match!“

X XX X

X Xx XXX

x

Method Raw Data Fuzzyfication Score Example result

Upscaling

Average

Equitable threat score

Fraction Skill

Score (Roberts and Lean, 2005)

Fractional coverage

Skill score with reference to

worst forecast

X XX X

X Xx XXX

x


Expected behaviour of scores

From Nigel Roberts (2005)


Application of scores to a perfect forecast

All scores should equal !

But, in fact, 5 out of 12 do not!


Requested theoretical properties of scores

Avoid « leaking » scores Use illustrative and understandable scores Score should give a real information of the forecast

quality on the different scales Monotonic behavior concerning

• scale (best values for large scales)• frequency of occurrence (best values for high

frequencies of occurrence) Represent some significant characteristics of the PDF

(obs and forecast)


Requested practical properties of scores

Agreement between subjective and objective judgment

Possible help in decision making Correspond to the needs of the users Should be able to provide a comparison between 2km

and 7 km models (also global models) Should not use a matching between prediction and

observation because it would not allow the generation of univocal products


Chosen scores

Our best candidates:

Upscaling and Fraction skill score

Corresponding products

• Upscaling mean around a point / station• Fraction skill score probability to exceed some

threshold in a neighbourhood


Sp

atia

l sca

le (

km)

Sp

atia

l sca

le (

km)

Fuzzy Verification: COSMO-DE – COSMO-EU

90

58

33

20

7

90

58

33

20

7

goodbad

Threshold (mm/3h) Threshold (mm/3h) Threshold (mm/3h)

- =

Fra

ctio

n sk

ill s

core

Ups

calin

g

=-COSMO-EU (7km)COSMO-DE (2.8km) Difference

COSMO-EU better COSMO-DE better

JJA 2007, Verification against Swiss Radar Composite, 3 hourly accumulations


Sp

atia

l sca

le (

km)

Sp

atia

l sca

le (

km)

Fuzzy Verification COSMO-2 – COSMO-7

90

58

33

20

7

90

58

33

20

7

goodbad

Threshold (mm/3h) Threshold (mm/3h) Threshold (mm/3h)

- =

Fra

ctio

n sk

ill s

core

Ups

calin

g

=-COSMO-7 (7km)COSMO-2 (2.2km) Difference

COSMO-7 better COSMO-2 better

JJA 2007, Verification against Swiss Radar Composite, 3 hourly accumulations


Monthly dependencycut-off 03h, accumulation 03h

COSMO-DE -

COSMO-EU

June

COSMO-2-

COSMO-7

July

August


Quarterly summaries of „Fuzzy“-scores

FSS Autumn 2007

U. Damrath


Monthly summaries of „Fuzzy“-scores

FSS July 2007

2020

Analysis of precipitation in boxesAnalysis of precipitation in boxes

Average number of stations in each area

( SON 2007)

X

We devised a verification methodology by aggregating observed and predicted precipitation in boxes of 1°x 1°(labelled boxes in the map)

The choice of the size and position of the areas has been performed according to different rules:

• boxes have to be enough large in order to contain a high number of observation points (ranging from 20 to over 100, depending on location and period of time considered)

• boxes have to be homogeneous as much as possible in terms of geographic-territorial characteristics

M.-S. TesiniM.-S. TesiniC. CacciamaniC. Cacciamani

2121

Box 2 aut2007Box 2 aut2007

25 mm/24

23 mm/24

19 mm/24

90th percentile of “climatological” pdf

2222

Consideration on “day-by-day” Consideration on “day-by-day” behaviour behaviour

• COSMO-I7 seems to be more realistic than ECMWF in reproducing the intra-box variability.

• However, COSMO-I7 presents both a large number of false alarms and high “spikes”. On the other hand, ECMWF presents a greater number of missed alarms, especially for high thresholds.

• According to most standard verification measures, COSMO-I7 forecast would have poor quality, but it might be very valuable to the forecaster since it provides information on the distribution and variability of the rain field over the considered region.


Neighbourhood method P. Kaufmann

• Cylindrical neighbourhood with fading zone

• Settings at MeteoSwiss: • COSMO-7 (6.6 km):

rxy= 5, rf= 5, rt=3

• COSMO-2 (2.2 km): rxy=10, rf=10, rt=1

• Effective radius:• COSMO-7: ~50 km• COSMO-2: ~35 km

x

y

t


12 July: high probabilities match well with precipitation pattern

24 h sum 06 – 06 UTC next day

Probability of 12 h sum above 35 mm

06 – 18 UTC

18 – 06 UTC


15 August: high probabilities match well precipitation pattern



06 – 18 UTC

18 – 06 UTC


17 July: completely missed event



06 – 18 UTC

18 – 06 UTC


Conclusionson verification of very high resoution models

• Results of Upscaling and Fraction skill score are reasonable.• Scores increase with box size, but it is difficult to extract

optimal size by looking at one single model.• Overall better results for very high-res models• This benefits of very high-res models is rather to see in

situations where precipitation variability is large: convection , orography, summer…

• …and at scales of 30 to 50 km• Products can be generated

• Regional means (not new)• Probability to exceed threshold in neighborhood• Or possibly the whole pdf?

Documents

Priority project « Advanced interpretation and verification of very high resolution models »