Transcript Michael Denhard

SRNWP-PEPS
a regional multi-model ensemble in Europe
April 2005: 19 NWS/ 21 forecast products
(1) Austria
ALADIN-LACE (9.6 km)
ARPEGE
(2)
Czech Repub
ALADIN-LACE (9 km)
ARPEGE
(3)
Croatia
ALADIN-LACE (9 km)
ARPEGE
(4)
Hungary
ALADIN-LACE (11 km)
ARPEGE
(5)
Slovakia
ALADIN-LACE (11 km)
ARPEGE
(6)
France
ALADIN (11 km)
ARPEGE
(7)
Belgium
ALADIN (15 km)
ARPEGE
(8)
Slovenia
ALADIN (9.5 km)
ARPEGE
(9)
UK
UM-EU/LAM (20/12 km)
UM-global
(10) Denmark
HIRLAM (16 km)
ECMWF
(11) Finland
HIRLAM (22km)
ECMWF
(12) Netherlands
HIRLAM (22 km)
ECMWF
(13) Spain
HIRLAM (22 km)
ECMWF
(14) Ireland
HIRLAM (16 km)
ECMWF
(15) Norway
HIRLAM (22/11 km)
ECMWF
(16) Switzerland
aLMo (7 km)
ECMWF
Sebastian Trepte [email protected]
(17) Italy
EuroLM (7km)
EuroHRM
Michael Denhard [email protected]
(18) Germany
LM (7 km)
GME
(19) Poland
Institute of Meteorology and Water Management
Internet: www.dwd.de/PEPS
Jean Quiby
März 2005
[email protected]
Deutscher Wetterdienst
1
Ensemble
Generation
PEPS grid with a
grid spacing of
0.0625° (~7 km)
covering Europe
Methodology
Pi ( x  T ) 
Numberof forecasts x exceeding T at i
Ni
where N i is the total number of forecasts at grid point i and T is a threshold
The ensemble size depends on location and every PEPS grid point
has its own probability distribution
März 2005
Deutscher Wetterdienst
2
Ensemble Products
1. Ensemble mean. Forecast periods +06...+30h (24 hours), +06...+18h and +18...+30h (12 hours)
• Total precipitation (accumulation), sum of convective and large scale precipitation
• Total snow (accumulation) ), sum of convective and large scale snow
• Maximum 10 m wind speed
• Maximum 10 m wind gust speed
• 2 m minimum/maximum temperature
2. Probabilistic products. Forecast period +06...+30h (24 hours)
• Probabilities of total precipitation
• Probabilities of total snow
• Probabilities of maximum wind speed
• Probabilities of maximum wind gust speed
Thresholds:
Thresholds:
Thresholds:
Thresholds:
> 20, > 50, > 100 mm
> 1, > 5, > 10, > 20 cm
> 10, > 15, > 20, > 25 m/s
> 10, > 15, > 20, > 25, > 33 m/s
3. Probabilistic products. Forecast periods +06...+18h and +18...+30h (12 hours)
• Probabilities of total precipitation
• Probabilities of total snow
• Probabilities of maximum wind speed
• Probabilities of maximum wind gust speed
Thresholds:
Thresholds:
Thresholds:
Thresholds:
> 25, > 40, > 70 mm
> 1, > 5, > 10, > 20 cm
> 10, > 15, > 20, > 25 m/s
> 10, > 15, > 20, > 25, > 33 m/s
4. Ensemble size per grid point (at least two members)
März 2005
Deutscher Wetterdienst
3
Maximum Ensemble Size
depends on main run and on meteorological parameter
März 2005
Maximum
Total
Ensemble Size precip.
Total Wind
snow speed
Wind gust
speed
Temperature
00 UTC
06 UTC
12 UTC
20
7
20
19
6
19
20
7
20
8
8
20
7
20
18 UTC
8
7
8
1
8
Deutscher Wetterdienst
4
Ensemble Mean
21/01/2005 00 UTC +06...30
März 2005
Deutscher Wetterdienst
5
probability forecasts
21/01/2005 00 UTC +06...30
März 2005
Deutscher Wetterdienst
6
Cut-off times
Model run
cut-off time
00 UTC
05.30 UTC
06 UTC
11.30 UTC
12 UTC
17.30 UTC
18 UTC
23.30 UTC
SRNWP-PEPS runs operationally since December 2004
(Distribution of forecasts to the contributing NWS)
März 2005
Deutscher Wetterdienst
7
The SRNWP-PEPS project
SRNWP-PEPS workshop
6th April 2005, ARPA-SIM, Italy
 products
 validation
 further developement
 rights of use
März 2005
Deutscher Wetterdienst
8
Workshop
products
 Mask of areas without sufficient models
 Wind gusts
provided by COSMO and some ALADIN countries
using different parametrisations
statistical estimation of wind gusts within PEPS?
 Statistics of availability of models
 Additional products
more sysoptic oriented parameters
indices of convectivity
 Precipitation
median instead of mean
lower thresholds
 PEPS-Meteograms (provided by Meteoswiss)
März 2005
Deutscher Wetterdienst
9
Workshop
validation
• Comparison with COSMO-LEPS
• Scoring probabilistic forecasts
- error measures
- FBI, POD, FAR, ETS, HSS, Odds Ratio
- BS, BSS, RPS, ROC
• Scale-/Object oriented techniques
- contiguous rain area method (Ebert &McBride)
• Severe weather Problem
- linear error in probability space (LEPS)
• Online verification
WG on Verifcation to coordinate verification with high resolution observations in
the contributing countries and to provide scientific expertise.
März 2005
Deutscher Wetterdienst
10
Workshop
further developement
Ensemble Calibration
Calibrated:
Intervals or events that we declare to have probability
P happen a proportion P of the time
Sharp:
Prediction intervals are narrower on average than
those obtained from climatology; the narrower the better
Dressing the probability distriubtion of the ensemble with
observational errors and give different weights to the ensemble
members
März 2005
Deutscher Wetterdienst
11
Workshop
further developement
Using Bayesian Model Averaging (BMA) to calibrate
forecast ensembles
Adrian E. Raftery, Fadoua Balabdaoui, Tilmann Gneiting and Michael
Polakowski
Department of Statistics, University of Washington, Seattle, Washington
p(y | ~
y1 ,..., ~
yn )   wk (ak  bk ~
yk ,  2 )
k
y
~
y
is the observed value
1 is the k th forecast
„The model is estimated from a training set of recent
data by maximum likelihood using the EM algorithm.
Good results with a 25-day training period.“
März 2005
Deutscher Wetterdienst
12
Workshop
further developement
BMA
work on precipitation is in progress
Software R package EnsembleBMA is available
Source
www. stat. washington. edu/ raftery
www. stat. washington. edu/ MURI
März 2005
Deutscher Wetterdienst
13
Workshop
further developement
The SRNWP-PEPS consits of different model grids with different
horizontal and vertical resolutions.
Question:
How can we account for these differences in an appropriate way ?

Statistical downscaling ?

Neighbourhood Ensemble ?
März 2005
Deutscher Wetterdienst
14
Workshop
further developement
Neighbourhood Ensemble ?
consider all gridpoints within a given distance of a point
spatial
temporal
Size of Area
t
Form of Area
x
Neighbourhood members from different grids should not have equal weights
Systematic errors (e.g. due to orography) should be corrected
März 2005
Deutscher Wetterdienst
15
Workshop
further developement
Concerning GLOBAL PEPS:
„According to most skill measures, these hybrid configurations outperform the
ECMWF-EPS at short range for most variables, regions and thresholds“
from:
Test of a Poor Mans Ensemble Prediction System for short range probability forecasting
Arribas, A., Robertson, K.B., Mylne, K.R.
Hybrid LAM-Ensemble ?
concatenate SRNWP-PEPS with other ensemble systems
• COSMO-LEPS
• INM Ensemble
• Meteo France PEACE Ensemble
• UK-Met Office LAM
• met-norway LAMEPS
März 2005
Deutscher Wetterdienst
16
Workshop
research projects using PEPS forecasts & products
Hydrological Ensemble Forecasts for the „MULDE“ catchment
Hybrid Ensemble
COSMO-LEPS (+120h)
SRNWP-PEPS (+48h)
LMK (2.8km)
"lagged average forecast"
Ensemble
(+18h)
consistent forecast scenarios of precipitation
for the Mulde catchment
up to +120h
International projects which use or may use SRNWP-PEPS forecasts
- EURORISK Prev.I.EW windstorms workpackage
- MAP D-Phase (Mesoscale Alpine Program)
März 2005
Deutscher Wetterdienst
17
Workshop
rights of use
Scientific use
products as well as individual forecasts
historic as well as live data
Request to DWD
DWD distributes the request
to the contributing NWS
NWS give their permission
Commercial use
products only
products have to be added to the ECOMET list with
permission of the NWS
März 2005
Deutscher Wetterdienst
18
Thank you
to all contributing Weather Services !
any questions or remarks ?
März 2005
Deutscher Wetterdienst
19