P

1
The ECMWF Seasonal Forecast System-3
Magdalena A. Balmaseda Franco Molteni,Tim
Stockdale Laura Ferranti, Paco Doblas-Reyes,
Frederic Vitart European Centre for
Medium-Range Weather Forecasts, Reading, U.K.
2
Overview

• Introduction to Seasonal Forecasts
• End to End Seasonal Forecasting System
• Importance of Ocean Initial Conditions
• ECMWF Seasonal forecasting system 3
• Overview
• Performance
• Web products
• Calibration of model output
• Multimodel (EUROSIP)
• Calibration Multimodel
• Summary

3
End to End Forecasting System
4
1997-1998 El-Niño forecast
Forecast
Initial Conditions
5
2007 La Niña
Initial Conditions
6
(No Transcript)
7
Impact on ECMWF-S3 Forecast Skill
In ECMWF S3, ocean Data Assimilation improves
forecast skill in the Equatorial Pacific,
especially in the Western Part
S2 S2ic_S3model S3
The impact of ocean initialization in the
prediction of SST is comparable to the impact of
atmospheric model cycle
8
The seasonal forecast System-3 (implem. March 07)

• COUPLED MODEL (IFS OASIS2 HOPE)
• Recent cycle of atmospheric model (Cy31R1)
• Atmospheric resolution TL159 and 62 levels
• Time varying greenhouse gasses.
• Includes ocean currents in wave model
• INITIALIZATION
• Includes bias correction in ocean assimilation.
• Includes assimilation of salinity and altimeter
  data.
• ERA-40 data used to initialize ocean and
  atmosphere in hindcasts
• Ocean reanalysis back to 1959, using
  ENACT/ENSEMBLES ocean data
• ENSEMBLE GENERATION
• Extended range of back integrations 11 members,
  1981-2005.
• Revised wind and SST perturbations.
• Use EPS Singular Vector perturbations in
  atmospheric initial conditions.
• Forecasts extended to 7 months (to 13 months 4x
  per year).

9
Rms error / spread in different ECMWF systems
Rms error of forecasts has been systematically
reduced (solid lines) .
10
(No Transcript)
11
(No Transcript)
12
ACC for seasonal-mean (1981-2005)
2m-T DJF from 1 Nov
2m-T JJA from 1 May
Precip DJF from 1 Nov
Precip JJA from 1 May
Doblas-Reyes
13
New products in the webocean reanalysis
http//www.ecmwf.int/products/forecasts/d/charts/o
cean/reanalysis/
14
New products from Sys-3 annual-range Nino indices
15
New products from Sys-3 tercile summary
16
New products from Sys-3 climagrams

• a) Teleconnection and monsoon indices with
  verification

http//www.ecmwf.int/products/forecasts/d/charts/s
easonal/forecast/
17
Climagrams area-averages of 2mT and rainfall
2m Temperature Amazones
Anomaly Correlation Temperature
Anomaly Correlation Precipitation
18
Climagrams area-averages of 2mT and rainfall
North-East Brasil
Anomaly Correlation Temperature
Anomaly Correlation Precipitation
19
Climagrams area-averages of 2mT and rainfall
South America Atlantic Coast
Anomaly Correlation Temperature
Anomaly Correlation Precipitation
20
Is the ensemble spread sufficient? Are the
forecast reliable?
Forecast System is not reliable RMS gt Spread
To calibrate the model output To sample model
error (multi-model) EUROSIP Both
21
Anomaly correlation of seasonal-mean rainfall
Franco Molteni
22
Can we predict tropical rainfall anomalies?
23
Prediction of All India Rainfall EOF filtered
fc. in JAS

CC .50
Franco Molteni
24
Prediction of All India Rainfall

JJAS CC .25 JAS CC .46
25
Prediction of East Africa short rains OND from
Aug.

Unfiltered fc. CC 0.04
EOF-filt. CC 0.42
Franco Molteni
26
Sampling model error The Real Time Multimodel
EUROSIP ECMWF-UKMO-MeteoFrance
27
TROPICAL CYCLONES
Forecasts starting on 1st June 2005 JASON
ECMWF
Met Office
Obs July-November
Meteo-France
Multi-model
Atl
W-Pac
E-Pac
1987-2004
2005
Frederic Vitart
28
MULIMODEL EUROSIP
But sometimes the spread with EUROSIP is too
large!!
ECMWF
MULTI-MODEL
29
Bayesian Calibration of the Nino Indices

• Based on the Forecast Assimilation Framework
• It will produce a revised mean and variance
• Specific Ingredients
• Take into account that error in the models can be
  correlated (remove correlation from errors, not
  from the signal, by doing SVD of error covariance
  matrix)
• Model for the errors
• Given the mean and variance, produce the
  individual plumes

30
EUROSIP Bayesian Combination
31
Sampling model error The Real Time Multimodel
EUROSIP ECMWF-UKMO-MeteoFrance
32
Conclusions

• The new ECMWF seasonal forecast system-3 gives
  improved predictions of tropical/summer
  variability respect the previous system.
• SST predictions are good in the tropical Pacific
  and eastern Indian Oc., but western Indian Oc.
  and tropical Atlantic are not better than
  persistence in NH summer.
• Difficulty in getting the correct rainfall
  variability over land. Predictive skill over land
  can be improved by exploiting teleconnections
  (calibration)
• The Multi-Model (EUROSIP) provides skilful
  predictions of tropical storms. In general it
  improves reliability, but sometimes the spread is
  too large
• Bayesian Calibration can improve the products,
  but attention should be paid to the estimation of
  the model error (sensitive to sampling size)

33
(No Transcript)
34
Climagrams monsoon indices / teleconnections
35
Prediction of All India Rainfall

JJAS CC .25 JAS CC .46
36
Can we make use of the larger scale signal?
37
Tropical storm annual frequency (1987-2004)
38
Examples of tropical storm tracks
ECMWF System 2
ECMWF System 3
39
Interannual variability of tropical storms in
EURO-SIP
Forecasts issued in June for the period
July-November
Correlation 0.72 RMS error 2.93
Frederic Vitart
40
Sampling model error The Real Time Multimodel
EUROSIP ECMWF-UKMO-MeteoFrance