MultimodelMultianalysis Mesoscale Ensemble MusE - PowerPoint PPT Presentation
1 / 1
Title:
MultimodelMultianalysis Mesoscale Ensemble MusE
Description:
Bayesian Model Averaging and methods based on non parametric distribution are ... Both SE and EM are better, on average, than the best model and in the worst ... – PowerPoint PPT presentation
Number of Views:25
Avg rating:3.0/5.0
Title: MultimodelMultianalysis Mesoscale Ensemble MusE
1
Multimodel-Multianalysis Mesoscale Ensemble (MusE)
P. A. Chessa, C. Dessy, G. Ficca , C. Castiglia
Servizio Agrometeorologico della
Sardegna, Viale Porto Torres 119, 07100, Sassari,
I email chessa_at_sar.sardegna.it
M. Marrocu, I. Di Piazza CRS4 Parco Scientifico e
Tecnologico POLARIS, Edificio 1, 09010 Pula (CA),
I email marino_at_crs4.it
- Deterministic approach
- Superensemble (SE)
- The coefficients found through a multilinear
regression - Bias removed mean (EM)
- All the model are unbiased before the calculation
of SE and EM
- Pre-operational tests
- Example for 2m Temperature
- 21 ground station over Sardinia
- Test of several training periods of different
lengths and variable position in the available
sample - Test period of 60 days with variable position
- Analysed 2m T over all domain
- Results
- Superensemble and bias removed ensemble mean
appear to work for parameters as temperature,
geopotential height, wind intensity and mslp. For
variables like rainfall a different approach has
to be used. - Both SE and EM are better, on average, than the
best model and in the worst cases (i. e. very
high spread) tend to be very close to it. This
make them the natural candidates as control
forecasts in a Multimodel context. Impossible to
say which is the best ! Minimization of scores
different from RMSE may be needed. - Better results may be obtained using an iterative
approach to calculate the regression coefficients
in order to ignore (for each grid point, variable
time step, etc) the model outputs associate to
negative values. This could also help for a
possible probabilistic interpretation.
- RATIONALE
- Support the operational activities of SAR (the
Sardinian MetService) - Provide early warning for unusual/severe weather
events - Study the predictability of important local
phenomena - Test the possibility to set up an operational
ensemble on the cheap using Linux Clusters and
developing a suitable GRID computing system
- Learning period
- Best trade off about 90 days
- Clearly flow dependent
- Pre-operational setup
- Models BOLAM - MM5 - RAMS
- I.C. and B.C. AVN 12Z - ECMWF 12Z
- Area 13.5W-34N/24.5E-54.5N
- Spatial Resol. 0.25
- Fct. time range 72h (step6h)
- Test period 15/10/2002 15/04/2003
- Implementation Cluster Linux with 16 nodes
bi-processor, Intel Xeon 3.06 Ghz 1MB
Cache L3
- The system is under-dispersive although the
spread-skill correlation is acceptable needs
calibration. - Bayesian Model Averaging and methods based on non
parametric distribution are under study for the
operational setup. - Calibration of precipitation almost impossible
over the typical learning period (60 to 90 days)
needs a different approach (perhaps reforcasting
can be one)
Better results can be obtained not considering
the model outputs associated to negative
coefficients.
- Future plans
- Design and assessment of probabilistic forecasts
products (under way) - Set up (6 to 9 months) of the operational
ensemble - Models BOLAM MM5 RAMS
- I.C. and B.C. 06Z and 18Z AVN 12h forecast of
the 12Z ECMWF run - Area 16W-30N / 32E-58N
- Spatial resolution 0.18 (nesting at higher
resolution for specific needs) - Forecast time range 96h.
- Extension of the ensemble size and increase of
spatial resolution (12-18 months) - Specific application
- Ship routing (Project WERMED)
- Grid computing test case (Project GRIDA3)
- Flash Flood Warning System (Project CEDRINO
BASIN).
Impossible to say which is best between SE and
EM.
2m T over all domain (analyzed data) an
example of the regression coefficient fields
