Scientific strategy for the implementation of a 3D-VAR data
assimilation scheme for a double-nested limited area model
Steluta ALEXANDRU, updated on
November 16, 2001
PRELIMINARY STUDIES:
The three-dimensional variational data assimilation (3D-VAR) scheme for
ALADIN was first implemented in 1999 in Toulouse Siroka and Horanyi, 1999).
Then in 2000 the scheme was successfully ported to Budapest (Horanyi, 2000)
for the Origin 2000 machine of the Hungarian Meteorological Service (HMS).
Recently AL13 code version of ALADIN is used for 3D-VAR taking into account
classical NMC statistics for the background term and SYNOP and TEMP
observations for the observational term of the cost function. It is mentioned
that 3D-VAR is acting for the upper-air meteorological fields and optimal
interpolation (CANARI) analysis is applied for the surface variables. Recently
new background error statistics were computed based on the lagged constant
coupling formulation of the NMC method (Horanyi and Boloni, 2001). Intensive
sensitivity experiments were performed with the lagged method in order to see
the effect of the different forecast differences and integration lengths on
the computed background error covariances. All the ingredients mentioned above
will serve as a basis for the work to be carried out in Budapest by Steluta.
MAIN OBJECTIVES:
Design and implementation of a 3D-VAR data assimilation system around the
ALADIN/HU (Horanyi et al, 1996) mesoscale numerical weather prediction model.
The initial version of the system should be not worse than the recently
operational dynamical adaptation system and should contain several promising
developmental areas for the further perfecting of the system. For the end of
the work in Budapest the developed 3DVAR system should be robust and superior
to the dynamical adaptation technique in terms of consecutive forecasts
originating from the 3D-VAR provided initial conditions. The system should be
capable for the easy and smooth transition towards 4D-VAR. The research and
development areas needed for 4D-VAR should be clearly distinguished and
possible solutions should be addressed.
MAIN DIRECTIONS OF RESEARCH:
1. Observations
Recently only SYNOP and TEMP observations are taken into account in the
course of the 3D-VAR scheme in Budapest. New data sources should be
investigated (e.g. satellite and/or radar data, AIREP data etc.) and impact
studies should performed. Alternatively observation information from Toulouse
can be taken and then only additional local information can be added to that
file. The recent version using SYNOP and TEMP measurements is considered to be
the reference and it should be carefully tested whether adding new data the
overall quality of the forecasts starting from 3D-VAR analysis is improved. It
means that this study should be performed in a successive way. A possible
direction of attack is as follows:
- wind data from Doppler radars (probably it is easy to try),
- investigation and use of AIREP data,
- introduction of raw radiances in the control variable (LTOVSCV switch in
the code),
- radar reflectivities (coordination with Toulouse activities)
It is noted that there are also several technical issues related to the
observation handling. Roger Randriamampianina will be in charge of supporting
Steluta in the technical problems related to observations.
2. Background error statistics
Several sets of background error statistics are available for the
ALADIN/HU domain. Choices should be done between classical (standard)
statistics and lagged statistics. In the latter case the best variant of the
lagged statistics should be applied. This part of the work will be carried out
together with Gergely Boloni, who would complete the started work on the
lagged statistics. It shouldn't be forgotten that there is a strong
inter-relation between the choice of background errors and first guess (for
instance in case of lagged statistics, when large scales are removed from the
covariances the first guess should contain information coming from the large
scales). See more details in the next point, where the different possibilities
for the first guess is explained.
3. First guess
It is not obvious at all what first guess to use in the context of
mesoscale data assimilation. It is rather clear for global models, because
there only the 6h forecast from a previous model state can be applied. In
mesoscale one of the following three possibilities might be chosen:
- Short range (6h) forecast from an earlier model run. It contains
information on the small scales (the scales of the model), but due to the fact
that it is a forecast it is not fully precise.
- Analysis of the global (regional) model. Only information is hold from
larger scale, but on the other hand the analysis is rather a precise mirror of
the atmosphere (so it is exact).
- Mixture of the previous two sources (blending), which is an optimal blend
of the small scale and large scale information. Simple blending algorithm
(spectral blending, where the small scales are taken from the ALADIN forecasts
and the large scales are taken from the ARPEGE analysis) should be tuned and
tried for obtaining first guesses for 3D-VAR. In case of combining blending
and 3D-VAR the order of blending and 3D-VAR is also to be studied (see the
work of Maria Siroka in Prague).
4. Role of initialisation
The first tests with initialisation and 3D-VAR was performed by Adam
Dziedzic (Dziedzic, 2000) in Toulouse. Adam found that DFI in its classical
formulation might destroy some useful signal of the 3D-VAR increments. At that
time it was decided to try to use the incremental version of DFI and see if it
is a more appropriate filter. Based on this work the following questions
should be addressed:
- Is there a need for initialisation after the execution of the 3D-VAR scheme?
- If yes, what are the possible initialisation methods?
- In case of application of digital filtering what is the best version of it
to be used (incremental DFI?)?
- What is the precise relation between the choice of first guess, background
errors and initialisation methods?
5. Effect of double nesting
It is of scientific interest to study the effect of double nesting in the
3D-VAR data assimilation scheme. Experiments can be performed with the
systematic comparison of 3D-VAR ALADIN, which is coupled directly from ARPEGE
and coupled from ALADIN/LACE. Sensitivity tests should be carried out with
respect to the domain size, resolution and coupling frequency.
6. Validation and tuning tools
It is of great importance to create an effective validation framework on
the testing of all the different options tried for ALADIN 3D-VAR. The
following main tools are proposed:
- Subjective evaluation: chagal and/or echkevo plots. It can be used mostly
for case studies.
- Objective scores (RMS, bias, etc.). This can be used for evaluating case
studies, and more importantly double suites.
- Posteriori evaluation tools: Jmin test, ratio of Jo and Jb (Sadiki, 2000,
Talagrand, 1999), generalised cross validation (Wahba et al., 1995), etc.
7. Towards 4D-VAR
At the end of the work on 3D-VAR some outlook should be given on the tasks
to be carried out for the 4D-VAR implementation.
MILESTONES, PRIORITIES, SCHEDULE:
Hereafter some schedule is planned to be drawn with special emphasis and
details on the first stay of Steluta.
1. 1st of November 2001 - 31st of May, 2002 (7 months)
- First adaptation in Budapest (administrative matters, email, users, etc.),
1 week
- Study of the existing informatic environment with the existing and working
scripts with special emphasis on the 3D-VAR scripts (including mandalay,
obsort, screening, 3D-VAR job - conf=131, surface canari, full-pos) and the
available evaluation tools, 4 weeks
- Creation of reference experiments with the existing tools, 2 weeks
- Creation and tuning of simple blending framework, 4 weeks
- Cycling experimentation with the different background error statistics,
first guesses, initialisation methods, together with their monitoring and
posteriori evaluation, 8 weeks
- Re-tuning of the version judged to be the best, 4 weeks
- Final implementation, double suites, 4 weeks
Outcome: Replacement of dynamical adaptation with a scientifically sound
3D-VAR scheme.
2. 1st of September, 2002 - 31st of March, 2003
- Double nesting versus simple nesting (blending parameters, comparison of
the different background error statistics: ARPEGE, ALADIN/LACE and ALADIN/HU).
- Improvement of the 3D-VAR scheme through observations: new observations
(radar, satellite, AIREP), better use of the already existing observations,
new observation operators and their adjoint, introduction of new observations
in the course of minimisation, etc.
- New formulation of the Jb term, relaxing homogeneity and isotropy
(co-operation with Simona Stefanescu, who supposed to work in that topic).
3. 1st of June, 2003 - 30th of November 2003
- Left others for 3DVAR
- Frequency of 3D-VAR, alternative algorithms toward 4D-VAR (potential usage
of asynoptic data)
- March towards 4D-VAR (initial experimentation)
- Redaction of a scientific paper (thesis)
COLLABORATIONS:
The work will be conducted by Andras Horanyi.
Strong cooperation with the local NWP team of HMS, especially with Gergely
Boloni for background error statistics related questions and Roger
Randriamampianina with observation handling related matters. Sandor Kertesz
will give some support on screening questions and Gabor Radnoti on blending
and code related issues. Afterall it is expected that a strong team (with
Steluta, Gergely and Roger in its core) will be out together and this team
will be capable to carry out together all the scientific and technical work
described above.
It is highly expected that frequent e-mail discussions should take part
with other ALADIN colleagues working on 3D-VAR. Hereafter the main subjects
and persons to discuss with will be summarised:
Subject
|
Person
|
All matters
|
Claude Fischer, Maria Siroka, Andras Horanyi
|
Strategic decisions
|
Jean-Francois Geleyn
|
Jb
|
Loik Berre, Gergely Boloni, Simona Stefanescu
|
Blending
|
Gabor Radnoti, Radmila Brozkova
|
Observations
|
Patrick Moll, Roger Randriamampianina
|
Screening
|
Parick Moll, Sandor Kertesz
|
Posteriori validation
|
Wafaa Sadiki
|
Code related matters
|
Gabor Radnoti
|
PARTICIPATION ON WORKSHOPS:
- Spring, 2002: Mid-term review of the ALATNET project, Brussels.
- Mini-workshop on 3D-VAR in Budapest (anticipated for May, 2002)
- October, 2002: EWGLAM meeting, the Netherlands
REFERENCES:
- Dziedzic, A., 2000: Geometrie et Initialisation dans le 3D-VAR/ALADIN. In
French. ALADIN internal note.
- Horanyi, A., 2000: 3DVAR workshop in Budapest. ALATNET Newsletter No. 1, 36-40.
- Horanyi, A and G. Boloni, 2001: Lagged constant coupling background error
statistics: preliminary results for the ALADIN model. Submitted to the
Proceedings of the 10th ALADIN workshop held in Toulouse (7-8 June, 2001).
- Horanyi, A., I. Ihasz and G. Radnoti, 1996: ARPEGE/ALADIN: A numerical
weather prediction model for Central-Europe with the participation of the
Hungarian Meteorological Service. Idojaras, 100, 277-301.
- Sadiki, W., 2000: Reglage des statistiques "lagged NMC" dans
3D-VAR ALADIN. In French. ALADIN internal note.
- Siroka, M. and A. Horanyi, 1999: The development of three-dimensional
variational data assimilation (3DVAR) scheme for ALADIN. ALADIN internal report.
- Talagrand, O., 1998: A posteriori evaluation and verification of analysis
and assimilation algorithms. Proceedings of the seminar on "Diagnosis of
Data Assimilation systems", pp 17-28. Available from ECMWF.
- Wahba, G., D. R. Johnson, F. Gao and J. Gong, 1995: Adaptive tuning of
numerical weather prediction models: randomised GCV in three- and
four-dimensional data assimilation. Monthly Weather Review, 123, 3358-3369.
OTHER RECOMMENDED LITERATURE:
- Jean Pailleux: Variational assimilation. Prepared for a summer school on
data assimilation in Toulouse, 1993.
- Courtier, Thepaut, Hollingsworth, 1994: A strategy for operational
implementation of 4D-VAR using an incremental approach. QJRMS, 120, 1367-1387.
- Lorenc, 1988: Optimal non-linear objective analysis. QJRMS, 114, 205-240.
- Thepaut and Courtier, 1991: Four dimensional variational data assimilation
using the adjoint of a multilevel primitive equation model. QJRMS, 117, 1225-1254.
- Lorenc, 1986: Analysis methods for numerical weather prediction. QJRMS,
112, 1177-1194
- Parrish and Derber: The National Meteorological Center's spectral
statistical interpolation analysis system. MWR, 120, 1747-1763.
- Berre, L., 2000: Estimation of synoptic and mesoscale forecast error
covariances in a limited area model. MWR, 128, 644-667.
- Courtier et al., 1998: The ECMWF implementation of three-dimensional
variational asssimilation (3d-var). Part I: Formulation. QJRMS, 124, 1783-1807.
- Derber and Bouttier, 1999: A reformulation of the background error
covariance in the ECMWF global data assimilation system. Tellus, 51A, 195-221.
- Hollingsworth and Lonnberg, 1986: The statistical structure of short-range
forecast errors as determined from radiosonde data. Part I: the wind field.
Tellus 38A, 111-136.
- Hollingsworth and Lonnberg, 1986: The statistical structure of short-range
forecast errors as determined from radiosonde data. Part II: the covariance of
height and wind errors. Tellus 38A, 137-161.
- Sadiki, Fischer, Geleyn, 2000: Mesoscale Background error covariances:
recent results obtained with the limited-area model ALADIN over Morocco, MWR,
123, 3927-3935.