3. Vincent GUIDARD : "Evaluation of assimilation cycles in a mesoscale limited area model"

1. Additional results on MAP IOP 14

(completion of the study described in the ALATNET Newsletter 5, "Combined use of 3d-var and DFI-blending on MAP IOP 14")

Some scores against observations were computed over precipitation forecasts. The forecasts were performed with constant coupling from several initial states valid at 12 UTC, 02 November 1999 :

- OPER : dynamical adaptation of ARPEGE analysis, digital filter initialization (DFI);

- DFI-blending : 24-hour DFI-blending assimilation, DFI;

- 3D-VAR : 24-hour 3d-var assimilation, incremental DFI;

- BLENDVAR : 24-hour Blendvar assimilation, incremental DFI;

- BLENDVAR+pseudo-prof. : 24-hour Blendvar assimilation, including pseudo-profiles of humidity in the last 3d-var analysis, incremental DFI.

Figure 1 shows the results for short-range forecasts, which reinforce the previous results : DFI-blending assimilation improves precipitation forecasts. The introduction of humidity pseudo-profiles implies some interesting modifications.

Figure 1 : Equitable Threat Score for short-range forecasts (ETS, the closer to 1. the better).

2. Biperiodisation & background error statistics : Impact on 3d-var analysis increments

"Single observation" experiments

Some experiments using only one observation have highlighted a drawback of biperiodisation : if the observation is too close to the C+I border, one part of the analysis increment (roughly 10 % of the whole increment) goes through the extension zone and appears on the opposite side of the domain (Figure 2).

Two solutions were considered :

- enhancing the width of the extension zone, but this implies a recomputation of some statistics and an increase of the computational cost (therefore not tested) ;

- rejecting all observations closer than 200 km to the C+I border, but a little part of the increment still goes through the extension zone and too many observations would be rejected.

Figure 2 : Analysis increment for temperature at 500 hPa. The observation is a simulated TEMP observation of temperature at 500 hPa, verifying : observation - first-guess = 1 K. Units : 0.01 K.

Similar experiments (not shown) were undertaken with a "band" of (real) observations instead of one (simulated) observation. The analysis increment generated by a 1000 km-wide band of observations in the southern part of the C+I domain is similar to a "single obs." increment. Neither rejecting the observations which are too close to the border, nor zeroing the s _b in a rim zone around the domain, solve the problem. The analysis is still worse than the first-guess (in comparison with the observations) in that part of the domain where there is no observation (i.e., in this case, in the northern part of the C+I zone). The next trial will consist of using the B-matrix auto-correlation functions as gridpoint filters (see 1d-case below).

Academic 1d model

An academic 1d assimilation model, with 300 gridpoints and no dynamics, is used to study in a simple way the gridpoint behaviour of the background error covariance spectrum and the impact of its modification. In order to damp the impact of an observation at long distance, a compact support has been implemented. It mimics ARPEGE "SUJBCOSU" routine, using a cosine-shape way to zero the gridpoint function. The first results are quite encouraging : the "undesirable" analysis increments are significantly reduced. This study has to go further : choice of the distance and the length of zeroing for various variables and vertical levels.

The next steps are expected to be :

- adapting ARPEGE "SUJBCOSU" to an ALADIN "SUEJBCOSU" and assessing the impact of compactly supported correlations in the ALADIN 3d-var analysis ;

- introducing dynamics into an academic model and studying the growth of forecast errors in a LAM and its coupling model (using a 1d shallow-water model, for instance).