Priority Project "COTEKINO"
COsmo Towards Ensembles at the Km-scale IN Our countries

Last updated: January 2014
See also: meetings

Project leader: Chiara Marsigli

Introduction

The COSMO Consortium has chosen already 10 years ago to share a consortium mesoscale ensemble system, COSMO-LEPS (Montani et al., 2011), which is running at ECMWF as time critical application thanks to the Billing Units made available by the COSMO Countries which are also ECMWF members (Germany, Italy, Switzerland and Greece). This experience has been followed by other consortium ensembles in Europe, both in LACE (ALADIN-LAEF, Wang et al., 2011) and in Hirlam-Aladin (GLAMEPS, Iversen et al., 2011).

COSMO-LEPS includes in its domain all the COSMO countries, with the notable exception of Russia, due to its spatial extension. Though, for the 2014 Winter Olympics, which are taking place in Russia, Sochi area, a re-location of the COSMO-LEPS system had been operated, leading to the development of the COSMO-S14-EPS ensemble (Montani et al., 2013), currently operational for the Olympic Games.

In the recent years, more and more emphasis is being devoted to high-resolution rapid-update-cycle (RUC) NWP applications, since weather forecast is focusing more on high-impact weather (HIW) and small-scale intense phenomena (e.g. convection), also aiming at different applications (assistance to aviation, hydrology, civil protection).

Motivation

For this reason, NWP development is concentrating on high-resolution modelling, O(1km), and this is reflected in the COSMO strategy, which has lead to focusing the consortium development on high-resolution data assimilation (KENDA PP), high-resolution modelling (UTCS PP), high-resolution verification (VERSUS2 PP).

Ensemble forecast was born to complement deterministic forecast, with products such as ensemble mean for medium-range predictions, ensemble spread to quantify the forecast uncertainty, meteograms for surface weather parameters, to present a spectrum of possible alternative scenarios. With the model resolution increase, the phenomena which are now described by NWP models are more and more stochastic in nature.

For moist convection prediction, skill is manifested through the statistical properties of the forecasted convection instead of by deterministic modelling (Fritsch and Carbone 2004). Therefore, the weight given to ensemble forecast is now greater than it was some 10 years ago, since for the convection-permitting NWP models it is crucial to be able to forecast not only the "best scenario" but ideally the whole pdf or, more realistically, a good representation of it.

The high-resolution of these systems prevents the possibility of running consortia ensembles, since it is prohibitive to cover a large domain with the required grid spacing. This is leading to the development of high-resolution ensembles on a national scale (Bouttier et al., 2012, Migliorini et al., 2011).

Status

Few years ago, DWD has started the development of a convection-permitting ensemble, COSMO-DE-EPS, which is operational over Germany at 2.8km horizontal resolution since May 2012 (Gebhardt et al., 2011). The system receives initial condition perturbations from the 4 global scale operational analyses of DWD, ECMWF, NCEP and JMA. The four operational global runs of these institutions provide also boundary conditions to 4 COSMO runs at 7km horizontal resolution which drive in turn the COSMO-DE runs. Model physics perturbations are also applied, by changing the value of few parameters of the physics schemes.

In the last year, also Italy and Switzerland, and now Poland, have started the development of convection-permitting ensembles in their countries. Furthermore, Russia is developing COSMO-RU2-EPS for assisting the Winter Olympics. Therefore, the need for a good strategy for convection-permitting ensemble has emerged at the Consortium level, requiring coordination among the activities taking place in the different countries, in order to share research and development work and to ensure good and timely exchange of information. It is proposed to realise this coordination through a new Priority Project, COTEKINO (COsmo Towards Ensembles at the Km-scale IN Our countries).

Needs

Nowadays it is still unclear how to best design convection-permitting ensembles. Several aspects should be taken into account and new perturbation methodologies should be developed or adapted to this scale.

BCs.

It has been proved that the role of boundary conditions is crucial also at the these spatio-temporal scales (Gebhardt et al, 2011; Vie et al, 2011especially if the domain is small. An important issue is the availability of perturbed boundary conditions with the appropriate spatio-temporal resolution.

DWD has chosen to downscale the information provided by the 4 operational runs of DWD, ECMWF, NCEP and JMA by nesting on each of them COSMO at 7km of horizontal resolution, which drives the COSMO-DE-EPS members.

ECMWF EPS is currently running with an approximate resolution of 32 km over central Europe, but it is planned to move to 20 km in 2015. This opens the possibility of using EPS BCs directly for driving the convection-permitting ensembles, provided that the resolution jump is not badly affecting the forecast. As for the temporal resolution, currently BCs are saved 3-hourly, while hourly BCs would be needed for km-scale applications on small domains. An exercise to deal with these issues is currently being made within SRNWP, thanks the hourly BCs provided by few experimental EPS runs at 16 km horizontal resolution. Finally, more EPS runs per day may be needed, while currently only 2 runs are performed (00 and 12 UTC).

Both COSMO-LEPS and the BC-EPS ensembles may enable an intermediate downscale, to cope with the resolution gap, and these datasets are available for testing to the COSMO partners.

ICs.

It is recognized that for the time scale dealt with by convection-permitting ensemble, data assimilation plays a crucial role . What dominates ensemble spread in the first six hours are perturbed initial conditions (Peralta et al, 2012, Vie et al, 2011), which should include both a good estimate of the initial condition(s) and a good representation of its (their) uncertainties.

Within the KENDA PP, a LETKF scheme for the km-scale has been developed, aiming at providing both deterministic analysis for driving high-resolution COSMO runs and a set of perturbed analyses to drive high-resolution ensembles. The scheme is available for testing to the COSMO community and tests are on-going in Germany, Switzerland and Italy. Though the KENDA development is included in the KENDA PP, how to set-up the system to provide ICs to a convection-permitting ensemble and how to derive ICs from KENDA is not part of the Project. Several options are possible, among them:

use analyses provided by KENDA directly as ICs for the ensemble. This may imply that the LETKF is run with the same horizontal resolution of the ensemble, or that an interpolation is performed.

use only KENDA-derived perturbations to the ensemble mean and apply them to a deterministic analysis obtained e.g. with nudging.

blend KENDA-derived perturbations with large-scale perturbations (e.g. derived from ECMWF-EPS or form BC-EPS)

...

Which option is best suited for a specific ensemble application may be dependent on the design of the ensemble itself (domain, number of members, which and how many observations are available over that domain, how the alternative analysis is computed, resolution on the systems, which boundary conditions are used, ...).

Attention should be paid to which kind of observations can actually be assimilated by KENDA. For example, for a clean comparison between the KENDA analysis and the nudging one, nudging should be run without LHN, since KENDA does not permit assimilation of radar data up to now. For this reason, a good link with the KENDA developments and with the KENDA PP should be organised and maintained.

Furthermore, also the set-up of the LETKF ensemble can be tailored for the specific application. For example, an important issue is how many large-scale ensemble members are needed to drive the LETKF ensemble and, if these are derived from a larger ensemble, how to select them.

For these reasons, it is believed that the outcome of the tests designed for defining the set-up cannot be automatically transferred from one country to the other, but appropriate testing should be carried on the different systems under development.

Therefore, it is proposed to limit the cooperation in this task to timely exchange of information and results as well as the exchange of developed tools, e.g. for member selection.

Model perturbations (atmosphere).

The importance of model perturbations in ensemble forecasting has been recognised since some years. It is found that model perturbations usually play a smaller role with respect to initial conditions (at the beginning of the run) and boundary conditions (later) perturbations, but they may become crucial in specific weather situation (Marsigli et al, 2009; Gebhardt et al, 2011) and a poor representation of the model error may lead to severe underestimation of the ensemble spread (Vie et al, 2011).

Several techniques have been developed to represent model error: perturbation of the physics schemes parameters (Gebhardt et al, 2011, Marsigli, 2009), random parameters (Bowler et al, 2008), SKEB (Shutts, 2005; Berner et al., 2009), SPPT (Buizza et al, 1999; Palmer et al, 2009), multi-physics, stochastic physics (Plant and Craig, 2008; Bengtsson et al., 2013). Interesting comparative studies have also been carried out (Berner et al, 2011; Hacker et al, 2011).

While there are some results about the applicability of these techniques for the O (10km) scale, it is still not assessed how to use them for the O(1km) scale. From the COSMO-DE-EPS experience it is clear that underdispersion severely affects the ensemble for surface variables, if only few physics parameters are perturbed.

The SPPT scheme of ECMWF has been recently implemented in the COSMO code, to test its capability of representing model error in the LETKF ensemble. This code can be tested also for the ensemble forecasting set-ups, though it is argued that this approach may lead to serious unbalances in the model. Also, the scheme has been originally developed for the IFS model at coarser resolution and should be adapted for use at the O(1km) scale. For this purpose, studies about the applicability of SPPT at 2.2-2.8km horizontal resolution are underway in Switzerland and planned in Italy.

On top of this, next autumn the SKEB scheme will be implemented in the COSMO model, opening the possibility of testing this new approach for providing model perturbations to the convection-permitting ensembles. Tests will be performed by MCH.

Following the 2012 ECMWF workshop on the representation of the model error and the 2013 SRNWP workshop on Physics and Ensemble, it is suggested to move towards "integrated" approaches, that is to physics schemes intrinsically stochastic. It is clear that this approach requires the work of model physics experts. For this reason, now on the one hand the link between model physics and ensemble forecasting is becoming tighter and the need for more frequent scientific discussions is consequently greater, on the other hand the roles are clearly better divided, the development of new model perturbation methods being in the hands of physics experts, while the evaluation of the resulting properties of the ensembles, where different perturbation techniques flow into, being in the hands of ensemble developers.

DWD plans to develop a new approach for model perturbation, which will be defined on the basis of a preliminary study, aiming at designing physically based model perturbations. This work will be coordinated with the other ensemble developments, to ensure timely feedbacks and possible contributions by other COSMO scientists, if available.

Model perturbations (lower boundary).

The sensitivity of moist atmospheric processes to soil conditions has been demonstrated in numerous studies (Hacker 2010 and references therein). Although the physical representation of soil moisture in land surface schemes is known to have an influence on the quality of atmospheric predictions, the parametrization of the processes of soil moisture physics is not straightforward and the uncertainty in the parameters are little researched (Cloke et al, 2012).

Some techniques have been proposed in the recent years: Cloke et al (2012) proposes a simple method, perturbation of 2 soil scheme parameters, in the ECMWF seasonal forecasting system; Sutton et al (2006) studied the impact of using two different soil moisture fields, estimated by 2 different LSM but with the same data and atmospheric forcing, showing that short-term precipitation is sensitive to the soil moisture, especially for a high-resolution run; anoncycling surface breeding method has been proposed by Wang et al (2010), where short-range surface forecasts driven by perturbed atmospheric forcing are used for generating the perturbation to surface ICs. A simple method is applied by Hacker (2010), who constructs perturbations of the soil moisture field that represent random, spatial correlated errors, to quantify the response to soil moisture perturbations. He recognises that this method is not suited for reproducing the characteristics of the actual soil moisture uncertainty, which varies locally with the properties of the soil, vegetation and background moisture itself (Hacker, 2010).

In order to partly account for soil moisture uncertainty, in the COSMO PP CONSENS a methodology for soil moisture perturbation based on Sutton and Hamill (2004) was developed by HNMS (COSMO Technical Report No. 22). The method was never tested in ensemble mode but it could be revived and tested on the new ensembles under development. Other methods can be tested, among those available in literature, or proposed, according to the need of the COSMO countries and on the available resources.

At DWD a simple method was developed to derive initial condition soil moisture perturbations from differences between COSMO-EU und COSMO-DE soil moisture. A long period test is now planned.

Also for lower boundary perturbation, the work should be tackled in close cooperation between soil experts and ensemble scientists.

Ensemble use and interpretation.

COSMO has a long and well established experience of operational ensemble forecasting. COSMO-LEPS is used operationally since almost 10 years and several applications are based on this system. Among them: hydrological models (EFAS hydrological ensemble for Central Europe, FEWS PO system for the Po river, systems on basins in most parts of Switzerland, ...), warning systems (operational at ARPA-SIMC and MeteoSwiss, MAP D-PHASE, Hymex SOP, ...), .

Soon we will benefit of the experience of the Winter Olympic Games with the COSMO-S14-EPS and COSMO-RU2-EPS ensembles. At DWD, the convection-permitting COSMO-DE-EPS ensemble is now operational since 1 year and studies have been made about which products are more useful for the forecasting activities and which products are more suitable for various application. It is proposed to capitalize from this experience, by initiating a cooperation between WG7, WG4 and WG5, aiming at defining how to improve ensemble interpretation and use.

Ensemble development evaluation.

It is clear that for the development of skilful convection-permitting ensembles, including the representation of several uncertainty sources and covering different COSMO countries, a great effort is required. In order to guarantee robust results from the experiments which can be useful for the whole COSMO ensemble community, great care should be devoted to evaluation of the different configuration options.

Therefore, it is proposed to dedicate some effort to develop common tools for evaluating the quality of the different configuration options, based on a common strategy for developing diagnostic tools (e.g. spectra) to evaluate the impact of the perturbations. This will decrease the risk of work duplication and will permit to maximise the exchange of software and results among the countries.

Actions proposed

1. Assess how to best derive initial conditions perturbations from KENDA in the different regions. This activity will be carried out in the COSMO Countries developing convection-permitting ensembles independently but in a coordinated manner. Therefore, it is proposed to activate a task to deal with organised and timely exchange of information about the results of the experiments as well as the exchange of developed tools, e.g.for a member selection.
   Also coordination with the KENDA Project is part of this activity.
2. Develop COSMO model perturbations appropriate for the O(1km) scale. The work is organized in three tasks: one carried out at DWD, where a new stochastic physics scheme is being developed, one carried out by Switzerland and Italy, where the SPPT scheme implemented into COSMO is tested for the 2.2/2.8km resolution, with the needed adaptations (and tuning), and one carried out by Switzerland, where the SKEB scheme will beimplemented in the COSMO model and tested. Good communication will be maintained in order to identify possible future contributions from other countries to the work of DWD, if needed.
3. Develop/implement lower boundary perturbation techniques. The task is organised in three phases: investigation of possible strategies for COSMO, sensitivity studies, implementation of one or more technique and their testing on the convection-permitting ensembles.
4. Work on interpretation and use of high-resolution ensemble forecast. No tasks are proposed concerning this issue for the moment, but the discussion will continue at the WG7 level.
5. It is not proposed to activate a task about the development of diagnostic tool at this stage, due to a lack of resources. (Possibility to use the KENDA diagnostic tool, NEFFprove ...)

Description of individual tasks

Task 1. IC perturbations derived from KENDA.

• 1.1: Communication about experiments and results. Exchange of dedicated tools. (0.2 FTE)
• 1.2: Coordination with the KENDA PP (0.1 FTE)

Deliverables:

• 1.1: Meetings between scientists working on this issue, in person at the GM and CUS and at least twice per year by web-conference (minutes).
• 1.2: Participation to the KENDA meetings and email exchange (minutes). Documentation of techniques to derive IC perturbations from KENDA, and related tools.

Task 2. Model perturbations.

• 2.1: Test SPPT at 2.2/2.8 km (1.2 FTE)
  • 2.1.1: ARPA-SIMC/CNMCA: Analyse the impact with respect to perturbed parameters and deterministic verification of perturbed members.
  • 2.1.2: MCH, Analyse the impact with respect to a non perturbed ensemble and further develop the scheme.
  • 2.1.3: RHM, Overview of the available methods for model perturbations. Test of SPPT in the Sochi 2.2 ensemble to assess the effect w.r.t. the current method.
• 2.2: Test SKEB in the 2.2 km ensemble at MCH (0.3 FTE)
• 2.3: Information about a new model perturbation strategy at DWD (PP 0.2 FTE)

Deliverables:

• 2.1: Definition of an optimal set-up of an advanced SPPT scheme for perturbing the COSMO model at 2.2/2.8km and assessment of its performance (report).
• 2.2: Assessment of the performance of the SKEB scheme for perturbing the COSMO model at 2.2 km. (report)
• 2.3: Ensure good communication to COSMO and timely information. Two meetings per year at DWD (minutes).

Task 3. Soil/surface perturbations.

• 3.1: Study of COSMO model sensitivity to lower boundary initial conditions (0.6 FTE)
• 3.2: Scientific investigation of the possible strategies to be followed: literature review, assessment of more suitable methods (0.3 FTE)
• 3.3: Develop techniques for lower boundary perturbation (1.8 FTE)
  • 3.3.1: test of perturbation of soil model parameters (RHM)
  • 3.3.2: selection and preparation of algorithms
  • 3.3.3: setting a (task-specific) test-bed, implementation and testing,
  • 3.3.4: analyse the impact with respect to a non perturbed ensemble

Deliverables:

• 3.1: Sensitivity test on the behaviour of different COSMO suites to different lower boundary initial conditions. (report)
• 3.2: Definition of one or more techniques for lower boundary perturbations to be developed for COSMO. (report)
• 3.3: Assessment of the performance of the proposed method(s) (report). The algorithm for lower boundary perturbation is implemented in one or more ensemble systems for testing.

Estimated resources

Estimated needs in FTE-years, total 5.0FTEs: 2013-2014: 3.0 FTEs and 2014-2015: 2.0 FTEs. See the following table for mre details:

Links to other projects or work packages

• The link with KENDA is explicitly addressed in task 1.
• The link with WG3a (stochastic physics development) is addressed in task 2.
• A link with WG3b needs to be maintained for the development of new lower boundary perturbations (task 3) and this will be pursued via email and through meetings.
• The link with WG4 and WG5 will be established during the project, following the lines described in the Science Plan.

Risks

The main risk of this project is in the outcome of the scientific research, since it is not a priori guaranteed that the proposed methodologies will prove successful for ensemble forecasting applications at the km-scale (KENDA ICs, SPPT, SKEB, soil perturbations).

Participants

• Chiara Marsigli (ARPA Emilia-Romagna)
• Nicola Loglisci (ARPA Piemonte)
• Riccardo Bonanno (ARPA Piemonte)
• Lucio Torrisi (CNMCA)
• Richard Keane (DWD)
• Ekaterina Machulskaya (DWD)
• Andrzej Mazur (IMGW)
• Grzegorz Duniec (IMGW)
• Witold Interewicz (IMGW)
• Andre Walser (MCH)
• Daniel Leuenberger (MCH)
• Marco Arpagaus (MCH)
• Daliah Maurer (MCH)
• Elena Astakhova (RHM)
• Dmitry Alferov (RHM)
• Inna Rozinkina (RHM)
• Gdaly Rivin (RHM)

References