Priority Project "APSU"
Ameliorating Perturbation Strategy and Usage of ensemble systems

Last updated: 15 Sep 2020

Project leader: Chiara Marsigli (DWD)

Project duration:

March 2018 to August 2020

FTEs (planned/used):

1.40 (1.65 originally, replanned) / 0.83 in COSMO year 2017/2018
3.60 / 2.85 in COSMO year 2018/2019
2.95 / 3.17 in COSMO year 2019/2020

Total FTEs planned:


Total FTEs used:



In March 2017, WG7 members identified the main issues on which the development in the ensemble field is needed in the COSMO Consortium. This led to the decision to propose a new Priority Project after the end of the PP SPRED, in order to continue working on the improvement of the Convection-Permitting (CP) ensembles in COSMO, especially with regards to model perturbations. This decision is based on the recognition that the perturbations applied to the model in the ensemble configuration do not provide yet a satisfactory representation of the model error, as evidenced by the spread-skill relation. On top of this, recently the physics of the COSMO model has been substantially developed, particularly with modifications in the turbulence scheme and in the land surface scheme, following the developments which took place in the ICON model. This has implied that some perturbations are now outdated and that new parameters can be subject to perturbations.

The research and development work of the present Project covers three main areas:

On top of these development areas, the transition from the COSMO model to ICON-LAM requires that tests of ICON-LAM in ensemble mode are planned in the COSMO countries (Task 6).

Note: The most widely used model perturbation techniques are shortly explained in Appendix 2, with the correspondent acronyms.



A short description of the status of ensemble developments and plans in the Consortium is here provided.

Ensemble systems at CP scale are now in place in several COSMO countries. The systems are listed below.

COSMO-DE-EPS, operational at 2.8 km horizontal resolution since May 2012. Domain is Germany, members are 20. The system has been upgraded in 2017. Now it receives Initial Condition perturbations from KENDA and Boundary Conditions from ICON-EPS. Model physics perturbations have also been reviewed, resulting in an extended set of parameter perturbations. In addition, the previous procedure of a fixed allocation of perturbations to members was replaced by a randomized method, i.e. a randomized selection out of a fixed set of predefined perturbed values of parameters is done for each start time of the COSMO-DE-EPS. Soil moisture and temperature perturbations are also applied to the initial conditions as part of the KENDA cycle. Developments to be operational in 2018: COSMO-D2-EPS with 2.2 km mesh size, 65 levels, extended model domain.
COSMO-E, operational since May 2016. Alpine domain, 21 members, 2.2 km mesh-size, ICs from KENDA at 2.2 km (same domain), BCs from IFS-ENS (the ensemble of ECMWF), model perturbations through SPPT (Stochastically Perturbed Parametrization Tendency, see Appendix 2), soil moisture perturbations included in KENDA. The COSMO model is run in single precision.
COSMO-2I-EPS (part of the Italian modelling system COSMO-LAMI, called COSMO-IT-EPS in its development phase), pre-operational since the end of 2017, 20 members, 2.2 km, domain is Italy. ICs are from KENDA analyses and BCs from the COSMO-ME-EPS ensemble of CNMCA (see here below), model perturbations to be added in 2019 (Parameter Perturbations possibly combined with SPPT), soil moisture perturbations included in KENDA.
TLE-MVE, operational since January 2016, 2.8 km horizontal resolution, 20 members. BCs and ICs provided through a time-lagged approach from the deterministic COSMO-PL (7km), model perturbation (Parameter Perturbations), soil perturbation by perturbing the c-soil parameter. Special dedicated random number generator for initialisation applied since January 2017.
only research using COSMO-Ru2-EPS (2.2 km, 10 members, ICs and BCs from COSMO-S14-EPS, SPPT for model perturbation) for Sochi region till 2018.

There are also two ensemble systems at coarser resolution with parametrised convection.

The Consortium ensemble COSMO-LEPS (operational since November 2002) is currently operational at 7 km of horizontal resolution, with 20 members. The COSMO model is run in single precision. ICs and BCs are from IFS-ENS and Random Parameter Perturbations (RPP, see Appendix 2) are applied. The forecast range is 132 hours and runs start every day at 00 and 12 UTC. Main planned developments are: explore the feasibility and impact of increasing the resolution from 7 to 5 km and apply SPPT as physics perturbation.

CNMCA runs COSMO-ME-EPS (part of the Italian modelling system COSMO-LAMI), operational since May 2014, Euro-Mediterranean domain, 40 members, 7 km resolution. ICs are from CNMCA-LETKF analysis, BCs are from the most recent IFS deterministic run perturbed using IFS-ENS, model perturbation are provided by SPPT. The forecast range is 72 hours and runs start every day at 00 and 12 UTC.


During the WG7 meeting held in Offenbach in March 2017, the main need recognised by the participants was to continue improving the CP ensembles, on the basis of the evidence provided by their spread skill relation, which is often not good enough in terms of near-surface variables. The shared opinion is that the spread of the ensembles in terms of near-surface parameters (temperature, wind) should be increased, in order to better match the forecast error. This should be done with care, since it is recognised that the ensemble spread cannot be increased up to exactly reaching the forecast error, because the latter includes also the systematic error, which should be tackled by improving the model, not by increasing the spread. On top, observation error should be considered in the computation of the forecast error.

What should be better represented in the ensembles is that part of the error which depends on the uncertainties in the determination of initial and boundary conditions as well as in the description of the physical processes in the model, which are mainly due to the variability of the model performance in dependence of the meteorological situation (usually called "error of the day" or flow-dependent error).

To reach this goal, it is needed to work on a better representation of the model error. The strategy we plan to follow is based on testing new approaches for model perturbation which have been recently developed or made available to the COSMO model, and to review and improve the already existing one. More specifically it is needed to:

Concerning the combination of perturbations, not all the combinations are meaningful and worth trying. In order to minimize the work, this activity should be regarded as a sanity check, to ensure that randomized combination of potentially all perturbations which will be included in one ensemble do not harm the system performance. It is not intended to carry out intensive tests of several different and specific combinations, but only those which are regarded as meaningful for a specific ensemble implementation.

Additional to improving the model error representation, further studies concerning the use of KENDA analyses to provide ICs to the ensembles are needed. In particular, it should be further checked the impact of KENDA analyses to initialize the CP ensembles in terms of quality of the forecast. In order to improve the usage of the KENDA analyses, it is also needed to develop and test of new methodologies for selecting an appropriate subset of KENDA members as ICs for the ensemble members. It is reminded that the development of KENDA itself and its "tuning" is addressed in the KENDA-O PP, with which a good exchange is already in place.

Ensemble are used for several operational duties, concerning high spatial and temporal resolution forecasts, therefore it is needed to continue working on ensemble post-processing and interpretation (also coordinating COSMO efforts within the EUMETNET project SRWNP-EPS II). In particular it is needed to:

These topics are cross-cutting with WG4 scope, with which cooperation will be enforced.

In general, for the present project to be successful, it is needed to establish and maintain an effective cooperation between WG7 and WG 4 and 5, as concerns the preparation of new products from the ensemble systems and their proper verification.

Finally, the need is recognised to start planning for the transition to the ICON-LAM model, which implies that tests of ICON-LAM in ensemble mode should also be carried out.

Actions proposed

On the basis of the above-mentioned needs, actions to be undertaken in order to address them are here identified.

In order to improve the description of the model error in the ensembles it is proposed to test and further develop new model perturbation methodologies and to update and improve the existing ones.

The new schemes for model perturbation now available in the COSMO model will be tested, specifically: the prognostic stochastic model of the model error developed at DWD (referred to as EM scheme, from the name of the developer E. Machulskaya), a perturbation method based on the Stochastic Pattern Generator (SPG) developed at RHM (Tsyrulnikov and Gayfulin, 2017) and a perturbations method based on adapted Random Number Generator (RNG) developed at IMGW.

At MCH resources have been made available shortly after the beginning of the Project to start testing different model perturbation methods within a Master Thesis: the iSPPT (SPPT where the tendency from each parametrization scheme is perturbed using an independent pattern) and perturbations based on analysis increments from the assimilation cycle.

An update of the Parameter Perturbation method will be carried out for the 2.2 km version of COSMO, with the purpose of updating the existing perturbations with respect to the new default values assigned to some physics parameters and the testing of the perturbation of the new parameters which have been introduced in the model thanks to the introduction of the ICON physics.

It is also proposed to further develop the surface/soil perturbation methods. In particular, a perturbation of the soil temperature depending on the soil characteristics and the use of the SPG for perturbing the soil variables will be tested.

For the action on model perturbation, collaboration with WG3ab is required. Collaboration with WG3a is needed for the selection of the physics parameters to be perturbed and their range. Collaboration with WG3b is needed on improving the perturbation of soil parameters and variables.

Collaboration with the CALMO-MAX PP is also required, as concerns the revision of the parameter perturbations (e.g. which parameters are most sensitive and how to set-up the perturbation range), since this activity is already included in a Task of the CALMO-MAX PP.

When testing new model perturbations, verification methodologies suited for the high resolution should be adopted (spatial verification methods). This does not require the development of new verification methods in COSMO, because the suitable methods already exist in the international literature and some of them are already used in COSMO countries, since long time at Arpae and DWD and more recently in other countries also thanks to the experience of the INSPECT PP. For the purpose of properly using these methods for the APSU PP, collaboration with WG5 for method selection and discussion of the results is required.

It is underlined that this action does not give birth to a dedicated Task, but subtasks addressing the verification activity needed for testing the developments are to be included into the appropriate Tasks.

It should be also underlined that in APSU different COSMO members will study and develop different model perturbation methodologies, following the work already on-going at their Centres and/or on the basis of available expertise and resources. Therefore results will be exchanged but this action will not end up in a comprehensive testing of all the methods over all COSMO domains.

Another needed action is to continue the development of post-processing methods for the CP ensembles, especially focused on selected phenomena and intense events (leading to high impacts). Algorithms for providing forecasts of fog and thunderstorms have been developed at CNMCA within the EUMETNET Project SRNWP-EPS II. They will be tested in the ensembles and dedicated verification methods will also be developed. In order to make the ensembles reliable in terms of probabilities, calibration of CP ensembles will also be carried out.

For this action, collaboration with WG4 is required.

As regards the need of improving the usage of KENDA analyses as ICs for the ensembles, communication with WG1 and KENDA-O will also be maintained.

If resources will be available, it is planned to start an action on improving the selection of KENDA analyses to initialize the CP ensembles. The same applies to the needed improvement of BCs provision to the CP ensembles, where actions will be activated when resources will be available.

Finally, the transition to ICON-LAM taking place in the Consortium requires to perform an action dedicated to test the ICON-LAM model in ensemble configuration. At present, plans about this topic in the COSMO Countries are not fully developed, therefore this action is included in the PP but detail of the Tasks will follow in the next year, with related assignment of FTEs. This activity will be coordinated with C2I, the PP on the transition to ICON-LAM which is being proposed to the COSMO bodies and which should start in April 2018.

Description of individual tasks

Task 1. New model perturbation methods

New methodologies under development for being used in the COSMO model will be tested and further developed.

Task 1.1 - Perturbations based on EM-scheme

At DWD, a prognostic stochastic model of the model error (EM-scheme) has been developed. The errors of the tendencies of different model variables are assumed to obey the Langevin-type stochastic differential equation, whereas they are correlated in space and time and stochastically driven by Gaussian noise. The weather-dependent values of the parameters of this equation are determined by predictors, whereas the dependencies of the parameters on the predictors are inferred from past data. The resulting model error is added to the dynamic equations of the model.

In accordance with the focus of PP APSU, the EM-scheme is expected to contribute to an improvement of ensemble properties like spread-skill relation for near-surface variables taking into account the actual meteorological situation (e.g. "error of the day").

First tests with an experimental set-up of the EM-scheme in COSMO-DE-EPS resulted in promising improvements of standard probabilistic scores, but it turned out that further development is needed towards an implementation with benefit for the operational forecast.

The work of this Task will be organized as follow:

Task 1.2 - Perturbations based on the Stochastic Pattern Generator

The Stochastic Pattern Generator (SPG; Tsyrulnikov and Gayfulin, 2017) will be applied for generation of model perturbations at RHM. The SPG has already been implemented in the COSMO code and will be used to perturb horizontal winds, temperature (accompanied with hydrostatically balanced 3D pressure perturbations), humidity, and cloud fields. The SPG will be used within a new AMPT (Additive Model error perturbations scaled by Physical Tendencies) scheme. In the AMPT, the model error perturbations are additive, their magnitude is computed online as a horizontal domain average of the modulus of the physical tendency in each perturbed variable at each model time step and each model level.

Numerical experiments with COSMO-Ru2-EPS (2.2 km horizontal resolution) over the Sochi region will be performed in order to tune the AMPT setup and to assess the effect of the new approach on the ensemble spread and skill. Merits and limitations of the new approaches will be identified and evaluated in terms of ensemble verification scores such as reliability diagrams, ROCA (Relative Operating Characteristic Area), and CRPS (Continuous Ranked Probability Score) for temperature, wind, and precipitation. For a description of the scores, refer to Wilks (2011).

Task 1.3 - Perturbations based on adapted Random Number Generator (RNG).

Model perturbations based on a new Random Number Generator (RNG), proposed and introduced at IMGW as the adaptation and an evolution of the Intel/DWD version will be tested.

The basic motivation for developing this RNG was presented during the PP SPRED, as a result of quasi-operational tests and verification of ensemble forecasts (see Duniec et al., 2016; Duniec et al., 2017; Mazur et al., 2017). The main issue arising when running standard (pseudo-) RNG, e.g. the one provided with the Intel© FORTRAN, is that in general it is based on a seed directly related to the machine time (milliseconds). When running a parallelized code on machine with many processors (cores), some of them may have identical seed producing, therefore, chain of identical perturbation amplitudes for different members. Consequently, obtaining new values of random numbers would be directly translated to the amplitude of the perturbation of the selected parameters/variables.

The effectiveness of this new scheme will be assessed through its application in the IMGW Time Lagged ensemble (TLE-MVE), with a model horizontal resolution of 2.8 km (CP scale) on a domain covering Poland and its close vicinity (area of approx. 800x700 km). Tests will be performed both in an operational configuration and in offline mode on a restricted area but for extended periods. Results will be verified in terms of skill (MAE of ensemble mean) vs. measurements carried out at Polish SYNOP stations in the relevant periods.

Task 1.4 - Perturbations based on SPPT with independent pattern (iSPPT).

Model perturbations based on SPPT are used in the operational convection-permitting ensemble of MCH and at RHM for research purposes and there are plans to use SPPT in other COSMO ensembles. SPPT is able to significantly increase spread in temperature and humidity in the lower troposphere in summer, but hardly in winter, since the sums of the parameterization tendencies are highest in summer, dominated by the turbulence scheme. Christensen et al. (2017) suggest SPPT with independent random patterns for the parameterization schemes (iSPPT). It allows to generate different perturbations in terms of amplitude and space-time correlation for the parameterization schemes which do not necessarily have the same error characteristics. Our plan is to implement iSPPT in COSMO-E focusing on independent perturbations for the turbulence, microphysics and radiation tendencies. The impact of this approach will be evaluated with case studies and if the results are promising in an e-suite with standard ensemble scores.

Task 1.5 - Perturbations based on analysis increments.

Piccolo et al. (2018) found that using analysis increments (difference between analysis and first guess) to represent model errors improves the spread and the reliability of ensemble forecasts. Analysis increments can take into account more possible sources of model errors than SPPT since they are not limited to the physical tendencies.

The main motivation for this task is to gain a deeper knowledge of the analysis increments in our KENDA data assimilation cycle and finally to estimate to what extent these analysis increments can be considered as a proxy for model errors in KENDA and COSMO-E for generating additional ensemble perturbations. Our statistical analysis includes seasonal mean and variance for each grid point of the model domain as well as diurnal variations of the analysis increments focusing on temperature, humidity and wind.


Estimated needed resources for Task 1: 3.60 FTEs (DWD, IMGW, RHM, MCH)

Task 2. Revision of the Parameter Perturbation method

A revision of the Parameter Perturbation method adopted in COSMO-2I-EPS (2.2 km horizontal resolution, over Italy) will be carried out.

The method for identifying the most suitable values of the physics parameters to be used as perturbations for the COSMO model at the 2.2 km resolution and over the domain of interest will be the same adopted in the SREPS and in the CONSENS PPs (COSMO Technical Report Nr. 13 and Nr. 22).

The work is organised in two subtasks.

Task 2.1 - Identification of the parameters to be perturbed.

First an identification of the outdated Parameter Perturbations will be performed, on the basis of the new defaults which have been assigned to some parameters of the COSMO model physics in the last few years.

Then, new schemes and parameters which can be subject to perturbation will be identified. This task requires the collaboration of WG3a and WG3b, whose scientists will be required to provide a list of parameters to be perturbed (with associated range of possible values). Since this work is already part of the Task 3.4 of the CALMO-MAX PP, first it will be checked what is already available from CALMO-MAX, then it will be decided if further work is needed. If more work is needed, it will be shared with CALMO-MAX and FTEs will be assigned only to one of the two projects, with agreements of the coordinators.

Task 2.2 - Test the new Parameter Perturbation

New parameter perturbation will be tested in a test suite implemented at ECMWF, renewing the already existing suite (called CSPERT) used for the SREPS PP. The suite consists of a selectable number of runs of the COSMO model with the same Initial and Boundary Conditions for all the runs, but with the possibility of assigning different values to one or more of the physics parameters.

The suite will be run for COSMO at 2.2 km over Italy (model configuration similar to the one used in the COSMO-2I-EPS ensemble), for two different seasons (spring and autumn).

Since a similar testing suite has been implemented at ECMWF in the framework of the CALMO-MAX PP, it will be considered how to benefit of it, possibly leading to a unification of the suites for future usage in the Consortium.

Then, the quality of each run will be verified separately, in terms of 2m temperature and dew-point temperature, 10m wind intensity and precipitation. For the verification of the precipitation, suitable methods for high-resolution verification (spatial methods; Marsigli et al., 2008) and phenomena-oriented verification (e.g. for thunderstorms) will be adopted. The collaboration of WG5 will be ensured for the selection of the methods and for the discussion of the results.

This work will be conducted over Italy only, therefore only one specific COSMO domain is addressed.


Estimated needed resources for Task 2: 0.3 FTEs (Arpae SIMC)

Task 3. Lower boundary perturbation

Methods for the perturbation of the model lower boundary (surface/soil) will be further developed, due to the potential they have shown for improving the spread-skill relation at the surface and in the boundary layer (COTEKINO PP).

It is recalled that for some of the current ensemble configurations, the perturbation of the soil moisture is performed in the correspondent KENDA data assimilation cycle, therefore Initial Conditions used by the ensemble members include already soil moisture perturbations.

Perturbation of soil variables and parameters is a very delicate issue for those model configurations where the soil is running free, because of the different time scales of the soil and of the atmosphere. Too large perturbations of the soil may lead to an increasingly large decoupling between soil and atmosphere. These perturbations may introduce errors in the soil fields, which cannot be cured by the data assimilation of atmospheric variables. This issue should be checked in the experiments as well as in the operational configurations.

Task 3.1 - Perturbation of soil surface temperature.

With regards to the method of soil surface perturbation developed at IMGW (Duniec and Mazur, 2014; Mazur and Duniec, 2015; Duniec et al., 2017), the correlation of the amplitude of the perturbation with the soil type and/or the land cover (using preceding research) will be evaluated.

Tests will be performed for model resolution of 2.8 km (CP scale), with model domain covering Poland and its close vicinity (area of approx. 800x700 km). Results will be analysed to verify an assumption that the amplitude of perturbation is lower for more "compacted" soil - e.g., for clay in comparison to sand. This issues will also be studied more extensively (in a climatologic sense), for five continuous years (beyond the scope of the PP).

On top, "energy conservation" question will be addressed, viz., how big perturbation(s) can be to avoid an unjustified increase of temperature. In the previous studies this problem was solved using a normalization approach: the total perturbation of temperature (the sum of  temperature increments/reductions over the entire domain) was set to zero. Again, the results of this procedure will be compared with the non-normalized ones for a long enough period.

Results will be verified in terms of skill (MAE of ensemble mean) vs. measurements carried out at Polish SYNOP stations in the relevant periods.

Task 3.2 - Surface perturbations based on the Stochastic Pattern Generator.

Surface perturbations generated using the SPG will be tested.

Multi-level soil moisture and temperature perturbations will be generated using the SPG-generated 4D spatio-temporal random fields. Non-linearly transformed SPG field values will be employed to simulate non-Gaussian probability distributions of soil moisture. Atmospheric and soil perturbations will be made mutually coherent in order to maintain meaningful surface fluxes.

Effects on the spread and on the performance of the COSMO-Ru2-EPS ensemble of SPG generated soil and combined soil/atmosphere perturbations will be examined and the optimal SPG configuration will be chosen. The ensemble forecasts will be verified against observations using standard probabilistic scores, with the emphasis on near-surface variables and precipitation.

Task 3.3 - Combination of soil and upper air perturbation.

A question about the quality of EPS generated by a combination of different perturbation methods will be a subject of assessment. For example, combination of perturbations of soil surface variables and parameters, in general do not necessarily produce a perturbation being an algebraic sum or a simple addition. Therefore combination of perturbations in different parts of the model will also be tested, e.g. soil temperature/moisture with collection efficiency coefficient as an upper air factor. Other parameters (e.g. soil porosity, LAI, root depth) will also be included in the tests. Especially the question of an impact of magnitude of soil porosity and its perturbation related to soil type on skill and spread is promising, due to its influence on heat and water transfer in the soil.

The tests will include comparison of ensemble forecasts with corresponding values measured at Polish SYNOP stations during the testing period. The outcomes will be also prepared for a publication, as it was done for the preliminary results for a six-month warm period only.


Estimated needed resources for Task 3: 1.05 FTEs (IMGW, RHM)

Task 4. Post-processing and interpretation of ensembles

The work of this task will address mainly the development of methods to make effective use of CP ensemble output, including the development of probabilistic products for selected weather phenomena, benefitting also of the work carried out in the SRNWP-EPS II project of EUMETNET, and calibration. The work requires collaboration with WG4.

Task 4.1 - Calibration

On the basis of the work done in the SPRED PP, calibration methods will be applied to the TLE-MVL ensemble outputs. The calibration methods are: linear regression, logistic regression, artificial neural network (ANN) method. The archive of the operational ensemble outputs will serve as a "learning set" for calibrating the new forecasts.

In the "standard" procedure an ensemble mean is calculated as an arithmetic (weighted) average of forecasts from all members, with equal weight for every member. Accounting for previous results and corresponding measurements (at Polish meteorological stations), different weights can be applied, thus producing (significantly) changed mean. Depending on the procedure of calculation of weights (ANN, multi-linear regression, logistic regression) the spread/skill differs considerably. This study will be carried out over an operational domain for Poland (horizontal resolution of 2.8 km, approx. 1150x1100km; see Duniec et al., 2017) and for operational (archived) forecasts. Therefore, the starting date for the study is July 1st 2016 and it will continue as the operational forecasts are prepared.

Preliminary results (for short periods covering - all in all - one year) were presented at COSMO General Meetings in 2016 and 2017 (see Duniec et al., 2016 and Mazur et al., 2017).

Task 4.2 - Specific products from ensemble outputs

In the SRNWP-EPS II Project of EUMETNET, algorithms have been developed at CNMCA for the post-processing of the ensembles to produce the forecast in terms of thunderstorms and fog.

The quality of the products generated with these methods will be verified, also using new type of observations. Verification will be performed over selected periods and events. Ensembles involved are TLE-MVE, COSMO-2I-EPS and COSMO-LEPS.

For fog, verification will be performed locally against SYNOP and metar data but the usability of satellite-based products will also be tested.

For thunderstorms, lightning detection network observations will be also used for verification.

It will be tested also the generation of blended products from the weighted combination of convection-permitting and convection-parameterised ensembles (COSMO-2I-EPS and COSMO-LEPS). This work test will be performed over selected periods and for the Italian domain.


Estimated needed resources for Task 4: 1.75 FTEs (CNMCA, IMGW, Arpae SIMC)

Task 5. Initial and lateral boundary Conditions for the CP ensembles

More long term activities (COSMO years 2018-2020) are foreseen in this Task:

These activities depend on the human resources needed to carry them out, which are not available or not confirmed at the moment. Nevertheless, they are included in the Plan as they will be activated if resources are available for the COSMO years 2018 (starting from Sep 2018) and 2019 (starting from Sep 2019).


Estimated needed resources for Task 5: ~ 0.6 FTE, starting from Sep 2018, of which only 0.1 FTEs are planned at the moment

Task 6. Transition to ICON-LAM

According to the Science Plan of COSMO, the Consortium is moving towards the new model of DWD, ICON. This means that the convection-permitting ensemble systems must be converted to the ICON-LAM model. In order to be able to perform a transition to ICON-LAM in the current ensemble configurations, two ingredients are needed: data assimilation for ICON-LAM based on LETKF and a model perturbation method for ICON-LAM. The first issue is addressed by the data assimilation projects and plans, the second is addressed in the present PP.

For tackling the issue of model perturbation with ICON-LAM, in addition to the technical adaptations, it has to be examined whether the previous approaches to the representation of the model uncertainties can also be used with ICON-LAM. Due to the change in the characteristics of the model, modifications of the applied methods are likely to be required. The migration of the CP ensemble systems to ICON-LAM will require: forecast experiments for different test periods to explore the properties of the new systems, identification of further development requirements, adaptation of ensemble generation methods.

At the moment the plan reads as follows:

test are foreseen in the period 2019 - 2020. The first version of the ICON-LAM-EPS will be tested without model perturbations.
first tests with  the ICON-LAM data assimilation are running on 2km grid (ICON-D2(-EPS)) as basic ETKF cycle (i.e. only short term forecasts as first guess for next filter step). Work on the implementation of parameter perturbations in analogy to the approach in COSMO-D2-EPS (“RPP” in Appendix 2) has started by identifying the suitable parameters of the ICON parametrizations and their reasonable range. Technical modifications of the current parameter selection within the ICON code have been implemented to enable the analogy to COSMO-D2-EPS. Possible modifications and extensions of the methodology may be required during the development. First experiments to assess the sensitivity of the ICON-D2-EPS to these parameter perturbations will start end of 2018. [Note: The carefully tuned method for parameter perturbations in  global ICON-EPS (40 km grid) and its nest over Europe (20km grid) will not be modified in this context.]

tests in ensemble mode will start after 2020, beyond the present scope of this Project

currently 7 km of horizontal resolution, in the second part of 2018 and in the first half of 2019 it is planned to implement a test ensemble system which uses ICON-LAM. With regard to the change of the model, a thorough revision of the values for the perturbed parameters will be also undertaken.

In a successive phase, the possibility to implement in ICON-LAM a parameter perturbation approach with ECMWF style (Stochastically Perturbed Parametrizations, SPP; Ollinaho et al., 2017), that is sampling a pdf of possible values for some parameters, will be also investigated if resources are available.

Coordination with the C2I PP on the transition to ICON-LAM will be maintained, starting from a meeting between the Project Leaders during ICCARUS 2018.


Estimated needed resources for Task 6: 0.9 FTEs (which will be increased when other plans will be available).



The tasks dealing with methodologies for model perturbation are prone to delays and modifications in the schedule, because they are dependent on the results obtained in the scientific investigation. A promising methodology can turn out to provide unsatisfactory results, this leading to the need of further and deeper research or even to the decision of changing the perturbation strategy. On top, the timing of the availability of the COSMO model versions can influence the schedule of the experiments.

The work scheduled in task 6 on the transition to ICON-LAM is dependent on the status of its implementation in the COSMO members.

Since many different topics are addressed in the Tasks, a challenge will be to avoid duplicates and to make people work together. This is even more true when cross-cutting issues between different WGs are concerned.



Christensen H. M., Lock S.-J., Moroz I. M. and Palmer T. N., 2017.
Introducing independent patterns into the Stochastically Perturbed Parametrization Tendencies (SPPT) scheme.
Quarterly Journal of the Roy. Met. Soc., 143: 2168–2181, July 2017 A. doi:10.1002/qj.3075.

Duniec G. and Mazur A., 2014.
COTEKINO Priority Project - Results of Sensitivity Tests.
COSMO Newsletter 14: 106-113.

Duniec G., Interewicz W., Mazur A. and Wyszogrodzki A., 2016.
Ensemble Prediction System at IMWM-NRI - Operational Setup and Preliminary Results of a Study in the Frame of SPRED Priority Project.
Presentation during COSMO General Meeting, 5-8 Sep 2016, Offenbach, Germany.

Duniec G., Interewicz W., Mazur A. and Wyszogrodzki A., 2017.
Operational setup of the soil-perturbed, time-lagged Ensemble Prediction System at the Institute of Meteorology and Water Management - National Research Institute.
Meteorol. Hydrol. Water Management, 5(2): 43-51, doi: 10.26491/mhwm/71048.

Marsigli C., Diomede T., Montani A., Paccagnella T., Louka P., Gofa F. and Corigliano A., 2013.
The CONSENS Priority Project,
COSMO Technical Report no. 22, pp 45. Available at

Marsigli C., Montani A. and Paccagnella T., 2008.
A spatial verification method applied to the evaluation of high-resolution ensemble forecasts.
Meteorol. Appl., 15: 125-143, doi: 10.1002/met.65.

Mazur A. and Duniec G., 2015.
Ensemble Prediction System (EPS)-based forecast prepared from perturbations of soil conditions.
COSMO Newsletter 15: 63-71. Available at

Mazur A., Wyszogrodzki A., Duniec G. and Interewicz W., 2017.
SPRED PP activities at IMWM-NRI.
Presentation during COSMO General Meeting, 11-14 Sep 2017, Jerusalem, Israel.

Ollinaho P., Lock S.J., Leutbecher M., Bechtold P., Beljaars A., Bozzo A., Forbes R. M., Haiden T., Hogan R. J. and Sandu I., 2017.
Towards process-level representation of model uncertainties: stochastically perturbed parametrizations in the ECMWF ensemble.
Quarterly Journal of the Roy. Met. Soc., 143, 702, 408-422, doi: 10.1002/qj.2931.

Piccolo, C., Cullen M., Tennant W. and Semple A., 2018:

Comparison of different representations of model error in ensemble forecasts.

Quart. J. Roy. Meteor. Soc., accepted. doi: 10.1002/qj.3348.


Tsyrulnikov M. and Gayfulin D., 2017.
A limited-area spatio-temporal stochastic pattern generator for simulation of uncertainties in ensemble applications.
Meteorol. Zeitschrift, 26(5): 549-566, doi: 10.1127/metz/2017/0815.

Wilks D., 2011.
Statistical Methods in the Atmospheric Sciences, Third Edition.
Academic Press, pp. 676.

Appendix 1:  Task table

Task Contributing scientist(s) FTEs
Start Deliverables Date of delivery

Grzegorz Duniec
Witek Interewicz
Andrzej Mazur




total 0.35


total 0.7


total 0.7

Mar 2018 Continuous assessment of RNG efficiency in a form of short reports and finally publication. Extended assessment of result of combination of various perturbation applications in a form of report, recommendation and publication. Aug 2018
Aug 2019
Aug 2020
E. Astakhova, D.Gayfulin, M.Tsyrulnikov 0.6 0.2 0.2 0.2 Mar 2018 Report on experiments with SPG. Recommendations on the SPG usage in additive and/or multiplicative modes in EPSs. Aug 2020
C. Gebhardt, E. Machulskaya 0.8 0.2 0.3 0.3 Mar 2018 Estimate parameters for the predictors of the EM-scheme; verification of the re-tuned EM-scheme; proposal of additional predictors Aug 2020
  L. Füzer, A. Walser 0.45   0.45   Oct 2018

COSMO-E case studies with iSPPT and evaluation of KENDA analysis increments as a proxy for additional model perturbations
(Master Thesis)

May 2019

N. N. 0.3   0.3   Mar 2019 Proposal of new parameter perturbations; verification results Sep 2019

Grzegorz Duniec
Andrzej Mazur




total 0.15


total 0.25


total 0.25

Mar 2018 Short reports, publication Aug 2018
Aug 2019
Aug 2020
E. Astakhova, D.Gayfulin, M.Tsyrulnikov 0.4     0.4 Sep 2019 Experiments with BEMS and SPG-generated soil perturbations Aug 2020

Grzegorz Duniec
Witek Interewicz
Andrzej Mazur




total 0.2


total 0.45


total 0.45

Short report.
Generation of new products.
Dedicated software and interface for results to be presented to specific end-users - algorithms, tests, implementations.
Aug 2020
Francesca Marcucci 0.5 0.25 0.25   Mar 2018 Short report Dec 2018
C. Marsigli
A. Montani




total 0.05


total 0.1

  Mar 2018 Generation of new products Jun 2019

(for now)
M. Arpagaus, A. Walser 0.1   0.1   tbd COSMO-E experiments with LBCs from ICON-EPS vs. IFS-ENS Aug 2019

E. Astakhova, D. Alferov 0.4   0.1 0.3 Mar 2019 Experiments with ensembles based on ICON-LAM Aug 2020
C. Gebhardt 0.3   0.2 0.1 Dec 2018

Implement an approach for parameter perturbations in ICON-D2-EPS. Assess the sensitivity of the forecasts to these perturbations. Perform and evaluate forecast test runs.

Aug 2020
A. Montani 0.2   0.2  

Jan 2019

Testing of ICON-LAM in COSMO-LEPS Aug 2019

Appendix 2:  Model perturbations.

The most widely used model perturbation methodologies are here listed and shortly described:

PP (Perturbed Parameters)
each member has a different value of one or several parameters, fixed during the integration
RPP (Random Perturbed Parameters)
each member has a different value of one or several parameters, fixed during the integration, but the value of the parameter is randomly chosen for each cycle and member
RP (Random Parameters)
each member has a different value of one or several parameters, fixed in space during the integration but varying in time; the value of the parameter is randomly chosen for each cycle and member
SPPT (Stochastically Perturbed Parametrization Tendency)
stochastically perturbed physical tendency, with spatial and temporal correlation

iSPPT (independent SPPT)

            as SPPT but the tendency from each parametrization scheme is perturbed using an independent stochastic pattern.

SPP (Stochastically Perturbed Parametrisations)
physics parameters are stochastically perturbed with spatial and temporal correlation
PSP (Physically Based Stochastic Perturbations)
Boundary Layer stochastic perturbations with amplitude based on information obtained from turbulence parameterization, with spatial and temporal correlation
different members use different physics schemes, fixed
different members use different models, fixed
Stochastic parametrisation
a scheme for parametrising a physical process in the model which is intrinsically stochastic