omsz
>>
  Hungarian Meteorological Service  founded: 1870
Research and development | Numerical Weather Prediction  | Analysis of the Atmospheric Environment | 
Felbontás: KicsiFelbontás: NormálFelbontás: KözepesFelbontás: NagyobbFelbontás: Nagy  

Short range ensemble prediction

Ensemble forecasts

Accurate forecasting of dangerous, hardly predictable events is one of the most important tasks of meteorology. We would not be satisfied if our models provided reliable predictions for common meteorological situations (simple frontal situation, anticyclone, etc.) but extreme events (wind storms, cyclones causing heavy precipitation, etc.) could be predicted only with low accuracy. That is why the developers of the numerical weather prediction systems are making compromise when setting up a system which ensures that the model provides accurate forecasts in both common and extreme situations.

To make forecasts more and more precise intensive research is being done in many fields of numerical weather prediction. One of these fields is the generation of initial conditions (when making a weather forecast one has to solve a set of partial differential equations for which initial and boundary conditions are needed), which is a very complex task. On the one hand measurements are made on a non-regular grid (so interpolation is needed to the grid of the model), on the other hand the number of measurements is very low compared to the model's degree of freedom. Finally we have to take into account the errors in the measurements.

It follows that the exact specification of the initial conditions is not possible. This can arise the question: how large problem these errors can cause? The system of equations which has to be solved when making a weather forecast is a set of nonlinear partial differential equations showing large sensitivity to initial conditions. This means that small errors in the initial fields can cause large errors in the forecast.

If that is the case, is there any solution to the problem? Fortunately the answer is yes. If the initial conditions could not be defined precisely then a set of initial conditions should be defined which are equally possible and only slightly differ from the "best" initial condition. Then forecast is made from each of them resulting an ensemble of forecast scenarios. That is why this technique is called ensemble method. The use of this ensemble enables us to associate a probability value to each forecasted weather event and to make conclusions on the uncertainty of the forecast. All these things make the ensemble technique a useful method especially when forecasting extreme events.

For making an ensemble forecast lots of methods can be used:

  • Perturbation of observations used in data assimilation
  • Perturbation of initial conditions in a random way, or by using special techniques like singular vector method or breeding method
  • Generation of initial conditions by using different assimilation techniques (OI, 3DVAR, 4DVAR)

Certainly not only the errors in the initial conditions can cause problems. Other sources of errors are the parameterization of small scale processes, the discretization in time and space and so on. These can also be handled for example by using different parameterization packages, different models and so on.


Methods of creating (short range) ensemble forecasts

Multi analysis ensembleThe analysis fields made at different centres (or at the same centre but using different assimilation methods) are collected. From each of them forecast is made. These forecasts will be the members of the ensemble system.
Multi model ensembleIn this case the forecasts made at different centres (or at the same centre but with different models) are collected. These forecasts will be the members of the ensemble system. The method is often refered to as "poor man's ensemble system", because each participating cetre has to create only one member (which is generally the operational forecast of that centre) of the ensemble.
Perturbation of observationsObservations used for data assimilation are perturbed (perturbations are of the order of observation error). This results in different initial conditions and finally in a set of forecasts. These forecasts will be the members of the ensemble system.
Generation of initial perturbations Perturbations used in this method are again of the order of analysis error. The two most frequently used techniques are singular vector method and breeding method.

Breeding method: As a first step initial conditions are randomly perturbed, then forecast is made from all these perturbed initial conditions. At the end of the forecast perturbations are re-scaled, the actual analysis is modified with these perturbations and the process continues. After 4-5 days this method leads to the selection (breeding) of fastest growing perturbations.

Singular vector method: This method requires the solving of an eigenvalue problem and has a strong mathematical basis. Using this method is very expensive in terms of CPU time.

Perturbation of physicsIn this case parameters of physical parameterization packages or the packages themselves are changed resulting in a set of different forecasts.
Downscaling of global ensemble forecastsThis is the easiest method of all, because all we have to do is to use the global ensemble members as initial and boundary conditions for our limited area ensemble system. The problem is that the perturbations generated for the global ensemble system are usually effective only on medium range and large scales therefore it is not sure that they are optimal for short range ensemble forecasts.

Another method is when only few members of the global ensemble system are used as initial and boundary conditions. For example: members of the global ensemble prediction system are grouped in different clusters. Within each cluster one representative member is selected according to the criterion that the representative member is closest to the members of its own cluster and most distant from the members of the other clusters. Finally all representative members are used as initial and boundary conditions for the limited area ensemble system.


Visualization of ensemble forecasts

It is not enough to prepare ensemble forecasts, they should also be displayed in a proper way. The visualization of ensemble forecasts is a little bit different from the visualization of classical (deterministic) forecasts because in the case of an ensemble system a set of forecasts (not just a single one) has to be visualized. In the meanwhile it is important that the amount of information has to be manageable for the forecasters (one can easily imagine how much time it would take to go through e.g. 51 forecast per day). Hereafter the possible visualization techniques will be presented. The plots used as examples are the products of our forecasting and visualization system.


Individual ensemble members

 

Like in the case of deterministic forecast ensemble members can be displayed one by one. However, in the case of an ensemble system with many members it is time consuming to go through all the individual members.

Ensemble mean

 

The mean of the ensemble members can also be visualized. However one should take care when using such a product: just imagine the case when half of the ensemble members indicate low pressure over Hungary, while the other half indicate high pressure. Their mean will be a featureless pressure field.

 

Example of ensemble mean

Figure 1.
 

Example of ensemble mean. 24h total precipitation forecast for the period 08 November 2004 06 UTC - 09 November 2004 06 UTC. During this 24h more than 30mm precipitation was observed in the middle of the country. This can also be seen in the forecast, although the area of most intensive precipitation is not exactly correct.

Stamp maps

 

For a given lead time and a given parameter all ensemble members are visualized, of necessity only in a small size. With the use of such a diagram it is very simple to realize large differences, but because of the small size details can be hardly seen.

 

Stamp maps

Figure 2.
 

Example of stamp maps. 24h total precipitation forecast for the period 08 November 2004 06 UTC - 09 November 2004 06 UTC for all ensemble members. Differences can be seen mainly in the western part of the country.

Plume diagram

 

For a given location and given parameter this diagram shows the variation of the chosen parameter in time. One can observe how predictability (and also uncertainty of the forecast) changes in time.

 

Plume diagram

Figure 3.
 

Example of plume diagram. Total precipitation forecast for Sopron started from 08 November 2004 00 UTC analysis. It is interesting to see that at 09 November 06 UTC some members are indicating 20mm/6h precipitation while some other predict only 1mm/6h.

Probability map

 

These kind of maps show the probability of a given event (e.g. temperature below zero).

 

Probability map

Figure 4.
 

Example of probability map. The probability of the event that 24h total precipitation would be more than 30mm is presented(values are in %).

Spaghetti diagrams

 

For a given lead time a chosen isoline of a given parameter is plotted for all ensemble members. While the plume diagram gives information about the uncertainty in time, the spaghetti diagram does the same but in space.

 

Spaghetti diagrams

Figure 5.
 

Example of spaghetti diagram. The 1015hPa isobar of mean sea level pressure is plotted for all 11 ensemble members. Forecast starting from the 08 November 2004 00UTC analysis. One can observe that over some parts of the domain the uncertainty is quite small, isobars are close to each other, while at some regions isolines are rather far, uncertainty is large.


Short range ensemble forecasts

In the last decade the research in the field of ensemble prediction systems was mainly concentrating on global scales and medium range. Nowadays the emphasis is more and more moving towards the short ranges and smaller scales with the main goal being the better understanding and prediction of local extreme events like heavy precipitation, wind storms, big temperature-anomalies. However, methods used in the medium range cannot be directly applied to short range forecasting. Research has already been done in the field of LAMEPS (Limited Area Ensemble Prediction System) and there are some operational or quasi-operational short range ensemble systems (e.g. SREF system at NCEP, PEACE system at Météo-France, the COSMO-LEPS system or the SRNWP-PEPS system).


LAMEPS research at Hungarian Meteorological Service

LAMEPS research started in 2003 at the Hungarian Meteorological Service with the final goal being the development of an operational system. This system will be based on the limited area model ALADIN (Aire Limitée Adaptation dynamique Développement InterNational). The ALADIN model is a spectral limited area model which was developped in an international framework. The model is used operationally at HMS. For making an ensemble forecast lots of methods can be used (e.g. multi-model, multi-analysis, perturbation of observations, singular vector method, breeding etc.). It is not known yet (especially at mesoscale) which method would provide the best forecasts. Therefore several methods will be tried and tested. At the moment we are testing two different methods and we are also taking part in the SRNWP-PEPS project. Besides, we are in connection with the experts of the ensemble technique: American researchers, Italian scientists (who are playing a key role in Europe in the development of short range ensemble forecasts) and of course with French colleagues (considering that the ARPEGE/ALADIN model system is being developped with their coordination).
Support for our LAMEPS related work is provided by different research projects.


Downscaling of global ensemble forecasts/1

We started our experiments with running ALADIN EPS coupled with PEACE (Prevision d'Ensemble a Courte Echéance) ensemble members. The PEACE system is now runnig operationally at Météo-France once a day (starting from the 18 UTC analysis). It is based on the global spectral model ARPEGE and has 11 members (10 perturbed and a control one). The initial perturbations of this global ensemble system are based on targeted singular vectors. We use all PEACE members as initial and boundary conditions for our limited area ensemble which means that our system also has 11 members.

In PEACE initial perturbations are based on targeted singular vectors. When computing singular vectors two very important features are target domain and target time. By choosing these parameters carefully, perturbations are believed to be efficient in the area of our interest. The target domain used in the PEACE system is covering Western Europe and the North Atlantic region. The target time is 12 h.

It was decided to carry out sensitivity experiments to investigate the impact of different target domains and target times used in singular vector computations and also to find out whether or not the PEACE provided initial and boundary conditions are convenient for the local EPS run, for a Central European application.

From the case studies and the experiment with downscaling the PEACE members it seems that the PEACE provided initial and boundary conditions are not really optimal for the local ensemble run, for a Central European application. It can be understood if we consider that the PEACE system was calibrated to Western Europe. Improvement can be made by changing the target domain and target time used in the global singular vector computation but this improvement is rather small. It looks that local perturbation will be needed inside ALADIN. One possible solution is the breeding method which is planned to be tried later this year.


Downscaling of global ensemble forecasts/2

As mentioned above the simplest way to generate a short range ensemble forecasts is to downscale all members of a global ensemble system. But this can be very time consuming if we consider a global ensemble system with for example 50 members. It can be asked whether or not it is possible to reduce the number of ensemble members without loosing too much information. A possible solution can be when only few members of the global ensemble system are used as initial and boundary conditions. For example: members of the global ensemble prediction system are grouped in different clusters. Within each cluster one representative member is selected according to the criteria that the representative member is closest to the members of its own cluster and most distant from the members of the other clusters. Finally all representative members are used as initial and boundary conditions for the limited area ensemble system. This technique is used in the COSMO-LEPS ensemble system. Also at HMS research has been started to develop such a system. It is planned to use ECMWF EPS forecasts as initial and boundary conditions for the limited area model ALADIN.


The SRNWP-PEPS project

Regional Modelling in Europe is organized in 4 Consortia: ALADIN (Aire Limitée Adaptation dynamique Développement InterNational), COSMO (COnsortium for Small MOdelling), HIRLAM (HIgh Resolution Limited Area Model), and UKMO (UK Met Office), each of them having their own limited area model. A huge variety of operational forecasts exist, which are produced on different domains with different grid resolutions using different model parametrizations or different data assimilation techniques. This makes the possibility of generating a Poor man's ensemble prediction system (PEPS). Each participating weather service only has to create one member of this system, which is usually the operational limited are forecast produced by that centre (so the workload of the participating services is not increasing).> With the coordination of DWD (Deutscher Wetterdienst) a project started with the main goal being the development of an operational European multi-model ensemble system for the short range. The members of the four above mentioned consortia take part in the project. The participating weather services (including HMS) send daily some of their forecasts to DWD and in return they receive EPS products.

Ensemble mean forecast
Figure 6.

Example of SRNWP-PEPS products. Ensemble mean forecast, 24h total snow (water equivalent) for the period 21 February 2005 06 UTC - 22 Fbruary 2005 06 UTC.

Probability map
Figure 7.

Example of SRNWP-PEPS products. Probability map, 24h total snow more than 5cm for the period 21 February 2005 06 UTC - 22 February 2005 06 UTC.


Publications (Post Script file format):