Statistical-dynamical downscaling of global climate simulations

The statistical-dynamical downscaling procedure links global and regional model simulations using statistics about large-scale weather situations in order to derive the regional climate corresponding to the global climate. A general assumption of this approach is that the climate of a given region (e.g. the Alpine region) is characterized by the frequency distribution of large-scale weather situations: In a changed climate the frequencies of the large-scale events might be altered, but their characteristics are assumed to remain unchanged.

The downscaling procedure

The procedure consists of the following steps:

1. Each day of a multi-year time series of global simulations (horizontal resolution: 250 to 500 km) is classified into one of about 100 defined large-scale weather situations. The selected weather classes should be characteristic for the regional model domain. Presently, classification is based on season, 850 hPa geostrophic wind direction and speed and the type of precipitation (stratiform or convective) .
2. Regional model simulations are carried out once for each weather class, being initialized with the vertical profiles of geostrophic wind, temperature and humidity corresponding to this class. The model domain covers the greater Alpine region (1200 x 700 km2) with a horizontal resolution of 20 km. Generally, the mesoscale simulations extend over a full daily cycle in order to permit the development of topographically induced thermal circulation systems.
3. The mesoscale model output is weighted with the frequency distribution of the classes to give the mesoscale climate (two dimensional fields of frequency distributions of climatological parameters like temperature, wind, precipitation).

Verification

Any downscaling procedure transmits uncertainties of the global climate simulation to the regional scale. Therefore, the verification of the procedure itself must be based on ``perfect'' large scale data. A set of 10 years (1981 - 1990) of daily homogenized ECMWF-analysis data on a Gaussian grid of T42-resolution (which is the same horizontal resolution as that of global climate simulations to which the downscaling is being applied) serves as input to the regionalization method, whereas the output is compared with SYNOP-data from 30 stations located throughout the regional model domain (values of temperature, pressure, wind, precipitation, 8 times a day for the same 10 years period).

For each weather class the regional model output (two dimensional fields of distributions of temperature, pressure, wind and precipitation) is compared with the distributions at the 30 stations obtained for the days classified into the corresponding weather type.

Finally the regionalized climate patterns are compared with the climatology of the station data.

FIGURE 1. Regional distribution of the mean wind speed (in m/s) at 10 m AGL for the winter half year.

FIGURE 2. Mean values of temperature, wind speed and precipitation at the location of twenty stations for the winter half year. Precipitation values are standardized. Dots: model data, triangles: observations.

Results

As an example for the resulting regional distributions of climatological parameters, Fig. 1 shows the mean surface wind speed field in winter. High values can be seen over mountains and over the Mediterranean sea. Also the high wind speeds near the Croatian coast are simulated.

Fig. 2 shows the comparison of mean temperature, wind speed and precipitation at the location of twenty stations. The difference in height of the model grid and the station location was taken into account with regard to temperature assuming a lapse rate of 6.5 K/km.

The problem of a verification based on local observations is that the value of one model grid is representative for a box with a volume of 20 km x 20 km x 50 m, whereas the regional extension for which the observations are representative is much smaller, especially for precipitation and wind measurements. E.g. the extension of representativity for a wind observation is about 100 m horizontally and 5 m vertically.

Nevertheless the agreement between model results and observations is generally quite good for temperature and wind speed fields. Absolute values as well as the regional distribution are fairly well reproduced.

The results for precipitation, which in fact is a much more critical parameter in models than e.g. temperature or wind speed, are not as satisfactory. It has to be noted that a much denser distribution of station data is needed for verification of precipitation fields.

In summary, the results of the verification show that in principle the statistical-dynamical downscaling procedure is able to reproduce the observed regional distributions of climatological parameters on a mesoscale, but that one cannot expect local features to be resolved.

References

Frey-Buness A., 1993: Ein statistisch-dynamisches Verfahren zur Regionalisierung globaler Klimasimulationen. DLR-Forschungsbericht DLR-FB 93-47.

Frey-Buness A., D. Heimann, R. Sausen,1994: A statistical-dynamical downscaling procedure for global climate simulations. Appl. Theor. Climatol. 50, 117-131.

Fuentes U., V. Sept, D. Heimann, R. Sausen,1995: Statistical-dynamical downscaling of global climate simulations. Proceedings of the 6th International Meeting on Statistical Climatology, June 1995, Galway, Ireland, 37-40.

MAP Data Centre - April '05 - MAP WebMaster