| Popis: |
The winter season 2013/14 has shown once again that mid-latitude windstorms have a strong impact on Europe in terms of potential damages and insured losses caused by gusts or indirectly by storm surges. To analyze windstorm occurrences and associated footprints, data sets or simulations are necessary, which have a high temporal and spatial resolution, cover reality as best as possible, and are able to predict wind or gust speeds for future climate conditions. In order to meet these requirements, this study combines data from General Circulation Models (GCMs), reanalyses, Regional Climate Models (RCMs), and observations with respect to the individual advantages. The first part relates large-scale reanalysis wind speeds and regional gust speeds simulated with the COSMO-CLM RCM by multiple linear regression models. These models are trained with a sample of 100 historical windstorms, which had a large influence on European onshore areas. By applying the obtained transfer functions to other large-scale data (GCM simulations), this kind of statistical-dynamical downscaling enables to estimate high-resolution gust speed footprints for Europe in a computationally cost-efficient way. A validation against observations of German weather stations shows that both the results for the combined downscaling approach and for pure dynamical downscaling with an RCM are in congruence with measurements and have similar deviations. The deviations from observations are caused by model biases due to local effects, which cannot be parameterized by the RCM. Thus, the second part aims at adjusting RCM simulations to observations. For this purpose, wind observations of 173 selected test sites and gust observations of 111 selected test sites of the German Weather Service as well as COSMO-CLM simulations are fitted by Weibull distributions. The observation-based distribution parameters are interpolated to the RCM grid, where they are related to the parameters of the simulations by probability mapping. This enables to correct the simulations while retaining the regular grid. Validation with grid points and nearby test sites shows an improvement of about 84 % and 64 % of the sites for wind and gust, respectively. Averaged over all test sites, for wind, 99 of the 100 selected events are improved in terms of more realistic simulations, and 88 events are improved for gusts. In the third part, the introduced statistical-dynamical downscaling is applied to a large ensemble of decadal hindcasts simulated with the earth system model of the Max-Plack-Institue (MPI-ESM-LR, baseline 1). The decadal predictability of peak wind speeds over Europe is analyzed using ten different realizations of yearly initialized decades between 1979 and 2010. Comparing original COSMO-CLM footprints and statistical-dynamically downscaled footprints of all days within the investigation period with respect to selected quantiles shows that the approach is appropriate for this large ensemble. The methodology is also able to identify periods with enhanced windstorm activity as exemplary shown for the years 1990 - 1993. The validation against uninitialized simulations using skill scores indicates that both the original hindcasts and the downscaled hindcasts are able to predict wind speeds and peak wind speeds for Europe on decadal scales. The performance is best for high quantiles equal or above 75 % and for the first years (1 - 4) after initialization. It is shown that the combination of the different data sets enables to increase the data resolution in a computationally inexpensive way, while small-scale features parameterized by the RCM are still included. This is important for large ensembles like given for decadal hindcasts and predictions. Additionally, the approach has the advantage of estimating peak winds as proxy for gusts, whereas most large-scale data sets only provide wind speeds. For wind speed, it is recommendable to adjust the simulations to observations by quantile mapping. The same technique delivers for gust speeds improvements only in some areas of the German investigation area. In general, adding or omitting parts of the approach introduced here is easy to implement as all techniques are applicable stand-alone and joined together. This enables to use the methodology also for other parameters. |
