Speaker
Description
At DWD exists a weather interpretation of model output from the global ICON, ICON-EU and COSMO-D2, which will be replaced by ICON-LAM in the near future. By applying a whole bunch of WMO compliant thresholds with complex relationships to all relevant model variables, e.g. temperature, pressure in different height levels, rain rate, wind gusts, cloud cover, etc., the goal of the weather interpretation is to derive the diagnostic elements for significant weather (ww). One fundamental difficulty is that precipitation components exists as sums of the last hours but all other components are instantaneous. On the one hand, this implies some peculiarities of the interpretation itself to avoid inconsistencies and misinterpretation. On the other hand, this affects the quality of the interpretation, e.g. fog which was present only in the past time interval but not at the current time. Finally, the model weather is verified against diagnostic weather from synop stations. Another crucial part is also the interpretation of thunderstorms and its verification.
To improve such a system, which is sensitively depending on many different thresholds, the idea is to apply machine learning algorithms to our weather interpretation in the future. However, there is no clear idea about how the machine learning can be adapted to our existing methods of weather interpretation. We want to use this workshop to collect first ideas about AI and deep learning itself and their potential application.