Kolloquium zur Masterarbeit: Gwenda R. Rode

Studentischer Vortrag (Kolloquium zur Masterarbeit/Master's thesis defense)
Vortragender/Speaker: Gwenda Roselene Rode
Ansprechpartner/Contact: Jun. Prof. Dr. Markus Schmidtchen, Dr. Simon Praetorius

Titel/Title: SynCast - Enhancing the Precision of Weather Forecasts for Wind Speed by Multi-Model Data Fusion via Graph Neural Networks

Zusammenfassung/Abstract:

Postprocessing of the results of numerical weather prediction is an essential part of forecasting, as it attempts to alleviate errors introduced due to limited numerical precision and imprecise initial conditions. One postprocessing technique is blending, which combines several forecasts and weights them according to statistical performance. 
The blending method "SynCast", which combines the results of two forecasts into one, is presented. It is based on a graph neural network, which applies spatial connectivity between grid points via message passing. SynCast' loss function is constructed in a way which verifies the results with radiosonde observations, while maintaining proximity to the original forecasts. The domain is constrained to northern Europe, and forecasts for the horizontal wind variables U and V in a height of 100m are used. SynCast's results are analyzed with several error metrics, as well as via their value distributions and connection to the values of the input forecasts.

SynCast increases the skill in comparison to the two input forecasts in various experiments. Its performance goes beyond a mere linear combination, but incorporates information gathered from the observations, while still maintaining some physical coherence provided by the structure of the forecasts in the output. With a low number of training samples and especially observations, SynCast barely reproduces extreme values. In a simplified setup, it is able to represent the occurrence of extreme high wind speeds. SynCast's abstract structure can be extended to other domains and variables, as well as different forecast and observation types.