Prof. Dr. Hans R. Künsch (ETH Zürich)
Joint work with Fabio Sigrist and Werner Stahel
The goal of postprocessing deterministic numerical
weather forecasts is to produce probabilistic forecasts
that are calibrated and sharp. In case of precipitation,
this is difficult because one has to deal with non-Gaussian
distributions and complicated space-time correlations.
I will present a method which uses a non-linear transformation
of a Gaussian space-time processes interpreted as
"potential precipitation". For the mean of this process, a
regression model with the deterministic forecasts as regressors
is used, and for the covariance a stochastic advection-diffusion
model. I will explain how this model is fitted and illustrate
the method with forecasts from the COSMO-2 model for northern
Switzerland.
