Task 2.2 Development of user-specific products for sectoral applications with statistical adjustment for reliable probabilistic predictions
Probabilistic decadal climate predictions need to be reliable in the sense that “they mean what they say”. This can be achieved with recalibration, i.e. learning a relation between ensemble hindcasts and observations. This relation aims at i) adjusting the mean of the forecasted probability distribution to minimize bias and drift, and ii) adjusting the width such that it reliably represents the forecast uncertainty. A non-homogeneous regression approach (e.g. Pasternack et al., 2018) allows adjustment for (conditional and unconditional) bias, drift and width of the distribution for annual values at individual grid points. The model developed by Pasternack et al. (2018) together with “boosting”, a powerful model selection strategy from machine learning in a cross-validation setting (Pasternack et al., 2021) has been developed with a focus on decadal predictions. Nonhomogeneous regression can be supplemented by logistic and quantile regression (Bentzien and Friederichs, 2012).
Task 2.2 will focus on: Recalibrating essential variables (temperature and precipitation) as well as non-typical model variables identified together with users, such as wind speed or radiation, including different time periods of averaging. Recalibration will be tested for dependence on the initialization state and for quantiles. For user-specific indices for sectoral applications, this goes beyond MiKlip (Mömken et al., 2021), which evaluated univariate indices. Here, we consider more complex indices, which require more than one variable, such as heat stress or drought indices. Computation of indices identified in Task2.1 will be implemented and calibrated, which requires both the individual variables as well as their interdependence. One method to restore dependence structure of the variables within the ensemble is the ensemble copula coupling (ECC, Schefzik et al., 2013) or Schaake shuffle (Schefzik, 2016).
The definition of products is based on intensive user interaction (Task 2.1). Derivation of userspecific products as well as the recalibration strategies will be realised as plug-ins for CODES that global model output as well as downscaled products can be used as input.
Contributors
Freie Universität Berlin
Prof. Dr. Henning Rust
MSc. Felix Fauer
MSc. Andy Richling
University of Bonn
Priv. Doz. Dr. Petra Friederichs
MSc. Philipp Ertz
Deutscher Wetterdienst
Dr. Clementine Dalelane
Dr. Andreas Paxian