Recalibrating counts of extreme temperature days in decadal predictions

Decadal prediction models mostly focus on predicting mean temperatures and precipitation on annual scales. For applications in agriculture and the health sector, indicators for heat stress and extreme temperatures appear to be more relevant than the mean temperatures. Those indices often involve maximum temperatures on a daily scale. Decadal predictions need to be recalibrated to reduce biases and adjust dispersion to match prediction uncertainty and hence increase reliability. In the frame of the research project "Coming Decade", funded by the German Ministry of Research, Technology and Space, we explore two different approaches to obtain recalibrated probability distributions for the annual counts of days with maximum temperatures exceeding a given threshold, i.e. Summer Days (T_max≥25°C) and Hot Days (T_max≥30°C).

(1) First, we obtain annual counts of Summer Days and Hot Days directly from decadal predictions of daily maximum temperatures. Subsequently, we recalibrate the distribution of counts from the ensemble forecast using a variant of the parametric Decadal Climate Forecast Recalibration Strategy (DeFoReSt) proposed by Pasternack et al. (2018) with distributions accounting for count data, i.e. Poisson or negative-binomial distribution.

(2) As an alternative approach, we apply a bias and drift adjustment of daily maximum temperatures using non-homogeneous Gaussian regression in the frame of generalized additive models. From the resulting adjusted daily temperatures we obtain counts for daily exceedances and aggregate them to an annual scale. We then recalibrate with the ensemble recalibration strategy (1).

We aim to compare these approaches for recalibrated Summer Days and Hot Days over Europe using a skill score for probabilistic forecasts like the CRPSS. We use decadal predictions from the operational decadal prediction system of the German Meteorological Service (DWD) based on the Max Planck Institute Earth System Model (MPI-ESM1.2-LR) and evaluate the performance with respect to the ERA5 reanalysis.