Ensembles estimate uncertainty in the future projections of Numerical Weather Prediction (NWP) models. They provide probabilistic spatio-temporal forecasts of different variables. The imperfection of these forecasts motivates statistical postprocessing.
The objective of this thesis was to implement and investigate univariate and multivariate techniques to statistically postprocess ensemble forecasts of vertical temperature profiles.
The postprocessing was done with ballon-carried radiosoundings of four station locations in Germany, which served as observations or response, and past model forecasts of the ECMWF-EPS, which served as the covariates.
Multivariate postprocessing was carried out with Ensemble Copula Coupling (ECC) by conserving the rank-order structure of the raw ensemble. But ECC did not calibrate locaton and scale, which were shown to be misspecified. Therefore two univariate postprocessing techniques, Non-homogenous Gaussian Regression (NGR) and Standardized Anomalies Model Output Statistics (SAMOS), were applied prior to ECC.
Discrete sampling from the calibrated probability margins was done randomly (ECC-R) and by quantiles (ECC-Q). Additionally the raw ensemble was rescaled with these margins conserving not only the rank-order structure but also the relative distances between the single ensemble member (ECC-S).
Models were selected and validated with the statistical criterions of log-likelihood, AIC/BIC, and univariate (CRPS and ignorance score (IGN)) and multivariate verification scores (energy score (ES) and variogram score (VS)) together with multivariate rank histograms (MVHs).
Univariate postprocessing with NGR resulted in a larger log-Likelihood, better AIC/BIC and CRPS and IGN, compared to SAMOS. The CRPS was significantly improved by NGR, as was the multivariate ES, but VS and the MVHs showed a worsening, compared to the raw ensemble. ECC, and especially ECC-S, further improved the ES significantly and improved the VS and the MVHs. ECC-S was the only model to significantly improve the VS of the raw ensemble, and that significantly.
Therefore, including also the distances between the ensemble members (ECC-S) in addition to their rank-order as previously done in ECC results in better multivariate forecasts of vertical profiles of air temperature.