Aspects of spatial postprocessing for global temperature forecasts

Feldmann, Kira

PDF
Diss Feldmann.pdf - Veröffentlichte Version
Download (3MB)

URN:	urn:nbn:de:bsz:180-madoc-636486
Dokumenttyp:	Dissertation
Erscheinungsjahr:	2022
Ort der Veröffentlichung:	Mannheim
Hochschule:	Universität Mannheim
Gutachter:	Schlather, Martin
Sprache der Veröffentlichung:	Englisch
Einrichtung:	Fakultät für Wirtschaftsinformatik und Wirtschaftsmathematik > Applied Stochastics (Schlather 2012-)
Fachgebiet:	510 Mathematik
Freie Schlagwörter (Englisch):	statistical postprocessing of weather forecasts
Abstract:	High quality predictions are essential for informed decision-making. This holds especially true in meteorology as weather phenomena can engender high socioeconomic cost. In the past decades, the paradigm in weather prediction has shifted from point forecasts to probabilistic forecasts providing a probability distribution aiming to capture the true uncertainty of the prediction. Operationally, these probabilistic forecasts are generated by ensembles consisting of multiple runs of numerical weather prediction systems that differ in model formulations and/or initial conditions. Despite best efforts, the ensemble forecasts can still be subject to biases and dispersion errors. Statistical postprocessing corrects these systematic shortcomings and releases the full potential of the ensemble. This work focuses on two aspects of statistical postprocessing: incorporating spatial dependency structure into the probabilistic forecast and the choice of an adequate training/verification set for the postprocessing model. Many real-world applications of statistical postprocessing benefit from modeling of dependencies – e.g. spatial, temporal or inter-variable. The majority of the pioneering postprocessing approaches did not address this need. Here, we extend the well-established postprocessing method Ensemble Model Output Statistics (EMOS) with a Gaussian random field that models global predictive errors. Indicated by the characteristics of the forecast errors, the covariance function of this random field is assumed to be non-stationary, accounting for land-water differences in predictive ability and correlation length. In case studies, we apply this spatial postprocessing methods to 2m temperature forecasts by The Interactive Grand Global Ensemble (TIGGE), as well as the European Centre for Medium-Range Weather Forecasts (ECMWF) ensemble, and compare their forecast skill to the reference standards ensemble copula coupling (ECC) and the Schaake shuffle. When employing statistical postprocessing methods, it is critical to choose appropriate verification data to train and assess the methods. Observation sites are non-homogeneously scattered across the globe, but yield truth data that are independent of the prediction system. Covering the entire Earth on a grid, (re)analyses combine past forecasts and observations and are available on the same spatio-temporal resolution as the forecsting model. Here, we contrast the benefits of postprocessing at observation sites to postprocessing against gridded reanalyses. In a case study, we apply EMOS to 2m ECMWF temperature forecasts, trained and assessed using both verification sets.