Original paper
Evaluation of a stochastic weather generator in simulating univariate and multivariate climate extremes in different climate zones across Europe
Dabhi, Hetal; Rotach, Mathias W.; Dubrovský, Martin; Oberguggenberger, Michael
Meteorologische Zeitschrift Vol. 30 No. 2 (2021), p. 127 - 151
79 references
published: Apr 22, 2021
published online: Sep 25, 2020
manuscript accepted: Jul 9, 2020
manuscript revision received: Jul 7, 2020
manuscript revision requested: May 26, 2020
manuscript received: Dec 16, 2019
Open Access (paper may be downloaded free of charge)
Abstract
Stochastic weather generators have been increasingly used as downscaling tools for climate change impact assessments. In spite of their widespread use, their potential to simulate climate extremes – especially multivariate extremes – is largely unexplored. The aim of this study is to assess the ability of the Richardson type six-variate weather generator SiSi to simulate the frequency of various univariate as well as multivariate extremes with focus on extremes related to the non-normally distributed weather variables relative humidity and wind speed. A total of 83 sites with different elevation and proximity to each other – thereby defining a European, a country (Austria) and a local (catchment) scale – and diverse climates across Europe are selected. Results show that SiSi is able to simulate univariate and multivariate extremes generally and equally well in all climate zones. The results depend on the nature of the individual variables involved in the extreme events. Among all the extreme events, the weather generator has a tendency to underestimate the extremes related to minimum temperature. The first-order auto-regressive (AR(1)) model used for modeling non-precipitation variables assumes the distribution of variables to be Gaussian. This assumption has been enforced in this study by transforming each non-precipitation variable to a normal distribution, but nevertheless the weather generator consistently underestimates the cold extremes. This is due to the multimodal nature of the distribution of temperature. The AR(1) model is not able to reproduce the multimodality of the distributions. The performance of SiSi does not depend on the climate type of a region or the proximity of sites to one another, rather it depends on the characteristics of a variable at an individual location.
Keywords
statistical downscaling • compound events • WGEN • Markov chain • fire weather index • wet/dry spells