For realistic assessment of the impacts of climate change, it is often necessary to downscale the outputs from climate models to the fine spatial scales that are relevant in determining system response. Statistical downscaling techniques range in complexity from rather simple techniques that easy to apply but limited in what they can achieve, through to more complicated methods that require more technical skill from the user but are potentially suitable for use in a much wider range of applications. Methods based on generalised linear models (GLMs) fall in the latter category, and have been shown in intercomparison studies^1,2 to perform comparably with other state-of-the-art techniques in reproducing a wide range of “weather” properties that are of interest in impacts applications. To date however, most applications of GLMs in this context have been univariate with the exception of Furrer and Katz (2007)³  who considered precipitation and temperature simultaneously.

This talk will present Rglimclim which is a multivariate, multisite weather generator based on GLMs. It is based on the Glimclim software package that has been widely used for univariate weather generation in the UK, Australia and elsewhere, but has been updated to allow for the simultaneous generation of multiple weather variables. A user interface in R (www.R-project.org) has also been written to simplify the processes of model fitting, checking and simulation. The flexibility of GLMs allows the structure of the generator to be determined by consideration of the physical relationships between variables, rather than by statistical convenience. This will be illustrated with an example of a weather generator constructed for the catchment of the river Thames in England, which has been guided by the structure of numerical weather prediction models.