|Working Group I: The Scientific Basis|
|Other reports in this collection|
10.6.2 Methodological Options
Weather generators are statistical models of observed sequences of weather variables (Wilks and Wilby, 1999). Most of them focus on the daily time-scale, as required by many impact models, but sub-daily models are also available (e.g., Katz and Parlange, 1995). Various types of daily weather generators are available, based on the approach to modelling daily precipitation occurrence, and usually these rely on stochastic processes. Two of the more common are the Markov chain approach (e.g., Richardson, 1981; Hughes et al., 1993, Lettenmaier, 1995; Hughes et al., 1999; Bellone et al., 2000) and the spell length approach (Roldan and Woolhiser, 1982; Racksko et al., 1991; Wilks, 1999a). The adequacy of the stochastic models analysed in these studies varied with the climate characteristics of the locations. For example, Wilks (1999a) found the first-order Markov model to be adequate for the central and eastern USA, but that spell length models performed better in the western USA. An alternative approach would include stochastic mechanisms of storm arrivals able to produce the clustering found in observed sequences (e.g., Smith and Karr, 1985; Foufoula-Georgiou and Lettenmeier, 1986; Gupta and Waymrie, 1991; Cowpertwait and O’Connel, 1997; O’Connell, 1999).
In addition to statistical models of precipitation frequency and intensity, weather generators usually produce time-series of other variables, most commonly maximum and minimum temperature, and solar radiation. Others also include additional variables such as relative humidity and wind speed (Wallis and Griffiths, 1997; Parlange and Katz, 2000.) The most common means of including variables other than precipitation is to condition them on the occurrence of precipitation (Richardson, 1981), most often via a multiple variable first-order autoregressive process (Perica and Foufoula-Georgiou, 1996a,b; Wilks, 1999b). The parameters of the weather generator can be conditioned upon a large-scale state (see Katz and Parlange, 1996; Wilby, 1998; Charles et al., 1999a), or relationships between large-scale parameter sets and local-scale parameters can be developed (Wilks, 1999b).
The more common transfer functions are derived from regression-like techniques
or piecewise linear or non-linear interpolations. The simplest approach is to
build multiple regression models with free atmosphere grid-cell values as predictors
for surface variables such as local temperatures (e.g., Sailor and Li, 1999).
Other regression models have used fields of spatially distributed variables
(e.g., D. Chen et al., 1999), principal components of geopotential height fields
(e.g., Hewitson and Crane, 1992), Canonical Correlation Analysis (CCA) and a
variant termed redundancy analysis (WASA, 1998) and Singular Value Decomposition
(e.g., von Storch and Zwiers, 1999).
An alternative to linear regression is piecewise linear or non-linear interpolation (Brandsma and Buishand, 1997; Buishand and Brandsma, 1999), for example, the “kriging” tools from geostatistics (Biau et al., 1999). One application of this approach is a non-linear model of snow cover duration in Austria derived from European mean temperature and altitude (Hantel et al., 1999). An alternative approach is based on Artificial Neural Networks (ANNs) that allow the fit of a more general class of statistical model (Hewitson and Crane, 1996; Trigo and Palutikof, 1999). For example, Crane and Hewitson (1998) apply ANN downscaling to GCM data in a climate change application over the west coast of the USA using atmospheric circulation and humidity as predictors to represent the climate change signal. The approach was shown to accurately capture the local climate as a function of atmospheric forcing. In application to GCM data, the regional results revealed significant differences from the co-located GCM grid cell, e.g., a significant summer increase in precipitation in the downscaled data (Figure 10.16).
Other reports in this collection