The Statistical Analogue Resampling Scheme (STARS) is based on a modeling concept of
Werner and Gerstengarbe (1997). The model uses a conditional resampling technique
to create a simulation time series from daily observations. Unlike other time
series generators (such as stochastic weather generators) STARS only needs a linear
regression specification of a single variable as the target condition for the
resampling. Since its first implementation the algorithm was further extended in
order to allow for a spatially distributed trend signal, to preserve the seasonal
cycle and the autocorrelation of the observation time series (Orlovsky, 2007; Orlovsky et al., 2008).
This evolved version was successfully used in several climate impact studies.
However a detaild evaluation of the simulations revealed two fundamental weaknesses of the
utilized resampling technique. 1. The restriction of the resampling condition on
a single individual variable can lead to a misinterpretation of the change signal
of other variables when the model is applied to a mulvariate time series.
(F. Wechsung and M. Wechsung, 2014). As one example, the short-term correlations
between precipitation and temperature (cooling of the near-surface air layer
after a rainfall event) can be misinterpreted as a climatic change signal in the
simulation series. 2. The model restricts the linear regression specification to
the annual mean time series, refusing the specification of seasonal varying trends.
To overcome these fundamental weaknesses a redevelopment of the whole algorithm
was done.
The poster discusses the main weaknesses of the earlier model implementation and
the methods applied to overcome these in the new version. Based on the new model
idealized simulations were conducted to illustrate the enhancement. |