Generating weather for climate impact assessment on lakes

Abstract

Lakes are driven in part by weather and they are affected by climate. The complex physical and biological processes that govern their behaviour makes it impossible to estimate their reactions to changed climatic conditions in a trivial manner. This work presents two weather generators (WGs), which are stochastic abstractions of observed meteorological time series, as tools for climate impact assessment on lakes. They enable to define "what if"-scenarios based on prescribed temperature changes, that are propagated to the rest of the generated variables, thus producing statistically balanced time series. For propagating the changes, linear and non-linear models of dependence were explored. The linear model consists of a Vector-Autoregressive process and the non-linear of a pair-wise copula construction method. Both methods are enhanced by phase randomization, a technique that uses the Fourier transform to generate "surrogate data" and helps to maintain longer-term dependencies in this context. The thesis also proposes a novel way to generate precipitation which works by estimating dryness probability from the state of non-precipitation variables and transforming the result to construct a time series without dry gaps. An upside to this treatment of precipitation is that it does not require a precipitation occurrence model and no different parameterizations for wet and dry states for the non-precipitation variables. The WGs were tested for their ability to extrapolate from colder towards warmer weather. While the more complex, copula-based method performed better than the simpler, linear method, it is shown that only relying on statistical relationships can mis-project the changes in dependent variables that accompany temperature increases. The two WGs are contrasted with a non-parametric K-Nearest Neighbors resampler, highlighting differences between those approaches and underlining specific weaknesses. The parametric WGs overestimate spread, while the resampler lacks variability resulting from a tendency to choose central values in both a uni- and multivariate sense.