9524770 Guttorp The problem of subgrid scale variability in general circulation models, and the generally poor quality of the precipitation part of such models, indicate the need for precipitation models based on large-scale atmospheric processes. A realistic stochastic model of precipitation can be based on meteorologically homogeneous weather states, each driving a simple stochastic model of precipitation. In order to determine appropriate weather states, a variant of canonical correlation analysis, appropriate for dependent data, is developed. Statistical theory allowing a space-time decomposition of atmospheric fields with attendant standard error assessment is developed. When applied to the leading canonical variates, such a method can isolate spatial and temporal scales of interest. A slightly different approach uses a hidden Markov model, driven by atmospheric data in a less explicit form than expressed above. This model is fairly accurate on relatively small spatial scales, and for temporally homogeneous parts of the year. This hidden Markov model approach is extended to a seasonal model, and its performance compared with the weather state approach, as well as with general meteorological forecast models. %%% The general circulation models used to assess climate change have relatively poor performance when it comes to precipitation. Furthermore, in order to assess the effect on regional hydrology of climate changes, models with a finer resolution than the circulation models are needed, since the hydrologic processes usually operate on a much smaller scale than the large-scale atmospheric processes that dominate the global climate models. Rather than developing precise deterministic models of rainfall, the researchers build a partly probability-based class of models, and compare it with meteorological forecast models and applied to both weather data and circulation model outputs. ***
