A novel scheme to derive optimized circulation pattern classifications for downscaling and forecast purposes

Many synoptic classifications that have been introduced share the problem of a large within-type variability. Therefore, there is still the need for refinement, especially when aiming to downscale climate model outputs. The objective scheme proposed here has been developed to derive circulation pattern classifications that optimally distinguish between different values of regional weather elements. In comparison with methods that similarly take local and large-scale information into consideration (such as classification and regression trees [CART]), computational demand is relatively low and human input is practically unwarranted. Recognition of individual cases is based upon composites of the large-scale conditions of several different meteorological parameters such as geopotential heights, temperatures and relative humidities. For each local weather element a so-called screening discriminant analysis is conducted selecting those large-scale synoptic variables yielding optimal selectivity. The procedure is basically analogous to a stepwise screening regression. At each stage the predictor field minimizing the RMSE between forecasts and observations together with the previously chosen predictor fields is selected. The procedure is terminated either when a maximal number of predictor fields is reached or when the inclusion of another field does not result in a further decrease of the RMSE. Similarity between mean large-scale conditions and the individual cases that are to be assigned is determined by means of a slightly modified version of the widespread Euclidean distance. Using daily data from 51 climate stations located in the Elbe river catchment in the northeastern part of Germany and the western part of the Czech Republic, the attainable reduction of climatological variances was calculated for a variety of weather elements. Results prove the ability of the scheme to effectively discriminate large-scale circulation patterns with respect to local weather parameters, especially temperatures (skill scores of approx. 80%). When solely accurate discrimination is of interest, a fuzzy variant can be utilized producing even better results. It yields the percentage of each circulation type for a specific case, i.e. day. However, in this case physical insights can not be obtained.