Identification of Highly Predictable Flow Elements for Spatial Filtering of Medium- and Extended-Range Numerical Forecasts

Abstract

To determine if some flow components are systematically forecast more accurately than others, 990 wintertime medium-range forecasts made at the European Centre for Medium-Range Weather Forecasts (ECMWF) are examined. It is found that forecasts skill of 500-mb extratropical large-scale heights tends to be a function of empirical orthogonal function (EOF) index, with those components that project onto the leading EOFs being markedly better forecast than the components that project onto the hailing EOFS. This is true for instantaneous forecasts of as long as 10 days? duration. Furthermore, by answering the question, Of all possible structures which structure on average is most accurately forecast? the potential for constructing a basis that is even more adept than EOFs at distinguishing well-forecast from poorly forecast flow elements is shown. Similarly, it is found that 10-day average ECMWF forecasts, as well as 29-day average forecasts produced by a general circulation model at the National Center for Atmospheric Research, can be effectively decomposed into components that on average are either easy or difficult to predict. Using the ability to make such a decomposition, spatial filters are designed that remove those components that are usually poorly forecast. These filters can markedly improve the skill scores of medium-and extended-range forecasts, though the more effective filters substantially reduce the explained variance of the forecasts. The filters are especially effective in the extended range. For example, one filter, by removing 43% of the variance, can improve the average anomaly correlation of verified 29-day average forecasts to 0.66 from an unfiltered skill of 0.46. Such filters are proposed as a means of enhancing the utility of extended-range forecasts.