Clustering Methods for Statistical Downscaling in Short-Range Weather Forecasts

Abstract

In this paper an application of clustering algorithms for statistical downscaling in short-range weather forecasts is presented. The advantages of this technique compared with standard nearest-neighbors analog methods are described both in terms of computational efficiency and forecast skill. Some validation results of daily precipitation and maximum wind speed operative downscaling (lead time 1?5 days) on a network of 100 stations in the Iberian Peninsula are reported for the period 1998?99. These results indicate that the weighting clustering method introduced in this paper clearly outperforms standard analog techniques for infrequent, or extreme, events (precipitation > 20 mm; wind > 80 km h?1). Outputs of an operative circulation model on different local-area or large-scale grids are considered to characterize the atmospheric circulation patterns, and the skill of both alternatives is compared.