CHAC: a weather pattern classification system for regional climate downscaling of daily precipitation

Abstract

A weather pattern clustering method is applied and calibrated to Argentinean daily weather stations in order to predict daily precipitation data. The clustering technique is based on k-means and is applied to a set of 17 atmospheric variables from the ERA-40 reanalysis covering the period 1979–1999. The set of atmospheric variables represent the different components of the atmosphere (dynamical, thermal and moisture). Different sensitivity tests are applied to optimize (1) the number of observations (weather patterns) per cluster, (2) the spatial domain size of the weather pattern around the station and (3) the number of members of the ensembles. All the sensitivity tests are compared using the ROC (Relative Operating Characteristic) Skill Score (RSS) derived from the ROC curve used to assess the performance of a predictive system. First, we found the number of observations per cluster to be optimum for values larger than 39. Second, the spatial domain size (~4° × 4°) was found to be closer to a local scale than to a synoptic scale, certainly due to a dominant role of the moisture components in the optimization of the transfer function. Indeed, when reducing the set of variables to the subset of dynamical variables, the predictive skill of the method is significantly reduced, but at the same time the domain size must be increased. A potential improvement of the method may therefore be to consider different domains for dynamical and non-dynamical variables. Third, the number of members per ensembles of simulations was estimated to be always two to three times larger than the mean number of observations per cluster (meaning that at least all the observed weather patterns are selected by one member). The skill of the statistical method to predict daily precipitation is found to be relatively homogeneous all over the country for different thresholds of precipitation.