An original way to evaluate daily rainfall variability simulated by a regional climate model: the case of South African austral summer rainfall

摘要

We discuss the value of a clustering approach as a tool for evaluating daily rainfall output from climate models. Ascendant hierarchical clustering is used to evaluate how well South African recurrent daily rainfall patterns are simulated during the austral summer (December to February 1970-1971 to 1998-1999). A set of 35-km regional climate simulations, run with the WRF model and driven by the ERA40 reanalysis, is chosen as a case study. Six recurrent patterns are identified and compared to the observed clusters obtained by applying the same methodology to 5352 daily rain gauge records. Two of the WRF clusters describe either a persistent and widespread dryness (65% of the days) or patterns similar to the seasonal mean rainfall gradient (13% of the days). The four remaining WRF clusters (approximate to 20% of the days) are wetter; they describe the weakening, conservation or strengthening of the average rainfall gradient. The WRF cluster rainfall patterns and their associated circulation match the observed clusters rather well, but their frequency of occurrence is greatly overestimated by WRF during dry events, and underestimated for near-normal rainfall conditions. The weak model biases found at the seasonal timescale conceal strongly biased intraseasonal rainfall variability. The WRF-simulated rainfall patterns are then temporally or spatially projected on to the observed clusters. Spatial projection proves to be the more useful of these two approaches in quantifying model skill by assessing both the temporal co-variability between WRF and observations, and the rainfall biases of the model with or without temporal dephasing. The WRF model simulates transient rainfall activity partially out of phase with observations, which induces large rainfall biases when temporal dephasing is not removed. Rainfall biases are significantly reduced, however, when temporal dephasing is removed. The clustering approach therefore proves its efficiency to highlight climate model strengths and deficiencies.