Stringent Evaluation of CMIP5 Precipitations and Development of Optimal Weighting Methods

Abstract:

Control simulations of precipitation by 23 CMIP5 coupled ocean-atmosphere general circulation models (CGCMs) have been evaluated in comparison with the Global Precipitation Climatology Project (GPCP) monthly precipitation data for the period of 1979-2005. Instead of looking at the first two moment statistics, mean and standard deviation, the performance of the CMIP5 models in reproducing the physical evolutions in the GPCP data by employing cyclostationary EOF (CSEOF) analysis. Each CSEOF loading vector (CSLV) and corresponding principal component (PC) time series derived from individual CMIP5 models are compared with those of the GPCP data for the first 10 modes. Only the first CSEOF mode (the seasonal cycle) was simulated with a reasonable accuracy; the average pattern correlation is 0.675 (0.446-0.764) and the average temporal correlation is 0.999 (0.997-0.999). The performance of the model quickly diminishes as the mode number increases, with the best pattern correlation less than 0.2 for modes higher than 4.

In an effort to improve the CMIP5 precipitation data for practical uses, optimal weighting scheme is developed in CSEOF space. In this new scheme, optimal weighting is conducted for each physical mode. For validation, the Jackknife method is used by filling a missing year in GPCP data using the 23 CMIP5 model datasets. This procedure is repeated until all of the 23 years are covered. Correlation and RMSE are significantly improved in comparison with those of the best CMIP5 models. A similar experiment is conducted using a conventional optimal weighting at each grid point; the performance of an optimal weighting scheme at each grid point is not better than that of the best CMIP5 models, suggesting that the multi-model optimal ensemble in CSEOF space is much better than that at each grid point.