Evaluations of three high-resolution dynamical downscaled ensembles over North America

Abstract:

Six 12 kilometer resolution WRF simulations are performed driven by three GCM outputs --- CCSM4, GFDL, and HadGEM --- in 10-yr or 15-yr historical period. The model fidelity is measured in terms of correlations, RMSEs, and probability distribution functions. The model evaluations are conducted over 20 subregions for North America. The 3D and surface variables considered in this study are precipitation, 2m-temperature, tmax/tmin, relative humidity, geopotential height, u-wind/v-wind, and sea-level pressure. We aim to find which downscaled GCM run showed the best skill over each subregion for all the variables. Depending on the variable and type of metric being examined, the output from each historical run will be primarily compared to North American Regional Reanalysis (NARR) and Variable Infiltration Capacity (VIC) data for the statistical analysis. It is important to both evaluate which high resolution ensemble member captures the climatological mean of the variables as well as to determine which run does the best at observing the extremes values, such as maximum/minimum temperature and large precipitation events. High RMSE values frequently occur because of errors in either the seasonal mean, which could indicate a specific bias in the model, or a failure to capture the tails of the PDF correctly, which indicates the model had trouble capturing the seasonal variability. This will give us valuable information at which GCM is the most effective at simulating historical periods using dynamic downscaling over specific regions of the United States and give us added confidence of their ability to forecast future time periods based on model performance.