Evaluation of Statistical Downscaling Methods for Regional Precipitation Over CONUS Using a K-Means Weather Typing Algorithm

Projecting the impacts of future climate change is crucially important to water managers planning to mitigate those impacts. Unfortunately, global climate model output is generally too coarse to provide the detailed, localized information necessary to prepare for those impacts, and the extreme computational cost to run more finely gridded models for a long enough time period can be prohibitive. Statistical downscaling methods are a relatively inexpensive way to refine those climate model outputs into a more useful form. Here, we assess multiple downscaling methods, including, LOcally Constructed Analogs (LOCA), the Ensemble-Generalized Analog Regression (En-GARD) model, and the Intermediate Complexity Atmospheric Research model (ICAR). These methods represent direct rescaling of climate model precipitation (LOCA), a statistical approach that incorporates meteorological information (En-GARD), and a simplified atmospheric model (ICAR). These methods have been applied to both historical and future CMIP5 precipitation, and the climate change signals can vary substantially between methods.

Here, we apply a K-means clustering algorithm to determine weather types for a dataset combining six CMIP5 ESMs and ERA5 over nine regions of CONUS over the period 1980-2005. The choice of variables differs for each region, with between two and seven variables chosen from the literature for each application. Because we use the combined CMIP5-ERA5 dataset to categorize rainy days into each weather type, this categorization can be applied to any downscaled data from that model (or from observations) that retains day-to-day weather information. The fidelity of precipitation for each weather type within each region is then evaluated for the downscaling methods through various statistical comparisons with observational datasets. By determining regions and meteorological situations where some or all the downscaling methods evaluated here perform well or poorly, we provide actionable information to decision makers as to which downscaling dataset or method is most appropriate for the application at hand.