Improving the CFSv2 Seasonal Precipitation Forecasts across the U.S. by Combining Weather Regimes and Gaussian Mixture Models

Abstract

While seasonal climate forecasts have major socio-economic impacts, current forecast products, especially those for precipitation, are not yet reliable for forecasters and decision makers. Here we developed a novel statistical-dynamical hybrid model for precipitation by applying Weather Regimes (WRs) and Gaussian Mixture Models (WR-GMM) to the National Oceanic and Atmospheric Administration’s Climate Forecast System Version 2 (CFSv2) precipitation forecasts across the continental United States. Instead of directly forecasting precipitation, WR-GMM uses observed precipitation from synoptic patterns similar to the future CFSv2 forecast. Traditionally K-means has been used to classify daily synoptic patterns into individual WRs, but the new GMM approach allows multiple WRs to be represented for the same day. The novel WR-GMM forecast model is trained on daily Climate Forecast System Reanalysis (CFSR) geopotential height and observed precipitation data during a 1981-2010 period, and verified for years 2011-2022. Overall, the WR-GMM method outperforms the CFSv2 ensemble forecast precipitation in terms of root mean square error, and Pearson correlation coefficient for lead months 1 through 4. Previous studies have used global climate models to forecast WRs in the Pacific and Mediterranean regions, usually with an emphasis on winter months, but the WR-GMM model is the first of its kind that promises great untapped potential to improve precipitation forecasts produced by CFSv2 across the continental United States.