Optimizing Downscaling Strategy through Selective Culling of NOAA Operational Forecast Ensemble for Cost-Effective Prediction

NOAA’s operational forecasts using the Climate Forecast System version 2 (CFSv2) present a prime candidate for driving dynamic downscaling seasonal predictions at finer resolutions. By employing the regional Climate-Weather Research and Forecasting model (CWRF), we not only achieve skill-enhanced downscaling climate predictions but also simultaneously provide agricultural forecasts by integrating its coupled dynamic crop growth module. However, downscaling proves to be computationally intensive, particularly challenging when employing the CWRF multi-physics ensemble approach. This challenge is magnified by utilizing almost 400,000 days of CFSv2 ensemble forecasts for a single year. As a result, generating downscaling predictions from the complete driving ensemble becomes formidable. Consequently, there arises a pressing necessity to optimize the downscaling strategy through the selective culling of the driving ensemble forecasts. We evaluated the efficacy of all possible subsets of the CFSv2 full ensemble forecasts in capturing spatial patterns and interannual anomalies of 500-hPa geopotential height, surface temperature, and precipitation. We have identified three culling strategies that effectively curate subsets showcasing superior performance compared to the complete ensemble. The initial strategy involves exclusively selecting forecasts initialized at 06Z, thereby reducing the ensemble size to one fourth. The second strategy entails excluding forecasts that lie within the spin-up period and exhibit weaker skills. The third strategy consists of choosing representative forecasts by start dates,further trimming the ensemble size by approximately a factor of five. By scrutinizing the trade-offs between predictive accuracy and ensemble size stemming from these subset strategies, we aim to glean insights to pave the way for more cost-effective CWRF downscaling prediction of CFSv2 ensemble forecasts.