Abstract

Predicting rainfall and temperatures on a sub-seasonal scale is a challenging task for meteorologists due to the loss of memory of initial atmospheric conditions, impacting forecast accuracy compared to short-range and seasonal outlooks. Accurate sub-seasonal predictions aid tactical adjustments ... Show morePredicting rainfall and temperatures on a sub-seasonal scale is a challenging task for meteorologists due to the loss of memory of initial atmospheric conditions, impacting forecast accuracy compared to short-range and seasonal outlooks. Accurate sub-seasonal predictions aid tactical adjustments to strategic decisions based on seasonal forecasts and facilitate timely reviews of prevailing seasonal conditions. In September 2020, NOAA NCEP implemented GEFSv12 to support stakeholders for sub-seasonal forecasts and hydrological applications. Consistent GEFSv12 reforecast data for 2000-2019 are initialized at 00 UTC once per day out to 16 days with 5 ensembles except on Wednesdays when the integration extended to 35 days with 11 members. Analyzing data size’s impact on deep-learning weather models is crucial for generalization, overfitting prevention, resource optimization, and guiding computational decisions. This study uses GEFSv12 Reforecast data to investigate sample size effects (15, 10, 5, and 3 years training periods) on deep learning weather prediction models on sub-seasonal rainfall and temperature (2015-2019) extreme predictions over CONUS. Show less