Regime Dependence of Predictability of Springtime Medium-Range Severe Weather over CONUS within Two GEFS Reforecast Ensembles

The practical predictability of severe convective weather events over the contiguous United States (CONUS) can vary dramatically across seasons, regions, and synoptic-scale regimes. In prior work, regimes conducive for convective storms were identified by training self-organizing maps (SOMs) to identify distinct patterns of shear and instability across the CONUS and then using the SOMs to cluster events. Here, we apply a similar technique to compare the skill of two different GEFS reforecast ensembles across the set of SOM identified regimes. The first GEFS ensemble is the original version that mimics the operational GEFS implemented in operations in 2012 (GEFS-old). The second GEFS ensemble is the latest reforecast that uses the GEFSv12 configuration that was implemented in September 2020 (GEFS-new). Specifically, we focus on Spring-time (March - May) Day 1 – 10 forecasts of CAPESHR (i.e., the product of convective available potential energy and 0 - 6 km wind shear) generated by GEFS-old and GEFS-new. The SOMs were trained with daily ERA-Interim CAPESHR reanalyses during the 1990-2016 period and lead to distinct clusters of events. Then, the Day 1-10 ensemble mean GEFS forecasts valid each day within each node were aggregated and verified to determine regime predictability using the equitable threat score (ETS) for a set of CAPESHR thresholds (e.g., ≥ 10,000). Only forecasts of events within the 2000–2016 period were aggregated, since this is the overlapping period between GEFS-old and GEFS-new.

In aggregate, the GEFS-new generated CAPESHR forecasts with larger ETS than GEFS-old. The ETS differences between the two forecast datasets were largest between Days 3–8. At short (Days 1–2) and long (Days 9–10) lead-times the two forecast datasets produced similar values of ETS. Overall, the patterns of skill among the SOM identified regimes are similar in GEFS-old and GEFS-new, with regimes that produced higher than average skill in the GEFS-old also leading to higher than average skill in GEFS-new. This provides evidence that the identification of predictable and unpredictable regimes is robust across model version changes. GEFS-old was better (i.e., had larger CAPESHR ETS) than GEFS-new across very few forecast hours, while in several regimes GEFS-new had much larger (>0.1) values of ETS. While using SOMs to identify enhanced severe weather predictability regimes and to perform regime-based forecast verification is promising, further interpretation is required to identify the dynamical mechanisms responsible for extended or reduced predictability among the regimes, as well as the reasons for the differences in the verification metrics between the GEFS versions among the regimes.