Ensemble Clustering applied to a Time-lagged Rapid Refresh Forecast System (RRFS) Superensemble: Evaluation during the 2024 Hydrometeorology Testbed Flash Flood and Intense Rainfall (FFaIR) Experiment

Ensemble clustering is an efficient ensemble post-processing approach that distills an ensemble forecast into its prevalent forecast scenarios by grouping similar ensemble members together. This clustering approach has progressed quickly through the R2O pipeline since its original implementation as a NOAA-CSTAR collaboration for forecasting nor’easters. Initially applied to a superensemble of GEFS, CMCE, and ECMWF members, this technique has since been adopted more operationally at the Weather Prediction Center (WPC), and as a centerpiece for the experimental ensemble visualization platform known as the Dynamic Ensemble-based Scenarios for IDSS (DESI).

However, the utility of an ensemble clustering technique for convection-allowing ensemble forecast systems remains largely unexplored, mostly due to the prerequisite of an ensemble system of sizable membership for clusters to prove useful. As such, we applied a k-means clustering methodology to a 28-member time-lagged Rapid Refresh Forecast System ensemble (RRFSe) system for the 2024 Hydrometeorology Testbed (HMT) Flash Flood and Intense Rainfall (FFaIR) experiment. Four clusters were computed daily from the full ensemble 6-h QPF field within a floater domain. Participants had the clustering output at their disposal for forecast purposes, and on the subsequent day, participants were asked to subjectively evaluate the performance of the clusters relative to observed precipitation accumulation and the full ensemble QPF field. 83.7% of responses during the experiment indicated that at least one of the 75th percentile QPF clusters outperformed the full ensemble 75th percentile QPF. When asked if the clusters simply grouped membership by initialization time, participants unanimously responded that different cycle times were well-represented amongst the clusters. This result not only bodes well for viability of the clustering approach on a time-lagged RRFSe, but also suggests that time-lagging could be a viable source of useful ensemble spread in future iterations of the RRFS ensemble system. These results and more insights gained from participants of the 2024 HMT FFaIR experiment will be discussed.