Better Utilization of Ensembles in Operations Through Clustering: An R2O Success Story

Operational model ensembles help forecasters determine the uncertainties of various weather phenomena, which need to be communicated to the public and various stakeholders. However, the standard ensemble mean and probability plots do not help forecasters communicate potential scenarios from the model ensembles. One solution to this issue is to cluster ensemble members into groups, which reduces the amount of data that needs to be assessed while providing the forecaster with a representative view of the ensemble data.

Using a NOAA-CSTAR collaboration over the Northeast U.S. Stony Brook University developed a clustering approach several years ago that utilizes ensemble-spread characteristics as represented by the leading Empirical Orthogonal Function (EOF) patterns of the spread. The EOFs are calculated across the model ensemble member dimension (GEFS+CMC+EC ensembles or 100 members), with the resulting modes showing dominant patterns of the difference between individual ensemble members and the model ensemble mean. The leading principal components (PCs) are the projections of the dominant EOF patterns onto the difference between each of the ensemble members and the ensemble mean. Since the leading PCs and the associated EOF patterns contain the main uncertainty information across the entire ensemble, they are used to perform cluster analysis. The first and second PCs for the ensemble members are used as input into a fuzzy clustering routine, which is utilized to group ensemble members with similar forecast scenarios. This effort was expanded to the Weather Prediction Center (WPC) in 2020 with the help of Joint Technology Transfer Initiative (JTTI) funding to include clustering of 500 hPa height to day 7 for three regions across CONUS and also Alaska. This WPC effort began as a basic clustering webpage of a few variables (2-m temperature, 24-h QPF, SLP) using these 500-hPa clusters, but given its success, Western Region STID and WPC engaged in a joint project in 2021 to enhance visualization of the cluster datasets. The resultant site - Dynamic Ensemble-based Scenarios for IDSS (DESI) offers numerous ways to interrogate the ensemble dataset via clusters. For example, one can create weather scenarios by examining the probability of each cluster spatially, and display many more variables (precipitable water, surface CAPE, cloud cover, etc) using histograms, plume diagrams, and box-and-whisker plots. DESI is now available at all NWS forecast offices.

This presentation will review the clustering approach with a storm event and highlight the new DESI software that forecasters have access to. Some of the benefits to forecasting and stakeholder communication will be highlighted as well as some potential future directions based on forecaster survey feedback and recent research into various clustering approaches and variables used for clustering.