Assessing the Spread/Error Relationship for East Coast Winter Cyclones

Ensemble spread provides forecasters a tool to assess the distribution (e.g., standard deviation) of a forecast variable. When the spread is large, there is significant uncertainty in the forecast and the potential for large error. Thus, a useful metric to quantify the dispersion of an ensemble is ratio of the standardized spread and error. Although ensemble spread is widely used by forecasters, the spread-error relationship has not been evaluated for many important weather phenomena in operational ensembles. This project aims to look at the relationship between the ensemble spread of the Global Ensemble Forecast System (GEFS) and the sea-level pressure error in the ensemble for winter (DJF) cyclones on the east coast of the United States. There also exists opportunities for better visualization of ensemble spread output, using an approach similar to the operational Situational Awareness Table (SAT) run by NOAA’s Weather Prediction Center. In particular, using a set of similar anomalous events from a reforecast climatological range, one can obtain a spread anomaly graphic to help the forecaster determine if the spread of an event is greater or less than previous events of similar magnitude.

To address these above objectives, 131 mid-latitude cyclones within DJF between 2007 and 2014 were identified located within a bounding box of 33N, 82W, and 48N, 62W for the 00 run cycle. The ensemble spread (standard deviation) was calculated using 20 perturbed members of the GEFS at a 1 degree by 1 degree resolution. The model errors were calculated using the ERA-Interim interpolated to 1 degree by 1 degree for forecast hours 24 to 168 (day 1 to 7) of each analyzed cyclone. A cyclone relative approach helped to provide more context regarding the structure of the ensemble spread and error, regardless of geographic location. To assess the spread and error about a cyclone, tracked data from the Hodges tracking algorithm as well as binary classification methods were used to identify centers of analyzed cyclones. The data were regridded to an 11 x 11 (latitude x longitude) array, centered about each cyclone, for each forecast hour. Preliminary results suggest that for all lead times there is reasonable spread-error correlation for weak to moderate spreads (r ~0.35), however the error/spread ratio is ~0.5 suggesting under-dispersion. Shorter lead times indicate a stronger relationship (r ~0.73). For longer lead times there is little correlation (r ~0.16). Furthermore, in general the location of spread and error is not co-located, with a larger discrepancy of spread to the north and east of the cyclone center, a magnitude bias with a larger center error, and a larger error to the south and west of the cyclone center. This presentation will also highlight how one can use historical spread and anomaly information from the SAT approach to derive a spread anomaly tool useful for forecasters.