Evaluation of a New Ensemble Mean Weighting Technique for the 2015 GFDL Hurricane Ensemble Forecast System

An ensemble forecast system based on the 2015 operational Geophysical Fluid Dynamics Laboratory (GFDL) hurricane model has been refined as part of the regional modeling effort for the Hurricane Forecast Improvement Program (HFIP). In this paper, we briefly describe the design of the ensemble forecast system and present results from the 2015 Atlantic and Eastern Pacific hurricane seasons. The primary focus of this paper is to introduce and evaluate two new ensemble mean formulation techniques that use (1) linear regression to remove systematic biases in track and intensity forecasts and (2) multilinear regression to weight individual ensemble members according to their relative performance in a four-year retrospective data set.

The GFDL hurricane ensemble has been running as part of the HFIP program since 2010 and provides real-time guidance on track and intensity to the National Hurricane Center (NHC) and Joint Typhoon Warning Center (JTWC). As in the previous years, most perturbations are applied by modifying the observed input parameters that define important characteristics of the bogus initial vortex structure (e.g., size and maximum tangential wind speed) and its local environment (e.g., sea-surface temperatures and vertical moisture profile). New member pairs were included in 2013 and 2015 that modified the model surface physics and planetary boundary layer depth, respectively. The goal of configuring such an ensemble has been primarily to create large intensity spread in the model's 5-day forecasts while also producing realistic possibilities of storm track and intensity evolution. Near-future work will be dedicated to developing storm track-specific perturbations with the intention of creating much higher realistic spread in our ensemble track forecast guidance.

While efforts still continue to develop a robust weighted ensemble mean technique, a simplified version has been developed and run in near real-time since August 2015. This version is essentially a simple linear average of bias-corrected ensemble member forecasts. Preliminary verifications for this bias-corrected mean show some intensity improvement over the control and uncorrected mean in each of the ocean basins presented herein, but it is evident that this particular method needs reformulation (e.g., explore advanced multi-linear regression techniques suitable for non-normal distributions). Interestingly, however, bias-corrected mean track guidance has shown solid improvement over the bias-corrected control and uncorrected mean in the Atlantic and Eastern Pacific at all forecast lead times in 2015.