Evaluating Methods for Generating Ensemble Members in the National Hurricane Center Wind Speed Probability (WSP) Model

The National Hurricane Center (NHC) Wind Speed Probability (WSP) model provides the foundation for several public-facing products that inform wind-related hazards associated with tropical cyclones (TCs). WSP comprises 1000 ensemble members (realizations) that provide the basis for probabilistic forecasts of 34-knot, 50-knot, and 64-knot winds. Although the WSP probabilistic forecasts are skillful, the underlying ensemble members—generated by perturbing track location, intensity, and wind radii parameters—often exhibit unrealistic temporal variability that results in jagged, "unsmooth" time series. The unrealistic, individual ensemble member behavior from the WSP model precludes the possibility of driving other TC-related hazard forecasts such as storm surge and ocean wave probabilities.

In this study, two methods are evaluated for smoothing the WSP ensemble members. The first method introduces additional predictors for the along- and cross-track errors in the ordinary least squares regression formulations to assess their impact on the fraction of variance explained. The second applies a post-processing method developed by Kepert (2014) to smooth the perturbations of track location, intensity, and wind radii. Kepert empirically attributed the jagged ensemble member track forecasts to a discontinuity in the first derivative of the perturbation correlation matrix. To alleviate this issue, Kepert modeled two key features—the perturbation correlations increase with time and the perturbation correlation function has fat tails. The former is addressed with a scaled log-transformation of the time variable that introduces two new parameters, and the latter is addressed by imposing a second-order autoregressive correlation function. The Kepert method is applied to the WSP ensemble member forecasts of track location, intensity, and wind radii for a set of 1000 members generated from the 2019–2023 climatology. Furthermore, individual ensemble member forecasts of intensity and wind radii will be applied to a regression model for generating forecasts of the radius of maximum winds.

A measure of smoothness is calculated for each ensemble member forecast parameter and compared with the OFCL forecast to quantify the degree of smoothing by the method. The moments of the distributions underlying the original and post-processed perturbation covariance matrices are contrasted to assess the conservative properties of the method, and ensemble verification measures are applied to the original and post-processed ensemble members to quantify dispersion and skill . Furthermore, two approaches to perturbing the wind radii are evaluated as candidates for an improved methodology in the WSP model, including an approach that has been applied to generating wind swaths for the Joint Typhoon Warning Center. The smoothed ensemble members will be discussed in the context of providing forcing to a hydrodynamical model for generating probabilistic storm surge forecasts.

The current WSP model perturbations are determined from the official track and intensity forecast error distributions and those of the simple wind radii model from the past five years. The track perturbations also have a weak dependence on the current forecast situation through the Goerss Predicted Consensus Error (GPCE), which uses the spread of a few dynamical track models and other forecast parameters to estimate the track forecast error. New methods for estimating forecast uncertainty using machine learning and dynamical model ensemble systems are being developed. Preliminary results incorporating these new methods into WSP will be presented.