Applying clustering and ensemble prediction concepts to consensus Tropical Cyclone Track Forecasting

Several groups are developing ensemble prediction techniques for tropical cyclone track forecasting. J. Goerss of the Naval Research Laboratory has developed a simplified, economical (free) consensus of up to three global or two regional model track predictions that has more skill in the average than any one of the individual dynamical models. Even though Goerss "ensemble" contains only a small number of members, each model is a sophisticated baroclinic model with small systematic errors, so that both perturbed initial conditions and "perturbed models" are included in the Goerss consensus approach. A simple extension of the Goerss technique has been to translate the 0600 and 1800 UTC regional model tracks to align with the 1200 and 000 UTC global model tracks, and vice versa. Thus, all five model tracks (if available) become potential members of a consensus track forecast, and a five-member consensus is shown to be superior to a three- or two-member consensus.

A sample of western North Pacific tropical cyclone tracks are used to calculate the consensus forecast track error (FTE) and four situations are considered: small spread and small error; small spread and large error; large spread and small error; and large spread and large error. A great reduction in FTE would be possible if one could always pick the model track with the smallest error, but no method has been developed to assist the forecaster in accomplishing this task. Rather, the approach proposed here is to reject erroneous tracks prior to calculating the consensus, which is called a selective consensus. Such an approach is shown to usually reduce the average FTE over a simple numerical consensus using all of the models all of the time.

This improvement of a selective consensus over a numerical consensus provides the motivation for studies of the sources of large dynamical model track errors. Such work has been completed for the 1997 western North Pacific season, and being extended to later seasons. Other studies are being completed for Atlantic and for Southern Hemisphere tropical cyclones to determine if the same conceptual models for detecting large errors apply in those regions. Reports of these studies are companion papers for the conference.