Tropical Cyclone Track Forecasts Using an Ensemble of Dynamical Models

TLDR

Tropical cyclone track forecasts from regional and global models are relatively independent, implying that a simple ensemble average may outperform individual models. The study compared forecast errors of a simple ensemble average of three models for the 1995–96 Atlantic hurricane seasons, and either three global or two regional models for the 1997 western North Pacific, against the errors of the individual models. The ensemble reduced mean 24‑, 48‑, and 72‑hour track errors by 16–23 %, lowered the 95th‑percentile error by 14–23 %, decreased error standard deviation, and provided a useful confidence spread, with comparable gains observed in the western North Pacific.

Abstract

The relative independence of the tropical cyclone track forecasts produced by regional and global numerical weather prediction models suggests that a simple ensemble average or consensus forecast derived from a combination of these models may be more accurate, on average, than the forecasts of the individual models. Forecast errors of a simple ensemble average of three models for the 1995–96 Atlantic hurricane seasons, and either three global models or two regional models for the western North Pacific during 1997, were compared with errors of the individual models. For the Atlantic, the mean errors for the joint ensemble were 120 km at 24 h, 194 km at 48 h, and 266 km at 72 h, which represent improvements of 16%, 20%, and 23% with respect to the best of the individual models. The joint ensemble also resulted in reduction in the standard deviation of the forecast error. The 95th percentile of forecast error for the ensemble was reduced 19%, 14%, and 23% with respect to the best of the individual models. The spread of the ensemble forecast was found to possess some potential for use by forecasters as a measure of confidence in the ensemble forecast. Similar results were found for the western North Pacific.