An Error Analysis of Atlantic Tropical Cyclone Track Guidance Models

View More View Less
  • 1 Hurricane Research Division, NOAA/AOML, Miami, Florida
  • | 2 National Hurricane Center, Coral Gables, Florida
  • | 3 Dept. Marine, Earth and Atmospheric Sciences, North Carolina State University, Raleigh, North Carolina
© Get Permissions
Full access

Abstract

Mean track forecast errors over the 6-yr period 1983–88 are compared for four tropical cyclone-track forecast models in use at the National Hurricane Center (NHC). The model types represented are statistical, statistical-dynamical, barctropic-dynamical, and baroclinic-dynamical. The statistical-dynamical model (NHC83) had the smallest mean errors at all forecast periods between 12 and 72 h while the baroclinic-dynamical model (MFM) had the largest errors at 12 and 24 h. NHC83 was the only model which had statistically significant forecast skill relative to the CLIPER model. When the forecasts were stratified by latitude, the MFM had significant skill at 36 and 48 h for the northern storms.

Further analysis of the forecast errors shows that most of the models have a bias towards the southwest. The MFM has a very large westward bias for low-latitude storms which results from a fast-speed bias in that region. NHC83 has relatively small speed and directional baises, while CLIPER and SANBAR have slow biases.

The MFM has been replaced (beginning in 1988) by a new baroclinic-dynamical model, the quasi-Lagrangian model (QLM). A comparison of the MFM and QLM for the 1988 season shows that the QLM has error characteristics similar to the MFM.

A barotropic-dynamical model is used to give some insight into the behavior of dynamical track prediction models. Results show that initial-position errors have a negligible effect on the average track forecast errors except at 12 h. Errors in the estimate of the initial storm motion affect the forecast errors out to 72 h. Several spatial filters are applied to the initial analyses which shows that the model storm track is the most sensitive to very large scales (3000–6000 km wavelengths).

Abstract

Mean track forecast errors over the 6-yr period 1983–88 are compared for four tropical cyclone-track forecast models in use at the National Hurricane Center (NHC). The model types represented are statistical, statistical-dynamical, barctropic-dynamical, and baroclinic-dynamical. The statistical-dynamical model (NHC83) had the smallest mean errors at all forecast periods between 12 and 72 h while the baroclinic-dynamical model (MFM) had the largest errors at 12 and 24 h. NHC83 was the only model which had statistically significant forecast skill relative to the CLIPER model. When the forecasts were stratified by latitude, the MFM had significant skill at 36 and 48 h for the northern storms.

Further analysis of the forecast errors shows that most of the models have a bias towards the southwest. The MFM has a very large westward bias for low-latitude storms which results from a fast-speed bias in that region. NHC83 has relatively small speed and directional baises, while CLIPER and SANBAR have slow biases.

The MFM has been replaced (beginning in 1988) by a new baroclinic-dynamical model, the quasi-Lagrangian model (QLM). A comparison of the MFM and QLM for the 1988 season shows that the QLM has error characteristics similar to the MFM.

A barotropic-dynamical model is used to give some insight into the behavior of dynamical track prediction models. Results show that initial-position errors have a negligible effect on the average track forecast errors except at 12 h. Errors in the estimate of the initial storm motion affect the forecast errors out to 72 h. Several spatial filters are applied to the initial analyses which shows that the model storm track is the most sensitive to very large scales (3000–6000 km wavelengths).

Save