Operational Hurricane Forecast Model Track Error Spikes in Hurricane Irma (2017)

Operational model forecasts of tropical cyclones show large variability in track errors from model run to model run. Errors may be due to individual model deficiencies, or difficult to forecast changes in the direction of a tropical cyclone. Variability may also be related to errors in data assimilation and model initialization, or to details in the atmosphere itself.

At times, periods of larger-than-average track errors occur in forecasts from several different models. In Fig. 1, we see the 48-hour track errors for Hurricane Irma (2017) from three operational models (National Centers for Environmental Prediction (NCEP) Global Forecast System – AVNO, NCEP’s Hurricane WRF model – HWRF, and the NCEP Global Ensemble Forecast System mean forecast track – AEMN) and one multi-model consensus model – TVCA (a blend of tracks from models run by the European Center for Medium-Range Weather Forecasting, the UK Meteorological Office and NCEP). All four models show a spike in track error for model runs initialized between 0000 UTC 01 September 2017 and 0000 UTC 02 September 2017. The five-year average track error is shown by the horizontal black line for reference. Not shown are similar plots for 12-36 hour forecasts, which show the same spike in errors for the same model initialization times. Also not shown are plots for longer range forecasts. The spikes persist through forecast hour 120.

During the first 48 hours of the model runs in this period, there were no major sudden changes to either the forecast track or the NHC best track, when the spike in track errors occurred. The Best Track from the National Hurricane Center shows a gentle curve towards the west-southwest. There were no big changes in track during this period. Since these track error spikes occur for more than one model (other models, including the Navy’s model and the UK Met Office model, show similar spikes in this period), the errors arise through the data assimilation/model initialization, and likely the state of the atmosphere itself.

Hurricane Gaston (2016) is another example. This storm moved towards the northeast from 0000 UTC 24 August 2016 past 0000 UTC 28 August 2016. A cluster of larger than average errors occurred in forecasts from model runs initialized between 0000 UTC 25 August 2016 and 1800 UTC 25 August 2016. The spike in forecast track error was not only present in 48-hour forecasts, but also in the 36-, 24-, and 12-hour forecasts. A recurvature of Gaston happened later in the forecasts, beginning with 72-hour forecasts initialized after 1200 UTC 25 August 2016. Hence, the model errors are not due to Gaston suddenly changing direction. As with Hurricane Irma (2017), these errors were present in multiple operational models. Thus, the errors cannot be due to model deficiencies.

In this presentation, a detailed analysis of the atmosphere during the track error spike for Hurricane Irma will be shown. In addition, other examples of this behavior will be presented, and the atmospheric patterns compared to determine a) if the atmosphere is similar in other cases, and b) if it is possible to anticipate, or at least recognize the patterns at their beginning, and thus provide forecasters with guidance on the quality of the model forecasts.