Evaluation of Dynamical Track Predictions for Tropical Cyclones in the Western North Pacific: Extensions to Other Years and Dynamical Models

Carr and Elsberry (1999; NPS Tech. Report) have described eight conceptual models that appear to explain cases of large ( > 300 n mi at 72 h) track errors by the Navy Operational Global Atmospheric Prediction System (NOGAPS) and the Geophysical Fluid Dynamics Lab (Navy version -- GFDN) models. This is believed to be the most comprehensive study of dynamical track prediction errors ever carried out, and goes far beyond the usual statistical summaries of track errors to search for causes. This paper will describe recent extensions of that research to other years and other dynamical models.

All previous large-error detection efforts for the NOGAPS and GFDN models have been done by the developer (LEC), who had extensive forecaster experience at the Joint Typhoon Warning Center. An elementary question is whether another experienced forecaster (GMD) could use the conceptual models and descriptions in Carr and Elsberry (1999) to detect the causes of large errors. That is, the first test is simply of the transferability of the knowledge base to an experienced forecaster. This was successfully accomplished by reading the report and with a few hours of discussions. Some cases are more difficult than others and required confirmation by the developer.

Whereas the NOGAPS and GFDN model-predicted wind fields (primarily 500 mb and 700 mb) and sea-level pressures were available for the 1997 development sample, only the tracks of the UK Meteorological Office (UKMO) model had been available. The UKMO kindly provided the necessary model fields for this test. The key question was if only a track is available, and it appears to be similar to another model track for which you have fields to detect the error cause, can you assume that the first model also has been affected by the same error mechanism? Unfortunately, the answer for the retrospective UKMO model test was that only in about half of the cases was the cause the same as had been inferred from having a similar-looking track. The important implication is that the forecaster needs the model fields (not just a track) to be confident in detection of likely erroneous tracks.

As indicated, the error-detection conceptual models were derived based only on the 1997 season plus one difficult 1998 case. The 1998 western North Pacific season was known to be one in which the NOGAPS and GFDN (and some other Asian) models provided less accurate track forecast guidance. The question was then whether the same conceptual models were applicable during the large-error scenarios, or if other conceptual models were required to explain these errors during the 1998 season. At least for the NOGAPS and GFDN model errors, the original conceptual models appear sufficient. Evaluation of the UKMO model errors is still in progress and will also be described in the conference preprint.