Developing a Systematic Approach to Tropical Cyclone Track Forecasting expert system-2: Results of a real-time prototype test

A tropical cyclone track forecasting expert system prototype based on the Systematic Approach concept (Carr and Elsberry 1994) has been developed. The key output of the expert system is a selective numerical model consensus (SCON) forecast track that excludes numerical models deemed by the forecaster to be significantly degraded, and that should be more accurate on average than a simple consensus of all available numerical model tracks (NCON). The expert system prototype, which has been named Systematic Approach Forecasting Aid (SAFA), was installed at both the Naval Postgraduate School (NPS) and the Joint Typhoon Warning Center (JTWC) in August 1999. A parallel real-time test of the SAFA prototype by an experimental forecast team at NPS and the JTWC forecasters (as permitted by operational workload) was conducted from 23 August to 30 September to evaluate the effectiveness of SAFA. Because of anomalously low TC activity in the western North Pacific during the test period, the NPS component of the test continued to the end of the 1999 season to provide a more statistically significant sample of SCON and NCON forecasts.

The objectives of the real-time test of SAFA by NPS and JTWC include: (i) evaluating the effectiveness of the various information display features; (ii) validating the error mechanism frequencies suggested by previous research (Carr and Elsberry 1999); (iii) determining whether error mechanisms could be correctly identified without the benefit of hindsight; and (iv) determining the circumstances under which a SCON forecast may be produced that is significantly more accurate than a NCON forecast. Because the parallel test commenced in late August, and the NPS component of the real-time test is still in progress, only preliminary conclusions concerning objectives (i)-(vi) can be made at this time. With regard to objective (i), JTWC personnel were impressed with the user-friendliness and information display capabilities of the SAFA prototype, although many helpful suggestions for improvement were made. Based on the small sample of SAFA forecasts thus far, the NPS experimental forecast team found that excessive direct cyclone interaction (E-DCI) was, as expected, the most frequent error mechanism. This result validates objective (ii) for E-DCI, and it is expected that similar results will be found for the other frequently occurring error mechanisms as the number of SAFA forecasts grows. In addition, the NPS team found that cases of E-DCI sufficient to cause a 72-h track forecast error exceeding 300 n mi were readily identifiable, thus tentatively validating objective (iii). Finally, the NPS team found that when only frequently occurring error mechanisms were occurring, it is possible to create SCON track forecasts that are much more accurate than NCON forecasts. In some situations where infrequent error mechanisms were the primary source of model error, or compensating and different error mechanisms were occurring in different models, the SCON forecasts did not improve on the NCON forecasts.