Applications of a Machine Learning Model for Estimating Tropical Cyclone Track and Intensity Forecast Uncertainty

In addition to deterministic track and intensity forecasts, the National Hurricane Center (NHC) is required to provide forecast uncertainty estimates, which are critical to mitigation activities. The first product to provide track forecast uncertainty information, the Hurricane Strike Probability Program, was implemented by NHC in 1983. Intensity and wind structure uncertainty were added in 2006 when the Strike Probabilities were replaced by wind speed probabilities (WSP). The companion time of arrival of tropical storm wind products was added in 2018. A significant limitation of NHC’s current operational forecast uncertainty products is that they are almost entirely derived from historical forecast error distributions. An exception is NHC’s WSP model, which has a weak dependence on the spread of a few track forecast models.

The Tropical Cyclone Artificial Neural-network Error (TCANE) model was developed to estimate situationally dependent track and intensity forecast uncertainty based on multi-model ensembles and TC environmental parameters. TCANE uses a fully connected neural network to predict the parameters of the sinh-arcsinh (SHASH) normal distribution for intensity errors, which accounts for skewness, and the bivariate normal distribution for track errors. A fundamental difference between this approach and existing operational techniques is that the uncertainty estimates are obtained as part of the network training, rather than from post-analysis of forecast errors. In addition, TCANE estimates the error distributions of the NHC official forecast, so it can be used as input for downstream applications such as the operational WSP model. TCANE was trained on data from 2013-2022 and verified on independent cases in 2023. Verification results will be shown for the independent sample. Prototype graphical products will also be shown and applications of TCANE to NHC’s next generation WSP model will be described.