Predicting Tropical Cyclone Track Forecast Errors Using a Probabilistic Neural Network

Uncertainty in estimations of tropical cyclone (TC) tracks up to five days are currently based on historical forecasting errors and are thus agnostic to specific TCs, i.e. they are constant over a season for each basin. Recently, a probabilistic neural network was developed for predicting the uncertainty in TC intensity using storm specific training inputs. Here, we present a similar approach for TC track uncertainty. For a given TC, the network outputs a bivariate normal distribution for each lead time up to five days. These bivariate distributions can be used to describe the probability of diverging from the TC forecast trajectory by an arbitrary amount.

The network is trained on storm specific inputs, including environment variables such as the TC center sea surface temperature and average vertical shear, as well as outputs from various models including DSHIPS and HWRF. The labels used in training are the difference between the best track reconstruction and the official forecast (for late predictions) or consensus forecast (for early predictions).

Several case studies will be presented for historical storms. Because of the probabilistic approach and storm specific inputs, the predictions returned by this network are a plausible alternative to the historically derived track uncertainty cones.