Predicting Tropical Cyclone Track Forecast Errors using a Probabilistic Neural Network

TL;DR

The paper addresses the need for forecast-specific tropical cyclone track uncertainty, criticizing static historical error methods used by NHC. It introduces a neural network that outputs the five parameters of a $\text{bivariate normal distribution}$ to describe track error, enabling dynamic, probabilistic uncertainty and landfall assessments. Calibration results using $PIT$ and interquartile range show well-calibrated predictions, with CRPS comparisons indicating improvement over the NHC cone and GPCE and parity with GEFS, all at negligible computational cost. This approach offers practical operational benefits, including probabilistic landfall probabilities and easy integration with new models and data streams, making it a strong candidate for real-time uncertainty estimation in tropical cyclone forecasting.

Abstract

A new method for estimating tropical cyclone track uncertainty is presented and tested. This method uses a neural network to predict a bivariate normal distribution, which serves as an estimate for track uncertainty. We train the network and make predictions on forecasts from the National Hurricane Center (NHC), which currently uses static error distributions based on forecasts from the past five years for most applications. The neural network-based method produces uncertainty estimates that are dynamic and probabilistic. Further, the neural network-based method allows for probabilistic statements about tropical cyclone trajectories, including landfall probability, which we highlight. We show that our predictions are well calibrated using multiple metrics, that our method produces better uncertainty estimates than current NHC approaches, and that our method achieves similar performance to the Global Ensemble Forecast System. Once trained, the computational cost of predictions using this method is negligible, making it a strong candidate to improve the NHC's operational estimations of tropical cyclone track uncertainty.

Figures (8)

• Figure 1: Schematic showing the network architecture and format of model predictions. In many of the following figures, these predictions are used to construct a two-dimensional cumulative distribution function, defined by the Mahalanobis distance Mahalanobis. Each labeled ellipse encloses the integrated probability out to that distance. Larger percentiles enclose more of the probability, thus, confidence that the truth falls within a percentile increases with percentile. Inputs are described in Table \ref{['tab:inputs']}.
• Figure 2: Interquartile range (IQR) versus error. Boxplots show the distribution of forecast error (filled area spans 25th to 75th percentile, whiskers out to 10th and 90th percentile) associated with the lower, middle, and upper tercile of IQR for each lead time. The IQR is a measure of the width of the predicted bivariate distribution.
• Figure 3: Probability Integral Transform (PIT) histogram for the Eastern Pacific (top panel) and Atlantic (bottom panel) for all forecast lead times. This metric describes how often the truth falls into each decile of the predictions (10th, 20th, 30th, etc.). A perfectly calibrated probabilistic model would have a uniform distribution of 0.1.
• Figure 4: Two examples of our predictions, as described in Figure \ref{['fig:prediction_schematic']}. The initialization time is fixed and the forecast every 12 hours out to five days is shown, along with the best-track reconstruction.
• Figure 5: Continuous Ranked Probability Score (CRPS) for the NHC cone and the bivariate predictions as a function of lead time. The median for each is shown as the solid lines, while the shaded area encloses the $10$th to $90$th percentile of CRPS values. A lower value of CRPS indicates a better prediction (a CRPS value of zero indicates a perfect prediction, with the entirety of the prediction weight at the truth, e.g., a delta function).