Learning a Prediction Interval Model for Hurricane Intensities

Predicting hurricane tracks and intensity are major challenges. Currently, track prediction models are much more accurate than intensity prediction models. The regression-based Statistical Hurricane Intensity Prediction Scheme (SHIPS), first proposed in 1994, is still the dominant model. In this paper we propose a new model called Prediction Intensity Interval model for Hurricanes (PIIH). Different from other models which only predict future intensities as a single value, PIIH is the only model which is also able to estimate localized prediction intervals. We model a hurricane’s life cycle as a sequence of states. States are discovered automatically from a set of historic hurricanes via clustering and the temporal relationship between states is learned as a dynamic Markov Chain. Using this Markov Chain possible future states of a hurricane are found and used to compute intensity predictions and prediction intervals. In addition PIIH also uses a genetic algorithm (GA) to learn optimal feature weights and a damping coefficients which minimizes prediction error of the model. For evaluation we use the same features as SHIPS for the named Atlantic tropical cyclones from 1982 to 2003. Performance experiments demonstrate that PIIH outperforms SHIPS in most cases, and obtains improvements of around 10% for predictions of 48 hours into the future. In addition, the estimated intensity prediction intervals are shown to be accurate.