Improvements in Hurricane Intensity Forecasts from a Multimodel Superensemble Utilizing a Generalized Neural Network Technique

Abstract

Abstract Forecasting tropical storm intensities is a very challenging issue. In recent years, dynamical models have improved considerably. However, for intensity forecasts more improvement is necessary. Dynamical models have different kinds of biases. Considering a multimodel consensus could eliminate some of the biases resulting in improved intensity forecasts as compared to the individual models. Apart from the ensemble mean, the construction of multimodel consensuses has always contributed to somewhat improved forecasts. The Florida State University (FSU) multimodel superensemble is one that, over the years, has systematically provided improved forecasts for hurricanes, numerical weather prediction, and seasonal climate forecasts. The present study considers an artificial neural network (ANN), based on biological principles, for the construction of a multimodel ensemble. ANN has been used for constructing multimodel consensus forecasts for tropical cyclone intensities. This study uses the generalized regression neural network (GRNN) method for the construction of consensus intensity forecasts for the Atlantic basin. Hurricane seasons 2012–16 are considered. Results show that with only five input models improved guidance for tropical storm intensities may be obtained. The consensus using GRNN mostly outperforms all the models included in the study and the ensemble mean. Forecast errors at the longer forecast leads are considerably less for this multimodel superensemble based on the generalized regression neural network. The skill and correlations of different models along with the developed consensus are provided in our analysis. Results suggest that this consensus forecast may be used for operational guidance and for planning and emergency evacuation management. Possibilities for future improvements of the consensus based on new advances in statistical algorithms are also indicated.