A Deep-Learning Driven Improved Ensemble Approach for Hurricane Forecasting
Tracking the path and forecasting the intensity of hurricanes are challenging. Dynamical models produce a significant model-measurement error. Accurate forecasting is very difficult to achieve after landfall. For track forecasting (where the storm is going to go), dynamical models are generally the best. For intensity forecasting, statistical models generally perform better. We can combine the advantages of both models using a machine learning ensemble approach. Machine learning models are computationally efficient and are currently used widely for forecasting and ensemble purposes. Deep Neural network (DNN) techniques comprise a popular and powerful class of machine learning methods. We developed a three-step DNN-based ensemble hurricane forecasting model using the output of eight dynamical hurricane models being used in ATCF system for forecasting hurricane track and intensity. We used all tropical cyclones in Atlantic Ocean from 2003-2016 and tested the model for those in 2017. The preliminarily results of our model show statistical advantages over NHC official forecasts. This presentation was given at the Earth Science Information Partners (ESIP) Winter Meeting in January 2019.