Combining Artificial Intelligence and Physics-Based Modeling Techniques to Improve Hurricane Track and Intensity Forecasting

This study describes an Artificial Intelligence (AI) based approach to correct the hurricane track displacement and intensity error forecasted by the current physical based forecast models such as GFS and HWRF. Keras Machine Learning (ML) framework is used to build a merged neural network composed of a Feedforward Neural Network (FNN) and a Convolutional Neural Network (CNN). The FNN is fed with latitude, longitude, maximum wind, and minimum sea level pressure information extracted from best track and HWRF forecast of cycles preceding times being forecasted. The CNN is fed with the HWRF 2D analysis fields of surface wind, pressure, temperature and TPW and U/V-Winds at 3 levels. The network is trained to estimate the latitude, longitude, maximum wind, and minimum sea level pressure error between the best track and forecast for 6, 12, 18, 24, 30, and 36 hour into the future. The model is trained over three hurricane seasons (2015-2017) for both Atlantic and Pacific basins and tested on storms from an independent hurricane season (2018). The result shows that the proposed merged network can improve the hurricane forecast predictions from the 2018 version of HWRF.