A Deep Recurrent Neural Network to Forecast the Intensity and Trajectory of Atlantic Tropical Storms

The National Hurricane Center (NHC) and National Oceanic and Atmospheric Administration (NOAA) provide predictions for storms trajectories, intensity, and size. They create these predictions based on models that can be classified into 3 groups: dynamical, statistical, and ensemble. Classifications also include relative compute time required to create an output grouped as either early or late and forecast parameters such as trajectory, intensity, and wind radii. The most accurate models are late models that take upwards of 6 hours to produce an output whereas models that can produce an output in seconds are called early. Early models are crucial for emergency scenarios because of their timeliness. The statistical baseline models such as OCD5 are based on multivariate regressors that can explain a significant amount of variance. The performance for these methods can be augmented by incorporating more advanced statistical methods from deep learning such as recurrent neural networks. In this study, we research and implement the domain of machine learning and deep learning into Atlantic storm forecasting for both trajectory and intensity and evaluate them against the NHC standards. Previous research into machine learning to forecast tropical Atlantic storms include a sparse recurrent neural network (Kordmahalleh, Sefidmazgi, & Homaifar, 2016) and an artificial neural network (Jung & Das, 2013) which achieved favorable results. Tropical system models created with deep neural networks can be utilized to develop more precise emergency planning. Because of this, there is a necessity for more accurate and timely models that can help reduce the amount of loss caused by tropical storms and hurricanes. The results of this study provide a reproducible bidirectional deep recurrent neural network implementing LSTM cells (BDRNN) forecasting both the trajectory and intensity of Atlantic storms utilizing training and testing data from the Atlantic hurricane database (HURDAT2). The model forecasts were evaluated based on 2017 data and indicated skill by performing better than the statistical baseline of OCD5. The BDRNN model can be used to provide improved decision support for emergency responses because accurate forecasts can be produced on demand. The study provides a promising framework for additional research that can incorporate satellite imagery and different domains such as the North Central Pacific.