Predicting Hurricane Genesis and Evolution with Deep Learning

Hurricanes and tropical storms sustain devastating damage especially if they are not anticipated with a high level of certainty. With the goal of improving the prediction of hurricane genesis and evolution, this work explore machine learning to analyze patterns that precede genesis more efficiently than traditional methods can. This study trained a series of deep neural networks on various observational/modeling data to identify and classify storm types at future times. These models were built on recurrent neural networks units that intake both spatial and temporal information in order to predict future storm genesis and evolution. They were able to predict wind fields, storm genesis and storm type with reasonable accuracy. Our approach is divided into two steps: a physical prediction step and a classification step, each handled by a deep learning model. These results demonstrate the effectiveness of utilizing machine learning methods to learn details about complex weather systems and to forecast major storms. Current models can predict wind field with an RMSE of less than 1% and classification accuracy approaching 90% for major hurricanes. With more data and computational power, we are improving versions of these models to reach even higher accuracies.