A Hybrid Dynamical/Machine Learning Approach to Hurricane Prediction

Abstract

This project takes a hybrid dynamical/machine learning approach to annual tropical cyclone (TC) prediction in the North Atlantic (NA) Basin. We train regression and classification models on synthetic data generated from a number of current and extreme climate simulations from the High Resolution Atmospheric Model (HiRAM). This emulation method has the dual benefit of increased computational efficiency compared to the dynamical approach and increased quantity and consistency of training data compared to the statistical approach. Each model is trained on inputs of Sea Surface Temperatures (SSTs) and the target of predicting annual TC counts in the NA basin. We find hat that the statistical models such as linear and Poisson regression perform better on experiments with less diverse training data and a shorter, more recent evaluation period. In contrast, Random Forest models tend to perform better in experiments with more diverse training samples and over the complete historical record. Additionally, train Random Forests on higher dimensional spatial SST data and create feature importance maps to improve the explainability of the model.