Using Evolutionary Programming to Generate Improved Tropical Cyclone Intensity Forecasts

The difficulty of forecasting tropical cyclones (TC) has lead to the development of many different TC models, which can be dynamical, statistical, or a combination thereof. While these models have made notable advances in forecasting track, improvements in forecasting intensity have been comparatively lacking, particularly for rapid intensification (RI) and rapid weakening (RW). The goal of this research is to use evolutionary programming (EP) to construct improved tropical cyclone intensity forecasts at lead times of up to five days. Previously, EP ensembles have been tested on 60-h minimum temperature forecasts as well as 500-hPa height forecasts and been shown to outperform traditional dynamical model ensembles and multiple linear regression (MLR) ensembles.

In this research, two similar but separate models are developed for each forecast lead time considered, one for the North Atlantic and the other for the East and Central Pacific basins. These models are developed using predictors sourced from the SHIPS developmental dataset, with data from 2000-2016 being divided to form training, cross-validation, and testing datasets. Preliminary results show that EP can potentially provide small but notable improvements to deterministic TC forecasts, especially at short lead times. Probabilistic forecasts for RI and RW are also generated, with results being shown at the conference. This presentation will summarize the development of the models and their performance on reserved independent testing data as well as for the 2018 Atlantic and East/Central Pacific hurricane seasons.