Using Evolutionary Programming to Generate Improved Tropical Cyclone Intensity Forecasts

The difficulty of forecasting tropical cyclones has led to the development of many different tropical cyclone forecasting models. These models can be stratified by those that are dynamical, or those which obtain forecasts by solving the primitive equations; statistical, or those which obtain forecasts through exclusively statistical means, or statistical-dynamical, or those which use output from dynamical models as input to statistical algorithms to obtain a forecast. While models have made notable advances in forecasting track in recent decades, improvements to intensity forecasts have been comparatively lacking. A significant contribution to intensity errors is the difficulty of predicting times of rapid intensification (RI) and rapid weakening (RW), the former of which is a National Hurricane Center priority given its potential significant societal impacts when near land.

The goal of this research is to use evolutionary programming (EP), a machine-learning method, to develop a statistical-dynamical model capable of provide improved tropical cyclone intensity forecasts to lead times of up to 120 h. An EP algorithm has a specified structure that represents a combination of coefficients, normalized variables, and mathematical operators. To begin, a set of initial algorithms is randomly generated and its performance on the testing data is evaluated. Based on the desired skill metric, a specified percentage of worst-performing algorithms is replaced by mutated versions of the best-performing algorithms. The process continues until the forecast skill gained between training cycles asymptotes to zero, at which time a specified number of best-performing algorithms (over the set of all cycles) is retained for further consideration.

For this research, two similar but separate models are developed: one for the north Atlantic basin and the other for the east/central Pacific basin. These models are developed using predictors sourced from the SHIPS developmental dataset from 2000-2016. The set of all observed tropical cyclones is stratified by lifetime maximum intensity (tropical cyclone, minimal hurricane, major hurricane) and assigned evenly and randomly to the training, cross-validation, and testing datasets. This ensures that an even distribution of storms by intensity is achieved in each dataset. Five sets of 2,000 algorithms are initialized, and 20% of the population of each tribe is killed off with each training cycle. The evolutionary process continues for 300 iterations per initialization, and a total of five initializations are used. The 100 best-performing models (as determined by performance on the cross-validation dataset) at each lead time for each basin are retained, from which a root-mean squared difference algorithm is used to identify a subset of five models with high skill and large differences from other algorithms. Bayesian model combination is then used to combine the five algorithms into a single model capable of providing skillful deterministic and probabilistic forecasts. This presentation will summarize model development and model performance in testing and in retrospective simulations of the 2017 hurricane season.