A transfer function model to predict hurricane intensity

A statistical model was designed to predict twelve and twenty four hours in advance the intensity of a tropical storm in the North Atlantic basin. The model is based on a transfer function and uses information from meteorological indices, and numerical models. Two types of data sets were developed to build the prediction scheme. The first data set was used to identify storms with similar characteristics to the current storm and the second data set uses steering and synoptic observations along the storm track to predict the hurricane intensity. An analog in this paper is a tropical storm that exhibits similar meteorological characteristics to the current tropical storm. The data sets included climatology, persistence, and synoptic information from 1948 – 2005. The major sources of information are Hurricane Best Track, the NCEP/NCAR Reanalysis data, and current analysis from numerical models.

The proposed methodology includes three major steps: 1) identify the upper air information to generate the initial conditions, 2) develop the upper air time series along the storm track, and 3) use upper air and meteorological indexes to predict the hurricane intensity.

Initial conditions. A competitive artificial neural network (ANN) and a classification algorithm are used to identify the analog hurricanes to the current storm. The analog is identified by implementing the following strategy. The first task consists of designing a competitive ANN and using the Kohonen learning rule. The competitive ANN assigns a code to each hurricane based on similarities identified for each variable at a given vertical level and at specific horizontal location with respect to the center of the storm and at a given point in time. The hurricanes that have the same code to the current storm indicate strong similarities and are named potential analogs. The second step consists of processing the first observations of the potential analogs by a second ANN. The hurricanes that fall in the same category of the current storm define the analog hurricanes. The next task involves filtering the analogs hurricanes by a classification algorithm whose criterion consists of minimizing the Euclidian distance between the current storm and the analogs. Thus, the storm that exhibited the minimum distance was called the analog hurricane to the current storm. Gridded upper air data from numerical models was obtained and include the following variables: geopotential height, air temperature, relative humidity, and wind vector components. These variables were extracted at the following levels: 1000, 850, 700, 500, 400, 300, 250, 200, 150 and 100 mb. These data were centered on the storm position with 10 degrees measured into the axial directions to East, West, South, and North. Spatial interpolation algorithm was used to obtain estimation at one degree of resolution. Thus, the extracted number of gridded points at each time interval was 441, i.e., the covered area at each time interval was a square of 21x21 degrees. A sequence of 30 points was defined as the initial conditions for the current tropical storm. The initial conditions have enough information to develop the stochastic transfer function model and perform prediction at 6, 12, 18 and 24 hours in advance.

Upper-air time series. The upper air for the current tropical storm was extracted from the outputs of numerical models. These are the same variables extracted at the same levels as described in step one. The time series are generated with information provided at every 12 hours along the storm track. The horizontal area and the interpolation algorithm are the same as described in step one. Dimensionality reduction was accomplished by using 7 principal components that represent about 90% of the total variance of the considered data set. In addition, the vertical wind shear, the maximum possible intensity, k-index, total-totals, and momentum were also computed along the storm track.

Predicting hurricane intensity. A transfer function model was used to represents the stochastic cause and effect relationship. A variable selection algorithm was developed to identify the best predictors that best explain the hurricane intensity. The algorithm is based on three regression concepts: parsimonious principle, stepwise selection, and avoiding multicollinearity problem. The principle of the algorithm is to divide the original set of predictors into smaller groups and perform variable selections in each group. The variables that best explain the underlying predictand are selected by using the stepwise technique in such a way that the multicollinearity problem is avoided. The Hooke and Jeeves and the maximum likelihood algorithms are used identify the structure of the transfer function model. Predictions are computed using the developed time series and the selected variables. The ensemble prediction is derived by computing the expected value of the individual predictions.