Extended Abstract (80K)14A.7
Upper air information and neural networks to estimate hurricane intensity
Nazario D. Ramirez-Beltran, University of Puerto Rico, Mayaguez, Puerto Rico; and A. Veneros
Upper air information and artificial neural networks (NN) are use to predict hurricane intensity in the North Atlantic basin. Hurricane Best Track, NCEP Reanalysis, and Radiosonde data are used to build a reliable data set to design and train a neural network model. The best track and NCEP reanalysis data provide historical information while radiosonde includes historical and current information. We are working with aircraft reconnaissance, and satellite data, which provide on-line upper air information to improve the working data set. The proposed prediction scheme uses historical data to identify the analog hurricanes. Self organized NN with the Kohonen learning rule is used to identify the storm analogs to the current hurricane. The analogs are based on: Julian date, Eastward and Northward displacement, passed intensity, and storm direction. The last 50 years of climatology and persistent data provide sufficient information to identify the hurricanes that best resemble the behavior of the current storm. Once the analog hurricanes are identified the historical NCEP reanalysis data are used along the storm track to develop the following synoptic variables: tropospheric vertical wind shear, maximum possible intensity, maximum momentum at 850 mb and 200 mb. Historical data from analogs and current radiosonde observations are used to estimate the synoptic variables of the current hurricane. Persistence, climatological and synoptic observations of the analog hurricanes and the current storm are combined to create the initial training set. The original variables are used to generate nonlinear transformations and an optimization algorithm is used to identify the variables that are best correlated with storm intensity. The best variables obtained from the optimization algorithm are divided into two parts. The first part is used to identify the optimal transfer function and the number of neurons in the hidden layer. The second part is used to identify an optimal initial point and to predict hurricane intensity.
It has been shown that the design of the NN requires a special procedure to obtain appropriate application, otherwise, the NN will provide misleading results. Preliminary results show that the proposed prediction scheme is a potential tool to increase the accuracy in predicting hurricane intensity.
Session 14A, Tropical cyclone intensity change III: Statistical-Dynamical Models
Thursday, 6 May 2004, 1:30 PM-3:15 PM, Le Jardin Room
Previous paper
Recorded presentation