Tropical cyclone intensity prediction using regression method and neural network

Abstract

Using the multiple linear regression method and the standard back-propagation neural network, tropical cyclone intensity prediction over the western North Pacific at 12, 24, 36, 48, 60, and 72 h intervals is attempted. The data contain a 31-year sample of western North Pacific tropical cyclones from 1960 to 1990 and eight climatology and persistence predictors are considered. The percent of variance explained by the neural network model is consistently larger than that explained by the regression model at all time intervals with an average difference of 12 %. The average intensity prediction errors from the neural network model are 10-16 % smaller, except at 12 h where the errors are nearly equal, than those from the regression model. This study clearly shows potential of the neural network in the prediction of tropical cyclone intensity.