Using the back-propagation neural network, a model for predicting tropical cyclone intensity changes in the western North Pacific at 12, 24, 36, 48, 60, and 72h is developed. The data used include the storm positions and intensities, the NCEP/NCAR reanalysis fields, and the sea surface temperature fields for western North Pacific storms occurred during a 14-year period of 1983-1996.
The predictors of the neural network model are selected based upon those of the multiple linear regression model. A regression analysis shows that the vertical wind shear predictor is consistently important over the prediction intervals. The average intensity prediction errors from the neural network model with climatology, persistence, and synoptic predictors are 7-16% smaller than those from the multiple linear regression model with the same predictors. Even the performance of the neural network model with only climatology and persistence predictors is slightly superior to that of the multiple regression model that includes synoptic predictors as well. It is revealed that the neural network model does not always improve upon the regression model for every year during the 14 years. However, the number of years that the neural network model is superior to the regression model is (much) larger than the number of years in the reversed situation, and appears to increase with decreasing prediction interval. Sensitivity experiments show that the average intensity prediction errors from the neural network model seem to be insensitive to the number of hidden layers or the number of units in hidden layer. However, there is some room for further improvement of the neural network model upon the regression model with a better hidden-layer structure for tropical cyclone intensity prediction. This study suggests that the neural network model that includes climatology, persistence, and synoptic predictors can be used as an effective tool in tropical cyclone intensity forecasts.