Prediction of tropical cyclone frequency with a wavelet neural network model incorporating natural orthogonal expansion and combined weights

A nonlinear wavelet neural network (WNN) model with natural orthogonal expansion (NOE) and combined weights is constructed to predict the annual frequency of tropical cyclones (TCF) occurring over the coastal regions of Southern China. Combined weights are obtained by calculating categorical weights, based on the particle swarm projection pursuit, and ranking weights, based on fuzzy mathematics, followed by optimization. The global monthly mean heights at 500?hPa and sea-surface temperature fields are used as two predictors. The linear and nonlinear information of the predictors with reduced dimensions is gathered through the NOE and combined weights, respectively, and treated as the input into the WNN model. This model is first trained with the 55-year (i.e., 1950?C2004) TCF data and then used to predict annual TCFs for the subsequent 5?years (i.e., 2005?C2009). Results show that the mean absolute and relative errors are 0.6175 and 9.34?%, respectively. The impacts of the combined weights, NOE and WNN as well as the traditional multi-regression approach on the TCF prediction are examined. Results show superior performance of the WNN-based model in the annual TCF prediction.