Accurate prediction models are expected for many disciplines. Considering the complicated linear and nonlinear relations among forecast objects and predictive factors

the natural orthogonal complement method and the projection pursuit of particle swarm optimization algorithm are used for the linear dimensional reduction and the nonlinear dimensional reduction

respectively. With this procedure

we project the high-dimensional nonlinear data to low-dimensional subspace and construct a genetic-neural networks integrated prediction model. The model is tested in the frequency prediction of landing-typhoon in southern China and then the model accuracy is compared with the result obtained by the regular regression statistical prediction method. The mean absolute error and the mean relative error of the five-year test prediction for the typhoon frequency are 0.81 and 13%

respectively

by using the new nonlinear prediction model proposed in this paper. The prediction results by the new model have been obviously improved

comparing to regular regression statistical prediction method. The results provide a new thinking and method for the prediction model study in other disciplines.