Artificial Neural Network Prediction Modeling Method Based On The Natural Orthogonal Launched

Upon using an artificial neural network(ANN) a new short-term climate forecast model with the monthly mean rainfall in June in the north of Guangxi as predictand is established making empirical orthogonal functions(EOF) to the 36 predictors ( 15 SSA predictors, 21 500hPa Height predictors ) with over 0.05 significant correlation level of previous 500hPa height and sea surface temperature(SST) field, and selecting the high relative principal components, at the same time, a new approach of constructing ANN learning matrix is developed. Predictive capability between the new model (principal components ANN model) and linear regression model for the same predictors is discussed based on the independent samples and historical samples. Evidence suggests that the prognostic ability of the new model with high stability, when hidden nodes changing nearby input nodes and training times changing at the certain extent, is significantly better than traditional step wise regression model mainly due to the new model condensing the more forecasting information, properly utilizing the ability of ANN self-adaptive learning and nonlinear mapping. But the linear regression technique only selects several predictors by the F value, many predictors information with high relative coefficients is not included. So the new model proposed in this paper is effective and is of a very good prospect in the atmospheric sciences fields.