A Numerical Prediction Product FNN Prediction Model Based on Condition Number and Analog Deviation

Aiming at the problem that the fuzzy neural network (FNN) technique itself does not provide the input matrix to the FNN prediction model, we present a prediction modeling methodology which combines the computation and analysis of condition number with FNN, and design the computation and analysis of analog deviation for the input matrix to choose samples close correlated with predictand as training samples, thus effectively reducing the scale of network and evidently enhancing the prediction ability of the FNN prediction model. Using the same CMA T213 and Japanese numerical prediction product (NPP) data, we performed the contrast experiments and analyses of the FNN prediction model for daily regional mean precipitation based on condition number and analog deviation against the condition number-FNN prediction model and the traditional stepwise regression prediction model, and results show that under the condition of the same number of selected predictors, the prediction accuracy of the FNN prediction model based on condition number and analog deviation is 12.6% higher than that of the stepwise regression model in the experiment of independent samples of 49 days.