神经网络和预报集成的若干问题及其对策

Some Problems in Neural Network and Prediction Ensemble and its Counter Measure

本文评述了神经网络在气象应用中存在的问题、产生的原因和网络结构设计的原则及经验,并提出了构造伪样本来诊断模型可能存在问题的方法;简述了预报集成的原理和原则,提出了构造因子子集差异法生成预报个体,进行集成,来解决小样本和因子选取问题的方案。

In this paper, the comments of problems in meteorological application of neural network and its causes of formation are given. The principles and experiences of neural network structure designs and also discussed. An approach named false sample verification method is proposed to diagnose modeling problems. Finally, the principles of prediction ensemble are briefly reviewed. It is suggested that by using the method of constructing the differences among factor subsets to generate individuals for prediction ensemble, then the problems of smaller samples and factor selection may be solved.