人工神经网络预报模型的过拟合研究

针对神经网络方法在预报建模中存在的“过拟合”(overfitting)现象和提高泛化性能 (generalizationcapability)问题 ,提出了采用主成分分析构造神经网络低维学习矩阵的预报建模方法。研究结果表明 ,这种新的神经网络预报建模方法 ,通过浓缩预报信息 ,降维去噪 ,使得神经网络的预报建模不需要进行适宜隐节点数的最优网络结构试验 ,没有“过拟合”现象 ,并且与传统的神经网络预报建模方法及逐步回归预报模型相比泛化能力有显著提高

STUDY ON THE OVERFITTING OF THE ARTIFICIAL NEURAL NETWORK FOR ECASTING MODEL

With the application of the artificial neural network (ANN) in the field of Atmo spheric Science, a "bottle-neck" was found while the artificial neural network model was applied for weather forecast: the fitting precision of training sample could not be definitely determined to make the model showing its best forecast ing capability. It was a key problem to be solved on the overfitting and genera tion capability of the ANN application theory area. Study on this problem is nec essary for the further operating application of ANN in the field of Atmospheric Science. A new forecasting model has been proposed for model establishment by m eans of making a low-dimension ANN learing matrix through principal component an alysis (PCA-ANN). The monthly rainfall of June?July and August were forecasted by using PCA-ANN, R-ANN(regression artificial neural network model)and SR(stepwise regr ession model) respectively. The results showed as followes:(1) The matrix built with active method, which could really reflect the internal relationship between the input and output, is a new method to develop the quali ty of generation capability and avoid or decrease the overfitting in ANN model. It also could be applied in other fields.(2) The PCA-ANN model could distill the main information of many factorials and remarkably decrease the dimension size of matrix. The noisy and overlap informat ion of the inputting factorials were also reduced because of the orthogonality o f each principal component.(3) The forecasting accuracy of PCA-ANN model was obviously higher than that of R-ANN model and SR model. Moreover, the PCA-ANN model needs not to test the bes t network structure and training precision. So it showed remarkable superiority co mpare with the method of finding suitable hidden nodes and the training course o f network structure.