大气中SO2浓度的小波分析及神经网络预测

应用小波分解和重构对SO2浓度年变化趋势进行分析,在此基础上,建立了一种分段BP神经网络预测模型,并对各段有针对性地设计了神经网络预测模型.采用主成分分析进行输入变量降维.在BP网络训练过程中,往往会出现过拟合的现象,为此,在训练过程中,将样本等间距地分离为训练集和验证集来防止这个问题.为了消除网络的权值初始化对学习系统复杂性的影响,采用了5个子网络输出取算术平均的神经网络集成的方法.预测结果表明,该模型应用于SO2浓度预测具有较高的预测精度和良好的推广能力,而且明显优于一般的神经网络模型.

Study on wavelet analysis and neural network prediction of SO2 concentration in air

Trend of SO_2 concentration is analyzed by means of the wavelet decomposition and reconstruction, a BP neural network predicting model with divided patter is firstly constructed and every pattern is pertinently designed. Dimensionality of input variables are reducted using Principal Component Analysis(PCA). In the training of BP neural network the over-fitting often appears which affects the result of forecasting. To prevent this problem the entire data set is divided into training set and validation set. Generalization ability is improved by using neural network ensembles. The results show that the divided BP neural network models of SO_2 concentration have good quality in terms of prediction precision and generalization, besides, it also indicates that this models are more approximate to realities in comparison with general BP neural network model.