基于经验模态分解与RBF神经网络的混合预测

Hybrid Prediction Method Based on Empirical Mode Decomposition and RBF Neural Network

为提高时间序列预测模型精度,根据各本征模态函数(intrinsic mode function,简称IMF)序列的变化特点,针对EMD-RBF神经网络隐含神经元数目及其中心数据选取问题,利用经验模态分解(empirical mode decomposition,简称EMD)的信号自适应处理能力和径向基函数(radical basis function,简称RBF)神经网络的非线性逼近能力,提出了一种基于EMD与RBF神经网络的混合预测方法。该方法将具有类似时频特性的本征模态函数分别建立RBF神经网络预测模型,采用基于统计分析的k-均值聚类方法自适应确定RBF模型参数,最后将各IMF-RBF神经网络预测结果进行重构得到最终预测结果。仿真结果表明,该方法充分考虑到各IMF本身的特性,增强了时序的可预测性,预测性能比传统反向传播(back propagation,简称BP)神经网络和小波BP神经网络更优越。将该方法应用在某装备温控系统性能监测中,其温度参数最大预测误差远小于传感器误差,说明将该方法在该装备故障预测中是可行的。

In order to advance prediction model precision of time sequence, according to the variety characteristics of the IMF(Intrinsic Mode Function), for EMD-RBF neural network’s number of hidden neurons and data selection, a new method is presented by using self-adaptive processing capability of EMD(Empirical Mode Decomposition) and nonlinear approaching capability of RBF(Radical Basis Function) neural network. First of all, RBF neural network prediction model is established from IMF with similar time-frequency characteristics, which adopts k-means clustering method based on statistical analysis to make RBF model parameters. Finally, the prediction results of IMF-RBF neural network are restructured to obtain the last prediction result simulation. Simulation results show that this approach enhance the predictability of time sequence, considering the characteristic of IMF itself. The prediction performance is better than traditional BP neural network and wavelet BP neural network. The method is applied to monitoring equip temperature control system, in which the maximum prediction error of temperature parameter is much less than the sensor error. Hence, this study is feasible in equipment fault prediction.