论文部分内容阅读
时间序列预测是一类非常重要的问题,但基本上局限于参数不可变问题的研究,而对实际问题中经常出现的更重要的参数可变系统的预测,由于构成几乎所有已有预测技术基础的Taken嵌入定理不再成立,所以这方面的研究成果极少.使用一种将(多)小波变换与反向传播神经网络相结合的新型网络结构——(多)小波神经网络,尝试对参数可变时间序列的预测.因为(多)小波神经网络的误差函数是一个凸函数,这在一定程度上可以避免经典神经网络容易陷入局部极小、收敛速度慢等问题.对著名的Ikeda参数可变系统的实验表明,多小波神经网络的预测性能较单小波神经网络要好,而单小波神经网络的性能较BP网要好.因此,该方法不失为时间可变系统预测的一种好的推荐.
Prediction of time series is a very important issue, but it is basically confined to the study of the immutable parameters. The prediction of the more important parameter variable systems, which often occurs in practical problems, is the basis of almost all existing forecasting techniques The Taken embedding theorem is no longer valid, so the research results in this area are very few.Using a novel network structure - (multi) wavelet neural network combining (multi) wavelet transform and backpropagation neural network, Because the error function of (multi) wavelet neural network is a convex function, it can avoid the problem of classic neural network such as easy to fall into local minima and slow convergence speed etc. To the famous Ikeda parameter The experiments of variable system show that the prediction performance of multiwavelet neural network is better than that of single wavelet neural network, while the performance of single wavelet neural network is better than that of BP network. Therefore, this method is a good recommendation for time-invariant system prediction.