连续小波神经网络收敛性能的探讨

Investigation on the Convergence Performance of Continuous Wavelet Neural Network

小波神经网络已经被应用于电力系统,并且表现出许多优良特性。但是人们对于小波神经网络的缺点却认识很少。首先介绍小波神经网络的原理,离散神经网络和连续小波神经网络的概念和特点,然后探讨了连续小波神经网络的收敛性能。研究指出,由于小波函数和S igm o id函数存在很大差异,因此当连续小波神经网络采用BP神经网络的初始化和训练算法时会出现收敛性差的问题,并通过实验进行了验证。最后本文给出了几种改进连续小波神经网络收敛性的建议。

WNN is used in power systems due to its many advantages, but little is known about its shortcomings. This paper introduces the principle of WNN, together with the characteristics of discrete WNN and continuous WNN, and then investigates the convergence performance of continuous WNN. The analysis reveals that because of the great differences between the wavelet function and sigmoid function, the continuous WNN shows poor convergence performance when the initialization and training algorithm of BP neural network are used. Experiment is performed to validate the analysis above. Finally some suggestions to improve the convergence performance of continuous WNN are given, and the necessity to explore suitable algorithms for continuous WNN is emphasized.