神经网络递推预测模型及其应用

本提出一种反向传播神经网络（ＢＰ网络）的改进算法，该算法可以克服常规ＢＰ网络的两个主要缺陷，明显提高学习速度。其主要特点是首先拟合ＢＰ网络各权系数的变化过程，然后根据拟合曲线来调整权系数，并在学习过程中动态调整学习速率和动量因子，此外，还采用了随机搜索法优选初始权重以克服局部极值问题。基于ＢＰ网络和上述改进算法，本提出一种递推预测模型，它包括绝对量值模型和相对量值模型两种，它以让神经网络用人们

THE RECURSIVE FORECASTING MODELS WITH NEURAL NETWORKAND ITS USE IN FORECASTING

This paper first sets.forth an improvedalgorithm of BP neurul network to overcome the twomain shortcomings of the normal BP algorithm. themajor chartcteristic of the improved algorithm is fittingIhe changing procession of the weights of BP networkand adjusting the weights according to the fitting curves,meanwhile, the learning rate and the moment factor arealso dynamically adjusted. In addition, random seurch-ing method id used to selecl good initial weights so as toavoid locul optimal problems.By using the improvedalgorithm, Ihe learning speed of BP neural network ismuch faster than the basic BP algorithm. Bused on HPneurul network and the improved algrorithm desribedabove, recursive models for forecasting are put forth.The basie idea of the conrtruction of the recursivemodels is that in many cases the value at time t can bededuced accordinq to the values before time t. So longas the BP neural network grasp the law of recureion. itcan be used for forecasting. The recursive forecastingmodels include absolute value bused model and relative value based model.They are suitable for not onty shortIerm forecasting, but also for long term forecasting. Isan example, Ihe recursive models are used for the fore-cast of electricity in China from 1993 to 2!2!.By com-paring with relevant reference ,the forecasting resultsare reliable.