摘要
样本构成、网络结构、学习算法是影响前馈网络应用的三大关键问题。本文提出了一种基于门限自回归TAR模型的非线性时间序列预测前馈网络方法,该方法利用TAR模型的门限值对训练样本进行分群,依据TAR模型的阶数、训练样本数等确定前馈网络结构,网络学习算法采用基于梯度法和共轭梯度法相结合的联合梯度算法。应用研究表明该方法有效地改善了网络的泛化性能,提高了预报精度,同时也缩短了网络的训练时间。
Multilayer feedforward neural networks have been used successfully to predict time series data. But the neural networks’performance is highly dependent on its structure,learning algorithm, training pairs, activation functions, and on the overfitting problem among other things. In this paper, we propose certain enhancements to go with the problem for feed forward neural works to become a practical forecasting tool. The basic idea is to use the threshold autoregressive model to determine the numbers of neural networks needed and the structure of each component. This process minimizes the data required and consequently the size of the network. A hybrid gradient method is used to train the network instead of backpropagation algorithm. Case studies with hydrological time series prediction are prestented. The improvement both in the prediction performance and convergence speed are found to be considerable.
出处
《水力发电学报》
CSCD
北大核心
1996年第3期24-32,共9页
Journal of Hydroelectric Engineering
基金
中国博士后基金
国家自然科学基金
关键词
前馈网络
时间序列
非线性
模型
水文预报
Feedforward neural network Time series Nonlinear TAR model Hydrological forecasting.