打折最小平方RBF网络及其时间序列预测研究

借用打折最小平方(DLS)原理构建了基于误差平方准则的DLS-RBF网络的学习算法.打折最小平方原理考虑了时间序列本身的结构性变化,较好地刻画了预测点与其他时刻数据的相关性,而这些恰恰是现有的径向基函数神经网络(RBF)在预测过程中所忽视的.实验表明DLS-RBF网络在非平稳方差时间序列和某城市自来水实际的月用水量预测中的效果明显,并优于RBF网络,但在混沌时间序列预测的实验中,因其自身的混沌特性,预测效果并不十分明显.

Discounted least square RBF neural networks with applications in time series prediction

The discounted least squares (DLS) principle is borrowed to construct the learning algorithm of DLS-RBF based upon squared error criterion. The principle of DLS formulates inherent structural changes and time correlation in time series itself precisely, while in the learning and predicting process of current radial basis function (RBF) network this correlation is neglected. Experiments show that DLS-RBF has better performance than RBF in non-stationary covariance time series and daily-life-water consumption prediction. But in the experiments of chaotic time series forecasting, the predictive effects are not prominent due to its chaotic characteristics.