基于支持向量回归神经网络的时间序列预测

Time Series Prediction Using Support Vector Regression Neural Networks

为了选择神经网络的最好结构以及增强模型的推广能力,提出一种自适应支持向量回归神经网络(SVR-NN)。SVR-NN用支持向量回归(SVR)方法获得网络的初始结构和权值,自适应地生成网络隐层结点,然后用基于退火过程的鲁棒学习算法更新网络结点参数和权值。SVR-NN有很好的收敛性和鲁棒性,能抑制由于数据异常和参数选择不当所导致的"过拟合"现象。将SVR-NN应用到时间序列预测上。结果表明,SVR-NN预测模型能精确地预测混沌时间序列,具有很好的理论和应用价值。

To select the ＇best＇ structure of the neural networks and enhance the generalization ability of models, a support vector regression neural networks （SVR-NN） was proposed. Firstly, support vector regression approach was applied to determine initial structure and initial weights of SVR-NN so that the number of hidden layer nodes can be constructed adaptively based on support vectors. Furthermore, an annealing robust learning algorithm was further presented to fine tune the hidden node parameters and weights of SVR-NN. The adaptive SVR-NN has fast convergence speed and robust capability, and it can also suppress the ＇overfitting＇ phenomena when the train data includes outliers. The adaptive SVR-NN was then applied to time series prediction. Experimental results show that the adaptive SVR-NN can accurately predict chaotic time series, and it is valuable in both theory and application aspects.