支持向量机的时间序列回归与预测

Support Vector Machines For Time Series Regression and Prediction

详细分析了支持向量机用于时间序列预测的理论基础。采用支持向量机、RBF和Elman神经网络模型,对仿真时序和工程滑坡变形时序进行了回归与外延预测。结果表明,在噪声水平较低时,SVR回归效果稍好,Elman与RBF网络的稳健性相对较差;随着噪声水平增大,两种神经网络的回归精度迅速下降。对于外延预测,两种神经网络仅限于短期的非线性模拟,而泛化性能更好的SVR在短期具有比较理想的效果,在较长的时间区间里也具有较高的预测精度(7步预测准确度控制在83.5%以上)。

A method for predicting time series based on support vector machines was proposed. The time series, including simulated data and landslide deformation data sets, were preformed for regression and prediction by support vector machine, RBF networks, and elman recurrent neural networks. A comparison of these three methods was made based on their predicting ability. The results show that： when noise level is lower in simulated experiment, support vector machine is perfect relatively, and the Elman and RBF network are of more instability, on the other hand, with the higher noise levels, the greater relative error of two networks models is made. For landslide data sets prediction, the neural networks are limited to predict short term nonlinear time series in terms of their accuracy, whereas support vector machine has a higher precision in the short term and long term.