两种基于支持向量机的时间序列数据建模

Two SVM Hybrid Models on Time Series Data

在统计网络传输数据建模上,平稳化的数据有利于预报建模。由于传输数据是非平稳时间序列,具有非线性、多尺度等特点,就如何削弱数据的随机性并构造计算模型进行仿真计算,本文实验建模了经验模式分解与小波分解组合支持向量机的两种计算模型。第一种建模方法是小波组合向量机建模,做法是先将数据流分解为长期趋势和随机扰动项,然后采用支持向量机对分解后的各分量预测,最后将各预测值相加得到最终预测结果;第二种建模方法是经验模式分解组合向量机建模,先将流量分解成不同频带本征分量,常规的做法是用向量机逐一对各分量进行预测,然后对预测值等权求和得到预测结果作为验证结果;新提出的做法是直接把各模式分量作为输入向量,与真实值建立预测模型。结果表明基于经验模式分解建模构造的新实验模型,相比小波组合模型在传输数据预报上更稳定可靠。

In building statistics model on network transferring data, smooth data is effective for prediction. Owing to the non-sta-tionary, non-linear and multi-scale characteristics of the transferring data, in this study two models that combining with SVM are proposed and compared. One is wavelet based model which separates the data into long-run vectors and random disturbance vec-tor and then using each vector as input for prediction, the final outcome is added by individual prediction. Another one is EMD based model, the data is decomposed by EMD into different smooth IMF components, and conventional method is using SVM to predict each component separately, the results are obtained by summing individual prediction with same weight, as checking results in experiment; the proposed model is computed directly using IMF component as input vectors for constructing SVM model with the original data. The computation results show that the proposed EMD computing model is obtaining higher predic-tion accuracy than wavelet model and the checking model. It is also proved more stability.