基于支持向量回归与地统计学的多维时间序列分析

Multidimensional Time Series Analysis Based on Support Vector Regression and Geostatistics

基于地统计学与支持向量回归,建立一种快速定阶、既反映样本集动态特征,又体现环境因子影响的高精度非线性多维时间序列预测方法(GS-SVR)。对带趋势时间序列平稳化后,先基于地统计学后效时间长度进行因变量快速定阶;再以支持向量机基于最小原则非线性筛选自变量,继以主成分分析消除自变量之间的信息冗余;最后以一步预测法检验GS-SVR的有效性。2个农业科学实例结果显示,GS-SVR在所有参比模型中预测精度最高,稳定性最好。GS-SVR能快速、准确实现模型定阶,是一种融合时间序列分析和回归分析的非线性多维时间序列分析方法,并具非线性、避免过拟合、避免局部最小、泛化能力优异等优点,在农业科学、生态学、经济学等多维时间序列预测领域有较广泛的应用前景。

To construct a novel nonlinear multidimensional time series analysis approach named as GS-SVR based on geostatistics （GS） and support vector machine regression （SVR）,which could provide a fast order determination and show the dynamic characteristics of dataset as well as the effect of environmental factors.Firstly,stabilizing the time series and estimating the order by geostatistics.Secondly,screening the variable by leave-one-out method based on SVR and eliminating redundant information of variables by principal component analysis （PCA）.Lastly,the reliability of GS-SVR was tested by two agricultural datasets with one-step prediction.The prediction results showed that GS-SVR had the higher prediction precision and stability compared with all reference models.As a novel nonlinear multidimensional time series analysis approach integrating time series analysis with regression analysis,GS-SVR had order determination quickly and accurately and higher prediction precision.It can be widely used in the prediction area of agriculture,ecology and economics.