支持向量回归建立成品汽油通用近红外校正模型的研究

Development of Universal Near Infrared Spectroscopic Calibration Models for Several Grades of Blended Gasoline by Support Vector Regression

针对目前采用偏最小二乘法建立成品汽油分析模型存在的问题,采用近几年新兴的支持向量回归方法建立了多种汽油标号通用的校正模型,其预测能力优于对应的偏最小二乘法,对汽油研究法辛烷值、烯烃和芳烃的预测标准偏差分别为0.37、1.28%和1.38%,可应用于实际的汽油管道自动调合近红外光谱在线分析。

On-line analysis technology is an indispensable part in automatic pipeline gasoline blending system. On-line near infrared spectrometry is one of the most appropriate methods for this application due to its rapid, convenience, and less maintenance. To overcome the non-linearity problem of partial least squares（PLS） in building universal calibration model for all grades of blended gasoline, a novel powerful nonlinear calibration method viz. support vector regression（SVR） was used to build this kind of calibration model in this paper. It was found that the prediction accuracy of the results obtained by the SVR universal model was high enough to meet the demand of the process control of gasoline blending. The root mean square errors of prediction（RMSEP） for research of octane number, olefin and aromatic were 0. 37, 1.28% and 1.38%, respectively. This calibration method can be applied to real online analysis of the gasoline automatic blending system by near infrared spectroscopy.