支持向量回归算法用于烷基苯若干热物性定量预测  

Support vector regression applied to the quantitative prediction of some physico-chemical properties of alkyl benzenes

作　　者：杨善升[1] 陆文聪[1,2] 陈念贻[2] 周应斌 唐卫东 

机构地区：[1]上海大学材料科学与工程学院,上海200072 [2]上海大学理学院 [3]金陵石化有限责任公司炼油厂,上海200444

出　　处：《计算机与应用化学》2005年第8期582-586,共5页Computers and Applied Chemistry

基　　金：国家自然科学基金项目(20373040)宁渡市重点博士(青年)基金项目(2003A61005)

摘　　要：烷基苯精馏分离是石油化工重芳烃加工的基本方法,各种烷基苯的热物性智能数据库对重芳烃加工过程优化控制有实用价值。本文研究了烷基苯系化合物若干热物性与化合物结构间的关系。采用新近提出的、特别适合于小样本多变量训练集的支持向量回归(support vector regression,SVR)算法总结了烷基苯系化合物已知物性的实验数据,建立了预报烷基苯系化合物若干物性的数学模型。47个烷基苯系化合物正常沸点、沸点汽化热、临界温度、临界压力和临界体积的SVR留一法(learing-one-out,LOO)预测的平均相对误差值(mean relative error,MRE)分别为0.370%,1.655%,0.791%,2.069%, 0.933%。结果表明,支持向量回归算法预测结果优于人工神经网络(ANN)和偏最小二乘(PLS)算法。Physicochemical properties of alkyl benzenes are essential to separate pure component from alkyl benzene mixture. QSPR （quantitative structure-property relationship） can be used to solve the problem of the prediction of physicochemical properties of the alkyl benzenes. In this paper, support vector regression （SVR） , a novel powerful machine learning technology based on statistical learning theory （SLT） was applied to QSPR study of 47 compounds of alkyl benzenes. Molecular structural parameters were considered as molecular descriptors. A floating search method based on leaving-one-out （LOO） cross-validation of SVR was introduced for parameters selection, In SVR modeling, some parameters （the type of kernel function, the regularization parameter C, and ε-insensitive loss function） were selected by using LOO cross-validation of SVR. The mean relative error （MRE） for the normal boiling point （bp） , heat of evaporation at the normal boiling point （Hvb） , critical temperature （Tc） , critical pressure （Pc） , and critical volume （Vc） of alkyI benzenes obtained by using SVR LOO were 0. 370% , 1,655% , 0. 791% , 2. 069% , 0. 933% , respectively. In a benchmark test, SVR models for bp, Hvb, Tc, Pc, and Vc were compared with several modeling techniques currently used in this field. The results showed that the prediction accuracy of SVR models was higher than those of back propagation artificial neural network （ BP ANN） and partial least squares （PLS） methods. The prediction accuracy of SVR models is high enough to meet the demand for chemical engineering simulation and design.

关 键 词：支持向量回归 QSPR 烷基苯 物性 定量预测 

分 类 号：TQ015.9[化学工程]