基于LM/SVM方法的二次反应清洁汽油辛烷值预测  

Prediction of Octane Number for Clean Gasoline Obtained from Secondary Reaction Based on LM/SVM Approach

作　　者：袁俊[1] 周小伟[1] 杨伯伦[1] 

机构地区：[1]西安交通大学能动学院化工系动力工程多相流国家重点实验室,陕西西安710049

出　　处：《高校化学工程学报》2010年第2期258-262,共5页Journal of Chemical Engineering of Chinese Universities

基　　金：国家重点基础研究发展计划(973)资助项目(2009CB219906);国家自然科学基金资助项目(20776117;20976144);高等学校博士学科点专项科研基金资助课题(20070698037)

摘　　要：提出了一种莱文伯格—马夸特(LM)算法和支持向量机(SVM)有机结合的LM/SVM新算法,并将其应用于基于集总模型的二次反应清洁汽油研究法辛烷值的预测。借鉴复杂反应动力学研究中的集总方法,将汽油研究法辛烷值看成汽油饱和烃集总、烯烃集总、芳烃集总的函数,并采用支持向量机表达该函数。针对支持向量机参数及核函数参数难以选择的问题,通过莱文伯格-马夸特算法搜索支持向量机中的参数,并采取把训练集分割成工作样本和检验样本的策略,从而解决了过拟合的问题。利用经典测试函数对LM/SVM算法的性能测试结果表明:LM/SVM算法不但精度优于文献报道的遗传算法与支持向量机相结合的GA/SVM方法,而且其效率也远高于GA/SVM方法。LM/SVM方法对二次反应清洁汽油研究法辛烷值预测的相对误差绝对值的平均值为0.71%。Novel LM/SVM approach coupling support vector machine （SVM） with Levenberg-Marquardt （LM） algorithm was proposed and applied in predicting research octane number （RON） of clean gasoline obtained from secondary reactions. The research octane number was considered as a function of saturation lump, olefins lump, and aromatics lump based on the lumping concept for complex reaction kinetics, and the function was expressed by using support vector machine. LM algorithm was adopted here to search parameters of support vector machine and kernel function so as to overcome the difficulty in parameter selection for support vector machine, and the training data were classified into work data and test data so as to avoid over-fitting. Classical test function was adopted to measure the performance of the proposed method. Testing results indicate that both the prediction precision and computation efficiency of LM/SVM method are higher than those of GA/SVM approach. Mean absolute relative error between predicted gasoline RON and the experimental data is 0.71%.

关 键 词：清洁汽油 辛烷值 集总 LM/SVM方法 

分 类 号：TE622.5[石油与天然气工程—油气加工工程] TP301.6[自动化与计算机技术—计算机系统结构]