用于FCC汽油辛烷值预测的非线性数学模型  

Nonlinear Mathematical Models for Octane Number Prediction of FCC Gasoline

作　　者：孙忠超[1] 山红红[1] 刘熠斌[1] 杨朝合[1] 李春义[1] 

机构地区：[1]中国石油大学重质油国家重点实验室,山东省青岛市266555

出　　处：《炼油技术与工程》2012年第2期60-64,共5页Petroleum Refinery Engineering

基　　金：中国石油大学(华东)研究生创新基金资助项目

摘　　要：依据汽油正构烷烃、异构烷烃、烯烃、环烷烃和芳烃(PIONA)的烃组成数据,将催化裂化(FCC)汽油单体烃组成分为37组,利用BP神经网络算法和支持向量机回归(SVR)分别建立了FCC汽油研究法辛烷值对37个变量的非线性数学模型。由MATLAB软件编写程序,利用Levenberg-Marquardt优化算法训练BP神经网络。支持向量机回归模型采用粒子群算法优化支持向量机参数及核函数参数,并采取交叉验证方法防止机器学习的欠学习和过拟合问题。计算结果表明:两种模型都能够较好地反映汽油单体烃组成与辛烷值之间的非线性关系;BP神经网络模型对辛烷值的预测性能好于支持向量机回归模型;增加样本数量,两种方法的预测准确性皆变好;针对40个样本的学习结果,两种模型预测的相对误差绝对值的平均值分别为0.148 7和0.167 4。With BP neural network （BPNN） and support vector regression （ SVR）, two nonlinear mathe matical models were established for research octane number prediction of FCC gasoline. Based on PIONA data of gasoline, the hydrocarbons in FCC gasoline were divided into 37 groups. The octane number was regarded as nonlinear function of these 37 variables. LevenbergMarquardt algorithm was used for training function of BPNN by the MATLAB software. The parameters of SVR and kernel function were selected by particle swarm optimization （PSO） and crossvalidation method was used for preventing the lesslearning and overfitting problems in SVR model. Calculation results show that both models are able to better reflect the nonlinear rela tionship between the octane number and hydrocarbon composition of gasoline. The performance of BPNN for octane numberrediction is better than SVR. Prediction accuracy of both methods is improved with increasing the number of learning samples. For 40 learning samples, the average absolute relative errors of prediction re suits for BPNbl and SVR are 0. 148 7 and 0. 167 4 respectively.

关 键 词：FCC汽油 研究法辛烷值 BP神经网络 支持向量机 粒子群算法 

分 类 号：TE626.21[石油与天然气工程—油气加工工程]