基于在线SVM的裂解炉燃料气热值软测量

Soft measurement for calorific value of cracking fuel gas based on Online SVM algorithm

针对裂解炉燃料气离线热值模型泛化能力差的问题,提出一种具有自适应能力的在线支持向量机(Online SVM)建模方法.该方法将增量式支持向量机(ISVM)与近似线性依靠(ALD)条件相结合,通过计算新样本与建模样本间的近似线性依靠值,选择满足ALD条件的独立新样本更新SVM模型.分析裂解炉燃料气热值的影响因素,并用Online SVM算法建立裂解炉燃料气热值在线软测量模型.该模型由离线训练模块和在线模型更新模块组成.离线训练模块基于离线数据训练得到初始热值软测量模型,在线更新模块通过使离线模型学习线性独立新样本来保证热值模型的在线预测精度.利用合成数据、Benchmark数据与裂解炉燃料气热值数据,将该方法与传统的支持向量机(SVM)与LS-SVM方法进行对比仿真研究.结果表明:该方法能够适应新的工况,具备自适应学习新样本的能力,可以用于具有慢时变特征的裂解炉燃料气系统热值软测量建模.

A newly adaptive online support vector regression machine（Online SVM）was proposed to improve the generalization ability of soft sensing model of calorific values of fuel gas in the cracker system that was constructed based on historical data.The approach combined the incremental support vector machine（ISVM）with approximate linear dependence（ALD）condition.New independent samples with ALD condition to update the SVM model were determined by calculating the approximate linear dependence（ALD）value between new samples and modeling samples.The influencing factors of calorific value of fuel gas of cracking furnace were analyzed,and an on-line soft sensing model of calorific values for fuel gas of the cracker system was established using Online SVM algorithm.This model consisted of off-line training module and on-line updating module.The off-line training module was mainly used to produce initially soft sensing model of calorific value based on historical data,and the on-line updating module was used to keep high predictive accuracy for on-line model of calorific value through making off-line training module to learn newly independent samples.A series of comparison simulation experiments were carried out between the proposed method and the conventional SVM and LS-SVM methods using synthetic data,benchmark data and calorific value data of cracking fuel gas.The simulation results show that the proposed method can adapt to new conditions with capability of learning new samples adaptively,and can be used for modeling of soft measurement for calorific values of fuel gas in cracker system with slow time-varying character.