摘要
针对多变量、非线性、时变的实际工业过程系统,提出了一种基于局部最小二乘支持向量机的潜空间广义预测控制方法。该方法通过偏最小二乘构建潜变量空间,从而将复杂的多变量系统转变成多个单变量子系统,然后在每个采样点利用即时学习选择相关数据样本,在潜空间内在线建立每个单变量子系统(SISO)的局部最小二乘支持向量机(LSSVM)模型,最后利用广义预测控制器对这多个子系统分别实施预测控制。利用即时学习剔除冗余数据样本,提升了LSSVM的鲁棒性,并且使其更适用于实时建模和控制。利用该控制器对四容水箱对象进行仿真研究,验证了算法的有效性。
A generalized predictive control approach based on local least squares support vector maehines(LSSVM) model in latent space is proposed for industrial process systems associating with multivariate, nonlinearity and time-varying characteristics. The method constructs latent variable space with partial least squares algorithm. Complicated multi-variable control system is converted into several single input single output subsystems(SISO). Relevant data samples at present are selected by Just-in-time learning(JITL) at every sampling instant. Local LSSVM model for each SISO is online constructed in latent space. Predictive control is implemented to these subsystems separately with generalized predictive controller. Redundant data samples are removed by JITL. Robustness of LSSVM is improved. It is much more applicable for real time modeling and controlling. The proposed predictive control is applied to a quadruple tank process forsimulation. The algorithm efficiency is verified.
出处
《石油化工自动化》
CAS
2017年第2期20-26,37,共8页
Automation in Petro-chemical Industry
基金
国家自然科学基金资助项目(61273160
61403418)
关键词
多变量
非线性
偏最小二乘
即时学习
最小二乘支持向量机
广义预测控制
multivariate
nonlinearity
partial least squares
just-in-time learning
least squares support vector machines
generalized predictive control