Research And Soft Sensing Applications Of Partial Least-Squares Regression

In petrochemical process, there are many important indicator variables that cannot be measured on-line. Such as purity, dry point of distillation product and distillation tray efficiency, etc. Soft sensor provides an effective way to solve these problems. This paper researches on Partial Least-Squares Regression and Partial Least-Squares Splines theory and establishes a soft sensing model of atmospheric distillation column aviation kerosene dry point by using this method. Secondary variables selecting method and model calibration method is proposed, research and simulation has been done for the implementation in the future.Firstly, soft sensing technology, engineering steps and background are described briefly. Partial Least-Squares Splines method is researched and analyzed mainly. According to the nonlinear modeling accuracy, the recognition of noise points and engineering applications, research and simulation has been done. Through the research of simulation cases and cross-validity, variables segmentation and component number selecting principle is proposed. Secondary variables selecting method based on PCA is researched, and then secondary variables selecting method based on Partial Least-Squares Regression is proposed. Secondary variables of aviation kerosene dry point are selected through these two methods. By using the data acquired and Partial Least-Squares Splines, aviation kerosene dry point soft sensing model is established. In order to analyze the characteristics of this method, RBF neural network and LS-SVM is used to establish soft sensing model with the same data. Then the simulation results are analyzed and compared. The simulation results show that soft sensing model based on Partial Least-Squares Splines obtains higher estimation accuracy and applies to fewer samples, more noise points situation. Finally, according to Partial Least-Squares Splines model, on-line calibration method is proposed.This paper provides a new nonlinear modeling method for implementation of soft sensing and makes it feasible to apply advanced control for oil refineries.