A new PLS approach with hybrid internal models

Abstract

The partial least square (PLS) approach is a very important statistical modeling method which is especially useful for the modeling of complex processes with high dimension of input variables, inadequate sample data, strong correlation between the input variables, and nonlinearity. In the paper, a new PLS approach with hybrid internal models is proposed which can choose the optimal model class for each internal model. Compared with the existing PLS approaches in which only a single kind of model class is chosen for all of the different internal models, the approach proposed in the paper can effectively improve both the approximation accuracy and the prediction stability of the final model. A real example of modeling the melt index in a polypropylene production process is given to show the efficiency of the method.