Neural Network for Combining Linear and Non-Linear Modelling of Dynamic Systems

The purpose of this paper is to develop a method to combine linear models with MLP networks, i.e. to find a method to make a nonlinear and multivariable model that performs at least as good as a linear model, when the training data lacks information. First, the MLP network for predicting the output from a dynamic system is described. Then two methods are proposed to combine linear and nonlinear modelling. The first method is the MLP network with linear path through, and the second method is a linear model with nonlinear error correction. Finally the two methods are tested. A thermal mixing process is used as a test system. This system is a multivariable and nonlinear process. The test is partly based on a simulation of the process and partly on data from a physical process. The results are given and discussed. >