About

This paper presents the concept of a model predictive controller (MPC) that can account for nonlinear process behavior by using process value estimation from a Neural Network. The controller contains a dynamic linear model prediction that can be adjusted using any nonlinear steady state estimator. The nonlinear prediction of future outputs is realized using an embedded neural network algorithm. The resulting prediction trajectory contains the final NN prediction and accounts for the linear dynamics that are captured in a step response model. This approach differs from others in that it adjusts process state variables rather than model parameters. It becomes possible because the dynamic matrix control (DMC) technique is used to control the process based on the prediction. In this particular case the process state becomes the explicit prediction matrix of future process outputs. The controller matrix is developed from a linear dynamic model. One of the advantages of such controller is its robustness. The original model may not be altered even if intermittent noise patterns falsely indicate model mismatch. This solution is easier to implement than known approaches where Neural Networks are used for nonlinear model estimates. The paper outlines the implementation of the concept in an industrial product and presents test results.

- Downloadable files require Adobe Acrobat Reader.