Dynamic matrix control algorithm based on interpolated step response neural models

This paper presents a nonlinear Dynamic Matrix Control (DMC) algorithm. A neural network calculates on-line step response coefficients which comprise a model of the controlled process. These coefficients are next used to determine the optimal control policy from an easy to solve quadratic programming problem. To reduce the number of model parameters (step response models usually need many coefficients) interpolated step response neural models are used in which selected coefficients are actually calculated by the neural network whereas remaining ones are interpolated by means of cubic splines. The main advantage of the step response neural model is the fact that it can be obtained in a straightforward way, no recurrent training is necessary. Advantages of the described DMC algorithm are: no on-line model linearisation, low computational complexity and good control accuracy.