MPC Algorithms with Neural Approximation

Abstract

This chapter discusses a family of suboptimal MPC algorithms with neural approximation the characteristic feature of which is the lack of on-line linearisation. A specially designed neural network (the neural approximator) approximates on-line the step-response coefficients of the model linearised for the current operating point of the process (such an approach is used in the MPC-NPL-NA and DMC-NA algorithms which are extensions of the MPC-NPL and DMC ones). Alternatively, the neural approximator calculates on-line the derivatives of the predicted output trajectory with respect to the future control sequence (such an approach is used in the MPC-NPLTNA algorithm which is an extension of the MPC-NPLT one). The explicit versions of MPC algorithms with neural approximation are also presented. They are very computationally efficient, because the neural approximator directly finds on-line the coefficients of the control law, successive on-line linearisation and calculations typical of the classical MPC algorithms are not necessary.