CONSCIENCE: Control and System Identification using Elements of Neural Network Computation Engineering
Hong-Te Su, Peter A. Minderman, Thomas J. McAvoy, John Wray · 1992
Neural networks are attracting a lot of interest as process models for model predictive control. This paper presents a neural network model predictive control algorithm (NNMPC). The optimal control problem is formulated, and it is solved using a feasible sequential quadratic program that handles position and velocity constraints. The process model is a recurrent neural network. In order to train a recurrent network, a more general learning law was needed. This learning law is presented. Further a significant computational advantage is realized in the model prediction control calculations by using a part of this general learning law. This benefit is discussed. Finally the NNMPC procedure is illustrated using a first principles representation of a multi-input, single-output industrial reactor.