Predictive neuro-control of uncertain systems: Design and use of a neuro-optimizer (Q1400310)

An integrated approach of adaptive neuro-modelling and predictive neurocontrol of uncertain stochastic time-varying systems is proposed. The controlled process is given by a difference equation to which the Gaussian white noise is added. At first the accurate estimates of the output (using the identifying neural network) are obtained. Then at each time step the neuro-minimizer (using another feedforward neural network) generates the optimal constrained control. Both neural structures are initially selected off-line by a statistical Bayesian decision procedure. A set of conditions under which the receding optimal neuro-controls ensure closed-loop stability of an uncertain stochastic time-varying system is provided. The actual and simulated experimental anaerobic digestion processes are selected to illustrate the capabilities of the proposed receding-horizon neuro-controllers.