Adaptive predictive control of nonlinear time-varying systems using neural networks

Predictive control of nonlinear dynamical systems in which the prediction is made using feedforward neural network is a plant model is discussed. The basic form of optimal predictive control is not tied with linear system parameters. Hence, it is a viable candidate for control of nonlinear objects. The control is based on a minimization of the sum of predicted squared errors over a prescribed range of prediction. A neural network, trained by backpropagation to mimic the behavior of a plant, is used to compute the squared error cost function for various candidates of controlling inputs. An optimal controlling input, by which the cost function will be minimized, is then numerically selected by, for instance, the simplex method. According to simulation tests on various nonlinear plants, the neural network adapts to time-varying plant dynamics as well as load upsets if the learning process is kept active through the control operation