Control of grinding plants using predictive multivariable neural control

This work investigates the use of a recently developed direct neural network (NN) multivariable predictive controller applied to a grinding plant. The NN controller is trained so that an estimation of the control error several steps ahead is minimized, which are given by a properly designed NN called predictor. An NN, which identifies the plant, is used to backpropagate the control error at present instant of time, as well as at various steps ahead. A linear, as well as a phenomenological (nonlinear), model of CODELCO-ANDINA grinding plant are used to simulate the proposed control strategy. The linear model was built from empirical data obtained from a real grinding plant around an operating point. The phenomenological model is based on a mass balance and power consumption of the mills containing 17 particle size intervals. Several tests are performed, driving the process to an operation point, and then, controlling it by training the NN controller on line. Finally, a comparison with other control strategies already applied at a simulation level is presented. These include classical and adaptive multivariable control algorithms. All the results presented in the paper are based on simulations.