Model Predictive Control in the Minerals Processing Industry

"Process control in the minerals industry is made difficult by factors such as the multivariable nature of the processes, noise, many disturbances, the changing nature of the processes, nonlinearities, process constraints, slow responses and large deadtimes. A good example of such a process is a milling circuit, where all these problems are experienced. Since model predictive control (MPC) effectively handles most of these difficulties, Mintek decided to develop, customise and test an MPC controller for milling control.Conventional MPC uses a dynamic model of the plant to calculate the future (predicted) plant outputs based on the movements of the manipulated variables in the past. The future moves of the manipulated variables can then be calculated using an optimisation algorithm that will minimise the difference between the setpoints and the predicted outputs. The least squares optimisation method is the quickest, but it does not cater for process constraints. Constraints can, however, be handled explicitly by using a quadratic programming optimisation method. This method ensures that the variables stay within the constraints at all times, but the calculation procedure is much more computationally intensive.• modifying MPC for the Minerals Processing IndustryIt was found that in order to obtain good, robust control using MPC in the minerals processing industry, the conventional MPC algorithm had to be adapted. The MPC controller was first tested on an extensive milling simulator, which was specifically developed for that purpose. Problem areas were identified and the MPC controller was specially adapted to accommodate for these difficulties. Some of the adaptations to the conventional MPC algorithm and special features added were as follows:Since milling plants contain many integrating processes (e.g. sumps), and integrating errors can cause offsets and poor control, it was essential that integrators be handled explicitly. This problem was solved by adapting the algorithm and also using an integrating error portion to compensate for integrating errors.Constraints on the inputs and outputs adversely affected the control. The controller using simple least squares optimisation was tuned with the constraints in place, while the maximum expected setpoint changes were made. This ensured that the process would be able to handle most setpoint changes up to these expected values without violating the constraints. A later improvement on this method was the incorporation of a quadratic programming algorithm to perform the optimisation. This method handles limits explicitly and was shown to perform better than the least squares methods (at the expense of a much higher computational cost)"