(369i) Arx-Based Model Predictive Control of Systems with Time Delays
Advanced control strategies such as Model Predictive Control have gained wide spread interest in many areas in the chemical industries, due to fast algorithms, a well established theory and growing number of successful industrial implementations. The main feature is that the optimal control signal is determined as a constrained optimization which utilizes future predictions of the plant behaviour. Hence the controller has a plant model embedded for state estimation. This is a clear advantage for systems with complex process dynamics such as non-minimum phase behaviour and dead times and furthermore for systems operated close to the process constraints. The achieved closed loop performance will be dependent on the quality of the future predictions. Therefore the necessity of an accurate noise model will be analyzed in relation to different control implementations. If an unmeasured constant disturbance enters the process, the effect of the disturbance can not be compensated by the standard controller. This type of input disturbance is common in the process industries when a feed source changes. Examples are refineries and cement industries where the composition of the crude oil or raw minerals may change significantly when feed is changed from one source to another. Offset can be eliminated in model predictive control by including an integrator on the input and output signals or by augmenting the system description with a disturbance state which can then be estimated by the state estimator. Alternatively the model predictive control algorithm can be combined with a recursive disturbance estimation algorithm. The closed loop performance of the system will depend on the nature of the disturbance and how the disturbance rejection is facilitated by the control algorithm. This contribution will analyse different methods for achieving offset-free tracking and there tuning in ARX-based model predictive control. The ARX model can be robustly estimated from process data for both single and multivariable systems; hence it constitutes an attractive model class in development of a methodology which brings the user from plant data to good closed loop performance. The different implementations and there tuning will be demonstrated on both a single and a multivariable control loop problem. The single loop example is an oil-gas furnace where the feed of fuel to the furnace is used to control the gas temperature. Disturbances will enter in the feed rate of the oil to the furnace. The multivariable case is a grinding process of cement clinker consisting of a ball mill and a cyclone separator. The manipulated variables are the feed rate of clinker and the separator speed. Controlled variables are the elevator power and the fineness of particles in the product. The main disturbance to this process is the hardness of the cement clinker which affects the product size distribution and quality.