Customized model identification and optimal model -based control strategies with application to paper machine cross -directional control problems

The paper machine is one of the key processes within paper making. It is the final unit operation prior to sale of the final product. Since experimentation on an actual paper machine is very costly, the work here is on a benchmark model of a paper machine. The simulation includes many new sensor and actuator technologies and dynamics not typical in this type of simulation. Due to its high value and constrained control space, the benefits of optimal control strategies for the Cross Directional (CD) control problem motivate this work. CD control refers to the regulation of paper properties in the direction perpendicular to the traverse of fibers within a paper machine.;Linear models for use in Model Predictive Control (MPC) are identified both in open-loop using a least squares parameter fit and in closed-loop with unconstrained MPC using Markov parameters. Resultant identified models are compared to nominal models in MPC and allow the controller to reduce product variations.;Control of a paper machine that undergoes machine speed changes motivates the development of a variable time delay compensation technique. The variable time delay compensator uses the structure of the Smith predictor but modifies the delay model using an approximation of the wave equation. The reduction of uncertainty in the variable time delay compensator allows the controller to be more aggressive as compared to its fixed delay predecessor.;MPC based on linear objective functions (LP-MPC) is also explored with respect to the large input-output space of the CD control benchmark. First, LP-MPC is analyzed for its closed-loop behavior using multi-parameric linear programming. The use of common tuning heuristics is found to minimize deadbeat and idle control behavior, once commonly assumed to be part of the nature of LP-MPC. Efficient LP-MPC formulations are also considered. Combinations of techniques such as coincidence points and efficient mathematics are considered for both their performance and efficiency. Finally, the benchmark is augmented to simulate some of the efficient algorithms while accounting for effects of input delay due to controller computation times.