Independent thesis Advanced level (professional degree), 20 credits / 30 HE credits
Model predictive control techniques are widely used in the process industry. They are considered methods that give good performance and are able to operate during long periods without almost any intervention. Model predictive control is also the only technique that is able to consider model restrictions.
Almost all industrial processes have nonlinear dynamics, however most MPC applications are based on linear models. Linear models do not always give a sufficiently adequate representation of the system and therefore nonlinear model predictive control techniques have to be considered. Working with nonlinear models give rise to a wide range of difficulties such as, non convex optimization problems, slow processes and a different approach to guarantee stability .
This project deals with nonlinear model predictive control and is written at the University of Seville at the department of Systems and Automatic control and at the department of Automatic Control at KTH. The first objective is to control the nonlinear Four Tank Process using nonlinear model predictive control. Objective number two is to investigate if and how the computational time and complexity can be reduced.
Simulations show that a nonlinear model predictive control algorithm is developed with satisfactory results. The algorithm is fast enough and all restrictions are respected for initial state values inside of the terminal set as well as for initial state values outside of the terminal set. Feasibility and stability is ensured for both short as well as for longer prediction horizon, guaranteeing that the output reaches the reference. Hence the choice of a short respectively long prediction horizon is a trade off between shorter computational time versus better precision.
Regarding the reduction of the computational time, penalty functions have been implemented in the optimization problem converting it to an unconstrained optimization problem including a PHASE-I problem. Results show that this implementation give approximately the same computational time as for the constrained optimization problem. Precision is good for implementations with penalty functions both for long and short prediction horizons and initial state values inside and outside of the terminal set.
2007. , 39 p.