High-performance Industrial Embedded Model Predictive Control: Efficient implementation of step response models and fast solvers
Abstract
The implementation of embedded Model Predictive Control (MPC) is treated in
this thesis with focus on new and challenging applications such as subsea production
and processing. The current technology for MPC in the petrochemical industry
relies on lower-level controllers and dedicated safety-systems. As a consequence, the
computer technology used for MPC and the software solutions do not necessarily
meet the oil and gas industry’s standard for safety and reliability in stand-alone or
highly automated operations. This thesis therefore aims at providing contributions
that enable the use of MPC on ultra-reliable industrial computer system hardware
such as programmable logic controllers.
Since such industrial hardware typically have limited computational resources,
efficient optimization algorithms for MPC form a natural part of this thesis. Furthermore,
the industrial focus of this thesis has motivated the combination of
academic advancements in the development of fast convex optimization problem
solvers with MPC problem formulations that reflect industrial practice. In particular,
the efficient implementation of step response model based MPC is the main
focus, and the outcome is an embedded step response MPC scheme that has low
computational complexity. A low complexity MPC scheme implies that accurate
control is achieved with minimal computational effort. For this reason, strategies
that enhance the computational efficiency and robustness of step response prediction
models were considered, and an implementation based on the recursive
representation of step response data was found to produce the most efficient and
reliable computations.
This thesis also facilitates the choice of a suitable solution method for the existing
(or traditional) way of formulating the step response MPC problem, which relies
on explicit output predictions. It is shown that both first-order and second-order
solution methods (or solvers) are capable of solving the traditional step response
MPC problem efficiently. However, the performance properties of different solvers
are affected in different ways by the MPC problem formulation. Different levels of
structure exploitation are achieved by different solver implementations, and careful
tailoring of the solution method to the problem structure is identified as a key
factor for high computational performance. Different tailored (sparse) solvers, implementing
a primal-dual interior-point method and a preconditioned primal-dual
first-order method, were used to solve the traditional step response MPC problem
efficiently. However, for problem setups with relatively small number of inputs,
the results suggest the use of condensing techniques to produce a dense problem,
suitable for dense solvers.
In addition, the explicit prediction strategy used in the traditional step response
MPC formulation is identified as the main limitation that rules out the possibility
of using some well established solution methods and highly efficient solver implementation
techniques. This limitation is removed by introducing a new multistage
framework for step response MPC. As a result, tailored Riccati and condensing algorithms
for interior-point methods are also proposed and discussed in this thesis.