Nonlinear Model Predictive Control and Dynamic Real Time Optimization for Large-scale Processes
In order to distinguish essays and pre-prints from academic theses, we have a separate category. These are often much longer text based documents than a paper.
This dissertation addresses some of the theoretical and practical issues in optimized operations
in the process industry. The current state-of-art is to decompose the optimization
into the so-called two-layered structure, including real time optimization (RTO) and advanced
control. Due to model discrepancy and inconsistent time scales in different layers,
this structure may render suboptimal solutions. Therefore, the dynamic real time optimization
(D-RTO) or economically-oriented nonlinear model predictive control (NMPC)
that directly optimizes the economic performance based on first-principle dynamic models
of processes has become an emerging technology. However, the integration of the firstprinciple
dynamic models is likely to introduce large scale optimization problems, which
need to be solved online. The associated computational delay may be cumbersome for the
We first derive a first-principle dynamic model for an industrial air separation unit (ASU).
The recently developed advanced step method is used to solve both set-point tracking and
economically-oriented NMPC online. It shows that set-point tracking NMPC based on the
first-principle model has superior performance against that with linear data-driven model.
In addition, the economically-oriented NMPC generates around 6% cost reduction compared
to set-point tracking NMPC. Moreover the advanced step method reduces the online
computational delay by two orders of magnitude.
Then we deal with a realistic set-point tracking control scenario that requires achieving
offset-free behavior in the presence of plant-model mismatch. Moreover, a state estimator
is used to reconstruct the plant states from outputs. We propose two formulations using
NMPC and moving horizon estimation (MHE) and we show both approaches are offsetfree
at steady state. Moreover, the analysis can be extended to NMPC coupled with other
nonlinear observers. This strategy is implemented on the ASU process.
After that, we study the robust stability of output-feedback NMPC in the presence of plantmodel
mismatch. The Extended Kalman Filter (EKF), which is a widely-used technology
in industry is chosen as the state estimator. First we analyze the stability of the estimation
error and a separation-principle-like result indicates that the stability result is the same as
the closed-loop case. We further study the impact of this estimation error on the robust
stability of the NMPC.
Finally, nominal stability is analyzed for the D-RTO, i.e. economically-oriented NMPC,
for cyclic processes. Moreover, two economically-oriented NMPC formulations with guaranteed
nominal stability are proposed. They ensure the system converges to the optimal
cyclic steady state.