Abstract
For process plants, physics-based steady-state simulators are used in design
because they encode physical structure and predict behavior across operating
ranges. In operation, however, models are identified from plant data
to capture local, control-relevant dynamics, including unmodeled effects.
Yet operational data are often collected near set points, limiting excitation,
while simulators can be biased and poorly calibrated. Despite their
complementary capabilities, weaknesses, and strengths, the two sources are
typically used in isolation and do not systematically inform one another.
This thesis develops ways to embed steady-state simulator information in
dynamical models by combining simulated equilibria with time-series measurements
to (i) learn about simulator mismatch and calibration of simulators,
and (ii) improve operational model prediction. A principled statistical/
optimal treatment of this embedding has not yet been treated systematically—
especially for steady-state simulators coupled to dynamical models.
Methodologically, the thesis proposes simulator-informed modeling tools:
prior-based regularization, data-driven weighting/tuning, feedback-loop extensions,
and bias-aware Gaussian-process fusion with calibration. The
methods are designed to remain compatible with linearization-based control
design and keep the resulting models interpretable. Results show that joint
fusion–calibration improves steady-state consistency and local prediction in
control-relevant regions when side information is handled in an uncertaintyaware
manner. Analytical steady-state information can improve feedback
loops, particularly when targeting slower transients. Overall, these tools
enable a more systematic and uncertainty-aware use of imperfect steadystate
simulators, particularly when simulator and measurement data are
sufficiently aligned. The methods enable the use of available information
without unsafe plant excitation or fully trusted dynamic simulators.