Large Scale optimization strategies for process operation have evolved sign
ificantly over the past two decades. Currently, continuous variable optimiz
ation problems (nonlinear programs) can be solved on-line Fur steady state
refinery models with several hundred thousand variables. Moreover, efficien
t NLP strategies have been developed for dynamic optimization problems. Sti
ll, the next step, on-line optimization of large dynamic chemical processes
, requires the tackling of a number of limitations and research challenges.
Many of the advances in NLP algorithms have taken place by recognizing and
exploiting the framework of Successive Quadratic Programming (SQP) algorit
hms. These are extensions of Newton type methods for converging to the solu
tion of the KKT (optimality) conditions of the optimization problem. Moreov
er, a number of innovations in algorithm design and problem formulation can
greatly improve performance. ks a result, very fast NLP algorithms can be
derived for data reconciliation, parameter estimation, nonlinear model pred
ictive control and dynamic optimization. Moreover, inequality constraints a
nd variable bounds can be treated through advances in interior point strate
gies. These methods preserve the particular problem structure and scale wel
l in performance for large-scale problems with many constraints. In particu
lar, we will consider the application of these strategics to problems in No
nlinear Model Predictive Control (NMPC). In parallel to the development of
efficient nonlinear programming algorithms, we also need to consider the in
corporation of nonlinear models that are accurate but also easily identifie
d and implemented. Here we consider DABNet (decoupled A-B neural network) m
odels which have very desirable approximation properties for a wide variety
of nonlinear systems. These models can be incorporated easily into the NMP
C framework and lead to very efficient online control algorithms. Finally,
the concepts developed in this paper will be illustrated by several case st
udies drawn from our previous work. These examples illustrate the importanc
e of nonlinear model predictive control and emphasize the need for efficien
t online computation.