You are on page 1of 10

Computers and Chemical Engineering 30 (2006) 1426–1435

Model predictive control—Building a bridge between theory and practice


J. Brian Froisy
Aspen Technology, Houston, TX 77042, United States
Received 28 February 2006; received in revised form 15 May 2006; accepted 16 May 2006
Available online 25 July 2006

Abstract
This paper describes the development of an MPC product at AspenTech. A brief historical view sets the stage, describing collaboration with
university researchers that formed much of the technical foundation. Gaps between theory and practice were addressed during product development.
Major effort focused on eliminating the need for practitioners to understand control theory, while ensuring that controllers satisfy theoretical
requirements. This allows control strategies to be designed and implemented via configuration and tuning in a domain familiar to a process
operations engineer. Terms like eigenvalue, covariance, detectability, controllability, etc. are never exposed. Robust numerical algorithms and
software suitable for real time execution must augment the control theory foundation. Best practices and expertise are embedded in the software
realization, thus reducing the skill level required for success while still allowing experts to push the envelope. The resulting technical scalability
facilitates a wide range of process control practice in a single framework.
© 2006 Elsevier Ltd. All rights reserved.

Keywords: Model predictive control; MPC; Industrial process control; State-space

1. Introduction available theory without being encumbered by such relics. Past


practice was retained when appropriate, however. The resulting
Model predictive control, MPC, has influenced process con- product is formally known as the APC State-Space Controller,
trol practice significantly during the past 20 years. Early MPC a module of the aspenONETM Advanced Process Control Solu-
technology (circa 1976) evolved principally in industrial set- tion. The term SSC is used in this paper for brevity.
tings, followed by copious numbers of academic papers ana- AspenTech and predecessor companies had supported vari-
lyzing and extending the underlying theoretical basis. Some ous university research consortia. Much of the theoretical basis
MPC applications were one-of-a-kind, implemented for specific for SSC can be traced to research from the Texas-Wisconsin
applications. Others were based on more generally applicable Modeling and Control Consortium, TWMCC, starting with its
commercial products. Two survey papers by Qin and Badgwell inception in 1993. This long-term relationship resulted in a
(1997, 2003) describe MPC products available during the past research thread that built incrementally on prior research. Incor-
10 years. These surveys summarize product features that relate porating new results was much easier than embracing new,
to important issues in process control practice. potentially outstanding, results that did not quite fit into this
Theoretical insight into MPC increased rapidly, but actual continuum. A hallmark paper by Muske and Rawlings (1993)
practice evolved more slowly. Practitioners, working with tools defines the starting point. Much of the product development
that were firmly rooted in the past, faced increasingly chal- involved filling theoretical and practical gaps from this initial
lenging applications that forced them to develop clever ad hoc point. Gap closure came from both academic research and inter-
solutions to address real-world issues. Unfortunate side effects nal AspenTech efforts.
of evolution, today’s MPC technology and products often con- The paper proceeds as follows. A short historical review of
tain “relics of the past” similar to unused genes found in liv- MPC provides an overall perspective. Then, the motivation for
ing creatures. Eventually, AspenTech began development of a a new product is described. This includes specifying the range
new product with the philosophy to start fresh, using the best and types of control applications to be addressed. In a sense,
this defines a certain view of process control practice and tools
needed to achieve successful applications. Gaps between theory
E-mail address: Brian.Froisy@aspentech.com. and practice are described in a separate section. An additional

0098-1354/$ – see front matter © 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2006.05.044
J.B. Froisy / Computers and Chemical Engineering 30 (2006) 1426–1435 1427

section is devoted to a few simple examples illustrating impor- integrators and various parameters to deal with the finite model
tant technical advantages compared with previous generation dynamic move plan calculation are examples.
products. No CPC paper would be complete without predictions A notable paper was published between CPC 4 and CPC 5
about future directions. These are given in a forward-looking (Muske & Rawlings, 1993). It provided a framework for infinite
section, which outlines areas for future research. Finally, con- horizon, constrained MPC that matched real-world needs fairly
cluding remarks are made including answering the question: well. For example, a structure for disturbance models that meets
why did it take so long? important requirements and demonstrates nominal stability for
constrained control was described. Another important CPC 5
2. Historical perspective paper by Wright (1997) shows how to formulate a general (lin-
ear) MPC optimization problem for efficient computation.
Early papers (Cutler & Ramaker, 1980; Richalet, Rault, There were no specific MPC sessions at CPC 6, but related
Testud, & Papon, 1978) document the generally accepted begin- papers appeared among other sessions. Perhaps this is a mark
ning of industrial MPC. The initial work was probably being of maturity. Eastman Chemical Company documented its
done at about the time of CPC 1 in 1976. A long sequence experience with infinite horizon state-space MPC (Vogel &
of MPC-oriented papers have appeared at every CPC since Downs, 2002). Their implementation was based on the 1993
then, including one on “Model Algorithmic Control” at CPC Muske and Rawlings paper. Several papers addressed MPC with
2 (Mehra, Rouhani, Eterno, Richalet, & Rault, 1981). nonlinear models, including one (Young, Bartusiak, & Fontaine,
Five years later, at CPC 3, a session was devoted to MPC 2002) that describes development of the NLMPC algorithm at
with focus on industrial applications. Although rather simple ExxonMobil Chemical Company in conjunction with university
by current standards, the applications were clearly economi- and private collaborators. Young describes the technology as
cally significant. Two papers (Froisy & Richalet, 1986; Garcia “by no means mature” but reminiscent of linear MPC of the late
& Prett, 1986) foretold of new directions for IDCOM and DMC, 1980s. This is probably a realistic assessment, suggesting that
the notable early MPC variants. An important theme was the maturation is now occurring. The rate at which this and compet-
need to address a wide range of real-world issues such as multi- ing nonlinear MPC packages evolve will certainly be a function
objective and economic optimization, dealing with hard and soft of the economic benefits they achieve. In today’s competitive
constraints, and robustness to modeling errors. environment, compelling economic benefits are essential to
There were two sessions on MPC at CPC 4 in 1991, obtain product and technology development funding.
attesting to its growing popularity. Some papers described The trajectory of MPC since CPC 1 has gone from initiation,
industrial applications. Others from the academic community through rapid growth impacting process control practice and
addressed theoretical issues such as stability of constrained theory in important ways. It may appear to have reached a steady
MPC (Zafirou, 1991). Two papers (Morari & Lee, 1991; Ricker, state—at least for linear MPC. However, a step change is on the
1991) dissected industrial MPC approaches, essentially reverse- horizon even for linear MPC.
engineering them in terms of traditional control engineering
theory. The analysis comparing theory and practice, notably with 3. A new MPC product—why?
practice coming first, was an important step. It exposed deficien-
cies both in practice and theory, thereby helping set new research Why would a company invest in a new linear MPC product?
directions. MPC was interpreted in a state-space framework, an And what does this have to do with process control practice? A
important step needed to meld MPC with established theory such few short answers follow.
as linear quadratic control. However, MPC practice was largely Despite being mature, existing MPC products are not suit-
formulated as a constrained optimization problem meaning that able for all applications or they do not capture all the benefits
significant gaps still existed between theory and practice. possible when they do apply. Compromises are inherent with
By the time of CPC 5 in 1996, the single MPC session had finite models and/or finite prediction and control horizons.
a “recent advances” paper and one on nonlinear MPC. Also The typical approach of reconciling model predictions with
notable was a survey and overview of industrial MPC (Qin measurements by simply biasing predictions is clearly not
& Badgwell, 1997) describing various commercially available optimal. Shortcuts in formulating the “dynamic move plan”
MPC products. These products which had largely evolved during optimization due to computational complexity result in subop-
the previous 10 years tended to be descendents of IDCOM and timal control, even if an optimal estimate of the future process
DMC. A few such as SMOC and RMPCT were not direct descen- evolution were available. Of course, these are the things that
dants. SMOC has the distinction of using state-space models and the academic community had been saying since first dissecting
feedback via a Kalman filter, unlike the other cited products. early industrial MPC technology.
More recent products, such as one from ABB, also use a state- There are important non-technical reasons to invest in a new
space formulation. The existence of several products confirmed controller. The high level of expertise required to implement and
the commercial significance and maturity of MPC technology. A maintain MPC is an impediment to more widespread utilization.
detailed look at the survey exposes the fact that these products Long-term total cost of ownership is an undeniable issue for both
had ad hoc features needed to deal with inherent limitations, vendors and end user companies. The need to achieve superior
such as those noted by the academic community. Lack of an performance, even with non-expert applications engineers, is a
explicit disturbance model, special parameters for models with driving force in today’s worldwide competitive environment.
1428 J.B. Froisy / Computers and Chemical Engineering 30 (2006) 1426–1435

The petroleum refining industry greatly influenced the evolu-


tion of MPC. There the emphasis was on control near operating
limits, often driven by a constraint-seeking internal optimiza-
tion or simply the objective to maximize production. Certainly
dynamic control was important, but not as strong a driving force.
Product blending, a normal industry practice, inherently reduces
the need to satisfy product quality on a minute-by-minute basis.
Some, but certainly not all, chemical manufacturing has simi-
lar characteristics. Short-term quality is often more important
than for petroleum refining. And quality becomes increasingly Fig. 1. MPC functional blocks.
important as manufacturing gets closer to the end user. An impor-
tant objective of the new MPC product was to enable improved inputs, U, and measured disturbances, F. The estimator com-
dynamic control for processes where product quality is empha- pares measured outputs, Y, with model predictions to update
sized, while retaining the ability to optimize and operate at the state estimate. Optimizer objectives, O, includes operating
constraints if necessary. limits and targets plus optional economic costs with priorities
There are three important technical characteristics of SSC: defining how conflicting objectives or infeasibilities should be
handled. The control block calculates the optimal move plan
• state-space model; based on the estimated state and feasible targets determined by
• state estimation; the target optimization.
• infinite horizon, constrained dynamic move plan. Control performance improvements are definitely expected
from this approach with explicit state estimation. A disturbance
These enabling technologies will help address an increas-
model that augments the basic plant model is a prerequisite.
ingly challenging set of industrial control problems. Physical
The (implied) disturbance model when simply updating out-
systems are naturally represented using the state-space model
put biases simply cannot do effective state estimation. Future
structure. Real processes have internal connections (states) and
predictions are inherently limited to knowledge of past inputs.
can exhibit a range of fast and slow dynamics, including inte-
Regardless of model structure, the internal process dynamic state
grating and unstable modes, all of which are naturally described
does not get feedback with bias-only updating.
in a state-space structure. State estimation provides the mech-
Improved performance from state estimation comes at a
anism to detect disturbances entering the plant. Once detected,
cost. The applications engineer is required to configure and
disturbances can be optimally rejected by the infinite horizon
tune a disturbance model. An important challenge was to
controller.
design an environment that facilitates this step when done by
From a high level, the major steps in all approaches to MPC
a process operations engineer lacking formal control theory
can be summarized with the following questions:
training.
• Where is the process now? The infinite horizon constrained dynamic move plan makes
• Where is it going, if there is no intervention? it possible to find the true optimal future trajectory, based on the
• Where do we want it to go? optimal state estimate. Prior industry practice always involved
• What is the best plan to get it there? compromise. Manipulated variables were calculated for a lim-
ited number of future points, often with some type of pattern
The characteristics of all MPC variants are effectively deter- that determined when changes would be allowed. This has been
mined by how they address these questions. referred to as MV blocking or compaction. It results in the
State estimation answers the first two questions. The estimate need to recalculate the move plan at each time step, even when
of current process outputs, given all available measurements, there are no changes affecting the process! Nominal stability
answers the first one. The future trajectory, where the process is was likely, if not guaranteed, when the move plan extended far
going, is calculated by evolving the state estimate into the future. enough and the final MV values were constrained to be at steady
A good state estimate is important because If you do not know state. Similar restrictions were sometimes placed on the future
where the process is and where it is going, your control plan samples where controlled variables, and hence the optimization
cannot be optimal. objective function, were calculated. These compromises are also
The fact is that commercially available products have largely relics of the past, forced by the need to simplify time-consuming
ignored this basic premise. For the most part, feedback enters by optimization computations.
applying an output bias to model predictions. There is no true An infinite horizon formulation largely eliminates compro-
state estimation. Why? This is a relic of the past. mises, but is computationally demanding. Vastly improved com-
Fig. 1 is a concise view of important functions required to puter power alone was not enough to make it practical. Exploita-
answer the four basic MPC questions. The Model and Filter tion of problem structure was the key. Part of our internal devel-
blocks perform state estimation. Note that the model is the aug- opment was to do this and extend published results to gain even
mented (plant + disturbance) model. greater benefit.
The estimated state, X, includes plant states and estimates of Control performance improvements can be expected with the
the unmeasured disturbances, D. The model “sees” manipulated “no compromise” infinite horizon move plan. Results for the
J.B. Froisy / Computers and Chemical Engineering 30 (2006) 1426–1435 1429

unconstrained case are identical to an optimal regulator—after • Is calculation of a “no compromise” infinite horizon, con-
all, this is the optimal plan. Current generation commercial con- strained move plan really feasible?
trollers simply do not have this property. • What special initialization steps are required?
Performance from state estimation comes at the price of • How do issues like asynchronous, missing or invalid data
defining a disturbance model and tuning the resulting filter. impact the nice theoretical basis?
Performance improvements from an infinite horizon move plan
comes with a computational price when compared with prior Of course, the real list was longer—but not initially evident.
shortcut approaches. On the other hand, dynamic controller con-
figuration and tuning is actually simpler. 4.1. Prototype—Phase 1
The answer to: why build a new MPC controller? Provide
higher performance and make it easier for non-experts to actu- A significant internal project commenced in late 1999 to
ally configure, tune and maintain real-world MPC applications. build a proof-of-concept prototype based mainly on published
The combination of state estimation and an infinite horizon con- theory and algorithms. The internal code name was MrC, for
strained move plan provides the technical foundation needed to Muske–Rawlings Controller. The objective was to answer some
meet these objectives. of these questions and, in particular, mitigate risk before pro-
The secondary question was how this all relates to the practice ceeding with all-out product development. State estimation and
of process control. Current practice with MPC is guided by the infinite horizon move plan calculation, the two main techni-
capabilities and limitations of available technology and prod- cal advances, received major attention. Results were encour-
ucts. Configuration and tuning parameters inherent in the tech- aging, at least to the “true believers”. This effort closed a
nology and/or product dictate, or at least influence, the project few theory-practice gaps, but mainly helped bring them into
workflow. Our product development, at an early stage, contin- focus. One notable gap was related to the usual assump-
ually addressed the question of how to expose the underlying tion that disturbances enter as random steps, regardless of
capability to the applications engineer responsible for config- their entry point. This does not characterize the real world.
uring and/or maintaining the controller. There was a continual The solution was to incorporate more general disturbance
tension between the desire for simplicity and that for extended dynamics.
functionality to satisfy an increasingly diverse class of control An important outcome in this phase was construction of a pro-
problems. Were we building this for the novice user with simple totype layer isolating the end-user engineer from control theory.
applications or for the expert capable of pushing the envelope? This demonstrated important “ease of use” ideas that persist in
The simple answer is: Yes! We were building it for both. SSC.
This dual objective was challenging. A guiding principle was Specification of the disturbance model illustrates how the user
to hide complexity, exposing as few choices and parameters as is isolated from theory. The augmented (plant + disturbance)
feasible. Moreover, these should be mapped into the domain of model must be detectable and there can be other restrictions
the process operations engineer whenever possible. In the case of depending on the rank of the state to output (C) matrix. The user
estimator tuning, the user does not enter disturbance and mea- builds a disturbance model by specifying “disturbance channels”
surement covariance matrices. These parameters are mapped corresponding to plant model inputs and outputs. Invalid com-
into “typical disturbances” and noise ratios by a layer isolating binations are blocked. Also, an indication of “how observable”
the user from underlying control theory. the specified disturbances will be is provided to assist the user.
Hiding complexity and theory was not enough. We also This is all done in a point-and-click environment appropriate for
wanted to provide a built-in workflow with a balance of structure a process operations engineer. Behind-the-scenes calculations
and flexibility to support a methodical approach to controller guarantee that the resulting estimator will be stable and offset-
configuration, tuning and maintenance. Beyond that, it was free in the outputs. Furthermore, offset-free control is guaranteed
desirable to provide smart tuning defaults often based on plant when there are sufficient degrees of MV freedom. This detail is
operating data. It remains to be seen how these concepts impact one of the gaps closed during the early development phase illus-
process control practice. Initial response has been encouraging. trating collaboration with the academic community (Muske &
New users are able to quickly develop control applications, with Badgwell, 2002).
limited training. A detailed math document was developed during this early
phase, and updated as necessary. It was based on TWMCC
results and other open literature. This exercise forced the disci-
4. Gaps between theory and practice pline needed to expose issues requiring attention and helped for-
malize assumptions when necessary. In retrospect, the approach
Theory and “toy problem” simulations clearly indicated that proved to be important in guiding the team throughout product
a new generation of MPC could frequently outperform available development.
products. Some important questions were:
4.2. Prototype—Phase 2
• Will the performance exist for practical applications?
• What surprises will emerge using infinite state-space versus A second prototype phase focused on closing gaps between
finite response models? theory and practice. A “textbook” fixed gain Kalman filter
1430 J.B. Froisy / Computers and Chemical Engineering 30 (2006) 1426–1435

had been used in the first phase, ignoring important issues optimization). A self-imposed requirement was that none of the
such as: online calculations would use Matlab. This was another step in
producing an industrially hardened system. The prototype was
• filter initialization; installed on a large process unit at a major chemical company.
• instrumentation failure; This was not a “toy problem”. The plant model had 15 inputs,
• asynchronous measurements, including lab data; 25 outputs and 120 states.
• operator intervention. The limitations of an output bias (i.e., a model that assumes
disturbances enter directly into the outputs) became apparent
These were ultimately addressed with a “dynamic Kalman during this stage. Input disturbances were available and eas-
filter”. It is initialized in a smart way to allow immediate use ily constructed. However, they are not always the right choice.
of state estimates. The gain is time-varying depending on mea- They are most useful when unmeasured disturbances are likely
surement availability. It will approach the steady state gain if to enter with dynamics similar to process inputs (manipulated or
all measurements are available synchronously at each sample feedforward variables). Detectability of the augmented model is
instant. theoretically required and offset-free output estimation is also a
Nice guarantees apply when all measurements are available. necessity for practical applications. These restrictions must be
Measurement loss due to operator action or instrumentation fail- imposed.
ure invalidates the guarantees. Is estimation still offset-free when The dilemma was how to specify state disturbances while
an output measurement is not available? The answer is yes for retaining the simplicity of output bias updates. Direct specifica-
the remaining outputs, but this was not immediately obvious. Of tion of state disturbances was not an option for two reasons. It
more concern, what happens if a measurement loss results in an would have been inconsistent with the guideline to hide com-
undetectable system? The online software (built at a later stage) plexity. Secondly, the use of empirical state-space models meant
checks detectability when measurement status changes. Critical that individual states do not have understandable physical mean-
alerts are sent to the operator when necessary. Non-critical loss ing. There would be no basis for choosing specific states as the
is handled more leniently. The net result is that the estimator can, location where disturbances entered. The solution was to find a
within reason, continue to operate without the full complement mapping from outputs to states, similar to the mapping between
of measurements. Also, changes in measurement availability do inputs and states used for input disturbances. As a result, the
not cause abrupt changes in the state estimate. estimator configuration step allows disturbances to be specified
Published results for efficient move plan calculation did not in terms of outputs with the map being automatically calculated.
include a direct pass-through term in the model. This was a short- A mix of input, state (based on the map) and output biases can
coming for the intended class of problems. A specific example be specified, subject to conditions that guarantee detectability
is the output of a PID controller when the setpoint changes. The and offset-free output estimation. This was significant gap filled
“proportional kick” results in immediate valve movement. It is during product development.
common for valve (in general, actuator) position to be a con- State estimation and infinite horizon control were the two
strained output in the MPC controller. Tedious re-derivation and main new areas, but by no means the only challenges. In a state-
programming was required. space framework, infinite horizon control requires calculation
The second prototype phase successfully closed the major of steady state, i.e. state values at steady state. A numerical
theory-practice gaps. The estimator proved to be the place where optimization is required in the general case. States are decision
most of the real-world issues had to be addressed. In retrospect, variables when a straightforward approach is used. This can
there were no real gaps in the theory. The discipline enforced result in a time-consuming calculation. This issue was addressed
while building a “real world” prototype provided the framework by reformulating the optimization to reduce the number of deci-
to adapt practice to theory, retaining the objective that applica- sion variables. This is another example of taking advantage of
tions engineers would not need a formal understanding of control problem structure to formulate a numerical solution that is both
theory. efficient and robust. Particular attention was needed for mod-
els with integrating states. The end-user never sees any of this
4.3. Prototype—Phase 3 detail.
During the phase three prototype, we began addressing
A third prototype phase began late 2002. The objective was to important modeling issues. Previously, it was assumed that a lin-
construct a system (not a product) suitable for actual industrial ear discrete state-space model was available. Ideally, the plant
application. The idea was to embed the new controller in a real- model can be identified directly as a state-space model. Vari-
time framework supporting such things as an operator interface, ous subspace model identification algorithms including a variant
DCS connectivity, data validation, etc. Some of the framework available in the DMCplus Model package could satisfy this
came from existing DMCplus® software and some was new, need. An alternative scenario is that there is a pre-existing finite
consistent with long term plans for a future real-time control response model. Software to convert these into state-space, using
and optimization environment. published algorithms, was constructed to address this need.
The early prototype software included a mix of Visual Basic Practitioners often use the term “model” to describe a specific
(for the user interface), Matlab (control technology) and For- input–output relationship. This could be a finite model repre-
tran/C++ (steady state optimization and dynamic move plan sented as a step response or possibly a SISO transfer function
J.B. Froisy / Computers and Chemical Engineering 30 (2006) 1426–1435 1431

fewer states can also provide more robust state estimation. One
particular algorithm was chosen after evaluation of several pub-
lished approaches. There was not any satisfactory approach to
model reduction for integrating and unstable states. This was
resolved with an internally developed algorithm for the unstable
partition.
Model reduction provides another example of how the engi-
neer is isolated from theory. Hankel singular values are used to
determine a cutoff point for state order. However, the engineer is
presented with a choice in terms of the number of states denoted
as: small, medium, large, and custom. The first three options
come from a specific range of Hankel singular values and Cus-
tom provides freedom when the pre-determined options are not
satisfactory. In any case, a user is presented with an overlay plot
of step responses for the original and reduced models to aid in
Fig. 2. Model construction workflow.
choosing the final size.
The technical issues encountered in the phase three proto-
model. An assembly of these (SISO step response or trans- type turned out to be mainly modeling related. In the end, it was
fer function models) describing all inputs and outputs is often possible to retain an input–output perspective for both plant and
used as the overall model. This input–output assembly does disturbance models. This required filling some important gaps
not reflect physical reality since multiple internal connections in published approaches. Significant effort went into numeri-
between inputs and outputs normally exist. A MIMO state-space cal algorithms, especially when dealing with integrating states.
structure is a more natural representation of the physical world. Additional software effort was required to host the new state-
Dynamic models based on physical principles have this struc- space controller technology in a real-time environment with
ture, although they are normally nonlinear. support for DCS data connectivity and a human interface.
It is possible to construct a MIMO state-space model from
an input–output assembly. Restrictions exist on the model real- 4.4. Prototype—Phase 4
ization. It must be a minimal realization to guarantee important
properties of the estimator and controller. The final prototype phase extended the previous work with
A flexible model construction workflow to obtain a plant an application at another chemical company. The plant model
model in state-space form was developed during the project. It is has 8 inputs, 20 outputs and 109 states.
summarized in Fig. 2. Individual sub-models (SISO or MIMO) Two new practical issues were addressed. One output (an
from various sources can be merged into a final MIMO state- important product quality, measured by an online GC) had a
space (minimal) realization. The engineering end-user continues nonlinear response. Linearization was applied using a log trans-
to work at the input–output level, with all the state-space detail form, thereby allowing the core model to remain linear. This is a
hidden. Wiener model. Hammerstein models (transformations on inputs)
The preferred path for empirical identification is via a sub- are also supported. This is typical industrial practice with vari-
space algorithm. FIR models are supported, allowing direct ous MPC products. Use of transforms with a state-space model
import of pre-existing finite response models. The model library does not introduce complexity beyond that for finite response
option provides a means to directly enter simple transfer func- models.
tion models. Physical dynamic models, from external sources, Control of calculated outputs is a second practical issue. The
are naturally in nonlinear state-space form. These can be used usual practice of performing identification directly on a calcu-
after appropriate linearization. lated quantity is not the best practice for state-space models.
Model assembly takes place in the “Merge” operation which State estimates are best done using directly measured outputs,
manipulates equations to construct the necessary state-space not calculations. The core plant state-space model should simply
matrices. Parallel, series and feedback connections are handled relate measured outputs to measured inputs in order to describe
automatically and exactly. the state equations most efficiently. Assuming the calculation
There was not any need to extend theory. The (significant) is a linear combination of measured inputs and outputs, calcu-
effort was to build a framework that allows the engineer to con- lated outputs can be appended to the state-to-output (C) matrix
tinue to work with an input–output mindset while supporting the and direct pass through (D) matrix. Calculated variables inherit
underlying state-space model. dynamics from the model in a natural way. Once the core model
The modeling story does not end once a minimal realiza- has been established, linear calculated outputs can be defined
tion describing the full input–output behavior is constructed. without having to repeat identification calculations or other
The state order can be too large for practical use. In our expe- “model construction” steps. Note that the estimator requires
rience, computational time for the infinite horizon move plan only measured outputs for the state update. The optimization
is approximately quadratic with respect to state order. Model and move plan employ calculated variables without any addi-
reduction is therefore needed for large models. Models with tional special configuration.
1432 J.B. Froisy / Computers and Chemical Engineering 30 (2006) 1426–1435

Fig. 3. Plant model.

We took advantage of the state-space structure during model


identification and construction in this prototype application.
Much of the identification was based on normal operating data,
greatly reducing the need for traditional step tests. This was pos-
sible for two reasons. The process had been manually controlled,
but with some stabilizing PID loops. Measured disturbances and
manual adjustments provide a suitable dataset for identification.
However, it was suitable only when process segments were taken
in isolation. Sub-models relating a few inputs and outputs could
be reliably identified. A few other sub-models were identified
based on specific step test data. All sub-models were then merged
into the main final model. This “merge” (Fig. 2) is effectively a
multivariable convolution plus feedback interconnections (due Fig. 4. Disturbance rejection comparison infinite horizon control.
to process recycles). This approach is not feasible with finite
models, especially when feedback connections are required.
Feedback performance is compared using different distur-
Another interesting model construction feature was exercised
bance models. The first assumes that disturbances enter directly
during the project. The process is open-loop unstable, as are
at the controlled output, y2 . The second assumes the entry chan-
many chemical reactor systems. Given normal data, it was pos-
nel is at the input, u. The third is also an input disturbance but
sible to construct sub-models with the stabilizing PID controllers
with the addition that the intermediate variable, y1 , is a mea-
mathematically placed in manual. An open-loop unstable plant
sured but uncontrolled output. The actual disturbance enters at
model (without PID controllers) was constructed via the merge
the input for all cases.
operation. This was used in analyzing process dynamics and PID
Simulation results are shown in Fig. 4. The top set of curves
loop tuning. While there was never any intent to allow the MPC
is the CV response, with MVs at the bottom. An unmeasured
controller to supervise the unstable system, it would have been
disturbance enters at time zero. Controller tuning is the same for
technically feasible. It is reasonable to expect that MPC control
all cases.
of unstable plants will occur when there is sufficient economic
The response with an output bias, denoted as curve 1, is
incentive to do so. Until now, this option simply did not exist.
sluggish as expected. The controller observes the disturbance
over time since its effect is dampened by process dynamics. MV
5. Examples movement is correspondingly slow.
The result with an input disturbance model (curve 2) is quite
Performance benefits from state estimation and infinite hori- different. The MV acts aggressively to reject the disturbance,
zon control are illustrated in this section. One example demon- with much better performance. Still, there is a significant tem-
strates improved disturbance rejection. It is well known that PID porary deviation from setpoint because of the process deadtime.
controllers can be tuned to reject disturbances more effectively The response is the best possible in the sense that the disturbance
than MPC that uses bias updates (Shinskey, 1994). Addition of model accurately characterizes the actual disturbance.
state estimation eliminates this performance difference and fur- The third case (curve 3) shows the response when the inter-
thermore applies to multivariable applications. mediate variable y1 is also measured with the input distur-
The second example demonstrates calculation of a complex bance model. This achieves almost perfect rejection since the
“no compromise” move plan. Subtle issues related to the infinite disturbance is detected earlier (before the deadtime) via the
horizon formulation are illustrated. intermediate measurement. Note that the MV response is less
aggressive than for case 2. Use of the intermediate variable gives
5.1. State estimation performance similar to traditional cascade control. Equivalent
behavior in an actual application on an extractive distillation
This example demonstrates controller performance with dif- process has been observed. For the general case, state estima-
ferent disturbance models and optional intermediate measure- tion using intermediate measurements results in “multivariable
ments. Fig. 3 is a block diagram of the plant. Physically, this cascade control”. This is a native capability that does not require
could correspond to a distillation column with steam (u) as additional user configuration.
the manipulated input, an intermediate uncontrolled temper- The second example illustrates behavior of the infinite hori-
ature (y1 ) and a composition (y2 ) as the primary controlled zon controller in a highly constrained scenario. This demon-
output. strates how a very complex control move plan can be calculated
J.B. Froisy / Computers and Chemical Engineering 30 (2006) 1426–1435 1433

Fig. 5. Infinite horizon move plan.

when required to obtain optimum performance. The plant is these hard constraints. Of course, the MV would not actually vio-
essentially two stirred tank reactors in series with four con- late the limit because the constrained horizon segment is always
trolled quality parameters plus operating limits on other outputs. guaranteed to satisfy MV limits.
Fig. 5 shows MV and CV responses that result from changing Note that the infinite horizon formulation differs from
the setpoint for quality variable CV9 from 450 to 400. A quick Muske’s original approach that forces MVs to reach the steady
transition to the new setpoint is desired, while maintaining CV10 state values at a finite horizon (Scokaert & Rawlings, 1996).
and CV11 at optimum values and other variables within limits. This restriction means that the resulting move plan is “less opti-
Grey regions visible in most subplots represent upper and lower mal” than the infinite horizon with a linear regulator. It is clear
limits. in the example that the MV trajectories would have to be signifi-
There are two responses for each variable. The thick line is cantly different if they were restricted to reach steady state after
the result when calculating 100 future MV values, correspond- 2 h. MV constraints include hard rate-of-change limits. These
ing to a 5-h constrained horizon. (The model sample period is are evident during the first 2 h when several MV responses are
3 min.) The thin line is the case with 40 future constrained values straight line segments. CV positional constraints are always soft
followed by the unconstrained infinite horizon linear regulator meaning that they can be violated if necessary to satisfy more
(u = −Kx). Forty samples correspond to 2 h, the first vertical grid important competing objectives. In the example, CV4 is con-
line. There is very little difference between the cases because strained near a low limit during the initial 2-h period.
few constraints are active after 2 h. A notable exception is for Fig. 6 zooms in to expose more detail, illustrating how the
MV3. It would violate the limit during the unconstrained infi- soft constraint is violated with the original (slow) tuning. The
nite horizon segment for about 2 h. The formulation guarantees dynamic optimization is penalizing this violation along with the
that MVs approach the feasible steady state eventually satisfying deviation of other CVs from their setpoints or beyond limits.
1434 J.B. Froisy / Computers and Chemical Engineering 30 (2006) 1426–1435

tant issue will be the ability to deal with MVs that have discrete
values. The underlying optimization is a mixed integer problem,
for which solutions exist. Still, issues unique to real-time control
optimization will have to be resolved, some of which are being
addressed (Bemporad & Morari, 1999). Also, it is reasonable to
expect that batch processes will require nonlinear models thus
adding to the challenge.
Despite the process industry focus, MPC can certainly be
applied to a wide range of processes. The technical foundation
is solid. That is not always sufficient. Highly regulated indus-
tries such a pharmaceutical or nuclear power are examples where
MPC has not, to this author’s knowledge, been applied. In fact,
Fig. 6. CV soft constraint response.
one might be able to make a case that the MPC mindset is exactly
the right approach to take these industries into the world of
advanced process control.
The additional response (dashed line) denoted as fast tuning A driving force in today’s competitive manufacturing envi-
shows that this particular violation was avoided by adjusting the ronment is the need for holistic solutions that are maintain-
dynamic optimization penalty weight. The user is not confronted able in the face of change. Stand-alone MPC applications can
with optimization weights, however. A visual interface depict- address specific operational problems, often with significant
ing a tuning dial that can be adjusted for “faster” or “slower” economic impact. More complex ones are not isolated and
response for any of the dynamic control objectives is provided must cooperate to achieve broad business goals. This is particu-
instead. larly true when economic optimization spans the process oper-
ations automation and business automation layers. There will
6. Future developments be an increasing need for model consistency as these domains
overlap.
This MPC environment marks a new beginning with control
theory and software to support a wide range of users and applica- 7. Conclusions
tions. The framework will facilitate incorporation of recent and
future control theory results for even better performance and The title contains the phrase “building a bridge between
increased ease of use. It is also suitable for extension to other theory and practice”. This describes much of the product devel-
classes of problems. For example, academic and practical results opment activity. A major objective was to develop an MPC
exist today that extend into the world of MPC using nonlinear environment, not just a controller, to enable engineers without
models. The software framework has been architectured to sup- formal control theory training to efficiently build, deploy and
port this kind of technology. The timing of new functionality maintain control applications. This included the objective that
will depend on market demand. the environment would be suited to a wide range of control prob-
Near-term enhancements are likely. The ability to tune a lems, especially in industries where the benefits of MPC have
Kalman filter using data from closed loop MPC operation based not been fully realized.
on the ideas of Odelson and Rawlings (2003) is one example. In retrospect, the theoretical basis existed or became available
This will make it possible for users, even non-experts, to effi- during the overall project. There was no need to compromise on
ciently enhance and maintain performance throughout the life theory to satisfy real-world issues. It did take significant effort to
cycle of a controller. understand and resolve the apparent gaps. Ultimately, the gaps
An obvious requirement for a successful MPC application is were filled by building a software environment that encapsulates
the model. Techniques for plant testing and model identification a great deal of theoretical and practical expertise. This is not a
are continually being improved. It is reasonable to anticipate single bridge; rather it is a continuum of interlocking bridges
that these will be integrated with the new environment so that connecting theory with practice.
models can be efficiently and accurately developed. One of the Several iterations of prototypes proved effective for finding
important challenges will be to provide the means to blend and resolving issues. Major risks were addressed first. It came
empirical and physical modeling. This can take the form of con- as no surprise that a solid numerical foundation would be a
strained identification (e.g., some steady state gains are known) critical need. Much of the linear algebra and control-specific
or parametric identification (a model structure is imposed, with software were readily available. Other software, especially for
unknown parameters). The general idea is to allow the user to optimization algorithms, was tailored to fit the specific problem
impose (incomplete and possibly uncertain) physical knowledge domain. The numerical foundation was incrementally improved
when fitting models to data. throughout the project, especially as stress testing demonstrated
MPC will surely be applied to batch processes, despite the subtle weaknesses.
continuous process orientation of current MPC theory and prac- Subspace algorithms were developed in parallel to support
tice. The concept of driving a process to a specific state at the end state-space model identification. An unexpectedly large effort
of a batch is all too natural for this to not be the case. An impor- was required to build an environment that makes it easy for the
J.B. Froisy / Computers and Chemical Engineering 30 (2006) 1426–1435 1435

user to assemble model fragments (input–output subsets) into a Mehra, R. K., Rouhani, R., Eterno, J., Richalet, J., & Rault, A. (1981). Model
final model and also reduce the model order when required. algorithmic control (MAC): Review and recent developments. In D. E.
Seborg & T. F. Edgar (Eds.), Proceeding of the engineering foundation con-
So, what is the answer to: “why did it take so long?”. The
ference (CPC2) (pp. 287–309).
project involved much more than building a new controller. It Morari, M., & Lee, J. L. (1991). Model predictive control: The good, the bad
was an entire MPC product, including a real time environment to and the ugly. In Y. Arkun & W. H. Ray (Eds.), Proceedings of the fourth
deploy and manage applications. The relentless push to make a international conference on chemical process control (pp. 419–444).
system suitable for non-experts without compromising the the- Muske, K. R., & Badgwell, T. A. (2002). Disturbance modeling for offset-
oretical foundation increased demands on numerical algorithms free linear model predictive control. Journal of Process Control, 12,
617–632.
and software. In the end, the guiding principle went beyond Muske, K. R., & Rawlings, J. B. (1993). Model predictive control with linear
trying to “hide complexity” to become “expose simplicity”, a models. AIChE Journal, 39(2), 262–287.
term coined by my colleague Tom Badgwell. While the result Odelson, B. J., & Rawlings, J. B. (2003). Online monitoring of MPC disturbance
may not be as simple as possible, it is certainly much simpler models from closed-loop data. In Proceedings of the American control con-
than it would have been without this mindset. And, the journey ference (pp. 2714–2719).
Qin, S. J., & Badgwell, T. A. (1997). An overview of industrial model predictive
continues—toward increased functionality and simplicity. control technology. In J. C. Kantor, Garcia, & B. Caranahan (Eds.), Pro-
ceedings of the fifth international conference of chemical process control,
Acknowledgments Vol. 93 (pp. 232–256).
Qin, S. J., & Badgwell, T. A. (2003). A survey of industrial model predictive
The work described in this paper represents a large team control technology. Control Engineering Practice, 11/7, 733–764.
Richalet, J. A., Rault, J. L., Testud, & Papon, J. (1978). Model predictive heuristic
effort, especially in the final year of product development. The control: Application to industrial processes. Automatica, 14, 413–428.
“true believers” that contributed during the early phases deserve Ricker, N. L. (1991). Model predictive control: State of the art. In Y. Arkun
special mention. These include my current and former con- & W. H. Ray (Eds.), Proceedings of the fourth international conference on
trol engineering colleagues at AspenTech: Tom Badgwell, John chemical process control (pp. 271–296).
Campbell, Dean Kassmann and Mike Keenan. Also, Dave Hein Scokaert, P. O. M., & Rawlings, J. B. (1996). Infinite horizon linear quadratic
control with constraints. In Proceedings of the IFAC ’96 world congress (pp.
deserves special mention for his untiring contribution to the 109–114).
initial and final online real time environment architecture and Shinskey, F. G. (1994). Feedback controllers for the process industries. New
software. It is safe to say that SSC would not have come to York: McGraw-Hill.
fruition without the individual contributions of this outstanding Vogel, E. F., & Downs, J. J. (2002). Industrial experience with state-space model
team. predictive control. In J. B. Rawlings, B. A. Ogunnaike, & J. W. Eaton (Eds.),
Proceedings of the sixth international conference on chemical process con-
trol, Vol. 98 (pp. 438–442).
References Wright, S. J. (1997). Applying new optimization algorithms to model predictive
control. In J. C. Kantor, Garcia, & B. Caranahan (Eds.), Proceedings of
Bemporad, A., & Morari, M. (1999). Control of systems integrating logic, the fifth international conference of chemical process control, Vol. 93 (pp.
dynamics and constraints. Automatica (special issue on hybrid systems), 147–155).
35(3), 407–427. Young, R. E., Bartusiak, R. D., & Fontaine, R. W. (2002). Evolution of an
Cutler, C. R., & Ramaker, B. L. (1980). Dynamic matrix control—a computer industrial nonlinear predictive controller. In J. B. Rawlings, B. A. Ogunnaike,
control algorithm. In Proceedings of the automatic control conference. & J. W. Eaton (Eds.), Proceedings of the sixth international conference on
Froisy, J. B., & Richalet, J. (1986). Industrial applications of IDCOM. In chemical process control, Vol. 98 (pp. 342–351).
M. Morari & T. J. McAvoy (Eds.), Proceedings of the Chemical Process Zafirou, E. (1991). On the effect of tuning parameters and constraints on the
Control—III (pp. 233–244). Elsevier. robustness of model predictive controllers. In Y. Arkun & W. H. Ray (Eds.),
Garcia, C. E., & Prett, D. M. (1986). Advances in industrial model-predictive Proceedings of the fourth international conference on chemical process con-
control. In M. Morari & T. J. McAvoy (Eds.), Proceedings of the chemical trol (pp. 363–394).
process control—III (pp. 245–293). Elsevier.

You might also like