Available online at www.sciencedirect.

com

ScienceDirect
IFAC PapersOnLine 55-17 (2022) 174–179
From-scratch development
From-scratch development and and
From-scratch
From-scratch
improvement development
development and
of aa problem-based
problem-based and
learning
improvement
improvement of
From-scratch
of a development
problem-based learning
and
learning
improvement
course: of
Nonlinear a problem-based
Model Predictivelearning
Control
course:
course: Nonlinear
improvement of
Nonlinear aModel Predictive
problem-based
Model Predictive Control
learning
Control
course:
for
course:
Nonlinear
for Chemical
Chemical andModel
and
Nonlinear
Predictive
Biochemical
Biochemical
Model PredictiveProcesses
Processes ⋆⋆⋆
Control
Control
for
for Chemical
Chemical and
and ∗Biochemical
Biochemical Processes
Processes ⋆
for Chemical and Biochemical Processes
Jose Matias Johannes Jäschke ∗ ⋆
Jose Matias ∗∗∗ Johannes Jäschke ∗∗∗
Jose Matias
Jose Matias ∗ Johannes Jäschke Jäschke ∗
∗ Jose Matias ∗∗ Johannes Johannes Jäschke ∗∗
∗∗ Chemical Engineering Jose Department,
Matias JohannesNorwegian Jäschke University of Science
∗ Chemical
and Technology,
Engineering Department, Norwegian
Sem Sælandsvei 4, Kjemiblokk 5,
University of Science
Trondheim, Norway
∗ Chemical
∗ Chemical Engineering
Engineering Department,
Department, Norwegian
Norwegian University of Science
and Technology,
∗ Chemical
Sem Sælandsvei
Engineering Department, Norwegian5,University
4, Kjemiblokk Trondheim,
University of Science
of Norway
Science
and
and Technology,
Chemical Sem
Engineering Sælandsvei
Department, 4, Kjemiblokk
Norwegian 5, Trondheim,
University of Norway
Science
and Technology,
Technology, Sem Sem Sælandsvei
Sælandsvei 4, 4, Kjemiblokk
Kjemiblokk 5, 5, Trondheim,
Trondheim, Norway Norway
and Technology, Sem Sælandsvei 4, Kjemiblokk
Abstract: This paper describes the development from scratch of a Nonlinear Model Predictive 5, Trondheim, Norway
Abstract: This paper describes the development th from scratch of a Nonlinear Model Predictive
Control course,
Abstract: This which
paper was designed
describes the for 5th
development year Process
from
from scratch Engineering
of master’sModel
of aaa Nonlinear students at the
Abstract:
Control
Abstract: course,Thiswhich
This paper was
paper describes
designed
describes the for
the development
5 th
development year Process
from scratch
Engineering
scratch of master’sModel
Nonlinear
Nonlinear studentsPredictive
Predictive
at the
Norwegian
Control
Abstract:
Control University
course,
This
course, which
paper
which of
was Science
designed
describes
was designed anddevelopment
the for
for 5
5
th
Technology
th
th year
year (NTNU).
Process
from
Process scratch After
Engineering
of
Engineering a taking
Nonlinear the Model
master’s
master’s course,
students
Model
students
Predictive
students
at the
Predictive
at the
Norwegian
Control University of
wasScience andforTechnology (NTNU). After taking the course, students
are able course,
Norwegian
Control
Norwegian to program
course,
which
University
which
University theirof
was
of
designed
own
Science nonlinear
designed
Science and
and for 5 th year
5model
Technology
Technologyyear
Process
predictive
(NTNU).
Process
(NTNU).
Engineering
controller
After
Engineering
After taking
taking
master’s
(NMPC) the
master’s
the
students
incourse,
MATLAB,
students
course,
at and
students
at
students
the
the
are able
Norwegian to program
University their of own
Science nonlinear
and model
Technology predictive
(NTNU). controller
After (NMPC)
taking the in MATLAB,
course, studentsand
have
are
are ablegained
Norwegian
able to a good understanding
to program
University
program their
their own
own nonlinear
of Science of the
and
nonlinear most important
model
Technology
model predictive
(NTNU).
predictive issues
controllerrelated
After
controller takingto the
(NMPC)
(NMPC) implementation
in MATLAB,
theimplementation
incourse,
MATLAB, studentsand
and of
have
are gained
able to a good understanding
program their own of the
nonlinear most important
model predictive issues related
controller to the
(NMPC) in MATLAB, andof
this
have
are
have type
gained
able to
gained of controller
a good
program
a in
their chemical
understandingown and
of
nonlinear the biochemical
most
model processes.
important
predictive issues The
controller course
related to
(NMPC) follows
the problem-based
implementation
in MATLAB, and of
this
have type
learninggained a good
of controller
principles.good It understanding
in chemical and
understanding
isin heavily
of
of the
focused the on
most
biochemical
most important
processes.
important
“doing”,
issues
whereissues coding
Therelated
course
related
assignments
to the
the implementation
to follows problem-based
implementation
a large part of
areproblem-based
of
of
this
have
this type
gained
type
learning of
of controller
a good
controller
principles. in chemical
understanding
chemical
It isinheavily and
of
and
focused the biochemical
most
biochemical processes.
important
processes.
on “doing”, processes. issues
where coding The
The course
related
course
assignments to follows
the
follows problem-based
implementation
areproblem-based
a large part of of
this
the type
course of controller
load. The student chemical
learning and isbiochemical
assessed by an The
unconventional course follows
combination of formative
learning
this principles.
typeprinciples.
ofload.
controller It is heavily focused on “doing”, where coding assignments are a large part of
learning
the course
learning
feedback principles.
(one-to-one
TheIt isinheavily
is
Itstudent chemical
heavily
session,
learning andisbiochemical
focused
focused
where we
on
on “doing”,
assessed
“doing”,
discuss
by processes.
their
where
an
where codingTheassignments
unconventional
coding
assignments
course
and
follows
combination
assignmentsthe areproblem-based
are
solution
aa of
large part
formative
large part of
strategies
of
of
the
the course
learning
course
feedback load.
principles.
load.
(one-to-one The
The It student
is heavily
student learning
session,learning focused
where we is
is assessed
on “doing”,
assessed
discuss by
by
their an
where
an unconventional
coding
unconventional
assignments and the combination
assignments
combination are a of formative
large
of
solution strategies part
formative of
of
the
the course
students) load. as The
well student
as learning
summative is
feedbackassessed by an
(systematic unconventional
grading of combination
their assignments of formative
and final
feedback
the course
feedback (one-to-one
load.
(one-to-one The session,
student
session, where
learning
where we
weis discuss
assessed
discuss their
by
their anassignments
unconventional and the solution
combination strategies
of formative of
the students)
feedback
exam). The
as well as
(one-to-one
student
summative
session,
perception where
of the
feedback
wecourse,
discuss their assignments
(systematic
which
grading of
assignments
was grading
assessedof
and
and theirtheassignments
athe
bytheir
solution strategies
and final
solutionquestionnaire,
feedback strategies of
of
the
the
exam).students)
feedback
students)
The as
(one-to-one
as
student well
well as
as summative
session,
summative
perception where
of thefeedback
wecourse,
feedback discuss (systematic
their was
(systematic
which assessed of
assignments
grading bytheir
and athe assignments
solution
assignments
feedback and
strategies
and
questionnaire, final
finalof
the
shows students)
that the as well
course as summative
adds value feedback
to them (systematic
by not only grading
helping of ofthemtheir assignments
write more and final
interesting
exam).
the
exam).
shows The
students)
The
that student
as
student
the well
course perception
as summative
perception
adds valueof
of the
the
to course,
feedback
course,
them by which
(systematic
which
not was
was
only assessed
grading
assessed
helping by
by
them aa feedback
their assignments
feedback
write more questionnaire,
and
questionnaire, final
interesting
exam).
Master’s The student
projects butperception
also byvalue of the course,
improving whichsetwas
theirbyskill forassessed by a feedback questionnaire,
the job market.
shows
exam).
shows that
The the
student course adds
butperception by value of theto them theirby
course, not
which only
was helping
for assessed them write
by a feedback more interesting
questionnaire,
shows that
Master’s that the
projects
the course
course alsoadds
adds improving
value to
to them
them not
byskill
notset only
only helping them
the job market.
helping them write
write more more interesting
interesting
Master’s
Master’s
Master’s
projects
© 2022 The
Copyrightprojects
shows that the
projects
but
course
but also
Authors.
but also adds
also by
by improving
This
value
by improving is an
to
improving their open
them their byskill
access
their skillnot set
article
skill set
for
only
set for under
for the
the
helpingjob
the
the job CC market.
them
job market.BY-NC-ND
market. write license
more interesting
(https://creativecommons.org
Master’s
Keywords: projects
Pedagogy butin also /licenses/by-nc-nd/4.0/)
by
control improving
engineering, theirchallenges
skill set for the job engineering
in control market. curricula,
Keywords: Pedagogy in control engineering, challenges in control engineering curricula,
problem
Keywords:
Keywords: basedPedagogy
Pedagogylearning, in programming
control
in control
control engineering, challenges
engineering, challenges
challenges in in control
in control engineering
control engineering curricula,
engineering curricula,
curricula,
problem
Keywords: based learning,
Pedagogy in programming engineering,
problem
Keywords:
problem basedbasedPedagogylearning,
based learning, in programming
control
learning, programming
programming engineering, challenges in control engineering curricula,
problem
problem 1. based
INTRODUCTIONlearning, programming but also to provide the right balance between fundamental
1. INTRODUCTION but also to provide the right balance between fundamental
1.
1. INTRODUCTION theory
but
but also and
to practicalthe
to provide applications.
1. INTRODUCTION
INTRODUCTION theory
but alsoand
also to practicalthe
provide
provide the
right
right balance
applications.
right balance between
balance between fundamental
between fundamental
fundamental
The scope of process 1. INTRODUCTION
control has been changing along but theory
theory
In also
this and
to
and
paper, practical
provide
practical
we the applications.
right balance
applications.
describe the between development
from-scratch fundamental
The scope of process control has been changing along In theory
this andand
paper, practical
we describe applications.
the from-scratch development
the
The years.
scope Typically,
of process it was
control limited
has to
been regulation
changing of key
along theory
and
In improvement practical applications.
of a nonlinear model predictive con-
theThe years.
The
process
the
scope Typically,
scope of process
of
variables
processit control
control
was limited
in theit control
presence
has been
has been
of disturbances
changing ofalong
to regulation
changing such
along
key and
that In this
In this paper,
paper, we
improvement
this paper, we describe
we describe the
the from-scratch
of a nonlinear
describe the from-scratch
model predictive
from-scratch
development
development
th development con-
the years.
The
process
the scope
years. Typically,
variables
years. in theit
of process
Typically,
Typically, it
was
was limited
presence
was has
limited
limited
to
to regulation
been
of disturbances
to changing
regulation
regulation such of
of
key
ofalong
key trol
that
key trol
and
In course.
improvement
this
and course.
and paper,
improvement
improvement
The we course
The course of a
describe was
of aa was
of thedesigned
nonlinear
nonlinear for
model
from-scratch
designed
nonlinear model
model
5 th year
predictive Chemi-
development
predictive
th
for 5predictive
year Chemi- con-
con-
con-
atherequired product specification isdisturbances
achieved safely and
aprocess variables
years. in
in the
Typically, presence
it presence
was limited of to regulation such that
of that
key cal/Process Engineering master’s students thatyear
the Chemi-
Norwe-
process
process variables
required product
variables in the
the presence
specification ofis
of disturbances
achieved
disturbances such
safely
such that
and trol
and
trol course.
improvement
course.
cal/Process The
The course
of
course
Engineering a was
was designed
nonlinear
designed
master’s for
model
for
students 5
5 predictive
th
th
at year
the con-
Chemi-
Norwe-
aconsistently
required
process (Pistikopoulos
product
variables in the et
specification
presence al., 2021).
is Consequently,
achieved safely the
and trol course. The course was designed for 5 year Chemi-
a required
required
consistently
asyllabus product
(Pistikopoulos
product et al., of
specification
specification isdisturbances
2021).
is achieved
Consequently,
achieved such that
safely
safely and
the gian
and gian
trol University
cal/Process
course.
cal/Process
University
cal/Process The ofcourse
Engineering
Engineering
of
Engineering
Science
Science was and
master’s
master’s
and
master’s
Technology
designed students
for
students
Technology 5 th(NTNU).
at
at the
year
the
(NTNU).
students atmethods
The
Norwe-
Chemi-
Norwe-
the Norwe- The
consistently
a required of control
product courses
(Pistikopoulos
specificationhas
et al., followed
2021).
is the same
Consequently,
achieved safely guide-
the
and aim
gian was
Universityto cover of the
Sciencemain and state-of-the-art
Technology (NTNU). for
consistently
syllabus
consistently of (Pistikopoulos
control courses
(Pistikopoulos et
hasal., 2021).
followed
et al., 2021). Consequently,
the same
such Consequently,
guide-the
the aim
the gian
cal/Process
gian wasUniversity Engineering
to cover
University of
of theScience
Science master’s
mainand
and students
Technology
state-of-the-art
Technology at methodsThe
the
(NTNU). Norwe-
The
for
lines.
syllabus
syllabus
lines.
syllabus
Forofexample,
Forof
consistentlyof
control
control
example,
control
when
courses
(Pistikopoulos
courses
when
courses
designing
ethas
hasal.,followed
designing
has 2021).
followed
followed
courses,
the
the same
such Consequently,
courses,
the same
same
it is con-
it isguide-
guide-
con-
guide- dynamic
aim
gian
aim was optimization
University
was to
to cover
cover of the
Science
the and
main
main control
and and how (NTNU).
state-of-the-art
Technology
state-of-the-art tomethods
(NTNU).implement
methods
The
for
The
for
sensus
lines. that classical toolsdesigning
such asfollowed
Laplace transformation, dynamic
aim was optimization
to cover the and maincontrol and how tomethods
state-of-the-art implement for
lines. For
syllabus
sensus
lines. For ofexample,
that
For control when
example,
classical
example, tools designing
courses
when
when hasas
such
designing
such
such courses,
Laplace
such the
courses, same it
it is
transformation,
courses, it is
is
con-
guide-
con- them
con- them
dynamic
aim
dynamic
dynamic
in practice.
was
in optimization
to cover
optimization
practice.
optimization the and
main
and
and
control and
state-of-the-art
control
control and
and
how
how
how
to
to
to
implement
methods
implement
implement for
closed-loop
sensus
lines.
sensusForthat analysis,
classical
example, and
tools lead/lag/PID
such as Laplace design should
transformation, be
closed-loop
sensus that
that analysis,when
classical
classical and designing
tools
tools such as
lead/lag/PID
such such design
as Laplace
Laplace courses, it is con-
transformation,
should
transformation, them
be dynamic
them
To
them
in
in
achieve
in
practice.
optimization
practice.
these
practice. goals, and
we control
designed and
this how to implement
course based on
the focus
closed-loop
sensus thatof analysis,
attentionand (Rossiter et al., 2019).
lead/lag/PID design should
should be
theclosed-loop
focus
closed-loop of classical
attentiontools
analysis,
analysis, and
and such as
et Laplace
lead/lag/PID
(Rossiter
lead/lag/PID al., 2019). transformation,
design
design should be To
be them achieve
problem-based
To achieve
these goals, we designed this course based on
in practice.
theselearning
goals, principles.
we designed Thethisstudents
course learn
based howon
the
the focus
focus of
closed-loop
However, of
newattention
analysis,
attention
developments(Rossiter
and
(Rossiter in et
lead/lag/PID al.,
al., 2019).
etoptimal design
2019).controlshould theories be problem-based
To
To achieve
achieve these
these goals,
learning
goals, we
we designed
principles.
designed Thethis
this course
students
course learn based
based howon
on
the
However,focus of
newattention
developments(Rossiter in et al.,
optimal 2019).control theories to
To code
problem-based
achieve a nonlinear
theselearning
goals, model
we predictive
principles.
designed Thethis controller
students
course (MPC)
learn
based howon
the
and focus
However, of attention
computational
new (Rossiter
methods
developments et
have
in al.,
optimal 2019).
been enabling
control large-
theories problem-based
to code
problem-based a learning
nonlinear
learning principles.
model predictive
principles. The
The students
controller
students learn
(MPC)
learn how
how
andHowever,
However, new
computational developments
new control methods
developments in
haveoptimal
in optimal been control
enabling
control theories
large-
theories from
to
to code
code scratch
problem-based a in
nonlinear
aa nonlinear MATLAB
learning model using
principles.
model three
predictive
The
predictive different
controller
students
controller methods
(MPC)
learn
(MPC)how
size
and practical
computational
However,
and practical new control
computational developments applications
methods
methods have
in
haveoptimal(Biegler,
been
been control 2018).
enabling
enabling Con- from
large-
theories
large-
to code
(Direct
from
scratch
Single
scratch
in MATLAB
nonlinear
Shooting,
inShooting,
MATLAB
model using
Multiple
using
three different
predictiveShooting
three
controller
different
methods
(MPC)
and methods
Orthog-
methods
size
and computational applications
methods have (Biegler,
been 2018).
enabling Con- (Direct
large- to
from
from code scratcha nonlinear
Single
scratch in
in MATLAB
MATLAB model predictive
using
Multiple
using three
Shooting
three controller
different
differentand (MPC)
Orthog-
methods
sequently,
size
and practical the process control scope has broadened and
size practical
sequently,
size practicalthe control
computationalcontrol
process
control
applications
methods
control have
applications
applications scope(Biegler,
been
has enabling
(Biegler,
(Biegler,
2018).
2018). large-
broadened
2018).
Con-
Con-
and onal
Con- onal
(Direct
from
(Direct
(Direct
Collocation
Single
scratch
Single
Collocation in
Single they
on Finite
Shooting,
MATLAB
Shooting,
on
Shooting, Finite
Elements).
Multiple
using
Multiple
Elements). Shooting
three
ShootingThen,
different
Then, and
andat methods
at
the end
Orthog-
Orthog-
the end
topics
sequently,
size suchthe
suchthe
practical
sequently,
topics
sequently,
as process
the
the minimization
as control
process control
control scope
applications
the minimization
process control
of energy
scope
scope
has
of (Biegler,
energy
has
consumption,
has broadened2018). Con-
broadened
consumption,
broadened
and
and of
and of
onaltheCollocation
(Direct
onal course,
Single they
Collocation on
Shooting,
on have
Finite
FiniteaMultiple
lecture
Elements).
Multiple
Elements).
Shooting
about
ShootingThen,
and Orthog-
implementation
Then, andat
at the
Orthog-
the end
end
or
topicsmaximization
such as the ofminimization
the productivity of have consumption,
been gaining onalthe course,
Collocation on have
Finitea lecture
Elements). about implementation
Then, at the end
or sequently,
topics
topics suchthe
maximization
such as process
as theof
the thecontrol
minimization
minimizationproductivity of energy
scope
of has
havebroadened
energy
energy consumption,
been gaining
consumption, and of onal
of
NMPC,
the
the
NMPC,
the
course,
Collocation
course,
course,
which
which they
they
they
is
on
is
complemented
have
Finite
have a
a
complemented
have a
lecture
Elements).
lecture
lecture
by
about
about
by
about
guest
guest
lectures
implementation
Then, at
implementation
lectures
implementationthe from
end
from
attention
or (Pistikopoulos
or maximization as theof the et al., 2021).
productivity Not
have only
been in gaining
theory, industry.
topics
attention
or such
maximization ofminimization
(Pistikopoulos
maximization of the et
the productivity
al., 2021).
productivity of energy haveonly
Not
have been
been in gaining
consumption,
theory, industry.
gaining of
of NMPC,
the
NMPC, course, which
which they is
is complemented
have a
complemented lecture
of NMPC, which is complemented by guest lectures from
by
about
by guest
guest lectures
implementation
lectures from
from
but also in(Pistikopoulos
attention practice (e.g., Rawlings et al. (2018)).
but also in(Pistikopoulos
or maximization
attention
attention practice
(Pistikopoulos the et
of (e.g., et
et
al.,
al., 2021).
productivity
Rawlings
al., 2021).
2021).
Not
et al.have
Not
Not
only
been
only
(2018)).
only
in
in
theory,
in gaining
theory, of
theory, industry.
NMPC,
industry.
Based
industry. which isfeedback
on formal complemented from the by guest
students lectures
(see from
Sec-
but
but also
attention
also
Therefore, in
in practice
(Pistikopoulos
practice
instructors (e.g.,
(e.g., etRawlings
al.,
Rawlings
teaching 2021).
process et
et al.
Not
al. (2018)).
only
(2018)).
control in theory,
courses in Based
industry. on formal feedback from the students (see Sec-
but also
Therefore, in practice
instructors (e.g., Rawlings
teaching process et al. (2018)). tion
Based 6), the
on students
formal have
feedback a good
from perception
the students of this(seehands-
Sec-
but also inlevels
graduate
Therefore, practice
have(e.g.,
instructors Rawlings
toteaching
face theprocess et al.control
challenge (2018)). courses in tion
of designing
control courses ina Based
Based 6), on theformal
on students
formal feedback
have a good
feedback from
from the
the students
perception
students of this (see
(see Sec-
hands-
Sec-
Therefore,
graduate
Therefore, instructors
levels teaching
have toteaching
instructors process
face theprocess challenge control courses
of designing
control courses in
a on
in on
tion
Based
tion approach.
6),
6), the
on
the They
students
formal
students enjoyed
have
feedback
have a
athe
good
from
good practical
perception
the part
students
perception of
of the
this
this coding,
(seehands-
Sec-
hands-
syllabus
graduate
Therefore, that
levelsattends
have
instructors the
to current
face
teaching the needs
challenge
process of theof
control process
designing
courses con-
ina tionapproach.
6), the They
students enjoyed
have athe
good practical
perceptionpart of the
this coding,
hands-
graduate
syllabus
graduate levels
that have
attends
levels have to
the face
current
to face the the challenge
needs
challenge of theof designing
process
of designing con- a despite
on
tion
a despite approach.
6),
on approach. the
the heavy
They
students
They workload
enjoyed
have
enjoyed athe
good
the associated
practical
perception
practical with
part
part of
of it.
the
this Several
coding,
hands-
trol
syllabuscommunity.
graduate that
levels The challenge
attends
have the
to current consists
needs notthe
of only of finding
of process con- a on approach.the heavyThey workload
enjoyed the associated
practical
the codespart
with of the
the coding,
it. Several
coding,
syllabus
trol
syllabus that
community.
that attends
attends theface
The challenge
the current
currentthe challenge
needs
consists
needs of
not
of the
theonly designing
process
of finding
process con-
con- students
despite
on approach.
despite
students
despite
usedheavy
the
the
used
the
the concepts
They
heavy
the
heavy
workload
enjoyed
workload
concepts
workload
andassociated
the
and practical
associated
the codes
associated
developed
with
part
with
developed
with
it.
of the
it.
it.
in the
Several
coding,
Several
in the
Several
the
trol right
syllabus
trolright course
community.
that
community. structure
The
attends challenge
the
The challenge
challengeto foster
current relevant
consists
needs
consists not
of technical
theonly
nottechnical
only of of
process skills,
finding
of findingcon-
finding course
students in their
used Master’s
the concepts projects.
and the Also,
codes one of the
developed students
in the
the
trol course structure
community. The to foster relevant
consists not only skills, coursedespite
students
students intheused
their
used heavy
the
the workload
concepts
Master’s
concepts andassociated
and
projects. the
the codes
Also,
codes with
onedeveloped
of the
developed it.students
Several
in the
incom-
the
the
trol
the right
right course
community.
course structure
The challenge
structure to
to foster
foster relevant
consists
relevant not technical
only
technical of skills,
finding
skills, that
course
students took
in the
their
used course
Master’s
the conceptshas been
projects.
and hired
the Also,
codesto
one work
of for
the
developed a
students
in the
the right course structure to foster relevant technical skills, that course tookin their
the Master’s
course projects.
has been Also,
hired toone of
work the
for students
a com-
the right course structure to foster relevant technical skills, course pany
that
course
in
that
took
in
their
the
their
Master’s
provides
course
Master’s model projects.
has predictive
been
projects.
Also,
hired
Also,
one
control
to
one
of
work
of
the
solutions,
for
the
students
a and
com-
students
⋆ The authors acknowledge financial support from the Norwegian that
pany
that took
took the
that provides course
the have
course modelhas been
predictive
has hired
been to hired
hired to
control
to inwork
work for
for a
solutions, a com-and
com-
⋆ The authors acknowledge financial support from the Norwegian several
pany
that
pany that
took
that others
provides
the have
course
provides been
model
model predictive
has hired
been to
predictive hiredwork
control IT
to inwork
control companies.
solutions,
for a com-
solutions, and
and
⋆
Research
⋆ The Council,
authors SUBPRO,
acknowledge grant number: 237893. several
pany thatothersprovides been
model predictive work
control IT companies.
solutions, and
⋆
Research
The
The Council,
authors
authors SUBPRO, financial
acknowledge
acknowledge grant number:
financial
financial
support
support
support
from
237893.
from
from
the
the
the
Norwegian
Norwegian
Norwegian several
pany
several that others
others have
provides
have been
model
been hired
predictive
hired to
to work
control
work in
in IT
IT companies.
solutions,
companies. and
⋆
Research Council,
The authors SUBPRO,
acknowledge grant number: 237893. several others have been hired to work in IT companies.
Research
Research Council,
Council, SUBPRO, financial
SUBPRO, grant
grant number:support
number: from the Norwegian
237893.
237893. several others have been hired to work in IT companies.
Research
2405-8963Council,
Copyright SUBPRO,
© 2022 The grant number:
Authors. is an open access article under the CC BY-NC-ND license.
This237893.
Peer review under responsibility of International Federation of Automatic Control.
10.1016/j.ifacol.2022.09.276
 Jose Matias et al. / IFAC PapersOnLine 55-17 (2022) 174–179 175

Therefore, we believe that the course has enhanced the backgrounds. The prerequisite is only coding experience,
capabilities of the students to meet the future challenges preferably in Matlab and Python. No a priori process
that they will face either in research and industry. modeling/system identification knowledge is required since
we provide all models used in the course. Moreover, we
2. BACKGROUND: MOTIVATION FOR only use simple models (for instance, see the bioreactor
DEVELOPING THE COURSE model in Bequette (2003) - Module 8) because we do
not want the students to struggle with complex system
We designed the course in 2019 to address four issues that equations, while the main focus should be on how to
students at our department faced at the time: formulate and solve dynamic optimization problems.

(1) Lack of experience with and knowledge about state-of

the art methods for process automation and control; 4. QUICK TOUR OF THE COURSE
(2) Every year, ca 15-20 students at our department
work on Master’s projects that involve some form of Figure 1 shows the syllabus of the course offered in 2021.
dynamic optimization and model predictive control; The lectures are designed to last for 90 min in total (two
However, in the Chemical Engineering curriculum, blocks of 45 minutes plus a 15-minute break).
there was no course providing the necessary prereq-
uisites, such that these students needed to learn the
basic principles on their own. Courses offered by other
departments were too detailed and included many
aspects that are were relevant for most of the work at
our department;
(3) Chemical Engineers are required to be increasingly
qualified in digital automation technologies, and pro-
gramming complex computer programs;
(4) Chemical engineering research has been largely dom-
inated by bioengineering in the last decades (Varma
and Grossmann, 2014), while control is still mainly
taught with focus on oil refineries and petrochemi-
cal plants (Pistikopoulos et al., 2021). The students
need to understand the capabilities and potential of
process control when applied to bioprocesses.
Faced with this scenario, we created the ”Nonlinear Model
Predictive Control for Chemical and Biochemical Pro- Fig. 1. Course Syllabus - Fall/2021
cesses control”. The course was designed with 3,75 credits
(approx. 15 hours of classroom teaching) and is taught as At the beginning of the course (Lecture 1 ), we give a
normal course module (1/2 semester). The main activities quick introduction to model predictive control (MPC),
in this course are focused on programming and implement- its capabilities, and implementation challenges. Several
ing an advanced model predictive controller in a bioreactor practical examples are given for contextualization. In the
case study. first week, we also offer a lecture/tutorial about CasADi
To address the four needs above, the course is heavily (Andersson et al., 2019), which is an open-source tool
focused on “doing”, where the concept was to give a mini- for nonlinear optimization and algorithmic differentiation.
mum necessary content of theory, and then let the students CasADi interface in MATLAB is used throughout the
work on implementing their own controller software. course for the coding of a nonlinear model predictive
controller.
Due to this slightly different approach from traditional
control courses, we decided to use an unconventional In Lecture 2, we review important concepts in uncon-
combination of formative and summative assessment of the strained and constrained optimization. They are necessary
student learning. The mandatory programming exercises for the implementation of different methods for numerical
count 70% of the final grade. Each exercise is designed optimal control. At this point of the course, we offer an
to solve a problem, and the complexity builds up during extra tutorial introducing IPOPT (Wächter and Biegler,
the course. After each exercise, the students receive one- 2006), which is the solver used during the course. We
to-one formative feedback, where we discuss the solution also introduce some MATLAB debugging concepts to help
approach, and the choices the students made. In addition, the students with their assignments, such as: how to add
each assignment is graded and the students receive written breakpoints to your MATLAB script, how to step into
feedback. The final oral exam (30% of the final grade) is functions and scripts, etc. The attendance to this extra
more focused on the theoretical background. tutorial is voluntary, but it was well received.
Lecture 3 focuses on Real-time Optimization (RTO),
3. INTENDED LEARNERS specifically on the two-step method proposed by (Chen
and Joseph, 1987). Despite not being strictly a model
Although the syllabus was developed for 5th year Chem- predictive control method, RTO is used to provide a sim-
ical/Process Engineering master’s students, the course plified optimization setting, which creates context about
is open for students from different departments and how the abstract concepts of unconstrained/constrained
176 Jose Matias et al. / IFAC PapersOnLine 55-17 (2022) 174–179

optimization of Lecture 2 can be applied to chemical en- show them how to use CasADi functionalities of interest.
gineering. Also, we use RTO to introduce the concept of Exercise 1 is also an “introductory” assignment. Here, the
repeatedly optimizing the process that is revisited in the students practice how to write an optimization problem in
last part of the course when the model predictive controller CasADi/MATLAB and how to characterize the optimal
is implemented. During the lecture, we present a brief solution in toy mathematical examples.
overview of common draw-backs of the traditional two-
step RTO, which are used as motivation to the relevance After the students have learned how to use the software
of optimizing the transient behavior of a system (dynamic with toy examples, we start to focus on NMPC-related
optimization). tasks. In Exercise 2, the students first develop a simple
bioreactor model and have to solve a steady-state economic
Coming to the solution of dynamic optimization (Lecture optimization problem, which is simpler to code. In compar-
4 - 6 ), we present three state-of-the-art strategies based ison to Exercise 1, the bioreactor provides a more realistic
on Nonlinear Programming (NLP). We present them in example to practice the coding of optimization problems.
order of complexity, i.e., direct single shooting, multiple The bioreactor example is used throughout the course,
shooting, and orthogonal collocation on finite elements. such that the students become familiar with it, and can
For more details about the methods, please refer to Biegler focus on understanding the new concepts in the exercises.
(2010). In the further course of this exercise, the students code
a version of the two-step steady-state RTO without the
Since single shooting and multiple shooting employ a nu-
model adaptation step. This provides them with a code
merical integrator, we start this block of lectures with a
framework for repeatedly solving a process optimization
quick overview of numerical integration methods. Also,
problem, which is used as a basis for repeatedly solving
in Lecture 5, we give a short introduction about direct
the NMPC problem in the last exercise.
and adjoint sensitivity methods as well as automatic dif-
ferentiation, which will be required for providing gradient As the course evolves (Exercise 3 to 5 ), The students
information to the NLP solver. Finally, in (Lecture 7 ), spe- to implement more and more sophisticated dynamic op-
cific cases of NMPC/MPC implementations are illustrated timization approaches to optimize the performance of the
with process examples. In addition to the guidelines for bioreactor in terms of its transient economic performance.
practical implementation, we also focus the discussion on They build on the methods of the previous assignments
problem formulation and computational issues. and use them as a baseline for discussing the performance
and implementation shortcomings. Finally, in Exercise 6
The course is complemented by one or two guest lectures
the students put all the pieces of the previous course
from industry (so far Equinor, Cybernetica AS, and Per-
together, and implement their own version of a Nonlinear
storp have presented), that demonstrate how the concepts
model predictive controller, that, at each sample time,
learned in the course are directly relevant in an industrial
queries the plant state, computes the optimal predicted
environment.
trajectory and implements the computed first input move.
4.1 Exercises For helping the students, we have 90-minute non manda-
tory help sessions for the assignments. They happen a few
As mentioned before, the course is heavily focused on days after the exercise is released. The idea is that the
“doing”. The main work is done by the student by solving students try to solve the problem before coming to the
problems in the exercises given the concepts introduced session. Then, during the classroom time, we only discuss
in the lecture. At the end of each exercise, we also added specific issues that the students are facing. Since students
1-2 questions for reflection, such that the students develop tend to have similar questions, we encourage the students
a deeper understanding of the matter. The exercises were to discuss the problems among themselves. The goal of
designed to be solved in about 3 to 4 hours each. The six these sessions is to engage the students and to create
assignments of the course are shown in Figure 2. an environment for not only clarifying doubts but also
sharing experiences in coding. Moreover, we believe that
these non-mandatory classes are essential for building up
a relationship between us and the students.
Table 1 shows an example of a week plan that was used
when we offered the course in 2021. Here, “Exercise - out”
indicates when the assignment becomes available to the
students, and “Exercise - in” the assignment due date.
Table 1. Example of course week plan.
Monday Tuesday Wednesday Thursday Friday
Lecture 1 Help session
Week 1 - - -
Exercise 1 - out Exercise 1
Fig. 2. Course Assignments - Fall/2021 Week 2
Lecture 2
Exercise 2 - out
Exercise 1 - in - -
Help session
Exercise 2
Exercise 2 - in
Lecture 3 One-to-one feedback Help session
By considering Figures 1 and 2, it can be seen that the Week 3
Exercise 3 - out
Exercise 1
grades out
-
session Exercise 1 Exercise 3
exercises are carefully designed to mirror the contents
presented in the theoretical lectures. The students have
one week to work on the assignments. However, before
the mandatory assignments, the students have access to
a CasADi example (in the form of a tutorial), where we
 Jose Matias et al. / IFAC PapersOnLine 55-17 (2022) 174–179 177

4.2 Evaluation/student learning assessment optimization and model predictive control. That is, the
activities and the learning outcomes were not aligned con-
The examination form is 70% the exercises and 30% the structively, as Julia was not the main topic of this course.
oral exam on the theoretical background. This is well
received by the students, who perceive it as fair due to 5.2 2019 (Regular module at NTNU)
the large amount of work required to code the programs
and to solve the problems. When we ran the course for the first time at NTNU
as a normal course module (1/2 semester), we chose to
During the course, the students were encouraged to work
change some aspects to help the students focus on the
together and help each other in understanding the prob-
essential part of the course: methods for setting up and
lems, but were required to hand in the exercises individ-
solving a Nonlinear Model Predictive Controller. The first
ually. The assignments are evaluated following a system-
modification was related to the programming language.
atic grading scheme, and formal written feedback on how
We changed from Julia to MATLAB/CasADi, a language
the students solved the problems is given individually.
that students in our department were more familiar with
In addition to this summative assessment, we also have
and that had a better documentation at that point. This
one-to-one formative feedback sessions after they get the
allowed us to align learning activities and the learning
written feedback. Here, we give them a more individual
outcomes. As a consequence, the students were able to
environment to them to ask questions about the concepts
quickly focus on the control part (instead of being slowed
taught in class and the coding. Despite the increase in
down by ”cryptic” error messages in Julia). We also
the workload, these one-to-one meetings are important to
changed the assessment from a single graded final project
understand how the student learning is evolving and if
to a different form, where the exercises count 70% of the
they are absorbing the contents taught in class.
final grade and an oral exam with the remaining 30%.
In the final oral examination (30% of final grade), the stu-
dents were asked generic questions about the approaches, 5.3 2020 (Regular module at NTNU)
as well as specific questions to the exercises that they had
handed in. This made it possible to assess the students In 2020, we added one-to-one formative feedback sessions
understanding of the topic in general, but also of their as described in Section 4.2. This one-to-one sessions helped
own code. foster the students’ motivation to really understand the
code that they handed in. Otherwise, in these sessions,
5. EVOLUTION AND CONTINUOUS it would become quite obvious if they had simply copied
IMPROVEMENT OF THIS COURSE someone else’s code, or did not understand what they did.
Additionally, at the end of the course, we have sent out a
This course has evolved based on the feedback of the stu- questionnaire to receive formal feedback on the course.
dents. Here is a timeline showing the course improvements.
The student feedback is shown in Section 6. 5.4 2021 (Regular module at NTNU)

5.1 2019 (Short course at the University of São Paulo) Based on the feedback forms, we found that the students
requested an earlier exam date and that the workload of
This course was first taught as an invited short course the assignments was considered heavy. This comes from
(1 week, Masters/PhD level) at the Chemical Engineering the students coding their dynamic optimization methods
Department of the University of Sao Paulo, Brazil. In from scratch.
the first version, the exercises were done in the Julia We adjust the slides and contents of some of the lectures
programming language (Bezanson et al., 2017). The course to solve this problem; the idea was to focus more on issues
concept was to give a minimum necessary content of that the students who took the course in the previous years
theory, and then let the students work on implementing had regarding the code. For example, in Lecture 6, where
their own controller software. The students were given we teach orthogonal collocation, the presentation style
coding exercises that were used to practice the approaches. relied on key concepts rather than on showing how to code
For the final evaluation, the students worked in pairs on the resulting system of equations in MATLAB/CasADi.
a project. The goals was to apply and compare two of the We adapted the slides as well as the notation to make a
approaches considered in the course on a larger case study parallel between the theory and the coding.
of their choice.
Remark: Both in 2020 and 2021 the course was held in
The course was well received, and 12 of 14 students person. Since the classes were small (13 and 4 students,
successfully completed it. This is very good, since the respectively) and the pandemic was relatively under control
course required a significant amount of work and was in Norway, we could safely comply with all NTNU Corona
not mandatory. It was communicated to us orally that regulations.
the students considered it useful for their work on their
masters and PhD theses. For example, one of the projects
6. STUDENT FEEDBACK
developed in this course led to a journal publication, see
de Oliveira et al. (2021).
The feedback questionnaire is shown in Table 2. It was
It was also communicated to us that some students strug- anonymous and composed by one qualitative question, ten
gled with learning a new programming language (Julia) at quantitative questions, and one open-ended question. The
the same time as learning complex topics from dynamic idea was to evaluate the workload, the students’ perception
178 Jose Matias et al. / IFAC PapersOnLine 55-17 (2022) 174–179

Table 2. Feedback Questionnaire

2020 2021
Question Possible responses
(average) (average)
1 The amount of work required for this course was: Adequate — Too much — Too low See Figure 3
Answer on the range of 1 (the lectures did not really prepare
To what degree did the lectures prepare you for the exercise work,
2 me for the exercise and the exam) to 5 (the lectures prepared 3.92 4.33
the exam, and reaching the learning goals
me well for the exercises and the exam)
To what degree did the exercises help me deepen my understanding Answer on the range of 1 (very little degree) to 5 (very high
3 4.83 4.67
of the topics? degree)
It was clearly communicated to me what was expected from me in
4 Answer on the range of 1 (disagree) to 5 (strongly agree) 4.17 4.67
the exercises and the exam.
Answer on the range of 1 (not useful) to 5 (it helped me
5 How do you rate the feedback on the exercises? 4.67 4.33
improve a lot)
I got the help I needed to do the required tasks and understand the
6 Answer on the range of 1 (I disagree) to 5 (I fully agree) 4.67 4.67
material
7 I feel that I achieved the requirement learning outcomes Answer on the range of 1 (I disagree) to 5 (I fully agree) 4.75 4.33
8 I would recommend this course to other students Answer on the range of 1 (strongly disagree) to 5 (strongly agree) 4.83 4.33
Answer on the range of 1 (strongly disagree) to 5 (strongly agree,
9 The quality of the course has met my expectations 4.67 4.67
exceeded my expectations)
The guest lectures helped me to put the learned material into Answer on the range of 1 (didn’t attend, N/A), 2 (strongly disagree)
10 4.25 3
context to 6 (strongly agree)
Other comments and feedback. Is there something we should do
11 differently next year? (If you want to, you can also let us know Open-ended question See Table 3
your name, then we can follow up)

about the alignment between the learning activities and

the lectures, and also their opinion in the effectiveness of
the feedback.
Note that the number of students changed drastically from
2020 to 2021. Hence, it is expected that the numbers
fluctuate due to the small sample in 2021. With that
disclaimer, the results in the table can be useful to guide us
for our thinking when offering the course again. In general,
the evaluations are positive and consistent between the two
consecutive years.
Regarding the practical part of the course, we want to
highlight Question 3. Here, we see that the exercises really
add value to the course, helping the students understand
the topics. Question 2 shows that the adaptation of the
slides in 2021 had a positive impact on preparing the
students for the exercise.
However, based on the open-ended question (see Table 3),
the students pointed out repeatedly that the course load
is heavy. Even though we promoted some changes in the
material that were well received, the students did not feel
Fig. 3. Feedback questionnaire: Qualitative question re-
that the overall workload was lighter. This should be the
sponse (Question 1)
key aspect for improvement in 2022, when the course will
be offered again. in this course are of value beyond the field of chemical
Another point of interest is the low score given to the guest engineering and process control.
lecture in 2021. In 2020, we had two guest lectures that
discussed practical aspects of linear and nonlinear MPC 7. CONCLUSION
implementations, which directly correlates with the course.
In turn, in 2021, we had a different lecturer that focused In this course, we present topics related to state-of-the-art
more on PID controllers. Although interesting discussion research on Nonlinear Model Predictive Control (NMPC),
about practical implementation took place during the but also try to cover topics of industrial practice. The
lecture, it was slightly out of context from the course. course was designed based on problem solving principles,
Consequently, the students justifiably gave a low grade where we used the lectures for conveying only the main
in the feedback questionnaire, despite the fact that the information regarding the implementation and solution of
lecture quality was high. NMPC, whereas the students work on assignments to gain
knowledge and skills related to the topic.
In addition to the formal feedback, colleagues have infor-
mally confirmed to us that they can do more interesting The course was created to attend four needs that the
projects with their students, because the students are able students of our department faced: (1) lack of knowledge
to understand and implement their own model predictive about state-of-the-art methods for process automation, (2)
control routines. This comes from doing the implemen- a course about these methods tailored for their needs as
tation as part of this course. Also, as mentioned, several Chemical/Process engineers; (3) improving their program-
students have been hired to work in IT companies. This ming skills; and (4) a deeper understanding of advanced
shows that the programming and automation skills learned control applied to biochemical systems.
 Jose Matias et al. / IFAC PapersOnLine 55-17 (2022) 174–179 179

Table 3. Feedback questionnaire: Open-ended Based on our experience developing this course, the as-
responses (Question 11) similation by the students of these complex topics can
be facilitate if we focus on the application first. For in-
Feedback (selected)
2021
stance, we could start by discussing robustness of the
The only things I want to emphasize is that I think the workload for this control design, highlighting why gain and phase margins
module is way to much, with the module only being 3.75 spt. are important as “safety nets” since we rarely know our
2020
Excellent course. The lectures were structured well, understandable and system perfectly. Then, present the Bode plot as a tool
had a clear connection to the exercises. One thing to consider would be for solving the problem. Finally, we tie it all together
to move the exam closer to the end of lectures, instead of in December.
It is quite difficult to use Casadi in the Single Shooting exercise since
with careful designed exercises, where the students learn to
it is quite different from MATLAB. Maybe we should have a lecture on solve the problem (here, control design robustness) under
how to use Casadi (basic syntax, etc.)? our guidance and mentoring. According to our experience,
Earlier exam date (closer to last exercise)
I enjoyed the practical part of the exercises, and it really helped the course the students will come up with their own questions while
just being a lot of a dump of theory. Also, while I put down adequate as solving the exercises and clearly see the meaning of the
response for exercises, along with the other module, the workload could
become a tad bit much, since these exercises and the course lectures happen
concepts presented to them, which, we believe, facilitates
in a relatively short timespan. Especially since matlab is relatively new for me, their own learning as a whole.
learning the details of that along the way added some difficulty.
So be careful not to shorten down the time to work on those exercises.
I want to praise the lectures and learning material, they were clear and REFERENCES
understandable, and going back through the lecture notes later is not a problem.
NMPC was the most enjoyable class this semester, and overall I find the module Andersson, J., Gillis, J., Horn, G., Rawlings, J.B., and
quality above most other courses I have taken in my 4th and 5th year.
Diehl, M. (2019). CasADi – A software framework for
nonlinear optimization and optimal control. Mathemat-
ical Programming Computation, 11(1), 1–36.
We believe that the course created satisfied these needs. Bequette, B.W. (2003). Process control: modeling, design,
First, the students that take the course learn how to solve and simulation. Prentice Hall Professional.
dynamic optimization problems using different state-of- Bezanson, J., Edelman, A., Karpinski, S., and Shah,
the-art methods with a focus on real-time control appli- V.B. (2017). Julia: A fresh approach to numeri-
cations (need 1 ). The control aspects taught in the course cal computing. SIAM review, 59(1), 65–98. URL
are more related to a chemical/process engineering per- https://doi.org/10.1137/141000671.
spective rather than discussing a Engineering Cybernetics Biegler, L.T. (2018). New directions for nonlinear process
aspects. For example, the assignments center on practical optimization. Current Opinion in Chemical Engineer-
implementation aspects rather than stability proofs (need ing, 21, 32–40.
2 ). To complete them, the students are required to develop Biegler, L.T. (2010). Nonlinear programming: concepts, al-
their own MATLAB scripts and understand how to code gorithms, and applications to chemical processes. SIAM.
abstract concepts with continuous guidance and feedback Chen, C.Y. and Joseph, B. (1987). On-line optimiza-
from the the lecturer and the teaching assistant team (need tion using a two-phase approach: An application study.
3 ). Finally, despite presenting several examples of appli- Industrial & Engineering Chemistry Research, 26(9),
cations in chemical engineering, the exercises are based on 1924–1930.
the control of a biochemical system (need 4 ). However, we de Oliveira, R.D., Guedes, M.N., Matias, J., and Le Roux,
still want to cover specific aspects of bioprocess control the G.A. (2021). Nonlinear predictive control of a bioreac-
next time we offer this course. tor by surrogate model approximation of flux balance
analysis. Industrial & Engineering Chemistry Research,
Regarding the pedagogical aspects, we tried to design 60(40), 14464–14475.
student activities (exercises) that give the students the Pistikopoulos, E.N., Barbosa-Povoa, A., Lee, J.H., Mis-
opportunity to become active in learning. For example, by ener, R., Mitsos, A., Reklaitis, G.V., Venkatasubrama-
solving problems (e.g. setting up and solving the dynamic nian, V., You, F., and Gani, R. (2021). Process systems
optimization problem), the student needs to think about engineering–the generation next? Computers & Chemi-
how the basic principles apply to this particular case. cal Engineering, 107252.
In face of complex and challenging coding exercises, the Rawlings, J.B., Patel, N.R., Risbeck, M.J., Maravelias,
students become engaged in the course as well as obtain C.T., Wenzel, M.J., and Turney, R.D. (2018). Economic
a deeper understanding of the tools presented during mpc and real-time decision making with application to
the lectures. According to the feedback received, this large-scale hvac energy systems. Computers & Chemical
approach has been well received by the students. They use Engineering, 114, 89–98.
the concepts learned during the course in their Master´s Rossiter, J.A., Zakova, K., Huba, M., Serbezov, A., and
thesis. Also, the set of skills developed while solving the Visioli, A. (2019). A first course in feedback, dynamics
exercises is useful beyond the context of their academic and control: findings from an online pilot survey for the
life. IFAC community. IFAC-PapersOnLine, 52(9), 298–305.
We believe that the same active learning principles can Varma, A. and Grossmann, I.E. (2014). Evolving trends in
be applied to more traditional control courses. It is pos- chemical engineering education. AIChE Journal, 60(11),
sible to transform the formal lectures into more engaging 3692–3700.
project-based and challenge-based learning. For example, Wächter, A. and Biegler, L.T. (2006). On the implemen-
in classical control courses, learning topics such as Bode tation of an interior-point filter line-search algorithm
and Nyquist plots can be extremely overwhelming for the for large-scale nonlinear programming. Mathematical
students. Questions such as “Why are we learning this?”or programming, 106(1), 25–57.
“When will we use this in the practice?” can easily appear.