Proceedings,

Proceedings, 10th
10th IFAC
IFAC International
International Symposium
Symposium on
on
Proceedings,
Proceedings, 10th IFAC
10th of
IFAC International
International Symposium
Available
Symposium on
online
on at www.sciencedirect.com
Advanced
Advanced Control
Control of Chemical
Chemical Processes
Processes
Proceedings,
Advanced 10th of
Control IFAC International
Chemical Symposium on
Processes
Advanced
Shenyang,
Shenyang, Control of
Liaoning,
Proceedings,
Liaoning, Chemical
China,
10th IFAC
China, Processes
July
International
July 25-27,
25-27, 2018
Symposium
2018 on
Advanced Control
Shenyang, of China,
Liaoning, Chemical Processes
July 25-27,
Shenyang,Control
Advanced Liaoning,
of China, July
Chemical 25-27, 2018
Processes2018
Shenyang, Liaoning, China, July 25-27, 2018
Shenyang, Liaoning, China, July 25-27, 2018 ScienceDirect
IFAC PapersOnLine 51-18 (2018) 566–571

Evaluation
Evaluation of
of Steady-State
Steady-State and
and Dynamic
Dynamic Soft
Soft Sensors
Sensors for
for Industrial
Industrial Crude
Crude
Evaluation
Evaluation of
of Steady-State
Steady-State and
and Dynamic
Dynamic Soft
Soft Sensors
Sensors for
for Industrial
Industrial Crude
Crude
Distillation
Evaluation of Distillation
Steady-State Unit
and
Unit under
Dynamic
under Parametric
Soft
Parametric Constraints
Sensors for
ConstraintsIndustrial Crude
Distillation
Distillation Unit
Unit under
under Parametric
Parametric Constraints
Constraints
Distillation
Andrei Unit* under
Torgashov***,, Anton
Andrei Torgashov
Parametric
Anton Goncharov
*,** Constraints
*,**,, Evgeny
Goncharov*,**
*,** Evgeny Zhuravlev
***
Zhuravlev***
***
Andrei Torgashov*,, Anton
Andrei Torgashov Anton Goncharov
***
Goncharov *,**, Evgeny Zhuravlev***
 *,**, Evgeny Zhuravlev***
Andrei Torgashov**, Anton Goncharov  *,**, Evgeny Zhuravlev***
Andreiof
*Institute
*Institute Torgashov
of Automation
Automation , Anton
and
and Goncharov
Control
Control  Process
Process , Evgeny
FEB
FEB RAS,
RAS, Zhuravlev
555 Radio
Radio St.,St., and
and
*Institute
*Institute of
of Automation
Automation and
and Control
Control

 Process
Process FEB
FEB RAS,
RAS, 5 Radio
Radio St., and
St.,Island
and
Far-Eastern
Far-Eastern
*Institute
Far-Eastern of Federal
Federal
Automation
Federal University,
University,
and
University, Bld.
Bld.
Control
Bld. 12(E),
12(E),
Process
12(E), Campus
Campus
FEB
Campus FEFU,
FEFU,
RAS,
FEFU, 5 Russky
Russky
Radio
Russky St., Island
and
Island
*Institute
Far-Eastern
Vladivostok, of Automation
Federal
Russia (Tel: and
University, Control
Bld.
+7-423-231-02-02; Process
12(E), FEB
Campus
e-mail: RAS,
FEFU, 5 Radio
Russky
torgashov@iacp.dvo.ru) St., and
Island
Vladivostok,
Far-Eastern
Vladivostok, Russia
Federal(Tel: +7-423-231-02-02;
University, Bld. 12(E), e-mail:
Campus torgashov@iacp.dvo.ru)
FEFU, Russky Island
Far-EasternRussia
Vladivostok, Federal
Russia (Tel:
(Tel: +7-423-231-02-02;
University, Bld. 12(E), e-mail:
+7-423-231-02-02; Campustorgashov@iacp.dvo.ru)
e-mail: FEFU, Russky Island
torgashov@iacp.dvo.ru)
Vladivostok,**JSC Russia“Honeywell”,
(Tel: +7-423-231-02-02; Kievskaya e-mail:
777 Kievskaya St., torgashov@iacp.dvo.ru)
Moscow, Russia.
Vladivostok,**JSC Russia“Honeywell”,
**JSC
**JSC (Tel: +7-423-231-02-02;
“Honeywell”,
“Honeywell”, 7 Kievskaya
St.,
Kievskaya e-mail:
St., Moscow,
St., Moscow,
Moscow,
Russia.
torgashov@iacp.dvo.ru)
Russia.
Russia.
***JSC
***JSC “Gazprom
“Gazprom**JSC “Honeywell”,
neftekhim
neftekhim Salavat”,
Salavat”, 7 Kievskaya
30
30 St., Moscow,
Molodogvardeitsev
Molodogvardeitsev Russia.
st.,
st., Salavat, Russia.
***JSC “Gazprom neftekhim Salavat”, 30 MolodogvardeitsevRussia.
**JSC “Honeywell”, 7 Kievskaya St., Moscow, st., Salavat,
Salavat, Russia.
Russia.
***JSC “Gazprom neftekhim Salavat”, 30 Molodogvardeitsev st., Salavat, Russia.
Abstract:
Abstract: The The***JSC
parametric
The parametric “Gazprom
parametric identification neftekhim
identification
identification problem Salavat”,
problem
problem for for 30 Molodogvardeitsev
industrial
for industrial crude
industrial crude st.,
distillation
crude distillation
distillation unitSalavat,
unit Russia.
(CDU)
unit (CDU)
(CDU) is
is considered.
considered.
Abstract:
Abstract: The parametric identification problem for industrial crude distillation unit (CDU) is
is considered.
considered.
We
We take take
take The
Abstract:
We the
the a
a priori
the parametric
a priori knowledge
priori knowledge
knowledge
identification of
of the
of the
the process
process
problem
process into account
into account
into account
for industrial crude by using
bydistillation
by a
using aa system
using system
system
unit (CDU) of constraints
of constraints
of constraints
is considered. for
for
for
Abstract:
We take
parameters The
theof parametric
a priori
soft sensorsidentification
knowledge models. of problem
the
The process for industrial
into
identification accountcrude
problem by distillation
isusing a
transformed unit
system (CDU)
of
into is considered.
constraints
a constrained for
parameters
We take theof
parameters ofa soft
soft sensors
priorisensors
knowledge models.
models. of theThe process
The identification problem
into account
identification problem isusing
by is transformed
a system into
transformed into a constrained
constrained
of constraints
a for
We take
parameters
optimization theof a priori
soft knowledge
sensors models. of the
The process into
identification account
problem by isusing a
transformed system of
into constraints
a constrained for
parameters ofproblem,
optimization
optimization problem, which
which
soft sensors
problem, which we
we solved
solved
models.
we solved using
using
The
using the
the active
active
identification
the active set
set
set method.
method.
problem
method. The
The
is
The static
static
transformed
static and
and
and dynamic
into a soft
dynamic
dynamic soft
soft sensors
sensors
constrained
sensors
parameters
optimization
are evaluated offor soft
problem, sensors
which
industrial CDU models.
we solved
located The
using
at identification
JSC the active
“Gazprom setproblem
method. is
The transformed
static and into
dynamic a constrained
soft sensors
are evaluatedproblem,
optimization
are for industrial
which CDU located
we solved at
usingJSC the active setneftekhim
“Gazprom neftekhim
method. The Salavat”
staticrefinery.
Salavat” refinery.
and dynamic It
It was
wassoftfound
found that
that
sensors
are evaluated
optimization
the evaluated
model
for
for industrial
problem,
performed which
industrialbetter
CDU
CDU located
wewhen
solved
located we
at
usingJSC
JSCthe
atused “Gazprom
active
“Gazprom
the setneftekhim
method.
neftekhim
proposed
Salavat”
The static
Salavat”
constrained
refinery.
and dynamic
refinery.
optimization
It
It was
was found
soft
found
approach
that
sensors
that
for
the
are
the model
evaluated
model performed
for
performedindustrialbetter
betterCDU when
located
when we
we atused
JSC
used the
“Gazprom
the proposed constrained
neftekhim
proposed Salavat”
constrained optimization
refinery.
optimization It approach
was found
approach for
that
for
are
the evaluated
model
identification for
performed
insteadindustrial
of better
robustCDU located
when
regression we at JSC
used
methods. “Gazprom
the neftekhim
proposed Salavat”
constrained refinery.
optimization It was found
approach that
for
identification
the model
identification instead
performed
instead of
of robust
better
robust regression
when
regression we methods.
used
methods. the proposed constrained optimization approach for
the model performed better
identification instead of robust regression methods.when we used the proposed constrained optimization approach for
identification
Keywords:
Keywords:
© 2018, IFACSoft
Softinstead
sensing,
sensing, of
(International robust regression
identification,
identification,
Federation methods.
constrained
constrained
of Automatic parameters,
parameters,
Control) optimization,
optimization,
Hosting by Elsevier distillation
distillation
Ltd. All columns.
columns.
rights reserved.
identification
Keywords:
Keywords: Soft instead
Soft sensing, of robust
sensing, identification,regression
identification, constrainedmethods.
constrained parameters,
parameters, optimization,
optimization, distillation
distillation columns.
columns.
Keywords: Soft sensing, identification, constrained parameters, optimization, distillation columns.
Keywords: Soft sensing, identification, constrained parameters, optimization, distillation columns.
and
and
and K-2
K-2
K-2 (Fig.
(Fig. 1).1). The
The plantplant is
is located
located at at JSC
JSC “Gazprom
“Gazprom
K-2 (Fig. 1).
1). The plant is
is located at
at JSC “Gazprom
1. INTRODUCTION 
1.
1. INTRODUCTION  and (Fig. The plant located JSC “Gazprom
1. INTRODUCTION
INTRODUCTION neftekhim
neftekhim
and K-2
neftekhim
and K-2
Salavat”
Salavat”
(Fig.
Salavat”
(Fig.
1).
1).
refinery
refinery
The
refinery
The
plant
plant
(Salavat,
(Salavat,
is
(Salavat,
is
Russia).
Russia).
located
located
at
Russia).
at
The
The
JSC
The
JSC
feed
feed
“Gazprom
feed flow
“Gazprom
flow
flow
flow
In the present, 1.
researchersINTRODUCTION
are paying more attention to the neftekhim
(crude) comes
comesSalavat” refinery
from refinery
the oil (Salavat,
oil desalting
desalting Russia).
unit The
and enters
enters feed
underflow the
In
In the
the present,
present, 1.
researchers
researchersINTRODUCTION
are
are paying
paying more
more attention
attention to
to the
the neftekhim
(crude) Salavat”
from the (Salavat, Russia).
unit and The feed
under the
In the present,problems
identification researchers of are
the paying
model more attentionofto soft
parameters neftekhim
the (crude)
16 th
th tray
th comes
ofSalavat”
from
K-1. The refinery
theoverhead (Salavat,
oil desalting products Russia).
unit and
of K-1The
enters feed
are under
gas flow
the
and
identification
In the present, problems
researchers of the
are model
paying parameters
more attention ofto soft
the 16 th tray
(crude) of
comes
tray of K-1.
from The the overhead
oil desaltingproductsunit of
and K-1
enters are gas
under and
the
identification
In the present,
identification
sensors for CDU
problems
researchers
problems
(Bolf et
of
ofal.,the
are model
paying
the2010;
model Dam
parameters
more andattention
parameters Saraf,
of
ofto
2006;
soft
the naphtha.
soft 16 th
(crude) of K-1.
th traycomes
The from
K-1. The
The
naphtha
overhead
theis oil
overheaddesalting
also
products
products
withdrawn
of
of K-1
unit from
and enters
K-1
the
are
are
top
gas
under
gas
of
and
the
and
K-2.
sensors for
identification
sensors for CDU
CDU (Bolf
problems
(Bolf et
etof al.,
the
al., 2010;
model
2010; Dam
Dam and
parameters
and Saraf,
Saraf, of2006;
2006;soft naphtha.
16 tray
naphtha.
th The
of
The naphtha
K-1. The
naphtha is
is also
overhead
also withdrawn
products
withdrawn from
of
from the
K-1
the top
are
top of
gas
of K-2.
and
K-2.
identification problems
sensors for CDU (Bolf of
et 2007;the model
al., 2010; Dam parameters
and of
Saraf, 2011). soft
2006; The 16 tray
naphtha. of
The K-1. The
naphtha overhead
is also products
withdrawn of
from K-1
the are
top gas
of and
K-2.
Macias-Hernandez
Macias-Hernandez
sensors for CDU (Bolf et
et al.,
al.,
et 2007; Napoli
Napoli
al., 2010; Dam and
and and Xibilia,
Xibilia, 2006; naphtha.
Saraf, 2011). The bottom
The bottom
bottomTheproducts
naphthaof
products
products ofis the
of also333 withdrawn
the
the strippers
strippers of
strippers of the
the column
offrom
the the top K-2
column
column K-2
of K-2. are
are
Macias-Hernandez
sensors forwidespread
Macias-Hernandez
The most CDU (Bolf et al.,
al.,et 2007;
et approach 2007; Napoli
al., 2010;
Napoli
for Dam
soft
and
andand
sensor
Xibilia,
Saraf, 2011).
Xibilia, 2006;
2011).
evaluation is naphtha.
The bottom
gasoline, The naphtha
products
kerosene cut ofis(KC)
also3 withdrawn
the strippers
and diesel offrom
thecut
oil columntop K-2
the(DOC). of K-2.
K-2 are
are
The
The most
The most widespread
widespread
Macias-Hernandez approach
et approach
al., 2007; for for
Napolisoft sensor
soft sensor evaluation
and Xibilia, 2011).
evaluation gasoline,
The
is gasoline,
is bottom kerosene
productscut
kerosene cut (KC)
of (KC)
the 3 and and dieselof oil
strippers
diesel oil cut
thecut (DOC).
column
(DOC). K-2The The
are
Macias-Hernandez
The most widespread et al., 2007;
approach Napoli
for soft and
sensor Xibilia, 2011).
evaluation is The bottom
gasoline, products
kerosene cut of the
(KC) 3 strippers
and diesel of the
oil column
cut (DOC). K-2 are
The
based
The
based on
on the
basedmost
on the partial
partial least-squares
widespread
the partial least-squares
approach for
least-squares (PLS)
(PLS)
(PLS) softmethod
sensor (Shang
method
method (Shang
evaluation
(Shang et
et al.,
et is main
al.,
al., main
main product
product
gasoline,
product
gasoline,
of
of
kerosene
of
kerosene
the
the
thecutcolumn
cut
column
(KC) and
column
(KC)
K-2
K-2
K-2
and
is
is
diesel
is
diesel
desired
oil cutcut
desired
desired
oil cut
cut
cut 111 (DC1),
(DOC).
(DOC).
(DC1),
(DC1),The
The
The
based most
on widespread
the partial approach
least-squares for
(PLS) soft sensor
method evaluation
(Shang et is
al., main
which product
is the of
mixture the column K-2 is desired cut 1 (DC1),
2015).
2015).
based The
The
on theintegration
integration of
of
partial least-squares the
the first
first
(PLS) principles
method (material
principles (Shang et and
(material al., main
and which product
is the of the of
mixture kerosene
ofcolumn
kerosene K-2cut
cutis(KC)
(KC)
desiredand
and diesel
diesel
cut oil
oil
1 (DC1), cut
cut
2015).
based
2015).
energy
The
on
Thetheintegration
partialand
integration
balances
of the
the first
least-squares
of
phase (PLS)
first principles
methodof
principles
equilibrium)
(material
(Shang
(material
inferential al., which
et and
and main
which
(DOC).
is the
the mixture
product
isDesired ofcutthe2 of
mixture of kerosene
column
kerosene
(DC2), K-2cut
which cutis
is
(KC)
desired
(KC)
the
and
mixture
diesel
and cut
diesel
of
oil
oil cut
1 naphtha
(DC1), cut
energy
2015).
energy balances
The integration
balances and
and phase
of
phasethe equilibrium)
first principles
equilibrium) of
of inferential
(material
inferentialand (DOC).
which
(DOC). isDesired
the
Desired cut
mixture
cut 2 (DC2),
of
22 of kerosene
(DC2), which
which cutis
is the
(KC)
the mixture
and
mixture of
diesel
of naphtha
oil
naphtha cut
2015).
energy
modeling The integration
balances
can be and
found of
phasethe
in first principles
equilibrium)
Chatterjee and (material
of
Saraf inferential
(2003),and which
(DOC).
and is the
Desired
gasoline. mixture
Thecut kerosene
(DC2),
atmospheric which cutis
residue (KC)
the and
mixture
(AR) diesel
is of oil
naphtha
withdrawn cut
modeling
energy
modeling can
balances
can be
be found
and
found in
phase
in Chatterjee
equilibrium)
Chatterjee and
and Saraf
of
Saraf (2003),
inferential
(2003), and
(DOC).
and gasoline.
Desired
gasoline. The
Thecut atmospheric
2 (DC2),
atmospheric whichresidue
is
residue the (AR)
mixture
(AR) is
is withdrawn
of naphtha
withdrawn
energy
modeling
Mahalec balances
can
and be and
found phase
in equilibrium)
Chatterjee and of
Saraf inferential
(2003), (DOC).
and
from theDesired
gasoline. The
bottom cut 2
part (DC2),
atmospheric
of columnwhich is
residue
K-2.the mixture
(AR)
The is
mainof naphtha
withdrawn
process
Mahalec
modeling can Sanchez
and be found(2012),
Sanchez (2012), and
and
in Chatterjee Fujii
andand
Fujii andSarafYamamoto
(2003), and
Yamamoto fromgasoline.
the bottom
The part of
atmospheric column K-2.
residue The
(AR) main process
is withdrawn
Mahalec
Mahalec
(2014).
and
can Sanchez
modeling Johansen
and be found
Sanchez (1996)
(2012), and
and Fujii
in considers
(2012), Chatterjee andand
Fujii
the andSaraf
more (2003), from
Yamamoto
Yamamoto
general and the
fromgasoline.
the of
variables
bottom
The
bottom
the
part
part
industrial
of
of column
atmospheric column
CDU are
K-2.
residue
K-2.
shown
The
Theinmain
(AR) isTable
main process
withdrawn
process
111 and
(2014).
Mahalec Johansen
and Sanchez (1996) (2012), considers
and the
Fujii more
and general
Yamamoto variables
from the
variables of
of the
bottom
the industrial
part
industrial of CDU
column
CDU are
are shown
K-2.
shown The in
in Table
main
Table processand
(2014).
Mahalec Johansen
and
(2014). Johansen
framework of Sanchez
integrating
(1996) (2012),
(1996)availableconsiders
and
considers a
the
Fujii more
and
the information
priori
general
Yamamoto
more general into from
may the
variables
be bottom
of the
considered part
industrial of column
CDU
informative are K-2.
inputsshown The
of in
the main
Table
soft 1 and
processand
sensor
framework
(2014).
framework of
Johansen
of integrating
integrating(1996) available
considers
available a priori
the
aa priori information
more
information generalinto
into may
may be
variables
be considered
of the
considered informative
industrial CDU
informative inputs
are
inputsshown of
of the
in
the soft
Table
soft sensor
1 and
sensor
(2014).
framework Johansen
of integrating(1996) considers
available the
priori more
information generalinto variables
may be of the
considered industrial CDU
informative are
inputsshown of in
the Table
soft 1 and
sensor
the
the identification
the identification
framework problem. available a priori information into model.
problem.
of integrating
identification problem. model.be considered informative inputs of the soft sensor
may
model.
framework problem. available a priori information into may
of integrating
the identification model.be considered informative inputs of the soft sensor
the identification
However, the problem.
methods developed from the previous research
research model.
the identification
However,
However, the
the problem.
methods developed from the previous model.
However,
do not the methods
completely methods deal
developed
developed
with
from
practical
the
the previous
fromobstacles previous
such
research
research
as small
do not
However,
do completely
the methods deal with
developedpractical
from obstacles
the such
previous as small
research
do not
not completely
However,
training the methods
completely
datasets
deal
deal
that
with
developed
withcover
practical
fromobstacles
practical
all the previous
obstacles such
such as
as small
research
small
training
do
do not
training
datasets
not completely
training datasets deal
completely
datasets
that
that with
deal
that with
cover
cover
cover all operating
all
practical
practical
all
operating
obstacles such
operating
obstacles
operating
points,
points,
as small
points,
such as small
points,
low
low
low
low
variability
training
variability ranges
variabilitydatasets
ranges for
ranges for
for key (informative)
keycover
thatkey (informative)
all operating
(informative) input
input variables,
points, low
variables, and
and
training
variability
an unknown datasets
ranges
feed that
for key
composition cover all
(informative)
(i.e. the operating
input
feed points,
variables,
distillation curveslow
and
an unknown
unknownranges
variability
an feed composition
feed composition (i.e. the
for key (informative)
(i.e. the feed
feed distillation
input variables,
distillation curves
curvesand
variability
an
as unknown ranges
feed for key
composition (informative)
(i.e. the input
feed variables,
distillation curvesand
as TBP).
an
as TBP).
unknown feed composition (i.e. the feed distillation curves
TBP).
an unknown feed composition (i.e. the feed distillation curves
as TBP).
as
In TBP).
order
as
In order to
In TBP).
In order
order to overcome
to
to overcome the
overcome
overcome the abovementioned
the
the abovementioned difficulties,
abovementioned
abovementioned
difficulties,
difficulties,
difficulties,
Torgashov
Torgashov
In order to
Torgashov et
et al.
et al. (2016)
al.overcome
(2016) has
(2016) has
hasthe proposed
proposed the
the use
abovementioned
proposed the use of
use of the
of system
system of
thedifficulties,
the system of
In order
Torgashov
parametric to
et overcome
al. (2016)
constraints. has
The the abovementioned
proposed
system of the use of
parametric the system of
difficulties,
constraints of
parametric et
Torgashov
parametric constraints.
al. (2016) The
constraints. The system of
has proposed
system ofthe parametric constraints
use of theconstraints
parametric system of
Torgashov
parametric
is et al. (2016)
constraints. has
The proposed
system of the use of
parametric the system
constraints of
is derived
parametric
is derived from
from the
derived constraints.
from the preliminary
the The system calibrated
preliminary
preliminary calibrated
of parametric
calibrated first
first principle
firstconstraints
principle
principle
parametric
is derived
(rigorous) constraints.
from the The system
preliminary of parametric
calibrated first constraints
principle
is deriveddistillation
(rigorous)
(rigorous) from the model
distillation
distillation model (Torgashov
(Torgashov
preliminary
model (Torgashov and
and
calibrated
and Zmeu,
first 2015).
Zmeu,
Zmeu, 2015).
principle
2015). In
In
In
is
the derived
(rigorous)
current from we
distillation
work, the consider
preliminary
model (Torgashov
the calibrated
extension and of first
Zmeu,
this principle
2015).
technique In
the
the current
(rigorous)
current work,
distillation
work, we
we consider
model
consider the
the extension
(Torgashov
extension and of this
Zmeu,
of this technique
2015).
technique In
(rigorous)
the
in current
the case distillation
work,
of CDU. we Wemodel
consider (Torgashov
introduce the extension
and and
solve Zmeu,
of
the this 2015).
technique
statements In
of
in
the
in the case
current of CDU.
work, we We introduce
consider the and solve
extension the
of statements
this technique of
in the
thethe case
case of
current
identificationof CDU.
work,
CDU. weWe
problems We introduce
consider
introduce
for static
and
and solve
solve the
the (steady-state)
extension statements
of this
the technique
statements
and dynamic
of
of
identification
in the case of CDU.
identification problems
problems We for static
introduce
for static (steady-state) and
and solve the statements
(steady-state) and dynamic of
in the case ofusing
identification
soft sensors CDU.
problemstheWe introduce
for
constraint and solve the
staticoptimization
(steady-state) and dynamic
statements
technique. dynamic of
soft sensors
identification
soft using
problemsthe constraint
for staticoptimization
(steady-state) technique.
and dynamic
soft sensors
sensors using
identification problems
using the
the constraint
for staticoptimization
constraint (steady-state)
optimization technique.
and dynamic
technique.
soft2.sensors using theCRUDE
INDUSTRIAL constraint optimization
UNIT DESCRIPTION technique. AND
soft2.
2. sensors using theCRUDE
INDUSTRIAL constraint optimization
UNIT DESCRIPTION technique. AND
2. INDUSTRIAL
INDUSTRIAL PROBLEM
CRUDE
CRUDE UNIT
UNIT
FORMULATION
DESCRIPTION
DESCRIPTION AND
AND
2. INDUSTRIAL PROBLEM
PROBLEM CRUDE FORMULATION
UNIT
FORMULATION DESCRIPTION AND
2. INDUSTRIAL PROBLEM CRUDE UNIT DESCRIPTION AND
FORMULATION
The PROBLEM FORMULATION
The crude
The crude distillation
crude PROBLEM
distillation
distillation unit
unit
unit considered
considered in
FORMULATION in this
this paper
paper is is Fig. Fig. 1.
1. The
The sequence
sequence of of industrial
industrial multicomponent
multicomponent distillation
distillation
The crude by
represented distillation
two unit considered
multicomponent considered in
in this
distillation this paper
paperK-1
columns:
is
is Fig. Fig. 1.
1. The
columns The sequence
sequence of
(CDU). of industrial
industrial multicomponent
multicomponent distillation
distillation
represented
The crude
represented by
by two
two multicomponent
distillation unit
multicomponent considereddistillation
in
distillation columns:
this paper
columns: K-1
K-1is columns
Fig. 1.
columns The(CDU).
sequence
(CDU). of industrial multicomponent distillation
The crude by
represented distillation unit considered
two multicomponent in this
distillation paperK-1
columns: is Fig. 1. The
columns sequence of industrial multicomponent distillation
(CDU).
represented by two multicomponent distillation columns: K-1 columns (CDU).
represented
2405-8963 by two
© 2018, IFAC multicomponent distillation
(International Federation columns:Control)
of Automatic K-1 Hostingcolumnsby(CDU).
Elsevier Ltd. All rights reserved.
Peer review©
Copyright
Copyright © under
2018 responsibility
2018 IFAC
IFAC of International Federation of Automatic
560
560 Control.
Copyright ©
© 2018
2018 IFAC
10.1016/j.ifacol.2018.09.364
Copyright IFAC 560
560
Copyright © 2018 IFAC 560
Copyright © 2018 IFAC 560
 2018 IFAC ADCHEM
Andrei Torgashov et al. / IFAC PapersOnLine 51-18 (2018) 566–571
Shenyang, Liaoning, China, July 25-27, 2018 567

We consider the plant with several measured inputs We use the determination coefficient (a number that indicates
u1,…,ui,,…,uN and one output y(). We use the measured the proportion of the variance in the dependent variable that
process variables (pressure, temperature, and flow) as inputs. is predictable from the independent variable)
A priori knowledge of the distillation process allows us to
select the most informative variables (Table 1) from the 1   i ( yi  yi ) 2
R2   (y i i  y a )2 , (3)
thermodynamic essence.
the root mean squared error (RMSE)
Table 1. Main process variables
 
1/2

No Notation Process variable, ui RMSE


M
i 1
( yi  yi )2 / M , (4)
1 TIС6 Temperature of the flow in the bottom
stripper of column K-2, °С the Akaike (1969) Information Criterion:
2 PIC2 Top pressure of the column K-2, MPa AIC  M In  y  y    2N  1  M InM  ,
i i
m 2
i
(5)
3 FIC13/ The ratio of steam to feed flowrate for the Schwarz Bayesian Criterion:
4
FIC5
TIC3
column K-2
Top temperature of the column K-2, BSC  M In  y  y    N  1InM   M InM 
i i
m 2
i
(6)
°С as identification criteria on a given time interval, where y i is
5 FIC9/ The ratio of steam to feed flowrate for the measured value of the output variable, yi is the value
FIC5 bottom stripper of column K-2 obtained based on the SS, y a is the mean value of the
6 FIC4/ The ratio of bottom pumparound to
measured output variable, and M is the number of output
FIC5 feed flowrate of column K-2
measurements. The model is more consistent the closer to
7 PIC1 Top pressure of column K-1, MPa
unity the value of the coefficient of determination R2 is or the
8 TIC2 Feed temperature of column K-2, °С
closer to zero the value of the RMSE is or then less the value
9 FIC10/ The ratio of the product of the top of the AIC and BSC are.
FIC5 stripper to the feed flowrate of column
K-2 The goal of the paper is to develop an approach for soft
sensor model identification based on the industrial data while
10 FIC11/ The ratio of the product of the middle
taking into account parametric constraints. These constraints
FIC5 stripper to the feed flowrate of column
can be derived from the rigorous modeling (Torgashov et al.,
K-2
2016). The introduction of the system of constraints also
11 TI1 Feed temperature of the column K-1,
allows us to overcome such difficulties as small training
°С
datasets and laboratory errors. The final boiling point (FBP)
12 FIC12/ The ratio of product of bottom stripper
of DC1 and final boiling point (FBP) of DC2 are considered
FIC5 to feed flowrate of column K-2
soft sensor outputs.

We consider the identification problem of the soft sensor (SS)
evaluation, which is best for predicting the quality of the
products of crude distillation process. We obtain the model
for the soft sensor in the form of a linear regression model
based on the following equation:
b0  b1u1    b2 u2   ...  bN u N   ,
y ( ) 
where bj is the j-th model coefficient, j=0,1,…,N, b0 is the
constant term, N is the number of input variables,  is the
irregular time points of output measurement 1,2,3,…,
 i   i 1   0   , i  2 ; 1  0   , 0 is the constant term; and
 is the random variable restricted in the given range.
The dynamic SS accounts for the influence the process
dynamics have on the quality of the products. The predictive
model is represented as a sum of convolutions of plant inputs
and a finite impulse response (FIR) hi (discrete analogues of
the first degree Volterra kernels):
y( )  h   1 h (k  1)u (  k )   2 h2 (k  1)u2 (  k )  ... (2)
n 1 n 1
0 1
k 0 1 k 0
...  kN0 hN (k  1)uN (  k ),
n 1
3. MAIN RESULTS
3.1 Steady-state model identification under constraints
Let u  [1, u1 ( ), u2 ( ),..., uN ( )]T be a combined vector of
the measured input variables and b = [b0 , b1 ,..., bN ]T be a
(1)
vector of coefficients of the same dimension, the components
of which reflect the contributions of the corresponding input
variables. Then the equation (1) takes the following form:
y  uT  b .
We form the vector Y of dimension q from the output value y
dataset as
Y  ( y ( 1 ), y ( 2 ),..., y ( q ))T
and the matrix U, containing the measured inputs uj,
corresponding to output value y from (1):
 1 u1 ( 1 ) u2 ( 1 ) ....uN ( 1 ) 
 1 u1 ( 2 ) u2 ( 2 ) ....uN ( 2 ) 
U 
      
 
 1 u1 ( q ) u2 ( q ) ...u N ( q ) 

We consider the multicollinearity case, which occurs when

where h0 is a constant term. there is an almost linear relationship between inputs. In this
case, the matrix C  U T U is close to singular, so it is the
smallest eigenvalue min  0 , the condition number is infinitely

561
 2018 IFAC ADCHEM
568 Andrei Torgashov et al. / IFAC PapersOnLine 51-18 (2018) 566–571
Shenyang, Liaoning, China, July 25-27, 2018

increased, and it causes the instability of the solution. If
min  0 , then it corresponds to the strict multicollinearity. In
order to obtain a stable solution, it is necessary to reduce the
condition number of the matrix C by, for example, adding
thereto a diagonal matrix B  k I (k> 0). Then we find the
solution in a class of ridge parameter estimates:
b ( U T U  k I ) 1 U T Y .
The quality, obtained using models (7), depends on the
number of available output measurements. The length of the
training sample is often insufficient to obtain reliable results.
Also, the available data contains significant measurement
errors in inputs and outputs, which are unmeasured
influences. Taking into account constraints on the model
coefficients bj allows us to avoid these problems. When
taking into account constraints on the model parameters, we
solve the problem of least squares with simple constraints on
the variables:
min  Y  Ub 
2
b min
bb max
. of search pk   AT  p FR , where tk is the number of active
of Lagrange multipliers corresponding to the lower and upper
active constraints.
Below is the algorithm for searching the minimum point b*
for each iteration k.
1. Find the starting point according to (7) in order to initialize
the method of the active set.
(7)
2. Verify the performance of the stop conditions. (Reaching
the performance errors of conditions (9), which are
constraints on the number of iterations).
3. Select a logic branch. Does it make sense to remove any
constraint from the list of the set of active constraints? The
condition of performance 3 in (9) is checked. If the
condition is not satisfied for some of the vector elements,
the constraint is excluded from the list of active
constraints.
4. Calculate the search direction pk. Equation (7) solves the
problem min  Y  Ubk  U FR p FR 2 . Calculate the non-zero
(8)  N  1  tk  , the dimensional vector p FR , and the direction
FR

The solution of the problem (8) is obtained by the active set constraints on k iterations.
numerical method (Gill et al., 1981). The given constraints 5. Calculate the step length αk. We calculate the diagonal
b p
are reduced to the form: matrix Ψ from  FR   Ψ  FR  
bˆ FR , bˆ FR consists of
 b
 FR   p
 FR 
Ab  bˆ , the elements b̂ , which aren’t active constraints. The
1 0  0 elements b̂ , which are opposite boundary values in (8) for
0 1  0 constraints in the active set, are excluded from bˆ FR . The
 
     bmin   k  min Ψii  is an available maximum positive step from
where  ˆ
1  , b   max  .

0 0 
A b  bk along pk. We remember the index j of minimum
 1 0  0
  positive diagonal element Ψ. If k  1 , then k  1 ;
 0 1  0
   
otherwise,  k   k .
  6. Add a constraint to the list of active constraints. If  k   k ,
 0 0  1
then j constraint bˆ FR becomes active and necessary to add
Constraint aTi b  bˆi is active in acceptable point b if aTi b  bˆi , to the list.
and inactive if aTi b  bˆi , a T is the i-th row of A. The sufficient 7. Approximate a recalculation. After bk 1 bk  k pk is
i

minimum conditions for simple constraints are as follows: calculated, return to step 2 of the algorithm.

1. b m in  b *  b m a x bmin * max 3.2 Dynamic model identification under constraints

FR  bFR  bFR

2. UFR ( Y  Ub ) 
T *
0
Let u [1, u1 ( ),..., u1 (  n1  1),..., uN ( ),..., uN (  nN  1)]T be the

3.λ min T
U min ( Y  Ub* ) λimin  0 i  1, , t min (9) combined vector of the measured input variables of dynamic
4.λ max U Tmax ( Y  Ub* ) , λ max  0 , i  1, , t max
i SS (DSS) with dimensionality q 1   kN 1 n k , where nk is a
5. UTFR U FR is the positive definite number of values of the k-th input variable and
h  (h0 , h1 (1), ..., h1 (n1 ), ..., hN (1), ..., hN (nN ))T is the vector FIR of the
where b* is the minimum point of the solution of problem (8); same dimension, the components of which reflect the
subscript FR indicates that in the vector and matrices, the contributions of the respective input variables of DSS. Then
elements and columns with index numbers corresponding to
the index numbers of b elements that have not met the y uT  h .
the equation (2) takes the following form: 
boundary values of (8) are used; the subscript min and max We write the vector Y of dimension q from the output value y
indicate that in the matrix, only the columns with index as
numbers corresponding to the index numbers of b elements Y  ( y ( 1 ), y ( 2 ),..., y ( q ))T
that take the appropriate minimum or maximum boundary and matrix U, containing the measured inputs uj,
value are used. tmin and tmax refer to the numbers of active corresponding to output value y from (2) as
upper and lower limits, respectively; min and max are vectors

562
2018 IFAC ADCHEM
Andrei Torgashov et al. / IFAC PapersOnLine 51-18 (2018) 566–571
Shenyang, Liaoning, China, July 25-27, 2018 569

1 u1 ( 1 )  u1 ( 1  n1  1)  uN ( 1 )   s max
j (1)  a3b max j

1 u ( )  u (  n  1)  u ( )   
  
U
1 2 1 2 1 N 2

        s max
j (fl(a1 n j ))  a3b maxj
.
  
1 u1 ( q )  u1 ( q  n1  1)  uN ( q )
 uN ( 1 ) 

uN ( 1  n N  1)) 
.
s max
j
 max
  s j (fl(a1 n j)  1) b max
j  
1  ( a3  1) 1  1/ (n j  fl(a1 n j ))  

 max 
 uN ( 2 )  uN ( 2  n N  1)) 
 
)  2) b max
 s j (fl(a1 n j  j  
1  ( a3  1) 1  2 / (n j  fl(a1 n j )) 

       
  
 uN ( q )  uN ( q  n N  1))  s max ( n )  b max
 j jj j 

Wee write the m matrix equationn as Y  U h . We introducce the

For b j  0 :
error function:
E  Y  Y  Y  Uh ,
 s min
j (1)  a3b j
min

 
wh here Y is the aactual measurrement of outtput, and miniimize   
thee following obbjective function:  s j (fl(a1 n j ))  a3b min
min
j

 
 2
Ψ E h)2 .
( Y  Uh (10) s min
j
 min
  s j (fl(a1 n j )  1) b min
j  
1  ( a3  1) 1  1/ (n j  fl(a1 n j )) 

 min 
The constraints on transient response comp
ponents are w
written  s j (fl(a1 n j )

 2) b min
j  
1  ( a3  1) 1  2 / (n j  fl(a1 n j )) 

as   
 s min ( n )  b min 
smiin  s  smax , (11)  j jj j 

wh
here , smin   s1min , ..., s min  ,
T
s   s1 (1), ..., s1 ( n1 ), ..., sN (1)), ..., sN ( nN )   s max 
T
N  j ((1)  0
 
s max   s1max , ..., s max
x

T
  
N 
. We obtain the constraaint (11) baseed on  s max a 
(fl( 1n j ))  0
,
j
thee value of bj ( j  0) of thee steady-state soft sensor m model 
y (
) b0  b1u1 ( )  ...  b j u j ( )  ...  bN u N ( ) . That s max
j
 max max
 
  s j (fl(a1n j )  1)  b j a2  1  a2  1/ (n j  fl(a1n j ))  

 max 
moodel is deriveed prior basedd on the steaady-state induustrial
datta of CDU. WWe select the parametric co onstraints for each 
max
 
 s j (fl(a1n j )  2)  b j a2  1  a2  2 / (n j  fl(a1n j ))  

FIR o a1, a2 and a 3.
R coefficient uusing the folloowing values of   
 s max
( n )  b max 
 j j 
a1 is a fraction of the FIIR length n j from whichh the
j

nvergence of elements of s min

con j
and s ma
j
ax
to bmin
j , bmjmax is wherre fl means roounding to thee nearest integ
ger number inn the
started. direcction of -∞.
An example
e for deetermining thee constraint sy
ystem in term
ms of
i a fraction off b min
a2 is j  0 ( b max
j  0 ) from
m which the sm
mooth max
s min
j and s j off s j under b j  0 is show
wn on the Figg. 2.
inccreasing (or ddecreasing) of o s min
o elements of j
(or s max
j
) We select
s the paraameters a1=0.55, a2=0.8, and
d a3=1.2.
beg
gins.

a3 is a fractionn of bmax
j  0 ( b min
j  0 ) from whichh the
mooth decreasiing (or increaasing) of elem
sm ments of s min
j
n
(or
s max
m
j
) begins.

Forr b j  0 , the ffollowing equuations are vallid for constraaints

 s miin
j (1)  0

 
  
 s min
j (fl
fl( a 1n j ))  0

 
s min
j
n
 min b min
  s j (fl(a1n j )  1)  j  
a2  1  a2  1/ (n j  fl(a1n j ))  

 min 
b min
 s j (fl(a1n j )  2) 

j  
a2  1  a2  2 / (n j  fl(a1n j ))  

   Fig. 2. The assign
nment of consstraints for eaach step respoonse
 s min
( n )  b min  function of the dyn
namic soft sennsor model.
 j j j 

563
2018 IFAC ADCHEM
570 Andrei Torgashov et al. / IFAC PapersOnLine 51-18 (2018) 566–571
Shenyang, Liaoning, China, July 25-27, 2018

The coefficients of the finite step response (FSR) s j are 4. Calculate the search direction pk. Use an equation like (7)
to solve the problem min  Y  Uhk  UZk pZ  . Calculate the
2
related to the components of the FIR hj by the relations

s j  k   i 1 h j (i ),
k
j  1, 2 ,  , N , k 1, , n j . (12) 
non-zero 1  kN1 nk  tk  dimensional vector pZ and the
direction of search pk  Zk pZ , where tk is the number of
active constraints on the k iteration.
The constraints (11) can be written as 5. Calculate the step length  k . From A  h  Ψp k   sˆ , the

Ah  sˆ , (13) diagonal matrix Ψ is calculated. The calculated

k  min Ψii  is a minimum non-negative available step
where from hk along pk, where i is the index number of the
1 0  0  element, which is not an active constraint in (13) or an
1 1  0   h1 (1) 
    
element of opposite boundary values (11) for constraints in
        the active set. Also, p k consists of elements of pk without
 ,  h1 ( n1 )  ,  s min 
0 0  1 sˆ   max  . the first element. We memorize the index j of the
A  
h    

 1 0  0   s  minimum positive diagonal element of Ψii. If  k  1 , then
   hN (1) 
 1 1  0     k  1 ; otherwise, k  k .
       
  h (n )
 N N  6. Add the constraint to the list of active constraints. If
 0 0   1  k  k , then j constraint sˆ FR becomes active and Zk is
recalculated.
The sufficient minimum conditions are as follows for the 7. Calculate the hk 1 hk  k pk and return to step 2.
current optimization problem (parametric identification of the
dynamic model): 4. INDUSTRIAL CASE STUDY
1. Ah  sˆ , A ACT h  sˆ ACT The CDU (Fig.1) is considered a case study for evaluating
2. Z T * U T ( Y  Uh* )  0 static (steady-state) and dynamic soft sensors based on the
(14) proposed identification algorithm under parametric
3. λ  A ACT A TACT  A ACT U T ( Y  Uh * ) , λ i  0 ,
1
i  1,  , t constraints. The final boiling point temperature of the DC1
4. ZT U T U Z is a positive definite, when we obtained a model on the training sample is
considered. The number of output observations in the training
where h * is the minimum point of the solution of problem sample is 70. The length of the test dataset is equal to 30
(10) with constraints (13); subscript ACT indicates that in the observations. Fig. 3 and Table 2 show the results of the
vector matrix, we only use the rows with index numbers performance of the static models on the test sample.
corresponding to the element index numbers of active Table 2. Results of the performance of the static models
constraint in (13); t is the number of active constraints;  is (test dataset)
the vector of Lagrange multipliers corresponding to the active
constraints; Z is the matrix of the columns that are the basis R2 RMSE AIC BSC
of the feasible direction of the search for equality constraints Without use
0,64 4,47 232,7 255,4
(13). The matrix Z is formed by the variable-reduction constraints
technique (Gill et al., 1981). With use constraints 0,84 2,99 175,5 198,2
The search algorithm of minimum point h * for iteration k is
360
as follows:
355
1. In order to start the method of the active set, it is necessary
350
to determine the starting point (using a solution of the
problem (10) without any constraints, with subsequent 345

correction of coefficients hi that do not fall under the 340

FBP,°С

constraints (13)). 335

2. Verify the performance of the stop conditions (reaching
330
the performance errors of conditions (14) and constraints
on the number of iterations). 325
3. Select a logic branch. Does it make sense to remove any 320 lab data
constraint from the set of active constraints list? The 315
w/t constrains
with constrains
condition of performance of a condition 3 in (14) is
29.06.2017 06.07.2017
checked. If the condition is not satisfied for some of the Time period
vector elements, constraint is excluded from the list of
active constraints, and it is necessary to recalculate Zk. Fig. 3. Comparative study of static soft sensor model’s
performance.

564
2018 IFAC ADCHEM
Andrei Torgashov et al. / IFAC PapersOnLine 51-18 (2018) 566–571
Shenyang, Liaoning, China, July 25-27, 2018 571

The improvement of the prediction quality by the criterion that the reduction of the root mean square error on the test
RMSE of the identified static model obtained with the sample can be more than 33%.
constraints on the parameters of SS is 100(4,47 – 2,99) /
7. ACKNOWLEDGEMENTS
4,47  33% compared to the case without constraints. In
order to investigate the influence of constraints on the This work was partially supported by the Russian Foundation
performance of the dynamic soft sensor model we obtained, for Basic Research (Project No. 17-07-00235 A) and by the
we compared solutions of the optimization problem (10) Ministry of Education and Science of the Russian Federation,
without constraints and optimization problem (10) with through Government Contract No. 02.G25.31.0173.
constraints (13). We use the same values of the ridge REFERENCES
coefficients. Fig. 4 and Table 3 show the results of the
performance of the dynamic models on the test sample for the Akaike, H. (1969). Fitting Autoregressive Models for
final boiling point temperature of DC2 when a model is Prediction. Annals of the Institute of Statistical
obtained based on the training sample. The number of Mathematics, 21, pp. 243-247.
measurements in the training sample is 280. The size of the Bolf, N., Galinec, G. and Baksa, T. (2010). Development of
test dataset is 120 for the case with the dynamic model of the soft sensor for diesel fuel quality estimation. Chemical
soft sensor. Engineering & Technology, 33, pp.405-413.
Chatterjee, T., and Saraf, D. N. (2004). On-line estimation of
Table 3. Results of the performance of the dynamic product properties for crude distillation units. Journal of
models (test dataset) Process Control, 14, pp. 61–77.
R2RMSE AIC BSC Dam M. and Saraf D.N. (2006) Design of neural networks
Without use using genetic algorithm for online property estimation of
0,40 3,79 6418,4 6482 crude fractionator products. Computers & Chemical
constraints
With use Engineering, 30 (4) pp. 722–729.
0,74 2,52 4453,2 4516,8 Fujii K. and Yamamoto T. (2014). Development of a
constraints
The improvement of the prediction quality by the criterion nonlinear soft sensor using a GMDH network for a
RMSE of the dynamic model obtained using the constraints refinery crude distillation tower, Electrical Engineering
on the parameters of SS is 100(3,79 – 2,52)/ 3,79  33,5 % in Japan, Vol. 188, No. 3. pp. 31-38.
compared to the identification without constraints. Gill, P.E., W. Murray and Wright M.H. (1981). Practical
Optimization. New York, USA: Academic Press.
Johansen, T. (1996). Identification of non-linear systems
in-line analyser
w/t constrains using empirical data and prior knowledge-an
190
with constrains optimization approach. Automatica, Vol. 32, No. 3, pp.
337–356.
185 Macias-Hernandez J. J., P. Angelov and Zhou, X. (2007).
Soft sensor for predicting crude oil distillation side
streams using Takagi-Sugeno evolving fuzzy models.
FBP,°С

180 Proc. 2007 IEEE Intern. Conf. on Systems, Man, and

175
170
04.05.17 05.05.17
Time period
Fig. 4. Comparative study of dynamic soft sensors.
6. CONCLUSIONS
The introduction of a system of constraints into the
identification algorithm improves the quality of derived
models, especially on the test dataset. The reason is related to
the integration of available a priori knowledge via the
constraint identification problem statement and solution.The
use of the active set method, taking into account constraints
on the model coefficients, can improve the quality of the
evaluated SS models, in particular in the case of small
training datasets containing lab errors. The test of the
proposed approach to solving the problem of obtaining a soft
sensor model for industrial crude oil distillation unit showed
Cybernetics, 7-10 Oct., 2007, Montreal, Canada, ISBN
1-4244-0991-8/07, pp.3305-3310
Mahalec, V. and Sanchez, Y. (2012). Inferential monitoring
and optimization of crude separation units via hybrid
models. Computers & Chemical Engineering, 45, pp. 15-
26.
Napoli G. and Xibilia M.G. (2011). Soft sensor design for a
topping process in the case of small data sets. Computers
and Chemical Engineering, Vol. 35. No. 11, pp. 2447–
2456.
Shang C., Yang F., Gao X. and Huang D. (2015) A
comparative study on improved DPLS soft sensor
models applied to a crude distillation unit. IFAC-
PapersOnLine, Vol. 48. Issue 8, pp. 234–239.
Torgashov A. and Zmeu K. (2015). Nonparametric soft
sensor evaluation for industrial distillation plant.
Computer Aided Chemical Engineering. Vol. 37, pp.
1487-1492.
Torgashov A., Skogestad S. and Kozlov A. (2016).
Comparative study of multicomponent distillation static
estimators based on industrial and rigorous model
datasets. IFAC-PapersOnLine. Vol. 49. Issue 7, pp.
1187-1192.

565