Proceedings of the 20th World Congress

Proceedings
The of
of the
International
Proceedings 20th
20th World
Federation
the Worldof Congress
Automatic Control
Congress
Proceedings
The of the 20th World Congress Control
The International
Toulouse, France,Federation
International Federation of
of Automatic
July 9-14, 2017 Available
Automatic online at www.sciencedirect.com
Control
The International Federation
Toulouse, of Automatic Control
Toulouse, France,
France, July
July 9-14,
9-14, 2017
2017
Toulouse, France, July 9-14, 2017
ScienceDirect
IFAC PapersOnLine 50-1 (2017) 1213712142
Analysis
Analysis and
and Modelling
Modelling of
of an
an Industrial
Analysis
Pressure and Modelling
Filtration using an Industrial
ofProcess 
Industrial
Data
Pressure
Pressure Filtration
Filtration using Process
using Process Data
Data 

F.D. Bahner
P.A. Santacoloma J. Abildskov
F.D.
F.D. Bahner
P.A.
Bahner
P.A. Santacoloma
Huusom
Santacoloma J.
J. Abildskov
F.D. J.K.
Bahner
P.A.
J.K.Santacoloma
Huusom J. Abildskov
Abildskov
J.K.
J.K. Huusom
Huusom

Process and Systems Engineering Center (PROSYS), Department of
Process and Systems Engineering Center (PROSYS), Department of
Process and Systems Engineering Center
Chemical
Process andand
Chemical and
Biochemical
Systems Engineering
Biochemical Center (PROSYS),
Engineering,
Engineering, (PROSYS),
Technical
Department
Technical University
Department
University
of
of
of
of
Chemical
Denmark,
Chemical and
and Biochemical
2800 Lyngby, Engineering,
Biochemical Technical
Denmark (e-mail:
Engineering, University
jkh@kt.dtu.dk).
Technical University of
of
Denmark,
Denmark, 2800
2800 Lyngby,
Lyngby, Denmark
Denmark (e-mail:
(e-mail: jkh@kt.dtu.dk).
jkh@kt.dtu.dk).
CP Kelco
Denmark, Aps,
2800Ved Banen
Lyngby, 16, 4623
Denmark Lille
(e-mail:Skensved, Denmark
jkh@kt.dtu.dk).
CP Kelco Aps, Ved Banen 16, 4623 Lille Skensved, Denmark
CP Kelco Aps, Ved Banen 16, 4623 Lille Skensved, Denmark
CP Kelco Aps, Ved Banen 16, 4623 Lille Skensved, Denmark
Abstract: In order to understand a series of pressure leaf filters located in the downstream
Abstract:
Abstract: In
In order production
to understand site,a series of pressure leaf filters
have located in the downstream
Abstract:
line of a In order
line of a bio-based to
to understand
order production
bio-based understand site,
a series
series of
historical
a historicalof pressure
pressure
process
leaf
leaf filters
process data
data filters
have
located
located
been
in
in the
been analysed. the downstream
analysed.
In general,
downstream
line
line of
changing
of a
a bio-based
raw materials
bio-based production
production site,
induce variability
site, historical
into the
historical process
pressure
process data have
profiles
data have been
and
been analysed.
thereby
analysed.cycleIn
In
In
general,
general,
durations
general,
changing
changing
of the raw
raw
manually materials
materials induce
induce
reinitialised variability
variability
dead-end into
into the
the
filtrations. pressure
pressure
The profiles
profiles
absence of a and
and
true thereby
thereby
steady cycle
cycle
state durations
durations
results in
changing
of the raw
manually materials induce
reinitialised variability
dead-end into the
filtrations. pressure
The profiles
absence of and
aa true thereby
steady cycle
state durations
results in
of
of the manually
uncertainty about
the manually reinitialised
the optimal
reinitialised dead-end
way
dead-endof filtrations.
running the The absence
filters,
filtrations.filters, and
The absence of
staff true
members
of members steady
a true steadyalter state
the results
operational
state in
results in
uncertainty
uncertainty
specificationsabout
about the
the
frequently.optimal
optimal way
way
It appears of
of running
that, in the
running the
somefilters, and
and
cases, staff
staff
this members
propagates alter
alter the
the operational
operational
disturbances rather
uncertainty
specificationsabout the
frequently.optimal
It way
appears of running
that, in the
somefilters, and staff members alter the operational
specifications
than ameliorate
specifications frequently. It
them. Statistical
frequently. It appears that,
analyses
appears that, in
are
in carriedcases,
some
some out tothis
cases,
cases, this propagates
propagates
illustrate
this propagates
disturbances
disturbances
the current situation
disturbances
rather
rather
and
rather
than
than ameliorate
ameliorate
especially allow them.
them. Statistical
Statistical
quantifying the analyses
analyses
extent of are
are
the carried
carried out
out
uncertainties. to
to illustrate
illustrate
Furthermore, the
the current
current
significant situation
situation and
and
correlations
than ameliorate
especially allow them. Statistical
quantifying the analyses
extent ofand are
the carried out
uncertainties. to illustrate
Furthermore, the current
significant situation and
correlations
especially
especially allow
allow quantifying
between process variables are
quantifying the extent
extent of
revealed
therevealed the
ofand uncertainties.
economically
the economically Furthermore,
motivated
uncertainties.motivated
Furthermore, significant
operational
significant correlations
objectives are
correlations
between
between
identified.process
process
Secondly,variables
variables are
are
working revealed
towards and
on-lineeconomically
predictions motivated
of filtrationoperational
operational
performance,objectives
objectives
a are
are
model
between
identified.process
Secondly,variables are towards
working revealed on-line
and economically
predictions motivated
of filtrationoperational
performance,objectives
a are
model
identified.
is presented.
identified. Secondly,
It is based
Secondly, working
working towards
on classical
towards on-line
filtration
on-line predictions of filtration
theory andofrequires
predictions filtration onlyperformance, a model
commonly available
performance, a model
is
is presented.
presented. It is based on classical filtration theory and requires only commonly available
is presented. It
measurements
measurements
is
is based
(pressure,
It(pressure, on
on classical
flow,
based flow, viscosity).
classical
viscosity).
filtration
filtration
The
theory
theory and
The generated
generated and requires
predictions
requiresare
predictions
only
are
onlyfound
found
commonly available
to be acceptable
commonly
to be available
acceptable
measurements
for many
measurementscycles,(pressure,
but in
(pressure, flow,
some
flow, viscosity).
cases fail
viscosity). due The
Theto generated
non-modelled
generated predictions
effects,
predictions are found
motivating
are found to be
further
to be acceptable
work.
acceptable
for
for many
many cycles,
cycles, but
but in
in some
some cases
cases fail
fail due
due to
to non-modelled
non-modelled effects,
effects, motivating
motivating further
further work.
work.
for many
2017, cycles,
IFAC but in some
(International cases fail
Federation due to non-modelled
of Automatic Control) Hostingeffects, motivating
by Elsevier Ltd. Allfurther work.
rights reserved.
Keywords: Biosystems and bioprocesses; Downstream processing; Parameter and state
Keywords:
Keywords: Biosystems
estimation; Biosystems
Data mining and
and bioprocesses;
ModellingDownstream
bioprocesses;
tools; Downstream processing;
processing;
and identification; PressureParameter
Parameter and
and state
leaf filtration state
Keywords:
estimation; Biosystems
Data mining and bioprocesses;
tools; Modelling Downstream
and processing;
identification; PressureParameter
leaf and state
filtration
estimation; Data mining tools; Modelling and identification; Pressure
estimation; Data mining tools; Modelling and identification; Pressure leaf filtration leaf filtration
1. INTRODUCTION is non-trivial from an operational point of view. Discrete
1. INTRODUCTION
1. INTRODUCTION is
is non-trivial
such asfrom
non-trivial from an
an operational point of view. Discrete
1. INTRODUCTION events
is non-trivial
events such from
as an operational
cleaning
cleaning
(the removal
operational
(the
point
pointof
removal of
of view.
offilter
view.cake)
filter
Discrete
and
Discrete
cake) and
To date, cake filtration is one of the workhorses of the events
events such
re-initialisation as cleaning
happen
such as cleaning (the
on
(thea removal
frequent
removal but of
but filter cake)
irregular
of filter and
basis.
cake)basis.
and
To
To date, cake
date, cake filtration
filtration is one
is one of the
of thetoworkhorses
workhorses of
of the
the re-initialisation
re-initialisation
By nature, thesehappen happen
happen on
on
disruptions a
a frequent
frequent
upset but
the butsteady irregular
irregular
state ofbasis.
the
biochemical
To date, cake industry
filtration when it comes
is one of theto the separation
workhorses of of
the re-initialisation on a frequent irregular basis.
biochemical
biochemical industry
industry when
when it
it comes
comes to the
the separation
separation of
of By
By nature,
nature,
surrounding these
these disruptions
line. disruptions
Buffer tanksupset upset the
have athe steady
steady
mitigating state
state of
of
effect, the
the
but
demanding slurries
biochemical industry with whena highit concentration
comes to the of suspended
separation of By nature, these disruptions upset the steady state of the
demanding
demanding slurries with aa highhigh concentration of suspended surrounding
surrounding
in reality oneline.
line. Buffer
Buffer
still finds tanks
tanks have
have
propagation a
a mitigating
mitigating
through effect,
effect,
the but
but
system,
solids. Due slurries
demanding slurries with
to its versatile
with a high concentration
applicability
concentration and of of suspended
robust sepa-
suspended surrounding line. Buffer tanks have a mitigating effect, but
solids. Due to its versatile applicability and robust sepa- in
in reality
reality one
one still finds propagation
filtration through
is pronethe to system,
solids.
ration
solids. Due to
properties,its versatile
both
Due to its versatile of applicability
which have
applicability and
been robust
demonstrated
anddemonstrated sepa-
robust sepa- which
in reality
which
is why
is one still
why still finds
finds propagation
intermittent
intermittent propagation
filtration
through
through
is prone
the
the
to
system,
induce
system,
induce
ration
ration
over many properties,
properties,
decades, both
both of
it of which
is which have been
have
an established been demonstrated
technology and which
variabilityis why
into intermittent
process and filtration
ultimately is prone
product. to induce
ration properties, both of which have been demonstrated which
variabilityis why
into intermittent
process and filtration
ultimately is prone
product. to induce
over
over
far frommany
many decades,
decades,
being replaced it is
it is an
in an established
theestablished technology
foreseeable technology
future. A trend and
and variability
In many into process
biochemical and
plants, ultimately
the absence product.
of a true steady
over many decades, it is an established technology and variability
In many into process
biochemical and
plants, ultimately
the absence product.
of aa true steady
far
far fromautomation
from being replaced
being replaced in the
in the foreseeable future. A trend
trend In many
state is anbiochemical
obstacle inplants,
the the absence
quantitative of
analysis trueof steady
down-
toward
far fromautomation
toward being replaced in in the foreseeable
in industrial
industrial
filtration is
foreseeable
filtration
future.
not new
future.
is not new
A - the
A trend
- the
In many
state is anbiochemical
obstacle inplants,
the the absenceanalysis
quantitative of a true of steady
down-
toward automation in industrial filtration is not new state is an
streamis processes.obstacle in
Thus, the quantitative
inquantitative
order to provide analysis of down-
a ofworking
economic
toward
economic
success
automation
success
of
of
automatically
inautomatically
industrial filtrationdischarging new - the
is nothorizontal
discharging
-
horizontal the state
stream an obstacle
processes. in
Thus, the in order to analysis
provide a down-
working
economic success
leaf filterssuccess
was shown of automatically
already in the discharging horizontal
1970s, seehorizontal
Rushton stream
basis,
stream it processes.
is necessary
processes. Thus,
to
Thus, in
reduce
in order
order to
to provide
uncertainties
provide to a
a a working
feasible
working
economic of automatically discharging basis, it is Currently
necessary to
leaf
leaf filters
et al.filters
leaf filters
wasEspecially
was
(1996). shown already
shown
wasEspecially
already
shown already when in in
in
the 1970s,
the
handling 1970s,
the 1970s,
see Rushton
see Rushton
pharmaceuticals
see Rushton
basis,
minimum.
basis, it
it is necessary
necessary to
is Currently thereduce
to reduce
incentive
reduce
uncertainties
uncertainties
for vast capital
uncertainties
to
to
to a
a feasible
a feasible
invest-
feasible
et
et al.
al. (1996).
(1996). Especially when
when handling
handling pharmaceuticals
pharmaceuticals minimum.
minimum.
ments in Currently
downstream the
the incentive
incentive
processes isfor
for vast
vast
limited, capital
capital
as invest-
invest-
competi-
or
et food
al. products,
(1996). closed
Especially systems are desirable for the minimum. Currently theprocesses
incentiveisforlimited,vast capital invest-
or
or food
food
reduced products,
products,
risk of closedwhen
closed
contamination. systems
systems
handling
are
Nevertheless,
pharmaceuticals
are desirable
desirable
many for the
for the
current
ments
ments in
in downstream
downstream
tivenessinpredominantly processes is
originatesisfrom limited,
a high as
as competi-
competi-
degree of
or food risk products, closed systems are desirable for the ments
tiveness downstream
predominantly processes
originates limited,
from a high as competi-
degree of
reduced
reduced
plants have risk of
of contamination.
contamination.
downstream processes Nevertheless,
Nevertheless,
where manual many
many current
current
operation tiveness predominantly
product-innovation. originates
Beyond this, thefrom aa high
long-term degree
ambition of
reduced riskdownstream
of contamination. Nevertheless, many current tiveness predominantly
product-innovation. originates
Beyond this, from
the long-termhigh degree
ambition of
plants
plants
(also on have
have downstream
decision-making processes
processes where
level) is where
regarded manual
manual
normal.operation
operation
There product-innovation.
of transitioning from Beyond this,
recipe-driven the long-term
open-loop ambition
production
plants have downstream processes where manual operation product-innovation.
of transitioning from Beyond this,
recipe-driven the long-term ambition
(also
(also
are
(also
on
on
variousdecision-making
decision-making
reasons
on decision-making for level)
level)
this, but
level)
is
isinregarded
regarded
particularnormal.
normal. There
There
persistently of
to
of transitioning
lean, automated
transitioning from requires open-loop
recipe-driven
fromplants
recipe-driven open-loop
sophisticated
open-loop
production
production
change
production
are
are various
various
changing reasons
reasons
process for this,
for this,
conditions butis
but
are inregarded
in particular
particular
challenging
normal. There
persistently
persistently
from a control
to
to lean,
lean,
management. automated
automated plants
plants
Undoubtedly requires
requires
a challenging sophisticated
sophisticated
endeavour, change
change
and
are various reasons for this, but in particular persistently to lean,
management. automated plants
Undoubtedly requires
aa challengingsophisticated
endeavour, change
and
changing
changing process
process conditions
conditions arecan
are challenging
challenging
generallyfromfrom aa control management.
it is for Undoubtedly
industrial practitioners challenging
and endeavour,
academia to and
work
engineering
changing
engineering
perspective.
process conditions
perspective.
They
They arecanchallenging
generally from a control
be attributed
control
be attributed
attributed
management.
it is for Undoubtedly
industrial practitionersa challenging
and endeavour,
academia to and
work
engineering
to short-lived perspective.
product They
cycles can
and generally
inconsistentbe feedstock it
it is
together,
is for
for industrial
recognising
industrial practitioners
and exploiting
practitioners and
and academia
progress
academia that to
has
to work
been
work
engineering perspective. They can generally be attributed together,
to
to
to
short-lived
short-lived
(weather,
short-livedregion,product
product
product
cycles
cycles and
pre-treatment,
cycles
and
andetc.).inconsistent
inconsistent
The necessary
inconsistent
feedstock
feedstock
adap-
feedstock
together,
made
together, in recognising
recognising
recognising
and
and exploiting
and exploiting
sensor technology, exploiting
progress
progress
monitoring-
progress and that
that
that
has
has been
has been
ultimately
been
(weather, region, pre-treatment, etc.). The necessary adap- made
made in
in
methodological sensor
sensor technology,
technology, monitoring-
monitoring- and
and ultimately
ultimately
(weather,
tivity
(weather, for region,
dealing
region, pre-treatment,
with these
pre-treatment, etc.).
uncertain The
etc.). The necessary
conditions adap-
is
necessaryisadap- pro- made in sensortools.
methodological technology, monitoring- and ultimately
tools.
tivity for keeping
dealing with these
theseinuncertain
uncertain conditions pro- methodological
Production tools.
processes
tivity
vided by
tivity for
for dealing operators
dealing with
with these the loop conditions
uncertain and would is
conditions is pro-
other-
pro- methodological
Production tools. that
processes that
have been running for decades
have
vided
vided
wise
vided
by keeping
by
require
by
keeping
keeping
operatorsautomation
operators
sophisticated
operators
in the
in
in
the loop
the
loopsolutions.
loop
and would
and
and
would other-
would
other-
other-
Production
have seen
Production processes
changes
processes for that
the
that betterbeen
have
have due running
been
been running for
for
to the involvement
running
decades
for decades
decades
wise
wise requirefiltration,
require sophisticated
sophisticated automation
automation solutions.
solutions. have
have
of seen
seen
capable changes
changes
engineersforfor
and the
the better
better
operators. due
due to
to
However, the
the involvement
involvement
these heuris-
In general,
wise requirefiltration,
sophisticated a seemingly
automation simple processing step,
solutions. have seen changes for the better due to the involvement
In
In general,
general, filtration, a
a seemingly
seemingly simple
simple processing step,
processing step, of
ticcapable
of capable
findings, engineers
engineers and
and
while robust operators.
operators. However,
However,
in their applicability, these
theseareheuris-
heuris-
often
InThe
 general, filtration, a seemingly simple processing step, of
ticcapable
findings, engineers
while and
robust operators.
in their However,
applicability, theseare heuris-
often

authors would like to acknowledge the Technical University tic
tic findings,
restricted
findings, to while
a local
while robust
modus
robust in
in their
operandi.
their applicability,
A first
applicability, stepare
are often
toward
often

of The authors
Denmark would
(DTU) like
wouldand to
to acknowledge
likeBIOPRO2. the
the Technical
The project receivedUniversity
financial restricted to a local modus operandi. A first step toward
 The
The
authors
authors wouldand
acknowledge
likeBIOPRO2.
to acknowledge
Technical
the Technical
University
University restricted
gaining a to a
deeper local modus
understanding operandi.
as well Aas first step
applying toward
quanti-
of Denmark (DTU) The project received financial restricted
gaining to a local modus operandi. A first step toward
support
of Denmark
of
from(DTU)
Denmark (DTU)
support
Innovation Fund Denmark
and BIOPRO2.
and BIOPRO2.
through
The project
The project
the BIOPRO2
received financial
receivedBIOPRO2
financial gaining
tative
gaining aaa deeper
deepertounderstanding
analyses understanding
any process is as
deepertounderstanding
as well
thewell
as well
as applying
as
derivation applying
as applying
quanti-
quanti-
of a reliable
quanti-
support from
strategic Innovation
research
from Innovation Fund
center (Grant Denmark
Fund number
Denmark through
through the
4105-00020B).the BIOPRO2 tative analyses any process is the derivation of a reliable
support
strategic from Innovation
research center Fund
(Grant Denmark
number through
4105-00020B).the BIOPRO2 tative analyses to any process is the derivation
tative analyses to any process is the derivation of a reliable of a reliable
strategic research center (Grant number 4105-00020B).
strategic research center (Grant number 4105-00020B).
Copyright
2405-8963 2017,
2017 IFAC 12643
IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Copyright
Copyright 2017
under IFAC
2017 responsibility
IFAC 12643
12643
Peer review
Copyright 2017 IFAC of International Federation of Automatic
12643Control.
10.1016/j.ifacol.2017.08.2152
Proceedings of the 20th IFAC World Congress
12138
Toulouse, France, July 9-14, 2017 F.D. Bhner et al. / IFAC PapersOnLine 50-1 (2017) 1213712142
process model. Classical filtration theory has its pitfalls
when it comes to modelling and ultimately designing in-
dustrial equipment, as delineated early by Ruth (1946).
Likewise, some of the more recent modelling approaches
are hard to match with industrial practice, where filter
cake properties change frequently and cannot be deter-
mined in recurring laboratory experiments or by utilising
exhaustive sensor technology due to the involved expenses.
This work encompasses a description of the complex oper-
ational aspects that are inherent to intermittent filtration
processes at hand of a comprehensible case study. Statis-
tical analyses of industrial process data are carried out,
important correlations identified and the results thought-
fully visualised. A predictive model for industrial scale
filtration using available process data only is presented. It
is largely based on classical filtration theory, and a number
of forecasts are found to miss, likely due to non-modelled
effects, motivating further work in this area.
2. FILTRATION THEORY
Scientific approaches to understanding filtration date back
to the 1850s, when Darcy noted the proportional relation
between pressure drop p and liquid flow Q through a
packed bed of solids (1). Physically arguing, the solid-
liquid suspension passes through channels that form in
the bed. Friction losses, depending on viscosity and flow
regime, occur when the medium passes over a surface.
Larger particles will start to accumulate on the upstream
side of the cake, which therefore grows over time. Smaller
particles gradually adhere to walls and cavities while pass-
ing through the channel-matrix. Pressure losses across the
cake are thus linked to the depth of the permeated media,
which furthermore influences the quality of a separation.
However, characteristics such as particle size distribution,
shape and morphology, surface charge, and pH will deter-
mine the pressure gradient throughout the bed (Tarleton
and Willmer (1997)). It is convenient to express the flow-
pressure drop relation by means of a cake resistance :
A p
Q= (1)
L
Here, L denotes the measure for the depth of the bed,
A the cross-sectional area and viscosity. In reality
is not constant throughout the cake, as nearly all filter
cakes exhibit compressibility - meaning that permeability
properties in the cake are not constant. For a limited range
of pressures, an empirical approximation is given by
(p, n) = 0 pn , (2)
where n is a compressibility index (0 denoting a non-
compressible material) and 0 a reference resistance. If a
more convenient expression is desired, one finds a remedy
in averaging over the cake depth, as documented e.g. in
Tien (2006), which then yields
= (1 n) 0 pn . (3)
In order to extend the descriptiveness of equation (1) to
an entire filtration cycle, the relationship between cake
growth and time must be established. If the concentration
of accumulating particles on the top-layer (c) and likewise
the cakes structural integrity are sufficiently constant,
c Q dt c V
L= = (4)
A A
12644
yields the desired equivalence, in which V denotes the
amount of processed filtrate. Thus, equation (1) with (4)
and ultimately (3) yield the necessary link between time
of filtration, processed filtrate, and pressure drop:
dt V Rm
= c 2 + (5)
dV A p Ap
In (5), a second term has been included to account for
the medium resistance across the membrane (Rm ). It is a
common assumption to view Rm as constant, furthermore
is the cake resistance typically larger by several orders
of magnitude. Reliable first principle models are scarce
and complex, thus it should be regarded best practice to
determine any , c or Rm experimentally in lab- or ideally
pilot scale experiments. However, one should keep in mind
that even then, cross-scale validity is usually an issue, as
discussed in Tarleton and Hancock (1997).
3. FILTRATION PRACTICE
Filtration units deployed in continuous production lines
are often controlled such that they yield a uniform filtrate
flow (constant rate filtration) throughout a cycle. Ideally,
a cycle is meant to come to an end when the maximum
operating pressure pmax of either pump or membrane is
reached. Equipment is designed such that large surface
areas can be accommodated small pressurised vessels,
normally by stacking many plates side-by-side, similar to
the layout of plate heat exchangers. The membranes alone
provide a weak separation, as a consequence high flow rates
can lead to penetration of the septum in the early stages
of a filtration. This risk is commonly reduced by applying
a precoat layer, using an inert material. Throughout a
filtration the cake then grows on the membrane surface,
and each two plates (leafs) are arranged such that they
form a cavity through which the filtrate is discharged.
In the case of highly compressible solids, filter aid can
be added during the filtrations as a body feed. This
has a positive effect on porosity, enhancing capacity and
longevity of the filter cycles. A rather detailed study on
the effect of filter aid in a bioprocess is presented by
Meindersma et al. (1997).
3.1 Operational Aspects
In the case of manually cleaned filters, personnel expenses
contribute a large part of the overall operational costs,
which has implications for the optimal utilisation of the
equipment. Two not mutually exclusive but dependent
objectives, subject to a soft constraint, could be identified:
Operate the units such that the number of total
manual filter cleanings n is minimised (objective)
Use manpower efficiently by maximising the amount
of cleaned filters per operator and shift (objective)
Process all upstream feed as steadily as possible by
making use of the buffer capacities (constraint)
Constraint: Firstly, in order to maintain plant capacity, it
is necessary that

Q = i Qi = Qplant , (6)
where Q is the average flow coming from all filters and
Qi = Qi (t) dt Tf 1
i describes the average flow per cycle
on a single filter. See (7) for a derivation of the equipment
 Proceedings of the 20th IFAC World Congress
Toulouse, France, July 9-14, 2017 F.D. Bhner et al. / IFAC PapersOnLine 50-1 (2017) 1213712142 12139
Fig. 1. Relation between flow rate, time, and final pressure
efficiency measure i . Buffer tanks are indispensable due to
the intermittent nature of the process and allow temporar-
ily altering flow rates without upsetting the surrounding
line. Thus, as long as tank capacity suffices, equation (6)
can be expressed as a soft constraint in averaged values,
which ideally allows scheduling the end times by manipu-
lation of flow rates. In the case study plant this is carried
out by the operators on a rule-of-thumb basis.
Objective 1 : The number of necessary manual actions is
minimised if the equipment efficiency measure for each
filter i, namely
T fi
i = , Tfi = f (Q, p, Z) , (7)
Tfi + Tcl
is maximised. Cleaning and reinitialising a filter is a
standardised procedure and takes a certain time Tcl . The
varying filter cycle durations are denoted by Tfi . Lastly,
Z indicates disturbances such as e.g. fluid properties. To
maximise i , one must assure that the maximum amount
of retentate (solids) is accumulated on the membranes
before being removed. Intuitively this suggests striving
for an even distribution of the workload between units; a
data-driven, quantitative relation between average filtrate
flow and cumulative amount of filtrate per filtration could
not yet be derived with sufficient reliability. Furthermore,
flow - and thus pressure - increases which could cause cake
blinding should be avoided. Lastly, a filter is cleaned more
often than necessary if pf < pmax on take out, as this leads
to a decrease in Tf (figure 1) and thereby .
Objective 2 : The operator load factor can be expressed as
N Tcl
operator = , (8)
Tshif t
where N denotes the number of manual actions. Thus,
an ideal filtration area would at all times be run such
that operators do not experience idle times between con-
secutive filter cleanings (figure 2a). In reality, long cycles
in combination with the flow rate constraint must result
in idle time, unless the plant can handle a flow rate in-
crease. However, scheduling uncertainty is so large that
non-productive operator time (figure 2b) can emerge (or
units are taken out prematurely) even in cases where plant
throughput cannot be maintained. This has strong eco-
nomic implications due to the involved production losses
and calls for a new, robuster operational paradigm.
3.2 Pressure Leaf Filtration - Case Study Setup
The filtration unit under study is located within one of the
pectin production lines operated by CP Kelco ApS in Den-
mark. Pectin is a gelling and thickening agent used pre-
12645
Fig. 2. Operator engagement in reinitialisations (grey bars)
dominantly in the food industry, and is generally leached
from citrus peels in acidic extractions. Intuitively, one
would expect a gelling agent to be viscous in solution, and
it is indeed a challenging task to separate the extraction
broth. A multi-step filtration is necessary, and initially,
first- and second stage solids are removed. Subsequently
the suspension is fed to the pressure filtration area, where
several manual and automatic units run in parallel. As
the manually cleaned filters are of a greater interest from
a process-optimisation point of view, the automatic ones
have been neglected in the analyses for now. All filters are
of identical design, however, two of the five have a smaller
membrane area. Precoating is necessary and filter aid is
supplied as a body feed throughout the filtrations. No-
tably, the precoat layer is applied using process fluid, thus
one should expect some inconsistencies in its composition.
The operators are instructed to terminate a filtration upon
reaching a pressure threshold. The vessel is then opened,
leafs and supporting rack are slid out, and finally the
membranes are hosed down with water in order to remove
the filter cake. Thereafter, the vessel is reinitialised, which
includes running in precoat-recycle until all filter aid is
evenly distributed on the membranes. In this procedure, it
is up to the operators to schedule the filtrations such that
no blockages arise. Due to the amount of information (mul-
tiple filters) and the non-linearity of the pressure profiles,
this is a complex task, and the quality of these decisions
is sometimes questionable. Pressure filtration can become
the bottlenecking step, and it is unclear whether this is due
to disturbances alone or if suboptimal operation is an issue.
Thus, in order to prevent the operators from involuntarily
propagating disturbances or scheduling in an erroneous
way, the desire for a more deterministic operating regime
with augmented decision-making has been expressed.
4. STATISTICAL ANALYSIS
Before attempting to model the filters, historical pro-
cess data have been analysed to gain an overview of the
modelling scenario, furthermore to investigate uncertain-
ties and correlations. Operational data from some of the
pressure leaf filters, logged at 1-minute intervals over the
course of a month, have been analysed. Due to the high
number of regarded cycles (>250), manual processing of
the pressure profiles was not feasible and an algorithm has
been developed to extract them from the time-series data.
4.1 Data Handling
The data are logged on a continuous time axis, and the
individual filtration cycles have been identified with an
algorithm that recognises the initiation of a new run
Proceedings of the 20th IFAC World Congress
12140
Toulouse, France, July 9-14, 2017 F.D. Bhner et al. / IFAC PapersOnLine 50-1 (2017) 1213712142

Filter #1 #2 #3 #4 #5 Cycle initiated

a: Average flow throughout cycle
14

Avg. Q (m3 h-1)

12

10

0 5 10 15 20 25 30
b: Duration of cycle

Cycle Duration (h)

Fig. 3. Pressure profiles on cycle time axes 30

20
a: Cycle Durations b: Take-out Pressures
40 10
Scores

Scores

20 50 0 5 10 15 20 25 30
Time (d)

0 0 Fig. 5. (Scaled) average flows and actual durations

10 20 30 3 4
Cycle Time (h) Pressure on take-out (bar)
Fig. 4. Distribution of cycle times and take-out pressures
based on the pressure profile. A new cycle is generally
characterised by a sudden decrease in pressure below a
(variable) threshold:
pi1 pi 1 | pi 2 , i = 1, ..., end (9)
All cycles could be uniquely identified with 1 = 1 bar
and 2 = 0.5 bar. However, it was necessary to eliminate
data-points between cycles, which can occur if data is
being logged even though a filtration has come to an end.
Checking for
pi pi+1
pi1 pi+1 3 | 4 , i = 1, ..., end ,
pi1 pi
(10)
with 3 = 2 bar and 4 = 0.01 allowed to remove all outliers
of this type. The check against 4 identifies whether the
gradient undergoes a change in sign during the transition,
another necessary condition for a new cycle.
4.2 Frequentist Analysis
Looking at figure 3, where the pressure profiles have been
stacked on their cycle time axes, the uncertainty becomes
apparent. There is some variability in the final pressures,
but first and foremost the cycle durations catch the eye,
as they stretch from few hours to more than a day. This is
put into a different representation in figure 4a, where one
can see the majority of cycles coming to an end between 8
and 18 hours with a tail spanning much further. Figure 4b
shows a histogram of the final pressures. The exhibited
profile - unlike the previous one - seems to somewhat
resemble a normal distribution. This could be explained
with the randomness induced by the operators when they
decide to take a filter out of production. Notably, the mean
of this distribution lies below the pressure threshold that
is specified in the operational guidelines.
5 6
initiated, but an unquestionable correlation between the
average flows and durations of the cycles across all filters
can be determined. From an operational point of view,
similar flow profiles are to be regarded as positive due
to the even distribution of the workload. However, it is
interesting to note that the scaled but representative av-
erage flow rates vary significantly with time. Furthermore,
flow rates are generally lower on filters 1 and 2 due to
their smaller membrane area, but it is surprising to see
that there is a distinct difference between the two, as they
are equal-sized. Upon inquiring, the operating crew stated
that Filter 2 is believed to perform worse than Filter 1.
However, looking at the averaged values, given as [mean
(st. deviation)] in table 1, the differences in cycle times
and ultimately the number cycles per month should be
noted. Durations on Filter 2 are significantly higher, thus
it is likely that the operators run the filter at lower flow
rates out of an erroneous assumption. Looking at the total
filtrate output per month, Filter 2 performs only at about
93% of Filter 1. This is a good example of hidden capacity
that can be revealed by systematically analysing process
data. A different approach at visualising the inter-filter
correlations is shown in figure 6. The centres of the black
circles denote a point in time at which a cycle has been
started, and it is obvious that periods of short or long
filtrations are experienced across the filters. The coloured
lines are normalised and scaled representations of the aver-
age flow rates of these cycles. One can see that higher flow
rates imply shorter cycles, however, looking closely, there
are exceptions to this rule. Lastly, a normalised and scaled
viscosity measure has been added. It is plotted on the
line corresponding to Filter 4 (where the measurement is
taken). However, one should expect the fluid properties on
the filters to be similar, as they are supplied from the same
tank. It is evident that, even though there is a notable
correlation between viscosity and filtration performance,
there are exceptions to this rule also. Summarising, figure
6 shows how filtration theory and industrial practice do not

4.3 Correlated Process Variables Table 1. Cycle statistics across filters

Filter Flow (m3 h1 ) Duration (h) Nr. of cycles
It is interesting to determine whether there are distinct #1 9.3 (0.44) 12.6 (5.1) 53
differences between the way each filter is operated and, #2 8.7 (0.41) 14.1 (5.5) 48
ultimately, performs. A first visual analysis can be carried #3 12.3 (0.69) 12.8 (5.4) 50
out at hand of figures 5a and b. There is some uncertainty #4 12.2 (0.55) 13 (5.7) 50
in the calculation of the points in time when a cycle is #5 12.2 (0.65) 13 (5.3) 50

12646
 Proceedings of the 20th IFAC World Congress
Toulouse, France, July 9-14, 2017 F.D. Bhner et al. / IFAC PapersOnLine 50-1 (2017) 1213712142 12141

Q on filter #1 #2 #3 #4 #5 Initiation

#5

#4
Filter

#3

#2
Fig. 7. Cycles with and without changing flow rates
#1

0 5 10 15 20 25 30 a: Cycle Durations b: Take-out Pressures

Time [d] 40

Scores

Scores
Fig. 6. Clustering of short cycles & viscosity effects 20 50

align at all times. Due to the equal number of total cycles

(table 1), the correlation matrix ij between the cycle 0 0
10 20 30 3 4 5 6
durations on the large filters can be calculated with ease. Cycle Time (h) Pressure on take-out (bar)
Beyond this, there is a correlation between the durations of
consecutive cycles on each filter (k,k+1,i ), but it is weaker
and applies to a lag of one cycle only. Remarkably, the Fig. 8. Distribution of fitted parameters
correlation is quite low on Filter 1. This could be related to
the fact that operators choose to adjust the flow rate such
that it is slightly higher than that on Filter 2 (which had
been classified as a problematic filter by the operators).
 
1 0.77 0.78
ij = 0.77 1 0.76 , i, j = 3, 4, 5
0.78 0.76 1
T
k,k+1,i = ( 0.22, 0.43, 0.42, 0.36, 0.48 ) , i, k

5. MODELLING THE FILTRATION CYCLES Fig. 9. Prediction misses trend despite of smooth profile
The general ambition of this section is to assess whether
it is possible to derive a predictive filtration model using a: Sensitivity to parameters b: Simulated pressure profile
p (bar) | Q (m3 h-1)

only process data. In this case, due to the high variability 5

10
between the cycles, the model is fitted independently for n
dy/d

each cycle instead of looking for an optimal parameter set 0 5

across cycles, or even filters. With dV /dt = Q and (3),
the filtration equation - foregoing the membrane resistance
-5 0
and solved for the pressure difference - reads 0 5 10 0 5 10
  1n
1 Time (h) Time (h)
V c
p = Q (n 1) . (11) Fig. 10. Local sensitivity measure for = 2, n = 0.4
A2
Cake resistance and compressibility index are to be fitted.
Lacking the necessary data, the solids concentration c can- dictability, and are expected due to model structure
not be uniquely determined, thus the lumped parameter and the limited amount of data available for fitting.
c = c is estimated. Beyond this, flow rates fluctuate (2) It is a parameter that can be iteratively adjusted
notably in a number of cycles. Thus, in order to reduce without upsetting the model outcome much.
the model error, the integral for V has been solved using
Looking at figure 7, the chosen model seems reasonable.
trapezoidal integration rather than assuming V = Qt
Despite of minor upsets to the flow (right-hand side), the
(constant rate filtration). For all affected cycles, this has
model fit is within acceptable bounds. For cycles with
improved the quality of the fit. However, in order to pre-
extreme fluctuations in the flow profiles, the model fit
dict performance, it needs to be a strict guideline that
will generally be worse, but this is not strictly applicable
flow rates remain unchanged unless absolutely necessary,
for all cases (see figure 12a, where a good model fit can
as changing flow rates, especially if not based on deter-
be achieved despite of an unsteady profile). Beyond this,
ministic decisions, rapidly deteriorate the precision of the
the fitted parameters are often highly correlated, for the
predictions. Instead of including a term for the membrane
two depicted cycles at around 0.9. Looking at all cycles,
resistance, the pressure profiles have been truncated by
this correlation is seen to move between 0.3 and 0.99.
an initial offset - which is later-on re-added to the model
Consequently one must deduce that this approach is likely
output. This has two major implications:
not able to yield a reliable prediction. The spread of the
(1) The number of variables to be estimated is reduced. fitted parameters (all cylces on one filter for the entire
This is positive, as high correlations imply low pre- month) is visualised in figure 8.

12647
Proceedings of the 20th IFAC World Congress
12142
Toulouse, France, July 9-14, 2017 F.D. Bhner et al. / IFAC PapersOnLine 50-1 (2017) 1213712142
10
Scores
5
0 10
-5 0 5
Prediction Error (h)
Fig. 11. Prediction errors, 75% of cycle, n = 0.45
Fig. 12. a: Example of a good prediction, b: Example of
failure due to non-modelled effects
5.1 Predictivity of the Model
To assess predictivity, the parameters are estimated at
a certain point before the end of a cycle. The predicted
model outcome is then compared with the actual values.
Under the premise of a flow guideline for the operators,
knowledge of the flow profile over the prediction horizon
has been assumed. The viscosity is set to the mean value
of the past 60 minutes before the prediction begins. In
figure 9, the outcome is based on knowledge of 75%
of the profile of the actual cycle. This cycle was not
chosen deliberately. It has been selected for its steady
flow profile, which implies a small modelling error. In
spite of this, the prediction misses the compressibility-
induced steep pressure increase toward the end completely.
This is seen throughout all cycles, and looking at the
unlike signs in the local sensitivity function (Sin and
Gernaey (2016)), based on the representative parameter
vales = 2, n = 0.4 in figure 10, one must conclude
that difficulties are to be expected in the simultaneous
estimation of both parameters. Thus, with figure 8b in
mind, the compressibility index is fixed at its maximum
likelihood value. However, in practice the predictions have
shown to be better when n assumes a slightly higher value.
Ultimately, it has been iteratively adjusted by looking at
the prediction errors across all cycles (restricted to one
filter). The error distribution for n = 0.45 and knowledge
of 75% of the cycle profile is plotted in figure 11. For most
of these predictions, the error is seen to be bounded by
approximately 2h. Figure 12a shows that good guesses
are possible also when the flow rate is unsteady. However,
it is to be expected that the quality of the predictions
would improve significantly under an operating regime
with constant / deterministically adjusted flow rates due
to the minimisation of non-modelled effects such as cake
blinding and particle rearrangement. From figures 12b
and 13 it can be concluded that the algorithm, while
performing rather well for a large number of cylces, is
not robust enough for rigorous scheduling. Furthermore,
it can be learned that there is only a weak correlation
(0.2) between absolute error and duration of a cycle.
Error (h) | Cycle time (h)
Prediction error
30 Duration of real cycle
20
10
0
0 10 20 30 40 50
Prediction nr.
Fig. 13. Correlation between pred. error and cycle duration
6. CONCLUSION
One should note the complexity arising in industrial-scale
filtrations due to the nature of the process as well as the
extent of regularly occurring disturbances. The situation
has been extensively elucidated and furthermore predi-
cated upon real process data. A predictive model poses a
first step toward a new operational regime that relies less
strongly on operator expertise, thereby reducing the risk of
human failure. It is found that the uncertain circumstances
and a series of irregular events cannot be entirely captured
by the model. This deteriorates the outcome of the pre-
dictions, but it is unlikely that the quality of the guesses
could be substantially improved by the use of a more
sophisticated model or wet lab experiments as many uncer-
tainties are believed to arise from operational disturbances
such as e.g. properties of the precoat layer. On the other
hand, correlations between consecutive cycles as well as
the interdependencies between the parallel units should be
acknowledged. They suggest an inferential superstructure
that weighs filtration theory based predictions and statis-
tically motivated values against each other. If this proves
to enhance robustness sufficiently, visualised predictions or
even a scheduling algorithm can be envisioned.
REFERENCES
Meindersma, G.W., Augeraud, J., and Vergossen, F.H.P.
(1997). Separation of a biocatalyst with ultrafiltration
or filtration after bioconversion. Journal of Membrane
Science, 125(2), 333349.
Rushton, A., Ward, A., and Holdich, R. (1996). Filtration
Process Equipment and Calculations. In Solid-Liquid
Filtration and Separation Technology, 397482. 1 edi-
tion.
Ruth, B. (1946). Correlating Filtration Theory with Indus-
trial Practice. Industrial and Engineering Chemistry,
38(6), 564571.
Sin, G. and Gernaey, K.V. (2016). Data Handling and
Parameter Estimation. In M.C. van Loosdrecht (ed.),
Experimental Methods in Wastewater Treatment, 201
234.
Tarleton, E. and Hancock, D. (1997). Using Mechatronics
for the Interpretation and Modelling of the Pressure Fil-
ter Cycle. Chemical Engineering Research and Design,
75(3), 298308.
Tarleton, E. and Willmer, S. (1997). The Effects of Scale
and Process Parameters in Cake Filtration. Chemical
Engineering Research and Design, 75(5), 497507.
Tien, C. (2006). Filtration Process Equipment and Cal-
culations. In Introduction to Cake Filtration, 1350. 1
edition.
12648