Novel decentralised control vs MPC for industrial high purity distillation

Research article
A comparison of a novel robust decentralised control strategy and MPC
for industrial high purity, high recovery, multicomponent distillation
Isuru A. Udugama a,b
, Florian Wolfenstetter a
, Robert Kirkpatrick a
, Wei Yu a,b
,
Brent R. Young a,b,n
a
Chemical & Materials Engineering, The University of Auckland, New Zealand
b
Industrial Information and Control Centre, The University of Auckland, New Zealand
a r t i c l e i n f o
Article history:
Received 19 July 2016
Received in revised form
19 January 2017
Accepted 10 April 2017
Available online 14 April 2017
Keywords:
High purity
Practical control
Distillation control
Override control
MPC
a b s t r a c t
In this work we have developed a novel, robust practical control structure to regulate an industrial
methanol distillation column. This proposed control scheme is based on a override control framework
and can manage a non-key trace ethanol product impurity specification while maintaining high product
recovery. For comparison purposes, a MPC with a discrete process model (based on step tests) was also
developed and tested. The results from process disturbance testing shows that, both the MPC and the
proposed controller were capable of maintaining both the trace level ethanol specification in the dis-
tillate (XD) and high product recovery (β). Closer analysis revealed that the MPC controller has a tighter
XD control, while the proposed controller was tighter in β control. The tight XD control allowed the MPC
to operate at a higher XD set point (closer to the 10 ppm AA grade methanol standard), allowing for
savings in energy usage. Despite the energy savings of the MPC, the proposed control scheme has lower
installation and running costs. An economic analysis revealed a multitude of other external economic
and plant design factors, that should be considered when making a decision between the two controllers.
In general, we found relatively high energy costs favour MPC.
& 2017 ISA. Published by Elsevier Ltd. All rights reserved.
1. Introduction and background
The determination of control system structure is an important
step in distillation control. In general, there are two main branches
of control structures available: a centralized multi-input multi-
output (MIMO) controller, or a set of single-input single-output
(SISO) controllers. Either one of these systems can be used to
control a distributed or decentralized process. In the chemical
industry, a decentralized type of control system is more common
than the centralized control system, as it has more captivating
advantages: it is easy to understand, uses uncomplicated hard-
ware, and employs simple working algorithms [1–3]. However,
decentralized control schemes lack the ability to operate a system
at optimal levels. Conversely, centralized types of controls, in
particular Model Predictive Controllers (MPC), have the potential
to operate systems at optimal levels. Despite this, there are lim-
itations to using centralised controllers, some of these limitations
are: higher maintenance cost, difficulties in operation, complicated
structure and lack of flexibility that can result in a fragile
controller that is not profitable [4–6]. In addition, the actual data
gathering process required to create an MPC model in an industrial
environment can be difficult and can be further complicated as it
might require techniques other than simple plant step tests [7].
However, in many situations plant step testing are still the most
convenient method to capture the dynamics of the process [8].
Once the relevant data is gathered, the design and tuning process
of MPC type controllers are also more complex and time con-
suming [9]. Even with recent advances in computer technology,
the field implementation of MPC still requires specialist high-end
computer platforms [10].
In methanol distillation a multi-component feed of methanol,
water and ppm levels of ethanol, is refined to achieve a tight high
purity product and bottoms (ppm levels) and high methanol
product recovery (β) (≥97.5%). For decades, operators and control
schemes only focused on maintaining product ethanol specifica-
tion (XD) below the industrial AA grade limit of 10 ppm [11], while
β was not actively pursued. As a safety buffer, the operators also
used excess reboiler duty to create a higher than required reflux
ratio which, in turn, produced product methanol with a XD of
∼4 ppm. This action provides the operators with a sufficient XD
buffer and can negate the non-linear behaviours of the column.
With plant management's goal of improving profitability, these
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/isatrans
ISA Transactions
http://dx.doi.org/10.1016/j.isatra.2017.04.008
0019-0578/& 2017 ISA. Published by Elsevier Ltd. All rights reserved.
n
Corresponding author at: Industrial Information and Control Centre, The
University of Auckland, New Zealand.
E-mail address: b.young@auckland.ac.nz (B.R. Young).
ISA Transactions 69 (2017) 222–233

distillation columns now need to operate at improved product
recoveries ( ∼99.5%) using “normal” levels of reboiler duty while
maintaining the AA grade product methanol specification. New
control structures will need to be developed to meet this objective
of increasing profitability.
To operate the column at high purity and recovery while
minimizing the use of reboiler duty requires a control structure
that can deal with non-linear behaviour, where a small deviation
from the operating point can have a significant influence on the
operability and process dynamics. As such, a linear process model
and linear SISO control structure might be insufficient to describe
the system dynamics and control of the column over changing
operating conditions [12,13]. Thus, a classic distillation column
control arrangement, such as DV control [14], can have difficulty
in controlling this type of process. In [15], these difficulties were
confirmed; DV control was unable to operate an industrial me-
thanol distillation process at required specifications. In contrast,
MPC, which was specifically developed to control complex multi-
variable processes [13] should be able to control the column at
new specifications. In recent years much research has been car-
ried out on implementing MPC on complex distillation units. In
[16] the authors had looked into the tuning parameters of MPC
for an industrial crude oil distillation unit. In [17], the authors
compared a MPC and decoupled PID control structure on a Me-
thyl Tert-butyl Ether distillation process and found MPC was able
to better control the process. In [18], the authors formulated a
multi-objective optimization procedure to operate a heavy crude
oil fractionator.
Despite these applications and improvements, the drawbacks
described previously means MPC might not be the best choice
commercially. Thus, there is an incentive in the methanol industry
to develop control structures that use decentralized control
schemes that can operate methanol distillation columns at high
recovery and high purity without the use of excess reboiler duty.
The conceptual framework of override control might be one al-
ternative to deal with the complex process dynamics of methanol
distillation. Override control is generally used to either protect
equipment from harmful conditions or to detect and correct faulty
process variables from effecting controllability [3,19,20]. In both
these variations, the override control “sits” above the regulatory
layer and uses logic operations in determining if an intervention is
necessary [20].
In this manuscript, we have developed a novel, robust, prac-
tical and easy to maintain control structure, that is able to meet
the tight specifications of industrial methanol distillation without
the use of excess reboiler duty. The proposed scheme has two
layers that consist of a PID based regulatory layer and a super-
visory layer inspired by override control. The regulatory layer,
together with supervisory control maintain both XD and β at their
tight specifications. In this control structure, the supervisory
layer addresses the “root cause” of the non-linear process dy-
namics and allows the PID regulatory layer to control the process.
For comparison purposes, a generic MPC with a discrete non
linear process model was also developed. Once completed, both
the proposed control strategy and the MPC were tested against a
range of process disturbances. To conclude, the economic im-
plications of choosing between MPC and the proposed control
scheme were discussed.
2. Set-up
In methanol refining, a multi-component feed (methanol, wa-
ter, ppm level of ethanol) is refined to achieve tight high purity
product and bottoms at high methanol recovery ratio (≥97.5%). In
this paper, we have looked into developing control schemes that
can operate the column at ≥99.5% recovery and below the 10 ppm
AA grade industrial methanol specification. The feed conditions for
the high-purity column are given in Table 1.
To produce AA grade methanol, the distillate needs to have a
methanol purity of 99.85 wt% and, more importantly, <10 ppm
ethanol [11]. From a fundamental and a process control point of
view, product ethanol is the “lower key” in this distillation process,
meeting this specification will guarantee the necessary product
methanol purity. There are also limits on other trace compounds
[11]; but these compounds are distilled out of the crude methanol
prior to its arrival in the main methanol refining column. Bottom
discharge mainly consists of water and must not contain more
than 10 ppm of methanol to satisfy waste water discharge re-
strictions. To accomplish such a high distillate and bottoms purity,
most of the ethanol entering the column needs to be taken out via
the fusel side draw. To satisfy both the operating stability of the
column and limitations on the disposal system, the mass flow of
this fusel stream must be kept within −2600 3700
kg
h
.
To maximize the profitability of the methanol production it is
important to recover as much of the methanol entering the col-
umn as possible in the product methanol stream. In general the
ratio of methanol in the feed to methanol in the product must be
at a high level (≥ )97.5% . In this paper we would call this the pro-
duct recovery ratio (β) and define it as:
β =
̇
̇
·
( )
m
m
100%
1
MeOH Dist
MeOH Feed
,
,
where the component mass flow rate of methanol in distillate is
̇mMeOH Dist, and the component mass flow of methanol in the feed
stream is ̇mMeOH Feed, .
Table 1
Industrial plant data.
Value Units
Feed Mass Flow 142,500 kg
h
Temperature 80 °C
Pressure 170.8 kPa
Methanol Content 82.5 wt%
Water Content 17.5 wt%
Ethanol Content 150 ppm
Distillate Mass Flow 115,150 kg
h
Methanol Content 99.99 wt%
Water Content 85 ppm
Fusel Mass Flow 3015 kg
h
Methanol Content 65 wt%
Ethanol Content 0.66 wt%
Bottom Mass Flow 25,485 kg
h
Methanol Content 5 ppm
Specifications Reflux Ratio 1.687 –
Reboiler Duty 88.25 MW
Methanol Recovery 98 %
I.A. Udugama et al. / ISA Transactions 69 (2017) 222–233 223

3. Modelling and validation
Process modelling and simulation are well recognized tools for
critical decision making, control design and optimization in pro-
cess engineering [3,21]. In this paper, the commercial simulation
software packages VMGSim and HYSYS were used to build and
validate dynamic models based on the industrial plant data. In this
instance the mass flow rate and reboiler duty of the industrial
distillation column was used to set the mass flow rate of the dis-
tillation column in the process simulation. The simulation
achieved a good match in all three compositions simulated, with
the all important ethanol profile being matched very closely even
down at the ppm levels. This biggest miss matches are in product
and bottoms water content, however, a closer looks reveals that
these compositions are in ppm levels, as a large % error translates
to a very small absolute error. Table 2 shows the comparison of
mass fractions between the real plant and the process simulation.
Once validated, the simulation was pushed towards 99.5 % re-
covery and was then revalidated based on plant trials and other
available data.
The vapour liquid equilibrium (VLE) in the distillation column
is described by an activity coefficient approach with a property
package using the Wilson model. In this model, the activity coef-
ficients (of each component in) the mixture were first determined
to calculate the fugacity of the liquids, while a Virial equation of
state was chosen to calculate the vapour fugacity. The column
solver uses equilibrium and enthalpy models paired with rigorous
thermodynamics in calculations. The column was simulated with
sieve trays, the overall tray efficiency of the column was set at 80%,
this figure was reached based on information provided by the tray
manufacturer who based this estimate of the specific tray installed
and the overall column loading. The Maximum flooding factor of
120% was set based on tray information received from the tray
manufacturer who based this estimate on tray hydraulics. The
basic PID controllers for pressure, flow and level control were also
necessary to operate the column at a stable point.
4. Control configurations
In this section, we will introduce both the proposed control and
MPC scheme designed for the high-purity methanol distillation
column. To run the plant at a maximum profit, methanol recovery
needs to be maintained at a high level, while satisfying the con-
straint of AA grade methanol in the distillate. In addition, the
control schemes must also be able to operate the column in a
stable manner. As such, both control schemes must meet the fol-
lowing requirements:
Maximizing/controlling methanol recovery
Handling the AA grade constraint on distillate ethanol content
Ability to reject disturbances in feed mass flow and composition.
4.1. Process dynamics
The requirement for high β and tight XD where non key ethanol
needs to be managed at ppm levels, creates unusual non-linear
process behaviours. To understand these non-linearities we have
plotted in Fig. 1 the response of XD to a change in β from a base
case of 99.6 %.
Fig. 1 shows that the process has a non-linear response for all
deviations in β. Analysing the + −/ 0.07% incremental recovery
curves, we can see that a positive change in β has a higher influ-
ence on XD in comparison to a similar reduction in β. This confirms
the non linearity of the system. The degree of non-linearity also
tends to increase with bigger changes in β. Fundamentally, this
behaviour is acceptable as larger positive deviations in β take the
system closer to (and above) 100% β at which point the imbalance
in the material and energy balance will result in a surge of XD. This
phenomena is further confirmed by the comparatively steep rise of
the +0.21% incremental recovery curve. Bigger negative deviations
in β, however, result in a smaller influence on XD as we are already
at ppm levels of ethanol. Further analysis shows that the process
itself has a relatively large time constant; even after 2000 min,
most of the plots have not yet reached steady state
From a practical perspective, we are more interested in the
time taken for XD to exceed the 10 ppm constraint if kept un-
checked as AA grade product methanol needs to contain 10 ppm
of ethanol. Even for a small increase in β of 0.14%, XD will exceed
the 10 ppm constraint over time. As the deviation is increased to
0.21 %, the time taken for XD to reach 10 ppm reduces dramatically.
4.2. Proposed control scheme
Control of a high purity industrial methanol distillation col-
umn is complicated due to the high purity product specification
xD, high product recovery requirements and need to minimize the
reboiler duty usage. In addition, the control scheme developed
needs to be industrially compatible. Standard control structures
available in literature are incapable of meeting all these re-
quirements. As such, in this work we developed a novel control
structure that can meet all these requirements. The proposed
control scheme in this paper was developed with the aim of
achieving high recovery and high purity while minimizing re-
boiler duty usage and in turn reducing the costs of operation. The
proposed control scheme comprises of two layers: a regulatory
primary layer and a supervisory secondary layer. The secondary
Table 2
Plant data Vs Simulation results.
Plant data Simulation Units % Error
Distillate Methanol 99.99 99.99 Mass Frac (%) 0
Ethanol 8 8 ppm 0
Water 85 56 ppm 34
Bottom Methanol 5 3.2 ppm 36
Ethanol 0 0 ppm 0
Water 99.94 99.94 Mass Frac(%) 0
Fusel Methanol 65 63.4 Mass Frac (%) 2.5
Ethanol 0.66 0.66 Mass Frac (%) 0.3
Water 34.4 35.6 Mass Frac (%) 3.5
Fig. 1. Process response of product ethanol specification for changes in product
recovery ratio (β).
I.A. Udugama et al. / ISA Transactions 69 (2017) 222–233224

layer consists of a recovery constraint controller (RC) and a dy-
namic reboiler controller (DR). We have used the control struc-
ture developed in [15] as a start point. We will call this proposed
control scheme by the acronym RCDR in the following sections.
Recovery Constraint Controller - RC. The RC controller acts as a
central supervisory controller that takes measurements of the
current recovery ratio and the ethanol content in the distillate
stream. The set point of the composition controller YSP QC, will be
set depending on the state of the system. To avoid sudden set
point changes of the composition controller, a rate limiter is added
to the output signal of the RC controller. The maximum rate of
change is set to 0.1
%
min
, this value was set experimentally to allow
for a relatively rapid yet smooth transitions between set-points,
even during quick product ethanol set point changes. The calcu-
lation of the set point is based on two different cases according to
the following equations.
⎡⎣ ⎤⎦
β β
β β= − · ·
−
≤
( )
−
Y Y MHigh recovery case: 10
%
if
2
SP QC SP H
SP
SP,
6
⎡⎣ ⎤⎦
β β
β β= − · ·
−

( )
−
Y Y MLow recovery case: 10
%
if
3
SP QC SP L
SP
SP,
6
Eq. (2) represents the high recovery case, where the current
recovery β is higher than the specified recovery set point (βSP). In
this case, the RC controller will take advantage of this beneficial
situation and produce higher grade methanol with a ethanol
content of 7.8 ppm. Accordingly, the set point of the composi-
tion controller will be lowered. This action also ensures that the
system is not running at recoveries of β 100% over long periods
of time, which would be infeasible. MH, the gradient in Eq. (2),
dictates the rate at which the set point will be increased for a
given increase in β. Where, a large MH represents an aggressive
controller. In a situation where β 99.6% we require the RC
controller to bring β back to out set point in a timely manner.
However, we do not require a very aggressive response as
β 99.6% is financially favourable (more methanol produced),
but needs to be managed as it can create potential control pro-
blems in future. Based on disturbance rejection tests carried out
MH is set to 2 to achieve timely response.
The low recovery case is represented by Eq. (3). If the system is
running at a recovery of β β SP
, the RC controller will try to in-
crease β by increasing the set point for the composition controller.
The set point can only be increased up to the maximum ethanol
content of 10 ppm. Unlike the high recovery case, the low recovery
case represents a situation that is economically unfavourable. This
requires the controller to intervene aggressively. To enable this,
ML, the gradient in Eq. (3), has to been set to high value to enable
an aggressive control response. Based on disturbance rejection
tests this value has been set to 6. Fig. 2 shows a control schematic
of the overall control scheme.
The low recovery and high recovery equations ensure that the β
is managed to be around 99.6 %. Even during sharp disturbances
(as discussed in Section 6), these equations will bring β back to set
point before the non-linearities can effect the column. As such,
these equations deal with the reasons for non-linearities rather
than their effects. Fig. 3 illustrates the overall structure of the RC
controller and discusses how these two controllers act together.
The flow rate of the fusel stream is also set by the RC controller
element. The necessary fusel flow rate is determined by Eq. (4). In
methanol distillation, 95 % of the ethanol entering the column
needs to be extracted at this side draw. If this term is not met for a
long period of time, there can be a major column upset due to the
build-up of ethanol in the tower.
̇ =
̇
( )
m
m
X 4
Fusel
EtOH Feed
EtOH Fusel
,
,
In this equation, fusel mass flow rate ( ̇mFusel) is determined by
the mass flow of ethanol in the feed ( ̇mEtOH Feed, ) and the con-
centration of ethanol in the fusel draw ( XEtOH Fusel, ). Depending on
the current feed composition, the amount of ethanol in the dis-
tillate and the fusel composition, an optimal fusel flow rate can be
calculated according to Eq. (4). This flow rate is implemented as
the set point of the fusel flow controller. Minimum and maximum
set point values of 2000 Kg/h and 3700 Kg/h respectively are en-
forced to reflect practical limitations. The express objective of
setting the side draw flow rate is to maintain column stability by
ensuring there is no major ethanol build up in the column and not
used as a manipulated variable for optimization.
To better illustrate the function of the controller, we have
constructed a control diagram of the RC controller in Fig. 3. The
RC controller consists of two parallel controller modules. The
top module deals with setting the XD controller set point, where
it uses the distillate methanol flow rate and feed methanol flow
rate are used as process inputs (PVs). It also requires the plant
operator to set the recovery ratio (β) and product ethanol (XD)
set point limits. In addition, a maximum ethanol limit also
needs to be specified. Generally, this can be the AA grade me-
thanol limit of 10 ppm. The dotted circle in this diagram shows
the logic switch and equations used to determine the relation-
ship between the recovery ratio and product ethanol specifica-
tion at a given process condition. The fusel flow rate set point
module is much simpler: feed ethanol mass flow and fusel
ethanol concentration, together with maximum and minimum
fusel flow limits, are used to determine the fusel flow set point.
In addition to these functions, the RC controller also sends the
current recovery ratio and recovery ratio set point information
to the DR controller, as discussed below.
Dynamic Reboiler Controller - DR. The DR controller comprises
of a feed forward controller and a feedback trim controller. The
feed forward controller increases or decreases reboiler duty based
on the feed flow rate. Eq. (5) has been empirically derived to
Fig. 2. Schematic of the proposed control scheme (including regulatory
controllers).

achieve a product recovery of 99.6% while achieving a 7.5 ppm
product ethanol specification. This equation was validated during
plat trials where the collumn was held at 99.6% recovery for a
short time period.
⎡⎣ ⎤⎦ ⎡
⎣
⎢
⎤
⎦
⎥
̇
= ·
̇
−
( )
Q m
h
kW
0.6707
kg
4200
5
Reb Feed
In cases where feed disturbances can be measured, the feed
forward controller can improve column stability. In addition, a
simple feed back trim controller is used to keep track of and
maintain β at set point by carrying out minor adjustments in re-
boiler duty. The feedback trim controller is a basic PI type
controller.
Fig. 4 illustrates the control architecture of the DR controller,
where a feed forward element together with a feed back trim
controller is used to set reboiler duty. The controller also takes
current recovery and recovery set point information from the RC
controller. The feed forward element in this instance acts as a
“coarse” controller, while the feed back trim acts as a “fine” con-
troller. This means a simple feed forward controller is sufficient
as the trim controller can deal with correcting offsets and re-
acting to unmeasured disturbances. The trim controller also al-
lows the DR controller to operate at different product ethanol
specifications and at different recovery rates. Since the trim
controller receives recovery information from RC controller, the
users only have to set the desired product ethanol and product
recovery ratio at the RC controller.
Overall Performance. From a chemical engineering perspective
the RC controller and DR controller manipulates the mass and
energy balance of the column respectively. This allows RC and DR
controllers to work together to achieve on specification product
recovery and ethanol specification. If a process disturbance influ-
ences the product recovery, the RC controller will actively
manipulate the product ethanol set point to bring the product
recovery back to acceptable values. At the same time DR controller
would change the reboiler duty to bring the recovery rate back to
set point. For example, a process disturbance which reduces pro-
duct recovery, the RC controller will relax the product ethanol
specification up to 10 ppm. This action allows the controller to
increase product recovery rate for a short time period without
breaking the product ethanol specification. If the DR controller is
switched off, the RC controller will remain at ethanol set point of
10 ppm and the recovery rate will drop down to a new steady
state, as the reboiler duty (energy) supplied to the column is not
able to maintain the desired product recovery rate at product
specification below 10 ppm. If we now switch on the DR controller,
it will now increase the reboiler duty to bring the recovery back to
the set point. As this happens the RC controller will slowly reduce
its set ethanol set point back to its steady state level (the level is
was at before the disturbances). In reality the RC and DR con-
trollers work at the same time and will simultaneously increase
product ethanol specification and reboiler duty to bring back the
product recovery back to specification as soon as possible. As time
progresses the RC controller will settle back down while the DR
controller will settle to a higher reboiler duty
Robustness and application. The relatively straight forward lo-
gic used in the supervisory control structure was built with the
ideology that if and when plant operations (control engineer/ su-
pervisor) can change the heuristically tuned parameters to fit
plant requirements at that time. It is authors experience that many
advanced process control solutions including MPC's tend to taken
off-line due to the relatively complex nature of them which re-
quires routine expert support to retune. Further, disturbance tests
performed in Section 6 illustrates the proposed controller is suf-
ficiently robust to handle large process swings.
4.3. Model predictive controller (MPC)
This section will give a more detailed overview of the im-
plementation of the model predictive controller(MPC). The choice
of tuning parameters for the MPC, as well as the tests conducted to
obtain the dynamic process model will be discussed. The methanol
distillation column can be controlled by a 2x2 multi-input multi-
output (MIMO) controller, with each input affecting both outputs.
The two inputs are the methanol recovery β and the ethanol
content in the distillate stream. The recovery is calculated ac-
cording to Eq. (1). The two outputs of the MPC are the set point for
the XD controller and the reboiler duty.
Process model. To build the process model, we carried out open
Fig. 3. Control Diagram of the RC controller illustrating the inputs, outputs and the information flow within the controller. Where (1), (2), (3) and (4) refer to Eqs. (1), (2),
(3) and (4) that are used in the RC controller calculations.
Fig. 4. Control diagram of DR controller illustrating the inputs out puts and the
information flow with in the controller. Where (5) refers to Eq. (5) used to in the DR
controller calculations.

loop step tests on our dynamic simulation, where we changed
reboiler duty and distillate flow rate and observed its effect on XD
and β. As expected, the process dynamics were relatively complex
and could not be accurately estimated by a simple first order plus
dead time model. Since industrial MPC modules allow the user to
input the actual set response (step response data points that de-
scribes a discrete transfer function), we normalised our responses
and created transfer functions that were then inserted into the
MPC module as the process model.The MPC module used in this
particular instance is a commercial system.
Tuning parameters and process variable. One distinctive fea-
ture of MPC is the large amount of tuning parameters and vari-
ables that needs to be set [22]. The first variable that needs to be
set is the control or sampling interval k. This denotes the time
interval in which the MPC will take actions. The large time con-
stant of the process means a long time interval between MPC
execution would be sufficient to regulate the process. The sam-
pling interval value k was first set to 100 min, as this is was 1/10
th of the time constant of the smallest process time constant.
However, this resulted in unfavourable results during the short
term disturbances. Thus a k value of 10 min has been set based on
short term cyclical disturbance tests carried out. From a practical
point of view, k of 10 min is still significantly long enough for the
MPC application to execute and subsequent real time measure-
ments needed (especially composition measurements) can be
updated within this time frame.
The prediction horizon P represents the number of sampling
intervals into the future where predictions are made. The predic-
tion horizon was set to P¼50 to account for the large time con-
stant of the process. The control horizon M is the number of
control moves into the future the MPC considers when predictions
are made. The default value of M¼2 was chosen, as this gave the
best responses in initial testing. The last parameters are the pro-
cess variable and manipulated variable weighting matrices Γy and
Γu respectively. We found that at low Γu values, the controller
took large controller actions that affected its stability, but was
better at rejecting sharp disturbances. Conversely, a high Γu re-
sulted in stable steady state performance, but with bad dis-
turbance rejection characteristics. Since we desire both stability
and disturbance rejection properties we have set Γy and Γu to
1 which gives equal priority to both.
Implementation. A schematic of the MPC scheme is shown in
Fig. 5. As previously described, the two inputs are the ethanol
content in the distillate XD and the methanol recovery β. The re-
covery is governed by the distillate stream which is adjusted by
the composition controller. The fast actions of this controller can
lead to a noisy signals for the MPC. In order to reduce the signal
noise, a rate limiter is used. The maximum rate of change on the
recovery input signal is set to 0.1
%
min
. A material balance is solved
at every sampling interval k following Eq. (4). The fusel flow in this
instance was kept at the default value of 3113
kg
h
5. Feed step tests
To understand the behaviour of the control schemes to process
disturbances, it was necessary to carry out disturbance step tests.
Based on analysis and operators experience, it was determined
that the feed flow rate and composition are the most likely process
variables to introduce process disturbances into the column.
5.1. Feed flow rate step test
To start off with, the feed flow rate was changed from
̇ =M 142.5Feed
t
h
to 147.5
t
h
. This represents an extreme scenario
where it is demanded that the collumn take extra feed flow in a
short period of time. The response of MPC and the proposed
control scheme to this step test are illustrated in Figs. 6 and 7.
Where the reboiler duty is the manipulated variable, while pro-
duct ethanol composition and product recovery are process
variables.
As expected, both the control schemes counter the increase in
feed flow rate by increasing reboiler duty. The proposed control
scheme reacts with an “impulse” like response due to the feed
forward controller. The feedback trim controller then reduces the
reboiler duty to achieve 99.6% recovery. In comparison, the MPC
controller has a far more gradual response and reaches a similar
steady state as the proposed control scheme. This behaviour can
be attributed to the ratio of Γy and Γu which is set at 1 (balanced
approach).
Both control schemes are able to tightly maintain XD and β,
while the MPC returns to set point quicker than the proposed
control scheme. It is also important to note the increase in XD
specification for the proposed control scheme is rapid. This be-
haviour can be attributed to the proposed control scheme trying to
“sacrifice” XD specification in order to maintain recovery at the
desired set point of 99.6%, This fact is further confirmed by
Fig. 5. Schematic of the high-purity methanol distillation column with model
predictive control (MPC).
Fig. 6. Reboiler duty response of MPC and proposed control scheme (RCDR) to
changes in feed flow rate.

analysing the results from Table 3.
Examining Table 3, we can conclude that, in terms of the re-
covery, the proposed control scheme does a better job at “arrest-
ing” the initial drop in recovery. As explained, the nature of the
proposed control scheme is to “sacrifice” XD in the short term,
which, in turn, enables β to quickly return to specified recovery.
However, this quick recovery comes at the cost of XD stability
where the total integral error for MPC controller is much lower
than the proposed control scheme.
Further analysing Fig. 6 illustrates that the rebiler duty ma-
nipulated by the proposed control scheme initially overshoots its
new steady sate settling point. From a industrial point of view the
overshoot in reboiler duty requires additional steam to be gener-
ated. In methanol production plants this additional steam re-
quirement can be generated with auxiliary natural gas firing at the
reforming section of the plant, which will allow the reboiler duty
to move above its new steady state during the transient periods.
5.2. Feed methanol content step tests
In this test, we have increased the methanol concentration of
the feed stream by 2wt%. To compensate, the water concentration
in the feed has been decreased by 2 wt%. This represents an ex-
treme situation where, a sudden change in feed gas or catalyst has
rapidly changed the methanol concentration.
Analysing Figs. 8 and 9 shows both the MPC and proposed
control scheme has increased the reboiler duty to deal with the
excess methanol concentration. Thermodynamically, this makes
sense as an increase in methanol concentration means a higher
mass of methanol needs to be distilled to keep the recovery ratio
at specification. Both control schemes have similar rise times
while the proposed control scheme seems to be increasing the
reboiler duty slightly more than MPC in the initial stages, but re-
turns to a similar steady state. In feed flow disturbance testing, we
looked at how the two control schemes react to a measured dis-
turbance, as the proposed control scheme had some degree of
“advanced” information through the feed forward controller. By
contrast, in the methanol content step test, both control schemes
get the same information through the process variable β. As such
both control schemes are on an “equal footing”.
In terms of ethanol ppm, the proposed controller seems to have
a bigger “swing” in ethanol concentration in comparison to MPC.
Table 4 confirms this observation, as MPC has a lower integral
absolute error in controlling XD. However, this is an expected
outcome, as the proposed control scheme sacrifices the short term
ethanol ppm stability (up to 10 ppm) to stabilize the recovery. This
is also evident as MPC has a bigger “swing” in recovery in com-
parison to the proposed control scheme and has a higher integral
error in controlling β. This test clearly shows that the proposed
control scheme can better stabilize the recovery (which is of fi-
nancial importance), while MPC is better at stabilizing the product
ethanol specification.
Fig. 7. Responses in β and XD of MPC and proposed control scheme (RCDR) to
changes in feed flow rate.
Table 3
Integral Absolute Error of XD and β for a feed disturbance (expressed as a percen-
tage of absolute integral error of the proposed control).
Variable Control Scheme Integral Absolute Error
XD Proposed Control 100 %
MPC 38.5 %
β Proposed Control 100 %
MPC 450 %
Fig. 8. Reboiler duty response of MPC and proposed control scheme (RCDR) to
changes in feed methanol content.
Fig. 9. Responses in β and XD of MPC and proposed control scheme (RCDR) to
changes in feed methanol content.
Table 4
Integral Absolute Error of XD and β for a feed methanol disturbance (expressed as a
percentage of absolute integral error of proposed control).
Variable Control Scheme Integral Absolute Error
XD Proposed Control 100 %
MPC 38.5 %
β Proposed Control 100 %
MPC 180 %

6. Disturbance tests
In order to compare the different control schemes in re-
sponse to real world conditions, their performance under cyclic
process disturbances was tested. The tests were designed to
reflect realistic disturbance scenarios in plant operation. The
main source of disturbances in this particular process is the feed
stream. Three different series of disturbance tests were per-
formed on each control system. They were feed mass flow var-
iations, feed methanol/water ratio variations, and finally, feed
ethanol content variations.
6.1. Disturbances in feed flow rate
To compare the ability of the two control schemes to handle
oscillating disturbances in feed flow rate, a sinusoidal variation of
̇ = −m 137500 147500Feed
kg
h
was introduced. To model both long
time period process variations and short time period plant up-
sets, two different time periods of =t 50minshort and
=t 2000minlong were examined. Fig. 10 shows the controlled
variables XD, β and the reboiler duty ̇QReb for the MPC and the
proposed control scheme.
For long time period fluctuations, it can be seen that there are
only minor differences in both the energy input and the
disturbance reactions of the MPC and the proposed control
scheme, as illustrated in Fig. 10a and c. However, for short time
period fluctuations, as shown in Fig. 10b, the reboiler duty for the
MPC is delayed by 15 min and the actions are very small compared
to the proposed control scheme. These observations are confirmed
in the amplitude of deviation results in Table 5, where all variables,
except the short time period reboiler duty reaction, are the same.
For short time period disturbances the proposed control scheme
has a 7 Â greater amplitude in its reboiler duty reaction compared
to the MPC controller. Closer observation of the tabulated results
Fig. 10. The effect of short time period and long time period feed flow rate fluctuations on reboiler duty, XD and β.
Table 5
Process and manipulated variable fluctuation for feed flow fluctuation.
Process Variable Control Scheme Long time
period
Short time
period
XD amplitude (ppm) Proposed Control 0.19 0.88
MPC 0.30 0.11
β amplitude (%) Proposed Control 0.05 0.34
MPC 0.36 0.37
Reboiler amplitude (kW) Proposed Control 4.19 3.28
MPC 3.12 0.30

also shows that the proposed control scheme is slightly better at β
fluctuations, while MPC is slightly better at controlling the ethanol
specification.
This behaviour can be explained by analysing the way the
two schemes work. The proposed control scheme uses a feed
forward element that adjusts the reboiler duty directly with
the feed flow rate. For this reason, the reboiler duty for both
short and long time period disturbances will always reflect the
change in feed flow rate for the proposed control scheme. MPC,
on the other hand, measures changes in recovery and ethanol
content in distillate to calculate the output. This means that
disturbances in the feed have to affect β and XD before the MPC
will take action. Since fluctuations in feed flow rate take some
time to affect β and XD, the response in reboiler duty has a
slight delay of 1.5 min, compared to the proposed control
scheme's 15 min. For short time period disturbances this effect
becomes visible. Also, the objective function of the MPC pe-
nalizes large control moves (changes in manipulated vari-
ables). Since the time period of the sine wave is only ten times
bigger than the sampling interval, the MPC does not have
sufficient time to carry out big changes in reboiler duty.
The response of the controlled variables are displayed in
Fig. 10c and Fig. 10d and tabulated in Table 5. For both long and
short time period disturbances, the recovery is almost constant
and very close to the set point of β = 99.6% for both MPC and
the proposed control scheme. The deviation of XD in the
distillate stream from set point is roughly the same for
MPC and the proposed control scheme under long time period
disturbances.
Since MPC takes smaller control actions in reboiler duty for
short time period disturbances, one would assume that MPC
would preform poorly in controlling XD in comparison to the
proposed control scheme. However, analysing Table 5 we can see
that proposed control scheme preforms 8 Â worse than MPC for XD
control and only slight better at β control. We can conclude from
this that the proposed control scheme slightly overreacts for short
time period cyclical process disturbances. To remedy this issue, we
can use a rate-limiter in the design of the DR controller, which
would slow down the rate at which the reboiler duty can change.
However, this type of a rate limiter will make the proposed control
scheme more vulnerable during sharp feed flow disturbances as
illustrated in Section 5.1. The average values of the variables in
Fig. 10 are summarized in Table 6.
From the average values in Table 6, it can be seen that the
reboiler duty used by both control schemes is the same. However,
since MPC causes a smaller fluctuation amplitude in the all im-
portant XD variable, we can look to increase the XD set point to
closer to the 10 ppm limit. Based on further analysis we have
found 9 ppm would be an acceptable and reasonable new set
point. The results and potential energy savings will be discussed
in Section 7.
In [15], we carried out similar feed flow disturbance tests on a
DV control configuration. These tests illustrated that a DV con-
figuration cannot maintain the recovery or meet the bottoms
methanol composition, which must be met in operations. For
comparison, both proposed control scheme and MPC controllers
are able to maintain this composition
6.2. Disturbances in feed methanol content
In this set of disturbance tests, the methanol content in the feed
stream was varied from =X 80%MeOH Feed, to =X 85%MeOH Feed, . This
disturbance was implemented as a varied methanol feed mass
flow ̇mMeOH Feed, . The mass flow of ethanol was unchanged, while
the water mass flow was adjusted to ensure a constant overall feed
mass flow of ̇ =m 142500Feed
kg
h
. The time periods tshort and tlong
were used for the sinusoidal disturbance. Fig. 11 and Table 7 show
the reaction of the proposed control scheme and MPC to short and
long time period feed methanol variations.
During the short term variations, both proposed control
scheme and MPC show minor fluctuations in reboiler duty. This
behaviour can be attributed to the rate limiter on proposed control
scheme and the Γu of the MPC which prevents rapid changes in
control variables. In general, both controllers are keeping XD spe-
cification and maintaining a normal recovery. During the long time
period disturbances we can see both the controllers changing re-
boiler duty accordingly. In terms of the XD specification and re-
covery, we can see more variation during the short time period
disturbances. Again, both controllers are maintaining specifica-
tions. A look at Table 7 also confirms that the key process and
controlled variables are not fluctuating too much and have similar
amplitudes of fluctuation. As expected, the proposed control
scheme is doing a better job at controlling β while the MPC con-
troller is doing a better job at controlling XD.
6.3. Disturbances in feed ethanol content
In the last test series, the ethanol content in the feed stream
was changed using a sinusoidal wave from =X 100 ppmEtOH Feed, to
=X 200 ppmEtOH Feed, . The mass flows of the other components
were unchanged, since variations in ethanol mass flow have no
major effect on the overall feed mass flow. Fig. 12 and Table 8 show
the reaction of the proposed control scheme and the MPC to short
and long time period feed ethanol variations.
During the short time period variations, both proposed con-
trol scheme and MPC show no control actions. This behaviour can
be attributed to the requirement for ethanol to accumulate/de-
plete for a long period of time to affect the column performance.
For the long time period disturbances, both control schemes react
in a similar manner, where reboiler duty is increased for in-
creases in feed ethanol content and decreased for decreases in
feed ethanol content. We can also see that both control schemes
have a 200 min lag.
7. Energy savings and economic benefits
In Section 6, tests illustrated that both MPC and the proposed
controller scheme have similar performance characteristics,
where MPC is able to control the XD tighter than the proposed
control scheme during process fluctuations, while maintaining
similar recovery. As such, the XD of the MPC controller can be
pushed closer towards the 10 ppm industrial AA grade methanol
Table 6
Average values of the variables displayed in Fig. 10.
Variable Controller type Average for tshort Average for tlong
Feed flow rate in
kg
h
MPC 142,538 142,515
Proposed Control 142,538 142,515
Reboiler duty in kW MPC 91,300 91,392
Proposed Control 91,126 91,549
XD in ppm MPC 8.0 7.9
Proposed Control 8.1 7.8
Recovery β in % MPC 99.62 99.61
Proposed Control 99.60 99.61

specification, which enables the column to reduce reboiler duty
usage. In further analyses we concluded that the XD set point for
MPC can be increased up to 9 ppm. Table 9 shows the net energy
benefit of using MPC over the proposed control scheme for this
particular case.
8. Economic implications of MPC vs the proposed control
scheme
So far in this manuscript, we have compared MPC with
proposed control scheme in terms of controller performance. But
this comparison does not cover the underlying economic im-
plications that inevitably would decide which control scheme
would be implemented in an industrial situation. Operating MPC
over operating the proposed control scheme allows energy savings
of ∼400 kW, even during steady periods of plant operations. In
general, the plant is likely to operate at steady state for 90% of the
time. In the other 10%, the plant might experience process dis-
turbances, in this instance the MPC would save ∼550 kW.
8.1. Monetizing energy savings
Assuming the plant operates throughout the year, we can cal-
culate the amount of energy we can save (ESaving) by implementing
MPC.
= ( · + · )· =
( )
E 0.4 MW 0.9 0.55 MW 0.1 8760
h
a
3635.4
MWh
a 6Saving
Before converting the calculated energy savings into saved costs
we need to consider the following factors:
Does the plant have an internal need for extra steam?
Can the plant reduce the gas usage? (Many methanol plants
require some auxiliary natural gas to supplement the steam
Fig. 11. The effect of short time period and long time period methanol concentration fluctuations on reboiler duty, XD and β.
Table 7
Process and manipulated variable fluctuation for feed methanol fluctuation.
Process variable Control Scheme Long time
period
Short time
period
MPC 0.17 0.10
MPC 0.24 0.39
MPC 1.99 0.32

generated by waste heat recovered at the reforming process).
Can the plant sell this energy to the electricity market? (By
converting excess steam to electricity in a steam turbine-
generator).
Can the excess steam be used to process more methanol?
Taking these options into consideration, we can come up with
two extreme scenarios. In the first scenario, the steam saved can
be used to make more product methanol, this is the most profit-
able scenario. In contrast, the worst case scenario can be where
the plant generates all its steam from waste heat and the savings
Fig. 12. The effect of short time period and long time period feed ethanol concentration fluctuations on reboiler duty, XD and β.
Table 8
Process and manipulated variable fluctuation for feed ethanol fluctuation.
Process variable Control Scheme Long time
period
Short time
period
MPC 0.11 0.04
MPC 0.17 0.11
MPC 1 0.07
Table 9
Comparison of energy consumption at different set points for the composition
controller.
Disturbance XD Set point ̇QReb for tshort
̇QReb for tlong
Flow rate disturbance 7.84 ppm 91,300 kW 91,392 kW
9.00 ppm 90,812 kW 90,867 kW
Energy savings 488 kW 525 kW
Methanol content disturbances 7.84 ppm 91,364 kW 91,454 kW
9.00 ppm 90,817 kW 90,841 kW
Ethanol content disturbances 7.84 ppm 91,532 kW 91,398 kW
9.00 ppm 90,871 kW 90,858 kW

in steam cannot be converted into additional revenue. All practical
scenarios fall in between these two extremes.
8.2. Cost of implementation and other considerations
When MPC is chosen as a control scheme, the following addi-
tional costs need to be considered:
Cost of equipment
Cost of setting up
Cost of maintenance
In some cases, the net benefit provided by MPC would be
“wiped out” by the cost of implementation and maintenance.
Again, the net cost per column will be reduced based on the size of
the methanol producer, as implementation of MPC on multiple
(similar) columns would reduce set up and maintenance costs.
Additionally, we also need to consider the following factors:
General resistance of plant operations to implement MPC
Implications of operating MPC during major process upsets/
catastrophic events
In general MPC would be favoured over the proposed control
scheme during times of high energy costs, not withstanding the
above. In current market conditions of low energy prices the
proposed control scheme will be favoured.
It is also important to note that both the proposed controller
and MPC would require additional process measurements (e.g
accurate composition of the feed), which will require costly gas
chromatograph to be installed. However, the increase in recovery
provided by both the control structures would more that offset any
additional costs of measurement.
From an industrial point of view, the conceptual ideology used
in the development of the proposed controller can be used as a
template and a road map to develop supervisory controllers for
other distillation applications where similar high recovery and
quality constraints are common. Especially in the chemicals in-
dustry where older generation plants are getting pushed to make
tight product specifications at highest possible recoveries.
9. Conclusions
In this manuscript we have developed a novel, industry friendly
control structure that is capable of operating an industrial me-
thanol distillation column at high product recovery, whilst main-
taining a tight product ethanol specifications and minimising en-
ergy usage. To facilitate the development and the testing process a
validated process simulation of an industrial methanol distillation
column was employed. A MPC with a discrete model was also
developed for comparison purposes. Both the proposed control
scheme and the MPC were subjected to a multitude of disturbance
tests, that included disturbance step tests, cyclical feed fluctua-
tions and set point changes. In all tests, the proposed control
scheme and MPC were able to maintain both recovery and XD
specifications using similar levels of reboiler duty. As such, both
control schemes can be potentially implemented in this applica-
tion. In addition, it was observed that MPC was able to control to
the XD specification tighter than the proposed control scheme. This
enabled the XD set point to be set closer to the 10 ppm industrial
AA grade methanol limit, allowing the average energy usage of
MPC to be reduced by ∼500 kW, which is ∼0.5% of total reboiler
duty. An economic analysis of the proposed control scheme and
MPC illustrated that many other factors, such as the value of en-
ergy savings and the costs of implementing MPC, need to be
considered in deciding between the two controllers. In general
however,we found that plants with low value for reboiler duty
(steam) savings would prefer proposed control scheme over MPC.
Appendix A. Supplementary data
Supplementary data associated with this article can be found in
the online version at http://dx.doi.org/10.1016/j.isatra.2017.04.008.
References
[1] Morari M, Evanghelos Z. Robust Process Control. Englewood Cliffs, NJ: Prentice
Hall.; 1989, 205–215. p. 359–389.
[2] Marlin TE. Process control: design process and control system for dynamic
performance.New York: McGraw Hill; 2000. p. 665–90.
[3] Svrcek WY, Mahoney DP, Young BR. A real-time approach to process control.
3rd ed. New York: John Wiley Sons Inc.; 2014. p. 185–213.
[4] Camacho EF, Alba CB. Model predictive control. New York: Springer; 2013. p.
127–75.
[5] Alford JS. Bioprocess control: advances and challenges. Comput Chem Eng
2002;30:1464–75.
[6] Van den Boom, TJ, Backx, TC. Model Predictive Control. Delft University of
Technology, 200, 5, 2005,.
[7] Hay G, Svrcek WY, Young BR. Nonintrusive frequency analysis for variable
pairing on a distillation column. ISA Trans 2005;44(2):295–303.
[8] Darby ML, Nikolaou M. MPC: current practice and challenges. Control Eng
Pract 2012;20(4):328–42.
[9] Luyben WL. Practical distillation control.New York: Van Nostrand Reinhold;
1992. p. 229–30.
[10] Taube M. (9th December 2015), Personal Communication with Michael Taube
from SD Consulting LLC, at The University of Auckland, New Zealand.
[11] Cheng W. Methanol production and use.Boca Raton: CRC; 1994.
[12] Fuentes C, Luyben WL. Control of high-purity distillation columns. Ind Eng
Chem Process Des Dev 1983;22(3):361–6.
[13] Garcia CE, Prett DM, Morari M. Model predictive control: theory and practice -
a survey. Automatica 1989;25(3):335–48.
[14] Green DW, Perry RH. Perry’s chemical engineers' handbook. 8th ed. New York:
McGraw-Hill; 2008.
[15] Udugama IA, Munir TM, Kirkpatrick R, Young BR, Yu W. High purity, high
recovery, multi-component methanol distillation control. Comput Aided Chem
Eng 2015;37:1613–8.
[16] Yamashita AS, Zanin AC, Odloak. Tuning the Model Predictive Control of a
Crude Distillation Unit. ISA Transactions: In press 2016.
[17] Iqbal SI, Aziz MN. Comparison of MIMO MPC and PI decoupling in controlling
methyl tert-butyl ether process. Comput Aided Chem Eng 2012;31:345–9.
[18] Wojsznis W, Mehta A, Wojsznis P, Thiele D, Blevins T. Multi-objective opti-
mization for model predictive control. ISA Trans 2007;46(3):351–61.
[19] Wade HL. Basic and Advanced Regulatory Control: System Design and Appli-
cation. 1st ed. North Carolina: ISA; 2004. p. 232–53.
[20] Romagnoli JA, Palazogl A. Introduction to process control.Boca Raton: CRC
Press; 2006. p. 227–9.
[21] Udugama IA, Munir TM, Kirkpatrick R, Young BR, Yu W. Advanced Dual
Composition Control for High-purity Multi-component Distillation Column.
Presented at ADCONIP: Hiroshima, Japan 2014.
[22] Bequette BW. Process Control: Modeling, Design and Simulation. 1st ed. New
Jersey: Prentice-Hall; 2003.

Novel decentralised control vs MPC for industrial high purity distillation

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Novel decentralised control vs MPC for industrial high purity distillation

Similar to Novel decentralised control vs MPC for industrial high purity distillation (20)

More from ISA Interchange

More from ISA Interchange (20)

Recently uploaded

Recently uploaded (20)

Novel decentralised control vs MPC for industrial high purity distillation