Professional Documents
Culture Documents
William L. Luyben (Auth.), William L. Luyben (Eds.) - Practical Distillation Control-Springer US (1993) PDF
William L. Luyben (Auth.), William L. Luyben (Eds.) - Practical Distillation Control-Springer US (1993) PDF
v
vi Contributors
Kurt Waller
Terry L. Tolliver Process Control Laboratory
Monsanto Co. Abo Akademi
St. Louis, MO 63167 20500 Abo, Finland
(Chapter 17) (Chapters 10 and 15)
This book is dedicated to Bea, Gus, and Joanne Luyben and Bill Nichol, four of
the most avid bridge players I have ever known. Run 'em out!
Preface
Distillation column control has been the the "Lehigh inquisition" and survived! So it
subject of many, many papers over the last has been tested by the fire of both actual
half century. Several books have been de- plant experience and review by a hard-nosed
voted to various aspects of the subject. The group of practically oriented skeptics.
technology is quite extensive and diffuse. In selecting the authors and the topics,
There are also many conflicting opinions the emphasis has been on keeping the ma-
about some of the important questions. terial practical and useful, so some subjects
We hope that the collection under one that are currently of mathematical and the-
cover of contributions from many of the oretical interest, but have not been demon-
leading authorities in the field of distillation strated to have practical importance, have
control will help to consolidate, unify, and not been included.
clarify some of this vast technology. The The book is divided about half and half
contributing authors of this book represent between methodology and specific applica-
both industrial and academic perspectives, tion examples. Chapters 3 through 14 dis-
and their cumulative experience in the area cuss techniques and methods that have
of distillation control adds up to over 400 proven themselves to be useful tools in at-
years! The collection of this wealth of expe- tacking distillation control problems. These
rience under one cover must be unique in methods include dynamic modelling, simu-
the field. We hope the readers find it effec- lation, experimental identification, singular
tive and useful. value decomposition, analysis of robustness,
Most of the authors have participated at and the application of multivariable meth-
one time or another in the Distillation Con- ods. Chapters 15 through 25 illustrate how
trol Short Course that has been given every these and how other methods can be ap-
two years at Lehigh since 1968. Much of the plied to specific columns or important
material in the book has been subjected to classes of columns.
ix
The use in this book of trademarks, trade names, general descriptive names, and
so forth, even if they are not specifically identified, should not be taken as
indication that such names, as understood by the Trade Marks and Merchandise
Act, may be freely used by anyone.
Contents
Preface ix
xiii
XlV Contents
Part 2 Methods 27
Index 529
1
Techniques and Methods
1
Introd uction
William L. Luyben
Lehigh University
3
4 Practical Distillation Control
, N,
Z
3
2
The mathematical approach to finding In Figure 1·1 we see that there are five
the degrees of freedom of any process is to control valves, one on each of the following
total all the variables and subtract the num· streams: distillate, reflux, coolant, bottoms,
ber of independent equations. This is an and heating medium. We are assuming for
interesting exercise, but there is a much the moment that the feed stream is set by
easier approach. Simply add the total num· the upstream unit. So this simple column
ber of rationally placed control valves. The has 5 degrees of freedom. But inventories in
"rationally placed" qualification is to em· any process always must be controlled. In·
phasize that we have avoided poorly con· ventory loops involve liquid levels and pres·
ceived designs such as the placement of two sures. In our simple distillation column ex-
control valves in series in a liquid filled ample, this means that the liquid level in
system. the reflux drum, the liquid level in the base
Introduction 5
of the column, and the column pressure tem. It will be discussed in many chapters in
must be controlled. When we say that the this book.
pressure is controlled, we do not necessarily
mean that it is held constant. If we mini-
mize it, we are "controlling" it. 1-2-2 Fundamental Variables for
If we subtract the three variables that Composition Control
must be controlled from S, we end up with 2 The compositions of the products from a
degrees of freedom. Thus, there are two distillation column are affected by two fun-
and only two additional variables that can damental manipulated variables: feed split
(and must) be controlled in this distillation and fractionation. The feed split (or cut-
column. Notice that we have made no as-
point) variable refers to the fraction of the
sumptions about the number or type of feed that is taken overhead or out the bot-
chemical components we are separating. So tom. The "fractionation" variable refers to
a simple, ideal, binary system has 2 degrees the energy that is put into the column to
of freedom; a complex, multicomponent,
accomplish the separation. Both of these
nonideal distillation system also has 2 de-
fundamental variables affect both product
grees of freedom. compositions but in different ways and with
The two variables that are chosen to be different sensitivities.
controlled depend on many factors. Some
common situations are:
Feed split: Taking more of the feed out of
the top of the column as distillate tends
1. Control the composition of the light-key to decrease the purity of the distillate
impurity in the bottoms and the compo- and increase the purity of the bottoms.
sition of the heavy-key impurity in the Taking more of the feed out of the bot-
distillate. tom tends to increase distillate purity and
2. Control a temperature in the rectifying decrease bottoms purity. On a McCabe-
section of the column and a temperature Thiele diagram, assuming a constant re-
in the stripping section of the column. flux ratio, we are shifting the operating
3. Control the flow rate of reflux and a lines to the left as we increase distillate
temperature somewhere in the column. flow and to the right as we decrease
4. Control the flow rate of steam to the distillate flow.
reboiler and a temperature near the top Fractionation: Increasing the reflux ratio (or
of the column. steam-to-feed ratio) produces more of a
S. Control the reflux ratio (ratio of reflux difference between the compositions of
flow to distillate flow) and a temperature the products from the column. An in-
in the column. crease in reflux ratio will reduce the im-
purities in both distillate and bottoms.
These examples illustrate that (a) only
two things can be controlled and (b) nor- Feed split usually has a much stronger
mally at least one composition (or tempera- effect on product compositions than does
ture) somewhere in the column must be fractionation. This is true for most distilla-
controlled. tion columns except those that have very
Once the five variables to be controlled low product purities (less than 90%).
have been specified (e.g., two temperatures, One of the important consequences of
two levels, and pressure), we still have the overwhelming effect of feed split is that
the problem of deciding what manipulated it is usually impossible to control any com-
variable to use to control what controlled position (or temperature) in a column if the
variable. This "pairing" problem is called feed split is fixed, that is, if the distillate or
determining the structure of the control sys- bottoms flows are held constant. Any small
6 Practical Distillation Control
changes in feed rate or feed composition (a) Manipulate coolant: A control valve
will drastically affect the compositions of changes the flow rate of cooling water
both products, and it will not be possible to or refrigerant. If an air cooled con-
change fractionation enough to counter this denser is used, fan speed or pitch is
effect. A simple example illustrates the changed. Note that the liquid in the
point. Suppose we are feeding 50 mol of reflux drum is at its bubble point.
component A and 50 mol of component B. Changes in coolant temperature are
Distillate is 49 mol of A and 1 mol of B; compensated for by the pressure con-
bottoms is 1 mol of A and 49 mol of B. troller.
Thus product purities are 98%. Now sup- (b) Vent-bleed: Inert gas is added or bled
pose the feed changes to 40 mol of A and from the system using a dual split-ranged
60 mol of B, but the distillate flow is fixed at valve system so that under normal con-
a total of 50 mol. No matter how the reflux ditions both valves are closed. The re-
ratio is changed, the distillate will contain flux must be significantly subcooled in
almost 40 mol of A and 10 mol of B, so its order to keep the concentration of
purity cannot be changed from 80%. product in the vent gas stream low.
The fundamental manipulated variables Changes in coolant temperature cause
(fractionation and feed split) can be changed changes in reflux temperature.
in a variety of ways by adjustment of the (c) Direct: A control valve in the vapor line
control valves that set distillate, reflux, bot- from the column controls column pres-
toms, steam, and cooling water flow rates. sure. This system is only useful for fairly
Fractionation can be set by adjusting reflux small columns.
ratio, steam-to-feed ratio, reflux-to-feed ra- (d) Flooded: Liquid is backed up into the
tio, and so forth. Feed split can be set condenser to vary the heat transfer area.
directly by adjusting distillate or bottoms Reflux is subcooled. If inerts are pre-
flows or indirectly by adjusting reflux or sent, the condenser must be mounted
steam and letting level controllers change horizontally to permit venting.
the product streams. (e) Hot-vapor bypass: A blanket of hot va-
por exists above a pool of cold liquid in
the reflux drum. Heat transfer and con-
1·2-3 Pressure Control
densation occur at this interface, and it
Pressure in distillation columns is usually is important to avoid disturbances that
held fairly constant. In some columns where would change the interfacial area. The
the difficulty of separation is reduced (rela- formation of ripples on the surface can
tive volatility increased) by decreasing pres- result in the cold liquid "swallowing"
sure, pressure is allowed to float so that it is the hot vapor and produce rapid drops
as low as possible to minimize energy con- in pressure. If the reflux drum is located
sumption. In any case it is important to below the condenser, a large vapor line
prevent pressure from changing rapidly, ei- connects the drum to the condenser in-
ther up or down. Sudden decreases in pres- let and a control valve in the liquid line
sure can cause flashing of the liquid on the below the condenser floods the con-
trays, and the excessive vapor rates can flood denser. If the reflux drum is located
the column. Sudden increases in pressure above the condenser, a control valve in
can cause condensation of vapor, and the the vapor line controls the pressure in
low vapor rates can cause weeping and the reflux drum (and indirectly the pres-
dumping of trays. sure in the column). Reflux is sub-
There are a host of pressure control cooled.
techniques. Figure 1-2 shows some common (f) Floating pressure: A valve-position con-
examples. troller is used to keep the cooling water
--f.#tJVv+----t:IiIEt-.. Coolant
(0)
.....::.;.;.-......L--901Et--~ Vent
(b)
o}' • ~
(c)
(d)
FIGURE 1·2. Pressure control schemes. (a) Coolant manipulation; (b) vent' bleed;
(c) direct; (d) flooded condenser; (e) hot vapor bypass; (f) floating pressure.
7
8 Practical Distillation Control
=lK
(I)
valve nearly wide open by slowly chang- base (or in the reboiler if a kettle reboiler is
ing the setpoint of a fast pressure con· used). These levels are controlled in very
troller. Dual composition control (or different ways, depending on a number of
some scheme to reduce reflux ratio as factors.
pressure is reduced) must be used to If the column is part of a series of units
realize the energy savings. in a plant, it is usually important from a
plantwide control standpoint to use the liq·
uid levels as surge capacities to dampen out
1-2-4 Level Control
disturbances. In such an environment, it is
The two liquid levels that must be con· usually preferable to control base level with
trolled are in the reflux drum and column bottoms flow and reflux drum level with
Introduction 9
columns have many trays and high reflux excess low-pressure steam, you would prob-
ratios (see Chapter 22). ably build a conventional column with many
High "K" columns: The separation is ex- trays (200) and high reflux ratio (14) and
tremely easy, resulting in a very sharp operate at 17 atm so that cooling water
temperature profile. This can lead to could be used in the condenser. The control
difficult control problems because of non- system would maximize recovery of propy-
linearity. lene by using as much heat input as possible
Sidestream columns: These complex config- (operating against a flooding or high-pres-
uration produce more than two products sure constraint) and controlling only the
by removing sidestreams from the col- distillate composition.
umn. Sometimes sidestream strippers or If, on the other hand, in your plant you
sidestream rectifiers are also used. have to produce incremental steam by burn-
ing more fuel, you would probably build a
The preceding partial list illustrates some vapor recompression system with fewer trays
of the generic columns that occur in distilla- (150), lower reflux ratio (11) and lower pres-
tion. However, despite the similarities, the sure (11 atm). The control system would try
individual columns within each of these to keep the column at the optimum distil-
classifications can be very different in both late and bottoms compositions (dual compo-
design and control for a number of reasons. sition control) that represent the best trade-
Some conditions that vary from column to off between compressor horsepower costs
column are summarized: and propylene recovery.
of trays by 10% and increase the capacity of will not vary excessively during distur-
the column, reboiler, and condenser by 20% bances.
to make it easier to control the column.
1-4-3 Effects of Contacting Devices
The type of contact devices used in the
1-4-2 Holdups in Column Base and
column can significantly affect the dynamics
Reflux Drum
of the column. Most columns with trays
From a steady-state economic standpoint, have more liquid holdup and therefore
these holdups should be as small as possible slower dynamic responses than columns with
so as to minimize capital investment. In- packing. The structured packing that has
creasing base holdup means that the col- become more widely used in recent years in
umn shell must be longer and requires that processes where pressure drop is important
the column be positioned higher above has faster dynamics than tray columns.
grade. Both of these effects increase capital These columns respond more quickly to
costs. Bigger reflux drums mean high costs changes in manipulated variables, but they
for both the drum and its supporting super- also respond more quickly to disturbances.
structure. Large holdups of liquids may also These considerations should be kept in mind
be undesirable because of safety considera- when designing control systems.
tions (if the material is toxic, explosive,
thermally sensitive, etc,),
1-4-4 Sensors
However, from the standpoint of dynam-
ics and control, the ability of the column to It is important for the operation of the
ride through disturbances is usually im- column to provide adequate sensors: enough
proved by having more liquid holdup avail- thermocouples located so that the tempera-
able. Thus there is a conflict between ture profile can be determined, differential
steady-state economics and controllability, pressure measurements to sense flooding,
and an engineering trade-off must be made. flow measurements of all streams, and so
Liquid holdup times of about 5 min are forth. It is important to include these mea-
fairly typical. This holdup time is based on surements in the original equipment design
the total material entering and leaving the when it is quite inexpensive and easy. Hav-
column base or reflux drum, not just the ing to make field modification to an operat-
stream that is being used to control the ing column can be very expensive and may
level. If the column is very large, subject to take a long time to accomplish.
many large disturbances, or forms part of a
series of operation units, somewhat larger
1-5 SPECIAL PROBLEMS
holdup times may be appropriate. If small
liquid holdups are required because of safety Throughout the chapters in this book there
or thermal degradation considerations, will be much discussion and many examples
somewhat smaller holdup time may be ap- of various types of columns and systems.
propriate. Several commonly encountered situations
But remember, one of the most common that present special control problems are
reasons why columns give control difficulties discussed briefly in the following text.
is insufficient liquid holdups, so make sure
the process designers do not squeeze too
1-5-1 High-Purity Products
hard for the sake of a few dollars. Safety
and thermal degradation problems can of- As noted earlier, distillation columns that
ten be handled by using external tankage produce high-purity products are very non-
for surge capacity so that liquid levels will linear and sensitive to disturbances. Very
not be lost or flow rate to downstream units small changes in the feed split can produce
12 Practical Distillation Control
14
Historical Perspective 15
An important implication of the preced- Today, overdesign safety factors are very
ing is that for optimum operation, one must small. Columns are designed to run much
let all variables vary somewhat except final closer to flooding and thermosyphon reboil-
product quality, and, sometimes, even that ers often operate close to choked-flow in-
may vary within prescribed limits. stability. Older columns with more safety
Classical control theory emphasizes (a) factor in design, and particularly with
rapid response to setpoint changes and (b) bubble-cap trays, were easier to control.
rapid return to setpoint in the face of dis-
turbances. These performance requirements 2-4 COLUMN TRAY AND
really apply only to product quality controls AUXILIARY DESIGN
and to some constraint controls. Some of
the successes of computer control have re- 2-4-1 Tray Design
sulted from restoring control flexibility that Circa 1950 we began to design columns with
was originally lost by excessive reliance on sieve trays instead of bubble-cap trays. For
fixed setpoint control of individual vari- a given load, a sieve tray column is smaller
ables. and cheaper. However, sieve tray columns
have more limited turndown: about 2: 1 in-
2-3 COLUMN DESIGN METHODS stead of about 8: 1. In addition they are
The reader is presumably familiar with subject to weeping and dumping. These
steady-state column design, so little more characteristics, together with tighter design
will be said about it. practices, increase control problems.
In the 1930s and 194Os, what chemical More recently, valve trays have become
engineers call "unit operations theory" popular. They have more turndown capabil-
made tremendous strides. The objectives ities than sieve trays, say 4 : 1, but have the
were twofold: same potential for weeping and dumping. In
addition, they commonly have another prob-
1. To ensure that design of piping, heat lem: inverse response.
exchangers, distillation columns, and so As noted by Rijnsdorp (1961), with an
forth, would have at least flow sheet ca- increase in boilup some columns show a
pacity. momentary increase in internal reflux, fol-
2. To ensure that this equipment would be lowed eventually by a decrease in internal
as little overdesigned as possible to mini- reflux. This is termed inverse response be-
mize capital investment. cause it causes a temporary increase in low
boiler composition at the bottom of the
For distillation columns, the McCabe- column, followed by an eventual decrease.
Thiele procedure for estimating the re- The mechanism was elucidated by Buckley,
quired number of theoretical trays was pub- Cox, and Rollins (1975), who found that
lished in 1925. It did not, however (at least most sieve tray columns demonstrated this
in my experience), achieve wide usage until effect at low boilup rates, whereas most
the early 1940s. During World War II a valve tray columns demonstrated it over the
group of us trying to increase production entire turndown. Thistlethwaite (1980) also
rates in a heavy chemicals plant found that studied inverse response and worked out
existing columns designed before 1930 occa- more details. Composition control via boilup
sionally had tremendous overdesign factors: at the base of a column is vastly more
they were 10 to 15 times larger than neces- difficult when a column is afflicted with in-
sary. verse response. Momentarily, the controller
By 1948, we were using a safety factor of seems to be hooked up backwards. For the
2 for designing distillation columns and heat control engineer this can be even worse
exchangers, and 25 to 50% for piping. than deadtime.
16 Practical Distillation Control
Since the mid-1980s packed columns, The large steam control valve may be
particularly those with structured packing, replaced by a much smaller valve for steam
have been used more extensively. They have condensate. Also, it is usually easier to mea-
about the same turndown as valve trays, say sure condensate flow rather than steam flow.
4 : 1, but do not suffer from inverse re- Flooded reboilers are, however, more slug-
sponse. On the other hand, changes in re- gish than the nonflooded ones, typically by a
flux flow can change column pressure drop factor of 10 or more. Flooded reboilers
significantly. sometimes permit use of low pressure, waste
steam. There may be little pressure drop
available for a steam supply valve.
2-4-2 Reboller Design
Thermosyphon reboilers, particularly of the 2-4-4 Column Base Designs
vertical type, are popular because of low
The size of the column base, the design of
cost. Several kinds of dynamic problems,
the internals (if any), and the topological
however, can result from their use (Buckley,
relationship to the reboiler are all impor-
1974). One is choked flow instability
tant factors in column control. Column base
(Shellene et aI., 1967), which occurs when
holdups may range from a few seconds to
the reboiler is sized too snugly with exces-
many hours (in terms of bottom product
sive heat flux.
flow rate). If there are no intermediate
Another problem is "swell." At low heat
holdups between columns or process steps,
loads there is little vapor volume in the
the column base must serve as surge capac-
tubes; at high heat loads there is much
ity for the next step.
more vapor in the tubes. Hence, an increase
Originally, column bases were very sim-
in steam flow will cause liquid to be dis-
placed into the column base, causing a tem- ple in design, but in recent years elaborate
porary increase in base level. If base level is internal baffie schemes have often been
controlled by steam flow, the controller mo- used. The objective is to achieve one more
stage of separation. Many of these schemes
mentarily acts like it is hooked up back-
make column operation more difficult and
wards. When "swell" is accompanied by in-
limit the use of the holdup volume as surge
verse response of reflux, controlling the base
level via heating medium may be impossi- capacity. Overall, their economics is ques-
ble. tionable; increasing the number of trays in a
column by one is relatively cheap.
Recent trends in thermosyphon design
Elimination of intermediate tanks can
seem to have had at least one beneficial
save a lot of money, both in fixed invest-
effect: critical aT is rarely a problem. Prior
ment (tanks) and working capital (inven-
to 1950, it could be observed fairly often.
tory).
This improvement apparently is due to much
higher circulating rates.
2-4-5 Overhead DeSign
In the chemical industry, particularly for
2-4-3 Flooded Rebollers
smaller columns, it is common to use gravity
Particularly with larger columns, it is com- flow reflux and vertical, coolant-in-shell
mon to use steam condensate pots with condensers. Instrumentation and controls
level controllers instead of traps. In recent can be much simpler than with pumped-back
years some reboilers have been designed to reflux. Regardless, however, of condenser
run partially flooded on the steam side. design, gravity flow reflux systems have a
Historical Perspective 17
dynamics problem if distillate is the con- Prior to 1940 most controllers were ei-
trolled flow and reflux is the difference flow. ther proportional-only or had manual rather
This is the well known problem of "reflux than automatic reset. A few vendors, how-
cycle." Its mechanism and several correc- ever, had begun to use primitive versions of
tive measures are discussed in a paper by automatic reset.
Buckley (1966). By 1945, the chief instruments available
When horizontal, coolant-in-tube con- for distillation control were:
densers are used, control is sometimes sim-
• Flow: Orifice with mercury manometer
plified by running the condensers partially
plus pneumatic transmission, or rotame-
flooded. This technique is especially useful
ters.
when the top product is a vapor. Flooded
• Temperature: Thermocouple or mercury-
condensers seem to be becoming more pop-
in-bulb thermometers.
ular, but like flooded reboilers, are more
• Pressure: Bourdon tube plus pneumatic
sluggish. For both, however, we have worked
transmission.
out mathematical models that enable us to
• Control valve: Air-operated with spring-
calculate the necessary derivative compen-
and-diaphragm or electric motor opera-
sation (Buckley, Luyben, and Shunta, 1985).
tor. Most valves, except for small ones,
were of the double-seated variety; small
ones were single-seated. Valve positioners
2·5 INSTRUMENTATION occasionally were used.
Before automatic controls and automatic By 1950 the picture had begun to change
valves became prevalent, control was dramatically. The so-called force balance
achieved by having operators observe local principle began to be used by instrument
gages and adjust hand valves to control pro- designers for pneumatic process variable
cess variables to desired values. According transmitters and controllers. This feature,
to stories told to me by old timers, some together with higher capacity pneumatic pi-
applications required such careful attention lots, provides more sensitivity, freedom from
and frequent adjustment that an operator drift, and speed of response. Mercury-type
would be kept at a single gage and hand flow transmitters gave way to "dry" trans-
valve. To help keep the operator alert, a mitters and mercury-in-bulb temperature
one-legged stool was provided. transmitters were replaced by gas-filled bulb
Gradually, this practice diminished as transmitters.
manufacturers developed automatic con- Piston-operated valves with integral valve
trols in which measurement, controller, and positioners appeared on the market. The
valve were combined in one package. Some- practicality of larger sized single-seated
times the valve itself, usually powered by air valves permitted less expensive valves for a
and a spring and a diaphragm actuator, was given flow requirement and also provided
a separate entity. Slowly, the practice of increased sensitivity and speed of response.
separating measurement, controller, and With improved actuators and positioners,
valve evolved. reliability improved, and many users aban-
The development of measurement de- doned control valve bypasses except for a
vices that could transmit pneumatic or elec- few cases. This saves money (investment)
tric signals to a remote location was well and in many cases reduces maintenance.
underway by 1940. This permitted the use In 1974, the Instrument Society of Amer-
of central control rooms and, together with ica (ISA) published new standards for calcu-
more automatic controls, a reduction in the lating flow through control valves. This was
required number of operators. one of the most valuable developments of
18 Practical Distillation Control
the past 25 years. It permitted much more insulation, and high input impedance elec-
accurate valve sizing, even for flashing, cavi- tronic circuitry.
tation, or incipient cavitation. In a parallel The first analog electronic controls ap-
development, valve vendors began to make peared in the mid-1950s. They had good
available more accurate plots or tables re- sensitivity and speed of response, but due to
lating C u' the valve flow coefficient, to valve the use of vacuum tubes, had somewhat
stem position. Together with nonlinear high maintenance. In addition, they were
functions available in new microprocessor far less flexible than pneumatics. Subse-
controls, these permitted the design of valve quent switching to solid state circuitry im-
flow compensators (Buckley, 1982), which in proved reliability, but flexibility still suf-
tum, permits much more constant stability fered. It was not until the early 1970s that
over a wide range of flows. one manufacturer produced an electronic
New pneumatic controllers with external line with all of the computation and logic
reset feedback appeared, which permitted functions available in existing pneumatics.
the use of anti-reset windup schemes. Circa In the meantime, the cost competitiveness
1952, one vendor marketted an extensive of pneumatics was improved by a switch
line of pneumatic computing relays, which from metallic transmission tubing to black
permitted addition, subtraction, and multi- polyethylene tubing (outdoors) and soft vinyl
plication by a constant, high- and low-signal tubing (in the control room).
selection, and so-called impulse feed- In the late 1970s a potentially momen-
forward. Decent pneumatic devices for tous change got underway-the introduc-
multiplying or dividing two signals did not tion of so-called distributed digital control
appear until the mid-1960s. Taken together, systems. By using a number of independent
these various devices permitted on-line cal- microprocessors, manufacturers were able
culation of heat and material balances. Cal- to "distribute" the control room hardware
culation of internal reflux from external re- and achieve better reliability than would be
flux and two temperatures also became obtained with a single, large time-shared
practical, and simple pressure-compensated computer (avoid "all eggs in one basket").
temperature schemes became feasible. Early versions were quite inflexible com-
The improved AP transmitters, origi- pared with the best pneumatic and elec-
nally developed for orifice flow measure- tronic hardware, but this situation has now
ments, permitted sensitive specific gravity changed. Some of the newest systems have
and density measurements. They could also more flexibility than any analog system and
be used to measure column AP, increas- lend themselves well to the design of ad-
ingly important for columns designed to run vanced control systems. Transmitters and
closer to flooding. control valves in the field usually may be
Parallel to developments in conventional either pneumatic or electronic.
instruments, there was a much slower devel- No discussion of instrumentation is com-
opment in on-line analyzers. In the late plete without some mention of computers.
1940s, Monsanto developed an automatic The first process control computer was put
refractometer, which was used in a styrene on the market in the early 1950s. A well-
monomer distillation train. In the early publicized application was a Texaco refin-
1950s, several prototypes of on-line infrared ery, where some optimization was per-
and optical analyzers were available. Prob- formed. Another optimization application
lems with reliability and with sample system of that era was a Monsanto ammonia plant
design plus high cost severely limited their at Luling, LA. Other machines came on the
applications. pH measurements, on the market in the next few years, but all of the
other hand, expanded more rapidly, as the early machines suffered from lack of relia-
result of improved cell design, better cable bility, slow operation, and insufficient mem-
Historical Perspective 19
ory. In the 1960s and 1970s applications known books are by Brown and Campbell
were chiefly as data loggers and supervisory (1948) and Chestnut and Mayer (1951). A
controls. Production personnel found that wartime text by Oldenbourg and Sartorius
CRT displays, computer memory, and com- (1948) was translated from German into En-
puter printouts were a big improvement over glish and became well known. Although the
conventional gages and recorders. As the theory required significant modifications to
memory and speed of computers have in- adapt it to process control, for the first time
creased and as the expertise of users has it permitted us to analyze many process
improved, we have seen a resurgence of control situations quantitatively.
interest in optimization. The technique involves converting pro-
The latest development in instrumenta- cess equations, including differential equa-
tion began in mid-1983 when one vendor tions, into linear perturbation equations. By
put a line of "smart" transmitters on the perturbations we mean small changes
market. These featured built-in micropro- around an average value. Linear differential
cessors to permit remote automatic calibra- equations with constant coefficients are eas-
tion, zeroing, and interrogation. These fea- ily solved by use of the Laplace transforma-
tures, together with reduced drift and tion. It is usually easy to combine a number
sensitivity to ambient conditions, and with of system equations into one. This equation,
more accuracy and linearity, improve plant which relates output changes to input
operation. They also facilitate and reduce changes, is called the transfer function. Time
maintenance. However, test data I have seen domain solutions are readily calculated, and
on such instruments indicate that they are if the uncontrolled system is stable, the
slower (poorer frequency response) than transfer function may be expressed in terms
their pneumatic counterparts. Eventual ex- of frequency response.
tension to instruments with multiple func- The transfer function approach permits
tions such as temperature and pressure is convenient study of system stability, permits
foreseen. Present indications are that within calculation of controller gain and reset time,
the next several years a number of manufac- and permits calculation of system speed of
turers will put "smart" instruments on the response to various forcing functions. In the
market. precomputer era, these calculations were
made with a slide rule and several graphical
2-6 CONTROL SYSTEM DESIGN aids. Today, the transfer function tech-
niques are still useful but the calculations
METHODS
usually may be made more rapidly with a
Prior to 1948, distillation control systems as programmable calculator or a computer.
well as other process control systems were Early applications were in flow, pressure,
designed by what might be called the instru- and liquid level systems. Although fre-
mentation method. This is a qualitative ap- quency response methods of calculation
proach based principally on past practice were most commonly used, some users pre-
and intuition. Many "how-to" papers have fer root locus methods.
appeared in the literature. A good summary Throughout the 19508 and 1960s the
of the state-of-the-art in 1948 is given by transfer function approach was used exten-
Boyd (1948). sively for piping, mixing vessels, heat ex-
In the period 1948 to 1952, there ap- changers, extruders, weigh belts, and, occa-
peared a large number of books based on sionally, distillation columns. Two books
control technology which had been devel- that appeared in the 1950s illustrated the
oped for military purposes during World application of transfer function techniques
War II. Commonly the title mentioned to process control (Campbell, 1958; Young,
"servomechanisms theory." Two of the best 1957).
20 Practical Distillation Control
The relatively simple modelling tech- As the speed and memory of technical
niques developed earlier were of limited use computers increased, distillation models
for systems with significant nonlinearities or have become more detailed and elaborate.
those describable by either partial differen- Tolliver and Waggoner (1980) have pre-
tial equations or by a large number of ordi- sented an excellent discussion of this sub-
nary differential equations. Gould (1969) ject.
explored some more advanced modelling Many newer control techniques have
techniques for such systems. been devised. Among these are "internal
The limitations of the simple transfer model control" and "dynamic matrix con-
function approach for distillation columns trol." These and other techniques are sum-
led some workers to try another approach marized and reviewed by Luyben (1990).
-computer simulation. The work of There has been some application to distilla-
Williams (1971) and Williams, Harnett, and tion control, as discussed in Chapter 12.
Rose (1956) in applying this approach is Although most process measurements are
well known. Originally, simulations were continuous, some are not, as, for example,
mostly run on analog computers. Then, once per day sampling and laboratory analy-
when improved digital computers and FOR- sis of column product streams. This led to
TRAN programming became available, an interest in the techniques of sampled-
there was a shift to digital techniques. Basi- data control. This permits the quantitative
cally, simulation consists of writing all of the design of discontinuous control systems.
differential equations for the uncontrolled There have been few applications to distilla-
system and putting them on the computer. tion control in spite of the fact that many
Various control loops can then be added, composition measurements are discontinu-
and after empirically tuning the controllers, ous. One application has been the use of
the entire system response can be obtained. Shannon's data reconstruction theorem to
In some cases real controllers, not simu- establish the relationship between a speci-
lated ones, have been connected to the fied sampling frequency and required pro-
computer. cess holdup, or between a specified holdup
In 1960, Rippin and Lamb (1960) pre- and required sampling frequency.
sented a combined transfer function-simu- These newer techniques require a lot of
lation approach to distillation control. The training beyond that required for transfer
computer simulation was used to develop function techniques, and industrial applica-
open loop transfer functions. Then standard tions have so far been limited. There is no
frequency response methods were used to question, however, that there exist ample
calculate controller gain and reset as well as incentives and opportunities for this newer
desired feedforward compensation. In our control technology.
own work, we have found this combination An interesting trend in recent years has
approach to be quite fruitful. been toward the development of standard-
Since about 1980, there has been enor- ized programs for control system design.
mous progress in small or personal comput- This is commonly called computer aided
ers (PCs) accompanied by a more gradual design (CAD). Although originally aimed at
evolution in software. The machines have larger machines, many useful programs exist
more memory and speed, and both ma- for small computers and even pro-
chines and software are more user friendly. grammable calculators. The portability and
Simple distillation columns are readily mod- permanent memory features of many of the
eled and programmed; see, for example, the latter render them very useful in a plant or
program illustrated in Luyben (1990). Mod- maintenance shop.
elling, including newer approaches, is dis- Since about 1970, there has been increas-
cussed by other authors in this book. ing interest in newer approaches to control.
Historical Perspective 21
The incentives have been twofold: egy, as mentioned earlier, has three main
facets: material balance control, product
1. Because flow sheet data are often incom- quality control, and constraints. What kind
plete or not exact, it is desirable to mea- of controls do we use to implement them?
sure process characteristics of the real Material balance controls are tradition-
plant after startup. Furthermore, process ally averaging level and averaging pressure
dynamics change as operating conditions controls. Although the basic concept dates
change. The increasing availability of back to 1937, averaging controls are still
on-line computers permits the use of widely misunderstood. The original theory
"identification" techniques. Information was expanded in 1964 (Buckley, 1964) and
thus obtained may be used to retune again quite recently (Buckley, 1983). To
controllers either manually or automati- minimize required inventory and achieve
cally. If a column is not equipped with maximum flow smoothing we use quasilin-
analyzers, identification techniques may ear or nonlinear proportional integral (PI)
be used to deduce compositions from a controllers cascaded to flow controllers. If
combination of temperature, pressure, we can use level control of a column base to
and flow measurements. A particular ap- manipulate bottom product flow and level
proach that was developed by Brosilow control of the overhead receiver to adjust
and Tong (1978) is termed inferential top product flow, then we can usually avoid
control. For binary or almost binary dis- the use of surge tanks between columns or
tillation, pressure-compensated tempera- other process steps. This saves both fixed
ture usually suffices. investment and working capital.
2. Many processes are plagued by interac- Product quality controllers are usually
tions. There are a variety of techniques also of the PI type. Overhead composition is
for dealing with this that may be lumped usually controlled by manipulating reflux
under the heading "modern control the- and base composition by manipulating heat
ory" (Ray, 1981). That aspect relating to flow to the reboiler. It is also increasingly
time-optimal control will likely be help- common to provide feedforward compensa-
ful in future design of batch distillation tion for feed rate changes. This minimizes
columns. Implementation of this technol- swings in boilup and reflux, and in top and
ogy usually requires an on-line computer. bottom compositions.
One of the most important developments
Some other methods of dealing with in- in control techniques is that of variable con-
teractions include relative gain array (RGA), figuration and variable structure controls.
inverse Nyquist array (INA), and decou- Although it is much more common to use
pIers. See Luyben (1990) for a good short fixed configuration and fixed structures, in
summary. reality either or both should change as pro-
cess conditions change. For example, the
steam valve for a distillation column re-
2-7 PROCESS CONTROL
boiler may, depending on circumstances, re-
TECHNIQUES
spond to controllers for:
In 1945 there existed no overall process
control strategy. As one process design Steam flow rate
manager put it, "instruments are sprinkled Column b.p
on the flow sheet like ornaments on a Column pressure
Christmas tree." Not until 1964 was there Base temperature
published an overall strategy for laying out Column feed rate
operational controls from one end of a pro- Column base level
cess to the other (Buckley, 1964). This strat- Column bottom product rate
22 Practical Distillation Control
The seven variables listed may also exert linear control, we recommend a book by
control on five or six other valves. We call Oldenburger (1966).
this variable configuration.
For level control, the quasilinear struc-
ture referred to earlier consists normally of 2·8 INFLUENCE ON DISTILLATION
PI control, but proportional-only control is CONTROL
used when the level is too high or too low.
We call this variable structure. To accom- Some years ago I was preseni: when one of
plish this automatically, it was proposed in our most experienced distillation experts was
1965 (Buckley, 1968) to use a type of control asked "What is the best way to control a
called overrides. In ensuing years a number column?" His answer: "Gently." When ap-
of other publications, (Buckley and Cox, proaching column control one should keep
1971; Buckley, 1969b; Cox, 1973) described in mind that conditions in a column cannot
the applications to distillation. The math be changed rapidly. Rapid changes in boilup
and theory are presented in a two-part arti- or rate of vapor removal, for example, may
cle (Buckley, 1971). cause momentary flooding or dumping. The
Fairly recently, there has been a growing developments since 1945 have therefore not
interest in "self-tuning regulators." One of been aimed at rapid control but rather at
the most prolific contributors has been smoother operation with better separation
Astrom (1979) in Sweden. So far, applica- at lower cost.
tions to industrial distillation columns are If we now look back at 1945 and move
not known, but they have been tried on forward we can see how the various devel-
pilot columns. opments previously discussed have inter-
Although little has been published, there acted to influence distillation control.
has been a certain amount of work with
nonlinear controllers. Unlike linear theory, 1. Switch to sieve and valve trays and tighter
which has a cohesive, well-developed column design. These have encouraged
methodology, nonlinear theory is essentially the use of minimum and maximum boilup
a bag of tricks. The best approach I have overrides. A high I1p override on steam
found is to run a large number of simula- is now universally employed by some en-
tions with different algorithms and for vari- gineering organizations. We also fre-
ous types and sizes of forcing functions. quently provide maximum and minimum
For averaging level control we found overrides for reflux.
(Buckley, 1986) that with linear reset and 2. Increased use of packed columns. As
gain proportional to the error squared, we mentioned earlier, packed columns, par-
got flow smoothing far superior to that ticularly those with structured packing,
achievable with any linear controller, have become more popular.
whether P or PI, and experiences where 3. Tighter design of heat exchangers, espe-
P-only controllers were replaced with this cially reboilers. To prevent choked flow
type of controller confirmed simulation re- instability we sometimes provide maxi-
sults. mum steam flow limiters.
For composition control, we tried nonlin- 4. Increased use of side-draw columns. One
ear gain to filter out noise. Sometimes it side-draw column is cheaper than two
worked well, sometimes not. conventional columns with two takeoffs.
Our experience with trial-and-error tun- Turndown is less, however, and the col-
ing of nonlinear controllers in a plant has umn is harder to control (Buckley, 1969a;
been negative. These applications should Doukas and Luyben, 1978; Luyben, 1966).
be simulated. For those interested in non- Such columns usually require more over-
Historical Perspective 23
Systems. Research Triangle Park, NC.: Instru- Continuous Distillation Columns. New York:
ment Society of America. Elsevier.
Campbell, D. P. (1958). Process Dynamics. New Ray, W. H. (1981). Advanced Process Control.
York: John Wiley. New York: McGraw-Hill.
Chestnut, H. and Mayer, R. W. (1951). Ser- Rijnsdorp, J. E. (1961). Birmingham University
vomechanisms and Regulating System Design, Chemical Engineer 12,5-14, 1961.
Vols. I and II. New York: John Wiley. Rippin, D. W. T. and Lamb, D. E. (1960). A
Cox, R. K. (1973). Some practical considerations Theoretical Study of the Dynamics and Control
in the application of overrides. Presented at of Binary Distillation. Newark, DE: University
ISA Symposium, St. Louis, MO, April 1973. of Delaware.
Deshpande, P. B. (1985). Distillation Dynamics Shellene, K. R., Stempling, C. V., Snyder, N. H.,
and Control. Research Triangle Park, NC: and Church, D. M. (1967). Experimental study
Instrument Society of America, 1985. of a vertical thermosyphon reboiler. Pre-
Doukas, N. and Luyben, W. L. (1978). Control of sented at Ninth National Heat Transfer Con-
sidestream columns separating ternary mix- ference AIChE-ASME, Seattle, WA, August
tures. In Tech. 25, 43-48. 1967.
Gould, L. A. (969). Chemical Process Control. Shinskey, F. G. (1977). Distillation Control. New
Reading, MA: Addison-Wesley. York, McGraw-Hill.
Luyben, W. L. (1966). 10 schemes to control Thistlethwaite, E. A. (1980). Analysis of inverse
distillation columns with side stream drawoff. response behavior in distillation columns. M.S.
ISA J. July, 37-42. Thesis, Department of Chemical Engineering,
Luyben, W. L. (1990). Process Modeling, Simula- Louisiana State University.
tion and Control for Chemical Engineers, 2nd Tolliver, T. L. and Waggoner, R. C. (1980). Dis-
ed. New York: McGraw-Hill, 1990. tillation column control: A review and per-
Nisenfeld, A. E. and Seemann, R. C. (1981). spective from the CPI. ISA Adv. Instrum. 35,
Distillation Columns. Research Triangle Park, 83-106.
NC: Instrument Society of America. Williams, T. 1. (1971). Distillation column con-
Oldenburger, R. (1966). Optimal Control. New trol systems. Proceedings of the 12th Annual
York: Holt, Reinhart, and Winston. chemical and Petroleum Industries Sympo-
Oldenbourg, R. C. and Sartorius, H. (1948). The sium, Houston, April 1971.
Dynamics of Automatic Controls. New York: Williams, T. J., Harnett, B. T., and Rose, A.
ASME. (1956). Ind. Eng. Chem. 48,1008-1019.
Rademaker, 0., Rijnsdorp, J. E., and Maar- Young, A. 1., Ed. (1957). Plant and Process Dy-
leveld, A. (1975). Dynamics and Control of namic Characteristics. New York: Academic.
2
Methods
3
Rigorous Modelling and
Conventional Simulation
Vincent G. Grassi II
Air Products and Chemicals, Inc.
29
30 Practical Distillation Control
E = U + PE + KE (3-2b) • Dift'crenIiaJ-AJaebraic:
'Index-Oor!
U=H- PV (3-2c) • Stiff I NOlI-Stiff
usually converge, dynamically, to a bona ous subsections within the distillation tower.
fide steady state as the integration pro- Finally, we will discuss the performance of
ceeds. This method of initialization is useful these models on various computer platforms
to get a dynamic simulation running quickly. and find how well the models match real
In fact, often it is numerically more stable plant towers.
to converge to the steady state dynamically
than to solve for the steady state alge-
3-2-2 Vapor-Uquld Equilibrium
braically.
The following simple first order Euler It is essential to have an accurate model or
integrator will be used to integrate the correlation for the vapor-liquid equilibrium
model: (VLE) and physical properties of the com-
ponents in the distillation tower. There are
dXI many useful sources for obtaining literature
X,H, = X, + dt at (3-3a) data and correlating equations to represent
your system (Dauber and Danner, 1988;
X( t = 0) = Xo (3-3b) Gmehling and Onken, 1977; Reid, Praus-
nitz, and Sherwood, 1977; Walas, 1985;
where X is a function of time (t). The Euler Yaws, 1977). The phase equilibrium of some
integrator is very easy to implement and is systems is essentially ideal. Raoult's law,
very effective at solving a system of nonstiff
differential equations. The Euler integrator YiP =xiP/ (3-4a)
works well for solving the distillation model
that we develop in this chapter. Models that represents ideal systems. Others may have
include vapor hydraulics and pressure dy- constant relative volatility:
namics are stiff. It is usually more efficient
to use an implicit integrator such as LSODE
(Hindmarsh, 1980) for such stiff systems. It
is important to remember the difficulty with
stiffness is that it slows down the simulation.
Because the computational speed of com- Some can be correlated simply as a polyno-
puters has dramatically increased over a mial in temperature:
short period, systems that were considered y.
a few years ago to be too stiff to solve K·I = ....!..
Xi =A·I + B.T + C-T 2 + ... (3-4c)
explicitly are now quickly solved by present I I
(3-5b)
aIn 1i
L
hmix = - RT
2
l;:Xi---aT (3-6c)
I
M, = A~T
, + B!-T2
, (3-6a)
Yaws gives constants for many common
hr = ArT + BrT 2 + llHt (O°F) (3-6b) components, or you can regress data to ob-
Rigorous Modelling and Conventional Simulation 37
tain the parameters. To compute the liquid the solution (Xl and x 2 ), and the function is
density of a mixture use evaluated at these two points (/1 and 12)' A
secant is then drawn between the two val-
ues. A third point ([3) is calculated where
(3-Sb) the secant intersects the line where I(x) =
x, and the function is evaluated at this third
point (/3)' If 13 is sufficiently close to x 3 '
where Pi = pure component liquid density
the solution has been found; if not, the last
Objft 3)
two points are used as the next set of initial
PL = mixture liquid density Objft3)
guesses and the procedure is repeated. The
ML = liquid mixture molecular
Wegstein method is very easy to implement
weight Ob lIb-mol)
in a computer program.
Mi = pure component molecular
The second convergence method that
weight (Ib lIb-mol)
is very useful for distillation problem is the
Tr reduced temperature
=
Newton-Raphson method. The Newton-
A, B = pure component density corre-
Raphson method requires the objective
lation constants
function to be written in the form I(x) = O.
This method requires the first derivative of
Vapor passing through a distillation tower
the function to be evaluated, preferably an-
tray aerates the liquid, creating a froth. The
alytically. This requirement adds some com-
clear liquid density, Equation 3-8b, is cor-
plication that the Wegstein method does
rected for frothing by correlations, usually
not have, but improves the rate of conver-
obtained from the tray vendor design guides
gence. Newton-Raphson is a very powerful
or generalized correlations (Van Winkel,
numerical method for thermodynamic cal-
1967). Computing the froth density as a
culations, such as the bubble point. The
function of vapor rate is necessary if you
analytical derivative is readily available for
want to study flow hydraulics. Changes in
froth density with vapor rate can lead to
inverse response behavior in bottom sump
x - I(x)
composition and level dynamics. There is no
need to compute the froth density continu- .
, ,\;f'
~'"
ously if you are not interested in froth dy- , '"
namics. Omitting this calculation at each 12
2
time step will improve the speed of the ,
I, I(x)
simulation.
'" '"
, '" '"
3-3 COMPUTER SIMULATION
'" '" '"
3-3-1 Algebraic Convergence Methods , '"
, '"
Our conventional simulation of a distillation
tower requires solving algebraic equations X, X2 X3
I(x) - 0
T
I(x)
LIquid
11 - I~
x y
o~----~~------~-------
ps1 = exp(c i .
,I
+~)
T+C.
3,1
1 -0
-1--1'
- I l=l-LYi=O
XI - X2 i
T
1. Calculate the equilibrium vapor compo-
sition and temperature from pressure and
liquid composition by a bubble point cal-
culation.
2. Calculate the actual vapor composition
from the Murphree vapor phase stage
LIquid
efficiency.
3. Calculate the vapor and liquid enthalpy
y
from their composition and stage tem-
x
perature.
4. Calculate the clear liquid density from
x·=-
,
YiP
'Yi P/
p'
,
= exp(c "T+
1 . +~)C .
composition and temperature.
3" 5. Calculate the liquid froth density at the
f= 1- LXi = 0 stage vapor rate.
6. Calculate the liquid rate leaving the stage
from the Francis weir formula.
7. Calculate the vapor flow leaving the stage
from the energy balance.
FIGURE 3-7. Equilibrium dew point calculation (bi- 8. Calculate the component and total mass
nary system at constant pressure). derivatives.
40 Practical Distillation Control
dM"
Material Balance dt OK F + L,,+l + v,,-l - L" - v"
Component Balance
d(h;M,,) L v L V
Energy Balance dt = hFF + h,,+lL,,+l + h,,-lv,,-l - h"L" - h"v"
dh L
- " =0
dt
Liquid Hydraulics
(Francis weir) Ln = Cp~wlenH~ (C is a constant for units conversion)
eo
p'
K. = 11,,, I,"
Vapor-liquid Equilibrium
',n P"
The stage model is the keystone of the 2. Calculate the tray temperature from en-
dynamic distillation tower model. For each thalpy and tray composition.
stage in the column, we repeat it in a loop. 3. Calculate the tray pressure from temper-
The stage model presented here can be ature and liquid composition.
modified easily to meet your specific mod- 4. Calculate the vapor rate leaving the tray
elling needs. For example, if precise froth from the tray pressure and tray vapor
density is not important for your needs, the flow hydraulics.
correlation can be replaced easily with a
constant frothing factor. Sidestream prod- Vapor hydraulics are orders of magni-
ucts are added easily by an additional term tude faster than liquid composition dynam-
for the sidestream flow leaving the stage. ics. This results in a very stiff set of differ-
We have treated stage pressure as a con- ential equations. It will require a small time
stant in this model. Variable pressure drop step, a fraction of a second, to integrate the
requires the addition of a pressure drop equations. It probably will be more efficient
relation expressed in terms of vapor flow. to use a stiff integrator, such as LSODE, to
This requires the energy balance to be solve this model. Fortunately, vapor and
solved as a differential equation. The change pressure dynamics are not usually important
in specific enthalpy is not neglected. for product composition control problems
and we do not have to resort to this model
too often.
1. Calculate the specific enthalpy and liquid We can use the stage model we have
composition by integration. developed here to help classify different
Rigorous Modelling and Conventional Simulation 41
types of distillation stage models. Five mod- model 3 if there is significant variation in
els are given in Figure 3-9. The simplest tray pressure drop, or if pressure dynamics
model (1) is appropriate when the molar are important. Distillation columns operat-
latent heats of the system components are ing under vacuum might require model 3.
very close, producing equimolar overflow. The differential equations become stiff when
The vapor rate through each section of the solving model 3, so model 3 is best solved
column is constant, eliminating the energy using an implicit integrator, such as LSODE.
balance. We integrate the differential mass Model 4 is rarely applicable to industrial
continuity equations, and the Francis weir columns. You need to use this model when
formula gives the liquid rate. there is significant vapor holdup, that is,
Models 2a, 2b, and 3 are the most widely pressure greater than 150 psia, and variable
applicable dynamic distillation tower mod- tray pressure drop is important. Variable
els used by the chemical industry. We have tray pressure drop is usually only important
developed model 2a here. I have found this in vacuum columns. Yet vacuum columns
model sufficient to solve 95% of the indus- have insignificant vapor holdup, so model 3
trial distillation problems that I have worked is usually used. Applications for model 4 are
on. The remaining 5% were solved using very rare, but should the need for it arise,
models 2b or 3. Use model 2b for high you will be ready. This model is probably
pressure (greater than 150 psia) systems solved best using a nonconventional simula-
where the vapor holdup is constant. Use tion, that is, solving the system of differen-
tial and algebraic equations simultaneously.
The differential-algebraic integration pro-
Modell
gram DASSL can be used to solve the
model.
· Equimolal Overflow
· Negligible Vapor Holdup
· Constant Pressure 3-3·5 Bottom Sump
· Single Liquid Phase
The model of the distillation tower bottom
sump is similar to that of the tower stage.
Model2a The bottom sump is modelled as an equilib-
· Negligible Vapor Holdup rium stage. There is heat input from the
· Negligible Specific Enthalpy Change reboiler. Material holdup is large and vari-
· Constant Pressure or Tray Pressure Drop able. Changes in specific enthalpy will not
be neglected. The dynamic model is pre-
Model2b sented in Figure 3-10.
The time derivative of the specific en-
· Significant, but Constant, Vapor Holdup
thalpy can be approximated numerically by
· Negligible Specific Enthalpy Change
· Constant Pressure or Tray Pressure Drop backward difference. The specific enthalpy
is saved at each time step. The current
value is subtracted from the previous value
Model 3
and the result is divided by the time step. It
· Negligible Vapor Holdup is necessary to compute the specific en-
· Significant Specific Enthalpy Change thalpy derivative so that the vapor rate leav-
· Variable Tray Pressure Drop ing the sump can be calculated from the
energy balance.
Model 4
""_--s_
B, XI>
dM
Material Balance - =L l - Vo-B
dt
d(Xb,jM)
Component Balance ---'-~--'-- = Xl ,.L l - YO Fo - Xb , jB
dt ' ,
tower, you need to verify that the model TABLE 3·2 Computer System Execution Speed
does mimic plant operation. RealTime
A good model will match the plant at Computer MFLOPS a Factor b
steady state and accurately track the pro-
Cray Y-MP (1 processor) 161 860
cess during dynamic upsets. A verification
HP9000/720 10 580
of the model includes a check of the overall IBM RS/6000-530 14 390
material and energy balances, steady-state DECstation 5000/200 4 250
temperature profile, and open loop dynamic VAX 8700 1 90
disturbance tests. A final test of closed loop Compaq 386/25 MHz 0.2 27
performance will ensure that your model is aDongara (1991).
accurate. Once you have verified many dif- bSimulation based on a 7.S ft, 38 tray alcohol-water col-
ferent columns, you will have confidence in umn with a setpoint change:
modelling a new design that cannot be veri- process real time
fied until startup. real time factor (RTF) = I . . .
e apsed sUDulatlon time
Figures 3-11a and 3-11b are two verifi-
cation plots completed for a real 38-tray
alcohol-water column. Figure 3-11a is the the engineer improved process insight and
steady-state temperature profile computed solves very significant problems.
using the simulation with plant data over- Table 3-2 contains the run time perfor-
laid. Figure 3-11b is a dynamic open loop mance, from various computing platforms,
test of a tray temperature for a step change of a dynamic distillation simulation. These
in reflux flow. These curves are illustrative data correspond to the 38-tray binary alco-
of my experience in successfully verifying hol-water distillation tower that we verified
dynamic distillation models with plant data. in the previous section.
The simulation can now be used to develop We measure execution performance in
new and improved process control strate- terms of a real time factor. This is the
gies. number of times faster than real time the
simulation runs. Computing performance is
growing geometrically each year. The desk-
top, laptop, or fingertop computer of tomor-
3-6 COMPUTATIONAL row will be fast enough that entire plants of
PERFORMANCE five or more towers will be added to the top
Computational performance and productiv- of Table 3-2.
ity have progressed well beyond our expec-
tations over the last 30 years. It is likely to
exceed our current expectations in the next
3-7 CONCLUSIONS
few decades. I have presented an approach to modelling
A decade ago, complex engineering prob- distillation dynamics that builds on engi-
lems were only discussed qualitatively. To- neering experience and judgment. The ap-
day, these problems are quickly quantified proach is based on modelling the physical
and solutions put into production. elements as modules and connecting the
We now have the power of supercomput- components together in the body of a simu-
ing at an affordable price, either on the lation. The modules chosen are those that
desk or tied to it through a network of are important to solve our engineering ob-
computers. System software and services jectives for the problem at hand.
have reached the point that engineers with- We developed a rigorous model for a
out extensive training in computer technol- general distillation tower. The model is
ogy can program steady-state and dynamic based on first principles. We implemented
models. The modelling activity itself gives the model in a computer simulation using
46 Practical Distillation Control
Gmehling, 1. and Onken, U. (1977). Vapor- Reid, R. C., Prausnitz, 1. M., and Sherwood,
Liquid Equilibrium Data Collection. Frankfurt, T. K. (1977). The Properties of Gases and
Germany: DECHEMA. Liquids. New York: McGraw-Hill.
Hindmarsh, A. C. (1980). LSODE: Livermore Van Winkel, M. (1967). Distillation. New York:
solver for ordinary differential equations. McGraw-Hill.
Lawrence Livermore National Laboratory. Walas, S. M. (1985). Phase Equilibrium in Chemi-
Luyben, W. L. (1990). Process Modeling Simula- cal Engineering. Boston: Butterworth.
tion and Control for Chemical Engineers, Yaws, C. L. (1977). Physical Properties. New
2nd ed. New York: McGraw-Hill. York: McGraw-Hill.
4
Approximate and
Simplified Models
Antonis Papadourakis
Rohm and Haas Company
John E. Rijnsdorp
University of Twente
48
Approximate and Simplified Models 49
TABLE 4-1 Purposes for Simple Distillation Models for a more thorough analysis of plant oper-
Understanding process behavior ability, controllability, and switchability, and
Optimization of process operation for operator training. In the latter case,
Startup, switchover, or shutdown speed of computation is more critical than
Assessing process controllability in the former case, which warrants model
Designing regulatory control structures simplification.
Composition estimation
Model-based control
Simulation, particularly of integrated plants 4-3 SIMPLE STEADY-STATE
(for operability
and controllability analysis and for MODELS
operator training)
As has already been indicated in the pre-
ceding section, simple models can be used
for tracing optimum operation points, pro-
Startup, shutdown, and major switchovers vided they are validated for the actualoper-
can only be analyzed by nonlinear models. ating region. A very simple model has been
For complete plants, consisting of many proposed by Shinskey (1967):
process units, order reduction is mandatory,
but not very easy. For an approximate anal- V/F = fJc In Sc (4-1)
ysis of minor switchovers, linear models are
sufficient and high order does not impose
any great difficulties. where V = vapor flow rate (mol/s)
There is a need to assess process con- F = feed flow rate (mol/s)
trollability in early design phases, when fJ c = column characterization factor
the flowsheet has not yet been frozen. Sc = separation factor, equal to
Here singular value decomposition (see
Skogestad and Morari, 1987a, for a brief
description and a worthwhile extension) has
proven to be a simple and valid method
(Perkins, 1990). This method requires a (lin- where LK = mole fraction of the light key
ear) frequency response matrix between component
process inputs and the variables to be con- HK mole fraction of the heavy
=
trolled. key component
The same model can be used for de- D = top product
signing the basic control structures, which B = bottom product
includes a separation between single-
input-single-output (SISO) and multiple- The column characterization factor is a
input-multiple output (MIMO) subsystems. function of the number of trays, the tray
For the latter, the block relative gain array efficiency, and the relative volatility of the
method (Arkun, 1987) has proven to be two key components. It is the parameter to
quite useful. be fitted to the operating region.
Real-time composition estimation and Equation 4-1 does not contain the influ-
model-based control are only practical when ences of feed enthalpy and feed tray loca-
the model order is low. In addition, in case tion. The first type influence can be ac-
of high-purity separations, nonlinearities counted for by defining separation factors
play an important role, which brings us again for each column section (Rijnsdorp, 1991):
to the problem of order reduction in nonlin-
ear models (see Section 4-6 for more detail). LR/D = fJR In SR (4-3)
Finally, in later process design phases,
simulation of complete plants is desirable Vs/B = fJ s In Ss (4-4)
50 Practical Distillation Control
As will be shown in the next subsection, 3. Vapor and liquid concentrations, associ-
this can be realized under relatively mild ated with temperature and composition
assumptions, with submodels for the follow- control.
ing clusters of process variables (see Figure 4. Condenser temperatures and heat flow
4-1): rates.
5. Reboiler temperatures and heat flow
rates.
1. Pressure and vapor flow rates, associated
with pressure and pressure drop control.
2. Liquid holdups and flow rates, associated The first three submodels have no inter-
with liquid level control in the top and actions, but there is an important action
bottom accumulators. from the first to the second and third, and
C
Tc
V.AP•• L.AM••
rectifyiq PAY. rec:tifyq L. P.t v.
IeCtial aectim
H
T"
from the second to the third. In fact, the Robinson and Gilliland (1950) have ex-
influences from the latter on the former tended the scope by introducing pseudo-
mainly go through the reboiler and the con- molecular weights. In this way, unequal sen-
denser submodels, which cause inherent sitivity coefficients for vapor and liquid can
regulation and deregulation, respectively. be equalized, The accuracy of the method
has been demonstrated for the design of a
column separating the nonideal mixture of
4-4-2 Assumptions
ammonia and water. Chen (1969) has shown
The assumptions are: that this method is also consistent for the
dynamic case.
• Equal (pseudo) molal overflow.
In actual trays, changes in liquid hold-
• Small influence from liquid tray holdup
up will influence the pressure drop across
on tray pressure drop.
the tray. For trays with downcomers within
• Tray pressure drop small compared to the
the normal operating range (where most of
absolute value of the pressure.
the liquid holdup variations appear in the
• Negligible heat loss.
downcomer), the liquid height on the tray
In distillation column design, a popular proper does not vary much, so the influence
assumption is that of equal molal overflow. of the vapor flow is most important. Of
In terms of the relations between specific course this no longer holds when the trays
enthalpies and compositions this assump- are overloaded, that is, when the downcom-
tion is equivalent to (where the summation ers start to overflow.
is over all except the least volatile compo- The third assumption even holds for vac-
nent) uum operation, possibly with exception for
trays near the top where the absolute pres-
HL(s,t) =Ho(s,t) - Eajxj(x,t) (4-13) sure is very low. Finally, the fourth assump-
and tion is warranted for well insulated big
columns and for small columns provided
Hv(s, t) = Ho(s, t) + Hc(s, t)
with a compensatory heating system.
- EajYj(s, t) (4-14) On the basis of the assumptions, the en-
ergy balance of the tray can be reduced to
the following, very simple balance for the
where xis, t) liquid mole fraction for
= vapor flows (Rademaker, Rijnsdorp, and
component j on tray s Maarleveld, 1975):
yis, t) = vapor mole fraction for
component j on tray s dP(s, t)
aj = sensitivity coefficient for B(s,t) dt =V(s-1,t)-V(s,t)
component j, assumed to
(4-15)
be equal for vapor and
liquid
Ho(s, t) = specific liquid enthalpy of
the least volatile compo- where B(s, t) = effective vapor capacity of
nent (J ImoI) the tray
Hc(s, t) = heat of vaporization of P(s, t) = pressure on the tray
the least volatile compo-
nent (J ImoI) The physical interpretation is that an in-
creasing pressure requires compression of
Evidently, in the enthalpy-composition hy- the vapor holdup and condensation of vapor
perspace this corresponds to parallel hyper- for heating the tray to the new equilibrium
planes for liquid and vapor enthalpy. temperature, both resulting in a difference
Approximate and Simplified Models 53
between the vapor flow entering the tray sures and the vapor flows. We just have to
and that leaving the tray. extend it with an equation for the tray pres-
The effect of resistance to mass transfer sure drop. According to the second assump-
between vapor and liquid can be accounted tion, the latter only depends on the vapor
for by a more refined version of Equation flow rate. Now Equation 4-16 can be ap-
4-15: proximated by
dP*(s, t) dP(s, t)
B*(s,t) dt +Bv(s,t) dt dP*(s, t) dP(s, t)
B*(s, t) dt + Bv(s, t) dt
= V(s - 1, t) - Yes, t) (4-16)
= fv{P(s - 1, t) - pes, t)}
of B(s, t)
B*(s, t) = remainder of B(s, 0
where f v{ } represents the relationship be-
P*(s, t) = pressure in equilibrium
tween vapor flow rate and pressure drop.
with the liquid tempera-
For n trays, Equations 4-17 and 4-18 lead
ture, determined by
to a set of state equations with the 2n tray
pressures [the pes, Os and the P*(s, t}s] as
dP*(s, t) [pes, t) - P*(s, t)] state variables. In analogy to electrical net-
B*(s, t) dt works it can also be interpreted as a resis-
Rp(s, t)
tance-capacitance ladder (see Figure 4-2)
(4-17)
with nonlinear resistances. Note that this
network does not tell the complete story. In
where R/s,O is the resistance to mass fact, the condenser reacts to the change in
transfer between the phases on tray s top vapor flow and so provides a feedback
(Pa s/mol). to the submodel. A similar influence occurs
at the bottom end via the reboiler.
4-4-3 Propagation of Vapor Flow and
Pressure Responses
4-4-4 Propagation of Liquid Flow and
The equations for the vapor flow balances
Uquld Holdup Variations
across the trays (see Equation 4-15, or
Equation 4-16 together with Equation 4-17) The responses of liquid holdups and flow
form the basis of the submodel for the pres- rates are determined by the total mass bal-
(ft/AP)-1
~T T
FIGURE 4-2. Electrical network interpretation of propagation of vapor flow and
pressure responses.
54 Practical Distillation Control
dML(s, t)
dt =L(s+l,t)-L(s,t) d[ ML(s, t)xj(s, t)]
dt
+V(s - 1,t) - V(s,t)
(4-19) d[ Mv(s, t)Yj(s, t)]
+ dt
With Equation 4-15 and with the hydraulic
= L(s + 1, t)Xj(s + 1, t)
equation for the tray,
-L(s,t)xis,t)
ML(s,t)
= fM[L(s, t), V(s - 1, t)] (4-20)
+ V(s - 1, t)Yis - 1, t)
we find
- V(s, t)Yj(s, t) (4-25)
dL*(s, t)
'TL(S,t) dt where xj(s, t) mole fraction of compo-
=
nent j in the liquid on
=L*(s + 1,t) -L*(s,t)
tray s
dP(s, t) Yj(s, t) = mole fraction of compo-
+ [1 - KVL(s, t)]B(s, t) dt nent j in the vapor on
(4-21) tray s
by a second-order-plus-deadtime model, 1 1 1
TO=--' T=-- T---
where the two dominant eigenvalues are 1 A1 ' 2 - A2'
used as the time constants and the rest of ""1
the eigenvalues are lumped as the dead-
time: d=
j=3
i:.(-~)-
Aj
E(-~)
""ji=2
(4-32)
(4-28) with Al » A2 > ... > An and ,.". > ""2 >
... >""m·
Celebi and Chimowitz (1985) provide a
with method for obtaining the rigorous transfer
function (Equation 4-30) for a distillation
1 1 column, but their method is useful only
T=-- T---
I A1 ' 2 - A' when considering the response of the col-
2
umn to changes in feed composition.
(4-29) In general, if the full linearized model of
a distillation column is available in analyti-
cal form in the Laplace domain, many or-
The transfer function given by Equation der-reduction methods can be used to pro-
4-27 has no numerator dynamics. Such duce an approximate model similar to the
transfer functions arise in staged systems one given by Equation 4-31. Surveys of these
when considering the response of a variable methods (i.e., moment matching methods,
at one end of a cascade to a disturbance cumulant matching methods, Routh approx-
introduced at the opposite end. In cases imations, impulse energy approximations,
where the transfer function of the column least squares approximations, etc.) are given
contains numerator dynamics (i.e., zeros), by Genesio and Milanese (1976) and by
the numerator and denominator dynamics Papadourakis, Doherty, and Douglas (1989).
of the column can be approximated sepa- However, the applicability of the preceding
rately (Kim and Friedly, 1974; Celebi and methods is limited because the full transfer
Approximate and Simplified Models 57
function model of a distillation column is the subscript f denotes the final steady state
seldom available in analytical form. reached by the column after a step change
The second effect of the realization that has been introduced and subscript initial
the nonlinear distillation dynamics are dom- refers to the initial steady state of the col-
inated by a time constant well separated umn.
from the rest is a large number of attempts The derivation of Te is based on two
to relate the dominant time constant to assumptions:
steady-state parameters. Many investigators
(Moczek, Otto, and Williams, 1965; Bhat Assumption I
and Williams, 1969; Wahl and Harriot, 1970; The flow dynamics are immediate (i.e., the
Toijala (Waller) and his co-workers, holdups and the flows change instantane-
1968-1978; Weigand, Jhawar, and Williams, ously).
1972; Skogestad and Morari, 1987b, 1988)
have tried to estimate this time constant, Assumption II
which appears to be equal to "the change in All the trays have the same dynamic respon-
column holdup of one component over the se, which corresponds to treating the col-
imbalance in supply of this component," umn as a large mixing tank.
that is,
•
Tc calculated by Equation 4·33 was found to
•
be 3.94 h (i.e., an error of only 8%). The
:I value of Tsc calculated by Equation 4·35 for
1 the same column was found to be 4.15 h
(Le., an error of only 3.5%).
• ., sa
4-5·3 linear Models In the Frequency
FIGURE 4·3. Binary distillation column used in Ex- Domain
ample 4-1.
For small variations about an equilibrium
point, Equation 4-18 can be linearized. Us-
1r----------------------------------------,
NORMALIZED LIQ MOL FRAC IN CONDENSER
0.8
0.8 Line 1
0.4
0.2
ing average values for B*(s, t), Bv(s, t), linearized equation for the pressure drop
Re(s, I), and the resistance value Rv'(s, I), across the tray:
and transforming to the frequency domain
gives SV(s - 1,jw)
SP(s - 1,jw) - SP(s,jw)
(4-45)
. jWTp,L ~P( .)
jWTp, V + . S, jW
1 + jWTe
(I
aL(l, jw)
SL*(m + 1,jw)
(4-43) (1 +jwTL)m
. 8V(m,jw) ]
+KVL [ SV(O,jw) - . m
where 8P(m, jw) = pressure variation (1 + }WTd
above the top tray
(Pa) +SP(jw)(1 - Kvd ( ~)
SPo = pressure variation
below the bottom tray
(Pa)
x [1 - __ 1_---::-]
(1 + jWT )
(4-48)
L n
J 1, J2 = roots of the charac-
teristic equation The tray parameter K VL has a strong im-
pact on the response of the bottom accumu-
lator level to the vapor flow. If it equals 1,
Tp,.L ])J+1=0 then the term KVL SV(O, jw) just cancels
1 + jWTe
the direct effect of the vapor flow variation,
(4-44) so the bottom level has to wait for the term
with SV(m, jw), which is delayed by n times
The transfer functions for the vapor flow the tray time constant Tv This makes bot-
responses can be found simply from the tom level control by vapor flow as slow as by
Approximate and Simplified Models 61
reflux flow. In practice this can force col- mass balance of the tray,
umn designers to increase the size of the
L[x(s + 1) - x(s)]
bottom holdup, and so make the column
more expensive. + V[y(s - 1) - y(s)] = 0 (4-54)
+jw[(e l - 1) 5x(s) + eo]B* 5P(s,jw) The most rigorous model for describing the
transient behavior of stagewise distillation
+ [xes + 1) - xes)] processes should include a large number of
x[ 5L(S: 1,jw) _ 5V(s ~1,jW)] variables as state variables (i.e., tray liquid
compositions, liquid enthalpies, liquid
holdups, vapor enthalpies, vapor holdups,
( 4-51) etc.). By employing certain assumptions,
which we will call simplifying assumptions,
where some of the state variables can be elimi-
nated, and the order of the model can be
V significantly reduced. Such simplifying as-
E = el- (4-52) sumptions are:
L
ML + elM v a. Negligible vapor holdup
7'
x
= L
(4-53)
b. Fast vapor enthalpy dynamics
c. Fast flow dynamics
and use has been made of the static partial d. Fast liquid enthalpy dynamics
62 Practical Distillation Control
Assumptions a and b are more or less stan- pseudocomponents with continuous ther-
dard in the literature of distillation dynam- modynamics for phase equilibrium calcula-
ics because interest is focused on composi- tions of "semicontinuous mixtures." Their
tion dynamics and control. Assumption c approach could be used for distillation
has been employed by many investigators modelling applications.
and is only reasonable if the measurement In general, the methods of continuous
lags in quality analysis are dominant (that is, thermodynamics are suited for mixtures
the flow dynamics are fast compared with containing very many components (i.e.,
the composition dynamics). In case the flow petroleum fractions, shale oils, etc.), but not
dynamics are significant, they can be treated for mixtures with a moderate number of
separately from the composition dynamics, components (i.e., less than 10). For the lat-
and then they can be superimposed on them ter cases, the lumping of some components
(see Section 4-4). Assumption d implies fast to pseudocomponents can be used to re-
enthalpy changes, turning the energy bal- duce the number of equations, but a lot
ances on each tray into a system of alge- remains to be done (i.e., development of
braic equations that must be satisfied at all lumping rules) before such an approach can
times (see Section 4-4). be successful.
Number of
Collocation Points Sj
Points n
M+ 1
1
2
M+1 ~
2 -2- ± V----u--1-2
-
3
M+1 M+1
-2-'-2-±
V 3M 2 - 7
20
The corresponding equations of the col- where the W functions are the Lagrange
location model for the module of Figure 4-5 polynomials
are
d j = 1, ... , n + 1
ML d/(Sj' t).= Li(Sj + 1, t) - Li(sj' t)
(4-58)
+ Vj(Sj - 1, t) - Vj(Sj' t)
(4-56)
j = 0, ... , n
for j = 1, ... , n, where the tilde denotes an Note again that So = 0 and sn+1 = M + 1.
approximate value. Therefore, the number The method is applicable for the solution
of equations describing the component ma- of both steady-state and dynamic models of
terial balances for the module is reduced multicomponent distillation columns with
from M (number of stages in the module)
to n (number of collocation points in the
module). The locations of the collocation
points S1"'" sn are the zeros of the Hahn Di.tillate product
polynomials (Hahn, 1949) with (a, b) =
(0,0). The zeros of the Hahn polynomials
for n ~ 4 are given in Table 4-3. Side product
Srivastava and Joseph (1985) have shown
that using Hahn polynomials with a = 0
and b > 0 leads to better approximations
Peed atream 1
where the error is more evenly distributed
throughout the column.
The compositions in each stage of the
module are given by Peed atream 2
n+l
xes, t) = E wh(s)x(Sj' t) Bouam. product
j=1
(4-57)
n
ji(s, t) = E Wjv(S)ji(Sj' t) nGURE 4-6. A complex distillation column with
j=O two feed streams and three product streams.
Approximate and Simplified Models 65
multiple feeds and/or product streams, and tive decomposition (scheme B) is shown in
it can accommodate columns with tray- Figure 4-7b.
dependent temperatures, liquid and vapor There is a trade-off between the two
flows, mixtures with nonideal thermodynam- decomposition schemes. Scheme A leads to
ics, trays with efficiencies less than 1, and so a smaller number of equations (Le., equa-
on. tions for modules plus equations for con-
One issue related to the implementation denser and reboiler), but the Jacobian of
of the method is associated with the way in the resulting model is a full matrix (Le., very
which a given column is divided into mod- few nonzero elements). In contrast, scheme
ules. Consider the column shown in Figure B requires more equations (Le., equations
4-6; it consists of two feed streams and for modules plus equations for condenser
three product streams. A modular decom- and reboiler plus equations for feed-prod-
position of this column is shown in Figure uct trays), but the corresponding Jacobian
4-7a (decomposition scheme A). An alterna- has a block diagonal form that facilitates
the solution of the system (Papadourakis,
1985). The superiority of scheme B, which
corresponds to treating all the feed-product
D trays as collocation points, was also recog-
nized (for different reasons) by Pinto and
Biscaia (1988). In addition, the modular de-
composition of a column according to
scheme B is easier to implement. The appli-
cation of the preceding method is illustrated
s in the following example.
D
Example 4-2
Consider again the binary column described
in Example 4-1. Applying the decomposi-
FI tion scheme B, the stripping section trays
S
(excluding the feed tray) constitute one
Modale 2 module, the rectifying section trays consti-
tute a second module, and the feed tray, the
FI condenser, and the reboiler are used as
collocation points by themselves. Applying
Module 3
F2 this decomposition scheme and the colloca-
F2 tion method of Stewart, Levien, and Morari
(1985), a reduced-order model of the col-
Modul.4 umn was developed.
The first equation is' the mass balance
around the reboiler:
B B
ScIIemeA ScbImeB d
(b)
MLBd/(Sl,t) = (L + F)i(Sl + 1,t)
<eI)
sponding to the reboiler and from Table 4-2 ing mass balance equations become
for n = 1 and M = 1, we get S 1 = 1.
d
Assuming that the dynamics of the strip- MLd/(S5,t)
ping section can be described by using only
two collocation points, the corresponding = Li(S5 + 1, t) - Li(S5' t)
mass balance equations become + V)i(S5 - 1, t) - Vy(S5' t) (4-63)
d
d M Ld/(S6,t)
M Ld/(S2' t)
= Li(S6 + 1, t) - Li(S6' t)
= (L + F)i(S2 + 1, t) - (L + F)i{S2' t) + VY(S6 - 1, t) - V)i{S6' t) ( 4-64)
where 54 is the collocation point for the The remaining variables in the preceding
feed tray. Again, from Table 4-2 we get (for equations are functions of the liquid and
n = 1 and M = 1) that S4 = 1. However, vapor mole fractions of the collocation
because there are 13 stages below the feed points as described by Equations 4-57 and
tray (reboiler + 12 trays of the stripping 4-58. For example, the variable i(S3 - 1, t)
section), we set S4 = 14. = i(9.16, t) is given by
Assuming that the dynamics of the recti- 4
fying section can also be described by using i(9.16, t) = L wh {9.16)i(Sj, t) (4-67)
only two collocation points, the correspond- j=2
Approximate and Simplified Models 67
where the w functions are the Lagrange the stage (collocation point) immediately
polynomials below the module.
By setting the derivatives of the coll-
ocation dynamic model equal to zero, an
j = 2, ... ,4 approximate steady-state model for the col-
umn was obtained. The steady-state compo-
sition profile of the column was calculated
(4-68)
by first solving the full model, and also by
solving the reduced-order model. The re-
Note that £(9.16, t) is the liquid mole frac- sults are shown in Figure 4-8. In the same
tion of a fictitious tray that is inside the figure, the results are shown from approxi-
stripping section (module). Therefore, and mate models developed by using 3 and 4
according to Equations 4-57 and 4-58, it will collocation points in each module (resulting
depend on the liquid mole fractions of the in 9 and 11 collocation points for the entire
collocation points of this module only, plus column). As can be seen in the figure, the
on the liquid mole fraction of the stage agreement between the full model and the
immediately above the module. Using the reduced-order model is excellent when 11
decomposition scheme shown on Figure 4- collocation points are used in the column,
7b, a tray immediately above or below a very good when 9 points are used, and poor
module is a feed or product stage and, when only 7 points are used for the entire
therefore, it constitutes the location of a column. It should be noted here that the
lone collocation point. solution that exhibits very good agreement
Note also that the vapor mole fraction of with the full solution (i.e., when 9 colloca-
a fictitious (or real) tray that belongs to a tion points were used for the entire column)
certain module depends only on the vapor results from the solution of 9 equations,
mole fractions of collocation points of the whereas the full model required the solu-
module and on the vapor mole fraction of tion of 25 equations.
0.8
0.6
0.4
.0.2
o~~~~~~--------------------------~
-0.2~~~~~~~~~~~~~~~~~~~~~
0123456789ronu~M~~rr~.~~HHM
0.888
0.88 t-U-.L.L.L..L.L.L."+I-...........r...u....I..I..j...L.U...L.UU-I..I"+L'U../..,u..L..L.4.L.L..1..L.L..I...&..I.J.+U..&..I.J...u.u
o 2.5 5 T.5 10 12.5
TIME (HOURS)
FIGURE 4-9. Transient response of the condenser of the distillation column of
Example 4-2 to a step decrease of the distillate flow rate by 1%. Une 1: full model;
line 2: reduced model with 9 collocation points; line 3: reduced model with 11
collocation points.
L
_I~Q~U~ID~M~O~L~E~F~R~AC~TI~O~N~IN~R~EB_O_I~LE_R________________-,
0.012.-
0.01 Line :3
0.004
0.002
O~ ...........r...u.~~~.L.L..I.+W~~~.L.L..I.LU~~.L.L.L..L.L..I.~~u"/"'~
o 2.5 5 T.5 10 12.5
TIME (HOURS)
FIGURE 4-10. Transient response of the reboHer of the distillation column of
Example 4-2 to a step decrease of the distillate flow rate by 1%. Une 1: full model;
line 2: reduced model with 9 collocation points; line 3: reduced model with 11
collocation points.
Approximate and Simplified Models 69
The method was also applied to obtain a control methods. Int. Chem. Eng. Symposium
reduced-order dynamic model for the col- Series 32,22-32.
umn. The predicted column response for a Bonvin, D. and Mellichamp, D. A. (1982a). A
1% reduction in the distillate flow rate is generalized structural dominance method for
shown in Figures 4-9 and 4-10. Again, the the analysis of large-scale systems. Int. J.
Control 35, 807-827.
agreement between the reduced-order mod-
Bonvin, D. and Mellichamp, D. A (1982b). A
els and the full model ranges from excellent unified derivation and critical review of modal
(model with 11 collocation points in the approaches to model reduction. Int. J. Con-
column) to very good (model with 9 colloca- trol 35, 829-848.
tion points in the column). Celebi, C. and Chimowitz, E. H. (1985). Analytic
The degree of order reduction achievable reduced order dynamic models for large equi-
for a given column (using orthogonal collo- librium staged cascades. AIChE J. 31,
cation) was dealt with in a paper by Srivas- 2039-2051.
tava and Joseph (1985). They developed an Chen, C. H. (1969). M.Sc. thesis, Polytechnic
index called order-reduction parameter that Institute of New York.
can be used to predict the extent of order Cho, Y. S. and Joseph, B. (1983a). Reduced-order
reduction achievable and to provide guide- steady-state and dynamic models for separa-
tion processes Part I. Development of the
lines for choosing an appropriately sized
model reduction procedure. AIChE J. 29,
reduced-order model. 261-269.
In general, reduced-order models ob- Cho, Y. S. and Joseph, B. (1983b). Reduced-
tained from the application of the previ- order steady-state and dynamic models for
ously mentioned orthogonal collocation separation processes Part II. Application to
methods can be used for both steady-state nonlinear multicomponent systems. AIChE J.
and dynamic simulation purposes with ex- 29,270-276.
cellent results. Use of these approximate Cho, Y. S. and Joseph, B. (1984). Reduced-order
models can lead to reductions in computa- models for separation columns-III. Applica-
tional times proportional to (Nc IN)2 where tion to columns with multiple feeds and
Nc is the total number of collocation points sidestreams. Comput. Chern. Eng. 8, 81-90.
in the column and N is the total number of Cotterman, R L. and Prausnitz, J. M. (1985).
equilibrium stages in the column (Pinto and Flash calculations for continuous or semicon-
tinuous mixtures using an equation of state.
Biscaia, 1988). Ind. Eng. Chem. Proc. Des. Dev. 24,434-443.
Cotterman, R L., Bender, R, and Prausnitz,
References J. M. (1985). Phase equilibria for mixtures
Arkun, Y. (1987). Dynamic block relative gain containing very many components. Develop-
array and its connection with the perfor- ment and application of continuous thermo-
mance and stability of decentralized control dynamics for chemical process design. Ind.
structures. Int. J. Control 46, 1187-1193. Eng. Chem. Proc. Des. Dev. 24, 194-203.
Bamberger, W. and Isermann, R (1978). Adap- Cotterman, R L., Chou, G. F., and Prausnitz,
tive steady-state optimization of slow dynamic J. M. (1986). Comments on "Flash calcula-
processes. Automatica 14,223-230. tions for continuous or semicontinuous mix-
Benalou, A, Seborg, D. E., and Mellichamp, tures using an equation of state." Ind. Eng.
D. A (1982). Low order, physically lumped Chem. Proc. Des. Dev. 25, 840-841.
dynamic models for distillation column con- Genesio, R M. and Milanese, M. (1976). A note
trol. Presented at the AIChE Annual Meet- on the derivation and use of reduced-order
ing, November 1982, Los Angeles, CA. models. IEEE Trans. Automat. Control AC-
Betlem, D. H. L. (1991). Personal communica- 21, 118-122.
tion. Georgakis, C. and Stoever, M. A. (1982). Time
Bhat, P. V. and Williams, T. J. (1969). Approxi- domain order reduction of tridiagonal dynam-
mate dynamic models of distillation opera- ics of staged processes-I. Uniform lumping.
tions and their relation to overall column Chem. Eng. Sci. 37, 687-697.
70 Practical Distillation Control
Hahn, W. (1949). Uber Orthogonalpolynome die Shibata, S. K., Sandler, S. I., and Behrens, R. A
q-Differenzengleichungen genugen. Math. (1987). Phase equilibrium calculations for
Nachrichten 2,4-34. continuous and semicontinuous mixtures.
Hawkins, D. E., Tolfo, F., and Chauvin, L. (1987). Chem. Eng. Sci. 42, 1977-1988.
Group methods for advanced column control. Shinskey, F. G. (1967). Process Control Systems.
In Handbook of Advanced Process Control New York: McGraw-Hill.
Systems and Instrumentation, L. Kane, ed. Skogestad, S. and Morari, M. (1987a). Effect of
Houston: Gulf Publishing Co. disturbance directions on closed loop perfor-
Kehlen, H. and Ratzsch, M. T. (1987). Complex mance. Ind. Eng. Chem. Res. 26, 2029-2035.
multicomponent distillation calculations by Skogestad, S. and Morari, M. (1987b). The domi-
continuous thermodynamics. Chem. Eng. Sci. nant time constant for distillation columns.
42,221-232. Comput. Chem. Eng. 11,607-617.
Kim, C. and Friedly, C. F. (1974). Dynamic mod- Skogestad, S. and Morari, M. (1988). Under-
eling of large staged systems. Ind. Eng. Chem. standing the dynamic behavior of distillation
Proc. Des. Dev. 13, 177-181. columns. Ind. Eng. Chem. Res. 27,1848-1862.
Levy, R. E., Foss, A S., and Grens, E. A, II
Srivastava, R. K. and Joseph, B. (1985). Re-
(1969). Response modes of a binary distilla-
duced-order models for separation
tion column. Ind. Eng. Chem. Fundam. 8,
columns-V. Selection of collocation points.
765-776.
Comput. Chem. Eng. 9, 601-613.
Moczek, J. S., OUo, R. E., and Williams, T. J.
Srivastava, W. E. and Joseph, B. (1987a). Re-
(1963). Proceedings of the Second IFAC
duced-order models for staged separation
Congress, Basle. New York: Pergamon.
columns-IV. Treatment of columns with
Moczek, J. S., OUo, R. E., and Williams, T. J.
mUltiple feeds and sidestreams via spline fit-
(1965). Approximation models for the dy-
ting. Comput. Chem. Eng. 11, 159-164.
namic response of large distillation columns.
Srivastava, W. E. and Joseph, B. (1987b). Re-
Chem. Eng. Prog. Symp. Ser. 61, 136-146.
duced-order models for staged separation
Papadourakis, A (1985). Stability and Dynamic
columns-VI. Columns with steep and fiat
Performance of Plants with Recycle. Ph.D. Dis-
composition profiles. Comput. Chem. Eng. 11,
sertation, University of Massachusetts,
165-176.
Amherst, MA.
Papadourakis, A., Doherty, M. F., and Douglas, Stewart, W. E., Levien, K. L., and Morari, M.
J. M. (1989). Approximate dynamic models (1985). Simulation of fractionation byorthog-
for chemical process systems. Ind. Eng. Chem. onal collocation. Chem. Eng. Sci. 40, 409-421.
Res. 28, 546-552. Stoever, M. A and Georgakis, C. (1982). Time
Perkins, J. D. (1990). Interactions between pro- domain order reduction of tridiagonal dynam-
cess design and process control. In Dynamics ics of staged processes-II. Nonuniform
and Control of Chemical Reactors, Distillation lumping. Chem. Eng. Sci. 37, 699-705.
Columns and Batch Processes. IFAC Sympo- Toijala, K. V. (1968). An approximate model for
sium Series 1990, No.7, J. E. Rijnsdorp et aI., the dynamic behavior of distillation columns.
eds. Oxford: Pergamon, 195-203. Part I. General theory. Acta Acad. Aboensis
Pinto, J. D. and Biscaia, E. c., Jr. (1988). Order (Math. Phys.) 28(8), 1-18.
reduction strategies for models of staged sep- Toijala, K. V. (1969). An approximate model for
aration systems. Comput. Chem. Eng. 12, the dynamic behavior of distillation columns.
812-831. Part II. Composition responses to changes in
Rademaker, 0., Rijnsdorp, J. E., and Maar- feed composition for binary mixtures. Acta
leveld, A (1975). Dynamics and Control of Acad. Aboensis (Math. Phys.) 29(10), 1-34.
Continuous Distillation Units. Amsterdam: El- Toijala, K. V. (1971a). An approximate model
sevier Science Publishers. for the dynamic behavior of distillation
Rijnsdorp, J. E. (1991). Integrated Process Con- columns. Part IV. Comparison of theory with
trol and Automation. Amsterdam: Elsevier experimental results for changes in feed com-
Science Publishers. position. Acta Acad. Aboensis (Math. Phys.)
Robinson, C. S. and Gilliland, E. R. (1950). 31(3), 1-25.
Elements of Fractional Distillation. New York: Toijala, K. V. (1971b). An approximate model
McGraw-Hill. for the dynamic behavior of distillation
Approximate and Simplified Models 71
columns. Part V. Comparison of theory with Toijala, K. V. and Jonasson, D. (1970). An ap-
experimental results for changes in reflux flow proximate model for the dynamic behavior of
rate. ActaAcad. Aboensis (Math. Phys.) 31(5), distillation columns. Part III. Composition re-
1-19. sponse to changes in reflux flow rate, vapour
Toijala, K. V. (1978). Simple models for distilla- flow rate and feed enthalpy for binary mix-
tion dynamics. Paper presented at the 86th tures. Acta Acad. Aboensis (Math. Phys.)
National AIChE Meeting, April 1-5, 1978, 30(15), 1-21.
Houston, Texas. Toijala, K. V. and Jonasson, D. (1971). An ap-
Toijala, K. V. and Fagervik, K. (1972). An ap- proximate model for the dynamic behavior of
proximate model for the dynamic behavior of distillation columns. Part VI. Composition re-
distillation columns. Part VII. Comparison of sponse for changes in feed flow rate for bi-
model with experimental results for con- nary mixtures. Acta Acad. Aboensis (Math.
trolled columns by the use of digital simula- Phys.) 31(11), 1-27.
tion. Acta Acad. Aboensis (Math. Phys.) 32(1), Wahl, E. F. and Harriot, P. (1970). Understand-
1-17. ing and prediction of the dynamic behavior of
Toijala, K. V. and Gustafsson, S. (1971). Process distillation columns. Ind. Eng. Chem. Proc.
dynamics of multi-component distillation. Des. Dev. 9,396.
Kemian Teollisuus 28, 113-120. Weigand, W. A., Jhawar, A. K., and Williams,
Toijala, K. V. and Gustafsson, S. (1972a). Pro- T. 1. (1972). Calculation method for the re-
cess dynamics of multi-component distillation. sponse time to step inputs for approximate
Comparison of theoretical model with pub- dynamic models of distillation columns.
lished experimental results. Kemian Teollisuus AIChE 1. 18, 1243-1252.
29,95-103. Wong, K. T. and Luus, R. (1980). Model reduc-
Toijala, K. V. and Gustafsson, S. (1972b). On the tion of high-order multistage systems by the
general characteristics of component distilla- method of orthogonal collocation. Can. 1.
tion dynamics. Kemian Teollisuus 29, 173-184. Chem. Eng. 58, 382.
5
Object-Oriented Simulation
B. D. Tyreus
E. I. Du Pont de Nemours & Co.
72
Object-Oriented Simulation 73
namic model is written from first principles ered. All this is put in the context of
and implemented in FORTRAN or BASIC creating dynamic simulation models for dis-
(see Chapter 3 for example). The models tillation systems.
are almost always tailored to the specific This chapter is, however, not detailed
problem at hand; seldom extensible to other enough to show you how to write your own
problems and almost never usable by oth- object-oriented dynamic simulator, nor does
ers. it give enough insight into how I have writ-
The tools we know today for creating ten mine. Instead I merely hope to remove
dynamic simulations are consistent with the some of the mystique around object-ori-
prevailing paradigm. Whether they are inte- ented programming and to spark enough
grators like LEANS, libraries of FOR- interest that you might consider experiment-
TRAN subroutines like DYFLO (Franks, ing with an object-oriented compiler next
1972), or equation-based numerical solvers time you write a simulation model. In case
like SPEEDUP (Perkins, 1986), these tools you do not like writing your own simulations
are aimed at the specialist who has consid- from scratch but rely on vendor tools, this
erable experience in using the tool, who chapter will hopefully interest you a little
knows how to model various processes in more in emerging simulation packages built
terms of their fundamental equations, and on object-oriented principles.
who is willing to spend considerable time in
entering code and data into input files,
which are compiled, edited, and debugged
before they yield results in terms of time 5·2 IDEAL DYNAMIC SIMULATOR
plots of pre specified variables over fixed Before delving into the principles of OOP,
time periods. we will take a closer look at a vision of a
I feel that in order to meet the broader suitable tool for creating dynamic simula-
need for dynamic simulations, a paradigm tions.
shift is required. Under a different para- In order to shift to a different paradigm,
digm, dynamic simulations would go beyond one has to ask the question: "What is it that
control system studies, and all process engi- we cannot do now that if it could be done
neers should be able to create and use such would have a significant impact on our busi-
simulations. I also feel that today's tools ness?" When I asked this question about
and techniques for generating dynamic sim- dynamic models and the simulation tools we
ulations will be inadequate under a differ- use, I found the following to be true:
ent paradigm. Instead, I suggest that new
software approaches, such as object-ori- • Most process engineers cannot create or
ented programming (OOP), will replace use a dynamic simulation because the tools
current tools in order to support a new and techniques are designed for special-
paradigm (Stephanopoulos, Henning, and ists.
Leone, 1990). I am also aware of a couple of • Dynamic simulators are not user-friendly,
simulation software companies that have preventing the casual user from consider-
reached the same conclusion. ing them.
The purpose of this chapter is thus to • Dynamic simulations are expensive to de-
familiarize you with an emerging software velop because it is generally difficult to
technology for dynamic simulations. I will reuse previously written code in new ap-
describe the concepts of objects and mes- plications, thereby forcing the developer
sages and contrast them to their more fa- to rewrite substantial sections of the new
miliar counterparts: data structures and model.
subroutines. The important features of in- • Dynamic simulations are not readily ac-
heritance and polymorphism are also cov- cessible because the development envi-
74 Practical Distillation Control
ronments and final models require rela- programming approaches supported by pro-
tively large, expensive computers. cedural languages, such as BASIC, FOR-
• Many simulators are not configurable and TRAN, Pascal, and C. In procedural pro-
require extensive programming and al- gramming, the focal point is data. Data are
most always compilation before the model transformed to results by letting procedures
can be run. or subroutines operate on them. For exam-
• Dynamic simulations are often slow. ple, in a simulation package, such as
• The user cannot interact with the simula- DYFLO (Franks, 1972), there are data
tion during a run. structures defined that specify how compo-
• Even for the sophisticated user, most dy- nent compositions, temperature, and pres-
namic simulation packages are limited be- sure are stored, how parameters enter the
cause they are not portable, extensible, or simulation, and how the results of chemical
maintainable. reactions are made available when they are
needed. DYFW then consists of a library
If instead we had simulation tools that were of subroutines that either take the data
configurable, user-friendly, flexible, port- structures as parameters or use the data
able, readily accessible, fast, and inexpen- directly from COMMON areas to produce
sive, we could dramatically change the way results. Similarly, when writing a Pascal pro-
we consider dynamics in process design and gram, the first thing that has to be done is
control. Such an ideal dynamic simulator to define the data types and variables that
would allow process engineers and control will be used by the procedures and func-
engineers to routinely develop rigorous, tions that get invoked in a predefined, pro-
nonlinear, dynamic simulations for most cedural fashion.
processes. It would allow us to study the If in procedural programming the key
impact of process design and controller words are data and subroutines, then in
structure on the operation of not just dif- object-oriented programming the key words
ferent unit operations, but entire processing are objects and messages (Pascoe, 1986).
units and even plants (see also Chapter 6). Although there are some similarities be-
The ideal dynamic simulator would allow tween data and objects on the one hand and
process engineers to explore and under- subroutines and messages on the other,
stand existing processes so the processes there are several significant differences. The
could be simplified and improved. It would most significant difference is between data
allow us to quickly validate our control sys- and objects. Data structures are passive
tems and to train operators with realistic, storage places for information that are
validated models. In short, the ideal dy- transformed by subroutine calls in the pro-
namic simulator would change the way we gram. Objects, on the other hand, are active
do business; it would help create a paradigm entities that exhibit behavior when they re-
shift. ceive a message. Objects are data structures
Once the vision of an ideal dynamic sim- and subroutines packaged into one unit (en-
ulator has been created, we need to know capsulations) where the data are globally
how such a tool could be developed. known to all the subroutines that belong to
Object-oriented programming provides an that object. A message is then the mecha-
excellent vehicle to take us to our goal. nism to trigger the appropriate subroutine
within the object. Because objects send
5-3 OBJECT-ORIENTED messages to each other and because the
particular object to receive a particular mes-
PROGRAMMING
sage can change during execution of the
To help explain the concepts of object-ori- program, there is no counterpart to the
ented programming, I will contrast this sequence of subroutine calls found in proce-
new software technique with traditional dural programs. Here lies part of the diffi-
Object-Oriented Simulation 75
DO 20 N=2,NF-1
20 DM(N) = UN + 1) + V(N- 1)- UN)- yeN)
C•••••••• DM FOR FEED TRAY AND REST OF COLUMN HERE
DO 100 J=1,NC
DO SO N=2,NF-1
SO DXM(N,J) = X(N + 1,J)*UN + 1) + Y(N-1,J)*V(N-1)- X(N,J)*UN)
-Y(N,J)*V(N)
c •••••••• DXM FOR FEED TRAY AND REST OF COLUMN HERE
100 CONTINUE
c •••••••• INTEGRATE A LA EULER
DO 200 N = 1,NT
200 M(N) = M(N) + DM(N)*DT
DO 400 J = 1,NC
DO 300 1,NT
XM(N,J) = XM(N,J) + DXM(N,J)*DT
X(N,J) = XM(N,J) I M(N,J)
IF (X(N,J) .GT. 1.0)X(N,J) = 1.0
IF (X(N,J) .LT. O.O)X(N,J) =0.0
300 CONTI NUE
400 CONTINUE
RETURN
END
c************************* VLE MODULE *************************
SUBROUTINE BUBPT(T,X,Y,P)
COMMON NC
COMMON ALPHA
DIMENSION ALPHA(S),X(S),Y(S)
SUM=O.O
D01K=1,NC
SUM = SUM + ALPHA(K)*X(K)
D02J=1,NC
Y(d) = ALPHA(d)*X(d) I SUM
RETURN
END
ys, alphas, Ls, and Ms. Typically, these cases, but how do we keep track of the
variables would be stored in single or multi- cases? If we decide to always use relative
dimensional arrays. volatilities (Equation 5-4) in the body of
As convenient as it might be to directly BUBPT, then that means that all simula-
code the equations into a program, we would tions using the tray section module will use
introduce a serious limitation in our code by relative volatilities for equilibrium. That is
mixing the tray equations with the details of not what we had in mind. We could of
the equilibrium calculations. If we later want course implement a different equilibrium
to change the way we calculate equilibrium, method in BUBPT, but even then, all simu-
we would have to go right to the heart of lations using the tray module would be
the program and make changes. Such edit- forced to use the same equilibrium calcula-
ing is always dangerous because there is a tion. For dynamic simulations, this is most
risk of breaking or invalidating a perfectly undesirable because different columns have
fine program. We could, of course, make a different modelling requirements. For ex-
copy of the program and keep the original ample, columns separating components with
as version} and edit the copy and call it low relative volatilities usually require many
version2. Such proliferation of versions has trays (80 to 120). These systems are almost
its own problems because it quickly leads to always ideal, meaning that relative volatili-
an unmanageable set of programs that not ties adequately describe their equilibrium.
even the owner can keep track of. A better Using Equation 5-4 in dynamic simulations
approach is to separate the programs into of such columns is very desirable because
different modules by use of subroutines. the calculation of equilibrium is explicit and
Equations 5-1 through 5-3, which describe fast. However, not all systems are ideal and
the composition dynamics on the tray, are they should therefore not be modelled by
not going to change much and should be relative volatilities. Wide boiling fluid sys-
kept in their own module. The equilibrium tems are best described by vapor pressures,
calculation (Equation 5-4), which is likely to possibly in combination with activity coef-
change more, should be in a different place. ficients. The bubble point calculation for
In order to communicate between the two liquids modelled by vapor pressures and
modules, we use a subroutine call such as: liquid activity coefficients is extremely time
CALL BUBPT (T,X,Y,P) in place of Equa- consuming and can easily make dynamic
tion 5-4. Strictly speaking, we only need X simulations of columns with many trays im-
and Y as arguments when we use relative practical. Fortunately, the separation of
volatilities, but it might be a good idea to wide boiling systems requires only a few
provide temperature (T) and pressure (P) in trays, making these columns as feasible to
case we later plan to perform bubble point simulate as the low relative volatility
calculations based on vapor pressures. columns provided we use the appropriate
The procedural implementation of the bubble point calculation in each case. How-
tray model now seems to be in good shape. ever, we just finished saying that all our
We have one module that describes the tray simulations will use the same version of
dynamics and another module that de- BUBPT, which implies that some of our
scribes the equilibrium (see Figure 5-2). We applications using the tray module will be
have eliminated the need to make changes incorrect or hopelessly slow depending on
in the tray section code and we have re- which method we use in BUBPT. To resolve
moved the multiple version syndrome. Or this dilemma, we could of course create a
have we? There are still some loose ends. In couple of different tray modules that make
return for removing Equation 5-4 from the calls to different BUBPT routines (BUBPTl,
tray section code, we had to permanently BUBPT2, etc.), but now we are back where
introduce the call to BUBPT (T,X,Y,P). We we started in version management and pro-
made the call general enough to cover many liferation of similar looking code.
78 Practical Distillation Control
Problems like these, which only magnify formally passed as parameters to the rou-
as we go beyond a single tray and into more tines. It is as if the data were present in
complex models, have prevented develop- invisible COMMON blocks in each of the
ment of the ideal dynamic simulator. It is message routines.
virtually impossible to come to grips with Next we specify a class for the tray itself.
these problems when using a procedural It might look like this:
approach. On the other hand, object-ori-
ented programming elegantly removes these class Tray
{
obstacles. To see how this is possible I need
to describe how a tray is modelled in an update(dt);
object-oriented style. Fluid* liquidIn;
Fluid* liquidOut;
Object-Oriented Approach Fluid* vaporIn;
In the object-oriented approach we start by Fluid* vaporOut;
creating two different classes. One class protected:
would model the tray itself and its behavior,
whereas another class would model the Fluid* holdup;
properties and behavior of the fluid system / /more data
on and around the tray. For example the };
fluid class might look like this: The Tray class would typically have several
messages that tray objects react to, but only
class Fluid one of these messages, update, is shown
{ here. In the protected data section, tray is
referring to a fluid object called holdup. In
the C+ + nomenclature F1uid* designates
bubblePointO; the variable type and holdup the variable
getXG); name. Our object-oriented model closely
getYG); mimics the real world; a tray is a physical
molesG); entity that holds a fluid phase. A tray also
getNumberOfComponentsO; knows about its environment such as the
liquid coming in from the tray above and
protected: the vapor leaving the tray. These are all
fluid objects that the tray object can refer
/ /data area to.
}; Now let us see what the update(dt) mes-
sage might cause a tray object to do.
where the messages bubblePoint(), getX{j) , update(dt)
getY{j), moles{j), and getNumberO/Compo- (
nents() would invoke specific behavior from holdup -+ bubblePointO;
all objects instantiated from class Fluid. The
nomenclature I am using to describe the
classes is similar to that of the object-ori- n = holdup -+ getNumberOfComponentsO;
ented programming language C+ +. The for all n do: (5-5)
data area where compositions, component holdup -+ moles(j) = holdup -+ moles(j) +
quantities, temperature, pressure, and phys- OiquidIn -+ moles(j) + vaporIn -+ moles(j) -
ical property parameters are stored is pro- holdup -+ getX(j)* L -
tected from outside use or inspection; only holdup -+ getY(j)*V)*dt;
the message routines have complete access
to these data even though the data are not
Object-Oriented Simulation 79
The message update(dt) takes one external X and Y as we need them. With data ab-
parameter dt. The tray object receiving the straction such a change does not impact
update message responds by first sending users of the Fluid class (e.g., tray objects).
the message bubblePoint() to its holdup ob-
ject [holdup -+ bubblePointO means "send
5-3-3 Inheritance and Polymorphism
the message bubblePoint to the object iden-
tified as holdup"]. The bubblePoint message We are now ready to tackle the problem of
requests holdup to calculate its bubble different equilibrium subroutines for dif-
point. Then update calculates the liquid (L) ferent column applications. First, assume
and vapor (V) flow rates leaving the tray that fluid objects calculate their bubble
(code not shown here). Next, update asks points from relative volatilities (Equation
holdup how many components it has and in 5-4). Furthermore, we have built a general
return informs holdup how each of these purpose distillation column from the tray
components will change as a result of inte- class. Now, we consider an application
grating the component balance. The compo- where vapor pressures and activity coeffi-
nent balance is derived from Equation 5-2 cients should be used to correctly calculate
integrated with a simple Euler integration the bubble point. This is where we got stuck
step dt. before with procedural programming. Ob-
ject-oriented programming allows us to pro-
Comparison of Approaches ceed. First we use the object-oriented con-
Equation 5-5 implements the object-ori- cept of inheritance to derive a new fluid
ented form for tray component dynamics, class from Fluid. Call the new fluid class
whereas Figure 5-2 shows the procedural VaporPressureFluid:
version. There are several important dif-
ferences between these implementations. In class VaporPressureFluid : public Fluid
Figure 5-2, we are explicitly declaring and {
referring to the mole fractions X and Y that bubblePointO;
are stored as part of the tray module. In };
Equation 5-5, the composition data are not
stored as part of the tray object but instead Creating this new class was simple enough:
as part of each fluid object (e.g., holdup). All we had to do was to give the class a
However, even in the fluid object, the data distinct name and mention from what other
are not explicit because we are not directly class it should inherit properties and behav-
referring to the mole fractions in holdup ior. A VaporPressureFluid object will thus
(e.g., holdup.x[j)). Instead, we are referring have all the properties and behavior of a
to the mole fractions in an abstract way by Fluid object. They will only differ in one
sending a message to holdup asking it to get aspect; they will calculate their bubble points
the appropriate mole fraction for us [e.g., using different methods. A Fluid object still
holdup -+ getXG)]. This concept is called calculates bubble points from relative
data abstraction and is another key idea in volatilities, whereas the VaporPressureFluid
object-oriented programming. Data abstrac- will use vapor pressures and activity coeffi-
tion allows us to change the data structures cients to complete its bubble point cal-
in classes without affecting any code that culation. The interesting part is that in both
uses objects instantiated from those classes. cases, the message is the same: bubble-
For example, initially we might have chosen Point(). We can thus keep our tray class and
to store the mole fractions X and Y as its update message intact because update
arrays in the fluid class. Later we can change was sending that very same message bubble-
our mind and decide not to store the mole Point() to its holdup. We simply have to
fractions explicitly but to store the number provide tray objects with Fluid objects when
of moles of each component and calculate we want to continue calculating bubble
80 Practical Distillation Control
points with relative volatilities or provide stantiated from a derived class (e.g., Vapor-
the tray objects with VaporPressureFluid PressureFluid) can replace an object from a
objects if we want the bubble point to be more basic class (e.g., Fluid) without telling
calculated using vapor pressures. This state- the user class (Tray) about the substitution.
ment might seem puzzling at first Tray objects send the messages as before
because the Tray class is declared with assuming that they are referring to Fluid
references to Fluid and not to Vapor- objects when in reality all those objects re-
PressureFluid. How then can we arbitrarily ceiving the messages are VaporPressure-
substitute a Fluid object with a VaporPres- Fluid objects and as such respond in their
sureFluid object? This is where the most own way to the same messages. Not only
powerful concept of object-oriented pro- can we perform this powerful substitution
gramming enters, namely, polymorphism. of base classes with derived classes when we
Polymorphism means many shapes and link our modules together, but we can also
refers to the principle that any object in- make the substitution while the program is
_ - - - -.... lIamer-l
so
ColumnS
- - - - -.... lIamer-2
Update(dl)
Tray object
executing. The run time assignment of ob- simulation systems that are very flexible.
jects from derived classes is called dynamic Before doing this, it is necessary to review
binding and is probably the most intriguing the principles of model structuring (Astom
aspect of object-oriented programming and Kreutzer, 1986; Mattsson, 1988).
(Cornish, 1987).
Figure 5-3 summarizes the concepts of
5-4-1 Structured Models
inheritance, polymorphism, and dynamic
binding as applied to the fluid objects The first principle of model structuring is
around the trays in a distillation column. that of hierarchical submodel decomposition,
Two different columns are part of a simula- which suggests that a complex system should
tion separating several components. Both be built from a set of small, self-contained
columns are instantiated from a generic units. The small units should in turn be
Column class built from the Tray class by built from yet smaller units until there is a
connecting Tray objects together. When a smallest logical unit that corresponds to
Column object is instantiated from a Col- some physical entity in the real world. In
umn class we specify how many stages it willthe other direction, a complex unit can be
have. This can be done by the user during part of an even more complex system. A
execution of the program and does not re- distillation column is an excellent example
quire separate compilation. Column A is of the potential use of hierarchical submod-
removing light ends from a mixture of mostlyels because a complete column with trays
isomers and requires only 10 stages. Col- and auxiliaries can be considered a complex
umn A should be simulated with a rigorous system. According to the hierarchical sub-
bubble point calculation involving vapor model principle, the column should be built
pressures and activity coefficients. Column from small pieces, in this case a tray section
B on the other hand separates the isomers module, a reboiler module, and a condenser
and needs 80 stages to do the job. For module. Furthermore, the tray section mod-
column B, relative volatilities are adequateule should be a collection of individual tray
to describe the vapor-liquid equilibrium. modules, which in turn contain different
The two columns can be simulated together fluid phase modules. In a different hierar-
using the same Column class with different chy, a column can be part of a larger, more
complex system, say a refining train or a
fluid objects on the trays. The different fluid
objects can be specified by the user, at runcomplete plant.
time, when the user requests two different The second modelling principle we need
instances of the same Column class. to consider is that of parametrization
(Mattsson, 1988; Nilsson, 1989). In design of
the modules for a hierarchical subsystem, it
5-4 DISTILLATION COLUMN is important that these modules be flexible
SIMULATION WITH and modifiable. If we can modify a module's
behavior based on parameters we give it, we
OBJECT-ORIENTED
are using parametrization. A simple exam-
PROGRAMMING
ple is the number of trays in a distillation
I have demonstrated how inheritance, poly- column tray section. A generic tray section
morphism, and dynamic binding help build of say 20 trays might be useful for many
a generic distillation column tray section simulations but not nearly as useful as a
that can be used with many different fluid generic tray section with a user-specified
systems and physical property calculation number of trays. The substitution of differ-
methods. I will now describe how these same ent fluid objects in the tray class can be
concepts and the notions of encapsulation viewed as another example of parametriza-
and data abstraction help us build complete tion (Nilsson, 1989). Here the tray is the
82 Practical Distillation Control
flexible module using the fluid system as a digm. So why do I advocate object-oriented
parameter to modify its behavior. programming as the preferred way of imple-
The concept of terminals is the third menting dynamic simulation systems? For
modelling principle we should consider the simple reason that object-oriented pro-
(Mattsson, 1989). A terminal on a software gramming makes it extremely easy to imple-
module is analogous to the antenna termi- ment and fully exploit model structuring
nal on a regular television set or the serial principles.
port on a personal computer. These termi- For example, there is a very natural map-
nals provide agreed upon formats for com- ping between hierarchical submodels and
municating information to or from the unit. classes. I have already indicated how the
The serial port on a personal computer Tray class was used to build a tray section.
might be the best example of a terminal. Similarly, Condenser and Reboiler classes
From the computer's point of view, we need can be used with the tray section to make a
not know what device is at the other end of complete column. The object-oriented prin-
the cable attached to the serial port. We ciples of encapsulation and data hiding en-
only need to know how to read from and. courage making these units self-contained
write to the serial port itself. Examples of and autonomous. Their only communication
terminals on a distillation column module with the outside world is through terminals
would be the feed nozzle, distillate nozzle, and connections that can be made very ver-
and bottoms nozzle. If these terminals have satile through data abstraction and poly-
the characteristics of a fluid, we can model morphism.
and simulate a column without knowing In this fashion I have used object-ori-
what is attached to the other ends of the ented programming to build an extremely
pipes connected to those nozzles. flexible dynamic simulator for distillation
The last modelling concept we will con- columns. One generic class represents the
sider is that of connections (Mattsson, 1989). tray section, one class represents a total
Connections are modules that connect units condenser, and one class represents a ther-
together via their terminals. For a distilla- mosiphon reboiler. These three classes are
tion column, the connections correspond to then connected to form a generic Column
the pipes that connect various submodules class. From the Condenser class, several
together. For example, the reflux pipe on a other types of condensers, such as partial
real column would be represented by a con- condensers, condensers handling inerts, and
nection module in a model of the column. spray condensers are derived. Because all
the derived condensers can take the place
of a total condenser by way of polymor-
5-4-2 Structured Models and
phism, different types of overhead systems
Object-Oriented Programming
can be realized with the generic column
Hierarchical submodules, parametrization, without ever having to modify or add to the
terminals, and connections are concepts that code describing the column. This way I did
have been used successfully in many process not have to concern myself up front about
simulation systems. They do not imply nor all the possible configurations for a column
are they synonymous with object-oriented because if I want to modify any part of the
programming. For example, both DYFLO column, I simply derive a new class or classes
and SPEEDUP have subunits that can be from the basic ones, slip them into the
connected together with terminals and generic column class, and the new behavior
streams. Parametrization is also used in appears almost as magic. There is never a
these programs to modify structure and be- risk of modifying and possibly breaking ex-
havior. However, neither DYFLO nor isting proven code or subunits. Because I
SPEEDUP uses an object-oriented para- use the same principle to model the fluid
Object-Oriented Simulation 83
streams in the column, I can use the generic surgery. This forced me to start from scratch
column to process literally every possible many times. On the other hand, as I devel-
fluid system with any physical property oped a new simulation, I always felt that
methods I care to consider. This even allows there were more than 50% common code
me to simulate reactive distillation systems and concepts from simulation to simulation.
with one and the same generic column By using model structuring principles and
model. After all, chemical reactions on the object-oriented programming, I have made
trays of a column are not a function of what it possible to reuse substantial amounts of
column we use but merely depend on the proven code. For example, when it used to
properties and the state of the fluid we take me a couple of months to write, debug,
process in the column. and run a dynamic model of a distillation
Flexible dynamic simulators do not come column, I can today put together and evalu-
without a price. The major disadvantage in ate the same simulation in a few days. Fur-
working with object-based systems is the thermore, I hand over an interactive model
lack of standard design principles for creat-to the plant personnel so they can make
ing classes. This is bothersome because the their own assessment of what needs to be
whole concept of object-oriented program- done in the process. The second reason I
ming critically hinges on the existence of a started to look at object-oriented program-
few powerful base classes such as the Fluid ming was the desire to have an interactive
class and the Tray class described earlier. simulator that engineers and operators could
Without these well thought out base classes, use to learn the principles of process dy-
object-oriented programming can turn into namics and process control. The object-ori-
a jungle of many small, specialized classes ented features I have described make it
that are hard to use and keep track of. It relatively easy to write programs that simu-
takes considerable time and effort, includ- late different unit operations or even a
ing many design iterations, to come up with whole process. After all, the concept of
a good set of base classes. However, once object-oriented programming grew out of
this front end loading effort is completed, the language SIMULA, which had been
the downstream possibilities are virtually specifically developed in the 1960s to deal
endless. with large scale dynamic simulations
(Magnusson,1990).
In relation to languages, it might be of
5-5 EXPERIENCE IN USING interest to mention which ones I have tried
and what I believe works. There are many
OBJECT-ORIENTED SIMULATION
different high-level languages that can be
FOR DISTILLATION
classified as object-oriented. Smalltalk is
The vision of the ideal dynamic simulator considered the purest object-oriented lan-
was created independently of any perceived guage. In Smalltalk everything is an object
relation to object-oriented programming. So and all computing is done by message pass-
when I originally started to explore object- ing. Because Smalltalk is an interactive lan-
oriented programming for dynamic simula- guage, there is no distinction between the
tions, I did it for very specific and less development environment and the run time
visionary reasons. The first reason was to simulation. New classes can therefore be
provide a better framework for reusing the defined at run time. This makes Smalltalk
code in the many custom simulations I had ideal for model building because hierarchi-
developed over time. Often new simulation cal models can be built and tested without
models were sufficiently different from pre- having to explicitly compile and link the
vious ones such that the old code could not program. Although run time interaction is
be used without substantial error prone the strength of Smalltalk, it is also the ma-
84 Practical Distillation Control
Cornish, M. (1987). What would you do with MaUsson, S. E. (1989). Modeling of interactions
object-oriented programming if you had it? between submodels. 1989 European Simula-
FYAI-T/ Technical Tips 1(10), 3-12. tion Multiconference, ESM'89, Rome, Italy.
Franks, R. G. E. (1972). Modeling and Simula- Nilsson, B. (1989). Structured modeling of chem-
tion in Chemical Engineering. New York: ical processes with control systems. AIChE
Wiley Interscience. fall meeting, San Francisco, Nov. 1989.
Luyben, W. L. (1990). Process Modeling, Simula- Pascoe, G. A. (1986). Elements of object-ori-
tion and Control for Chemical Engineers, 2nd ented programming. BYTE Aug., 139-144.
ed. New York: McGraw-Hill. Perkins, J. D. (1986). Survey of existing systems
Magnusson, B. (1990). European perspective. 1. for dynamic simulation of industrial process.
Object-Oriented Program. 2, 52-55. Modeling, Identification and Control 7, 2.
Mattsson, S. E. (1988). On model structuring Stephanopoulos, G., Henning, G., and Leone, H.
concepts. 4th IFAC Symposium on Com- (1990). MODEL.LA., a modeling language
puter-Aided Design in Control Systems, P. R. for process engineering. Comput. Chem. Eng.
China. 14(8), 813-869.
6
Plantwide Process Control
Simulation
Ernest F. Vogel
Tennessee Eastman Company
86
Plantwide Process Control Simulation 87
Feeds
Feeds
Tank 1
'---...L.-------+I Tank 2 I - - - - - - - - o t
Feeds
pense of implementation, process downtime equipment. Such surge tanks are essential
or hardware and software expenses. Dy- for interfacing batch and continuous pro-
namic simulations showing how both the old cesses. Simulation provides a way to deter-
and the new strategies perform for typical mine how large surge tanks need to be and
process disturbances often can accomplish where they should be located to sufficiently
both of these tasks. attenuate harmful disturbances.
Reactions:
A+B-+R
2B .... S
Feeds A&B
A B N
C
§
8
needed? How much might they help? How simulated dynamic behavior is easily ex-
frequently are composition measurements plained, but counterintuitive. The mind does
needed? not easily predict dynamic responses by su-
Are SISO PID controllers satisfactory? perimposing all of the simultaneous dy-
Are there any severe interactions? Are they namic effects in a chemical process, particu-
bad enough to justify a more advanced algo- larly if the process contains one or more
rithm? How much will an advanced algo- recycles.
rithm help? The explicit results from process control
Which disturbances have the most effect? simulation tend to be more qualitative than
What happens if feed A or feed B is in quantitative. Application of process control
excess? simulation often results in a decision as
illustrated in the example above, for exam-
Process Design ple, use control strategy B rather than strat-
Is a surge tank needed between the egy A, buy a continuous analyzer for use in
columns? The simulation can justify it if a feedback or feedforward strategy, a surge
needed or prove that one is not necessary if tank is necessary, and so forth. To contrast,
that is the case. process design simulations often provide ab-
Is the reflux drum on the first column big solute numbers, for example, reboiler heat
enough so that disturbances in the column duty, distillation column diameter, and so
are sufficiently attenuated and do not upset forth.
the reactor through the recycle?
make good boundaries for simulations be- mic reactor and determining how high the
cause most disturbances are significantly at- temperature will go during an upset and
tenuated at those points. Also, portions of a how quickly it will rise, accuracy of the
process that are not connected by recycles model is very important.
can be segregated.
The accuracy of a simulation is not easily
measured. In some cases, the plant does not 6·5 BUILDING A PLANTWIDE
yet exist. For existing processes, perturbing PROCESS SIMULATION
the plant to observe dynamic responses is
usually not desirable. Generally, to obtain There are several simulation environments
more accuracy, the simulation must use an engineer can use to build a dynamic
more rigorous or complex models of unit flowsheet simulation. Three possibilities are
operations, valves, and so on. described in the following subsections.
In control strategy design, dynamic simu-
lation is typically used for making relative
performance comparisons between alter- 6-5·1 Programming Environment
nate control strategies. When making rela- In this case, the engineer writes source code
tive comparisons, a high degree of accuracy for the differential and algebraic equations
is not as important as when determining that describe the process to be simulated.
absolute values. If two control schemes are The programmer must provide numerical
significantly different, simulations will dem- methods and interface routines as well.
onstrate that fact, even if the model re- Simulations built in this environment tend
sponse does not exactly match the process to be customized for simulating one particu-
response. If the differences between two lar process and are not easily modified for
schemes are small, other criteria should be other problems. Writing and debugging sim-
used to choose between them, for example, ulations is likely to be time intensive. This
ease of operator understanding and ease of approach offers the most flexibility, but re-
implementation and support. quires a high level of expertise in both mod-
As an accuracy guideline for control elling and programming.
strategy design, match the simulation's
steady-state base case to the plant opera-
tion or to the steady-state process design
6-5·2 Equation Solving Environment
conditions to within 5%. These conditions
include flows, temperatures, pressures, com- The equation solving environment is a
positions, levels (inventories), and vessel higher level than the programming environ-
volumes. Note that with good estimates of ment. Similar to the programming environ-
vessel volumes, and flow rates, residence ment, the engineer must provide the differ-
times (time constants) will be similar to ential and algebraic equations that describe
those of the process. Thus, validation of the process to be simulated. However, the
dynamic responses is generally not neces- numerical method and interface routines
sary for control strategy design. If the pro- are provided. This approach is similar to the
cess conditions are matched as previously programming approach in that simulations
described, the process dynamics are usually tend to be customized and writing and de-
sufficiently accurate for control strategy bugging the simulation can be time consum-
comparisons. ing. Although this environment may not be
Note, however, that for other applica- quite as flexible as the programming envi-
tions model accuracy is much more impor- ronment, the total time required to build
tant. For instance, when evaluating an the simulation and the programming exper-
emergency shutdown system on an exother- tise required will likely be less.
Plantwide Process Control Simulation 93
egy design simulations involve relative per- variable is perturbed, the simulation is re-
formance comparisons. However, some pro- converged to steady state and the changes
cesses contain very unique unit operations in the controlled variables are calculated.
and no matter how complete the unit opera- The derivative is then approximated as
tion library in a ftowsheet simulator, there ac/am.
will be problems that require custom model- From steady-state gains, the simulator
ling. can then calculate several interaction mea-
sures as described in Chapter 8: relative
6-6 FEATURES OF A PLANTWIDE gain array (Bristol, 1966; McAvoy, 1983),
singular value decomposition (Forsythe,
PROCESS SIMULATOR FOR
Malcolm, and Moler, 1977), Niederlinski's
CONTROL STRATEGY DESIGN stability index (Niederlinski, 1971), and
There are several tools in a plantwide pro- block relative gain (Manousiouthakis,
cess control simulator that set it apart from Savage, and Arkun, 1986).
other more general ftowsheet simulators. Also from steady-state simulation, sensi-
A key feature is the capability to perform tivity plots as illustrated by Luyben (1975)
both steady-state and dynamic simulation. can be generated. These sensitivity plots
Steady-state simulation facilitates the devel- have information similar to the steady-states
opment and initial screening of candidate gains, but the plots also reveal the nonlin-
control strategies by providing a quick way earities of the process. Figure 6-6 shows an
to calculate several steady-state perfor- example of a sensitivity plot for the first
mance measures. column of the example process (Figure 6-5).
A common first step in control strategy This plot shows how the bottoms composi-
design is the calculation of steady-state gains tion varies for three disturbances while a
for selected controlled (c) and manipulated proposed control strategy holds the con-
(m) variables. Steady-state gain may be de- trolled variables constant. The curves are
fined as iJc/iJm. To calculate the steady-state generated by varying each disturbance vari-
gains, each manipulated variable is per- able one at a time and reconverging the
turbed one at a time from an initial base simulation to steady state with each varia-
case condition. After each manipulated tion. In this case, the study shows that the
0.0040 --r-------------------,
Disturbance Variables
A-Feed Rate
B-Feed Temperature
C-Feed Composition
(Weight Fraction
Component R)
0.0035
B
A
C
0.0030 +----------.-----------1
10 Percent Base Case 10 Percent
Decrease Value Increase
Disturbance Variable
D. E. Rivera1
Arizona State University
problems because they are not parsimo- allowing a user to concentrate on making
nious (i.e., compact) system representations. the design decisions and judgments involved
Overparametrization and lack of indepen- in system identification.
dence among the model parameters results The chapter is organized as follows. Sec-
in a poorly conditioned estimation problem tion 7-1-1 describes the general class of
and estimates with high variance result when transfer function models on which the iden-
conventional least-squares estimators are tification methods covered in this chapter
used. To compensate for this problem, prac- are based, whereas Section 7-1-2 presents
titioners are forced to perform long dy- the overall iterative methodology of system
namic process tests to generate large quan- identification. More detailed descriptions of
tities of data for the estimation. In some each step in the process are provided in
plants, performing very long dynamic tests subsequent sections: Section 7-2 examines
may be acceptable. Nevertheless, an overall the design of perturbation signals to ob-
goal of system identification is to obtain the tain information for single-input-single-out-
required models with the minimum of dis- put (SISO) and multi-input-multi-output
ruption to normal process operation. There- (MIMO) systems; Section 7-3 covers topics
fore, we present two alternative approaches associated with identifying model structure
for obtaining step or impulse models, and estimation of model parameters, which
namely, estimation of parsimonious low- includes a discussion of bias-variance
order transfer function models (that are trade-offs in system identification (Section
converted after estimation to the required 7-3-1), nonparametric and parametric iden-
step or impulse form) and the use of biased tification (Sections 7-3-2 and 7-3-3), con-
least-squares estimators (for example, ridge trol-relevant identification (Section 7-3-4),
regression and partial least squares). The identification in the closed-loop (Section
use of biased estimators is one way to pro- 7-3-5), and the treatment of nonlinearity in
vide more efficient estimation if impulse or distillation systems (Section 7-3-6). Model
step response models are to be identified validation is discussed in Section 7-4, some
directly from process data. practical aspects of conducting dynamic tests
We assume in this chapter that model on a process are discussed in Section 7-5,
estimation is performed off-line using batch and an example is provided in Section 7-6.
data collected from a dynamic process test
conducted over a fixed period of time for
7-1-1 Discrete Transfer Function
the purpose of designing a fixed-structure
Models for Distillation Systems
controller. We do not consider on-line re-
cursive estimation because the resulting We assume that the system to be identified
adaptive controllers, particularly for multi- is linear and time-invariant and can be rep-
variable constrained systems, have not resented by a transfer function model of the
proven reliable enough to be considered as form
practical tools for distillation control at this
time. y(t) = Go(q)u(t) + v(t) (7-1)
Finally, we would like to emphasize that
the daunting programming task previously
associated with getting started in system
where
identification has been eliminated with the
availability of a number of commercial soft- subscript 0 = the true (unknown) system
ware packages for system identification. yet) = a system output (e.g., a tem-
These packages provide a wide range of perature or composition mea-
data analysis and identification tools, imple- surement from an on-line an-
mented with efficient numerical procedures, alyzer, or transformations of
98 Practical Distillation Control
the orders of the polynomials and time the problem of how to perform identifica-
delays. tion when the end-use of a model is for
Step 3. Parameter estimation. The parame- process control. Topics specific to distilla-
ter estimation step involves the selection tion are discussed where they arise.
of a number of design variables: the nu- Again, we emphasize that commercial
merical procedure to obtain estimates for software packages for system identification
the model parameters, the form of the are available today that greatly facilitate
objective function (e.g., sum of squared Steps 2 to 4.
prediction errors), the number of steps
ahead in the prediction-error equation,
and whether to use a prefilter, and if so, 7·2 PERTURBATION SIGNAL
its form. A complicating factor in this DESIGN
step is the potential for numerical prob- 7·2·1 Discussion
lems in the parameter estimation algo-
rithm (e.g., ill-conditioning, multiple lo- A primary goal of system identification is to
cal minima). be able to discriminate among competing
Step 4. Model validation. Having obtained a models within a specified model set. This
model, its adequacy must be assessed. discrimination (and the related operation of
Among the issues to consider are: parameter estimation) will be efficient only
What information about unmodelled dy- when performed using data obtained from a
namics is resident in the residuals (the process test in which certain identifiability
time series resulting from the difference conditions have been satisfied. These condi-
between the predicted model output and tions are usually expressed as the require-
measured output) and how this informa- ment that a perturbation signal be per-
tion assists in the iterative identification sistently exciting (e.g., see Ljung, 1987,
procedure? Chapter 14). In practice, this requires a
Does the step response of the estimated perturbation signal that is sufficiently rich in
model G(q) agree with physical intu- frequency content to excite a system up to
ition? some specific frequency, above which the
Although these issues are important, the dynamics are considered unimportant. For
ultimate test of model adequacy is example, for control purposes, we are often
whether it meets the requirements of the not interested in modelling the fastest modes
intended application (which in this case of a system (e.g., the dynamics of valve
is process control). motion). A number of questions arise:
process of identification. In theory, the de- The sequence repeats itself after T = NTel
sign of a control system should also be in- units of time, where N = 2 n - 1. For Tel =
corporated in this iterative procedure be- one sampling period, the power spectrum of
cause controller design and tuning affect a PRBS signal is approximately equal to 1
the modelling requirements for process con- over the entire frequency spectrum. For val-
trol and therefore influence signal design, ues of Tel > 1, the spectrum drops off at
that is, the frequency range of interest. higher frequency and has a specific band-
However, in practice, including controller width useful for system excitation. The
design as an iteration is not necessary; suf- power spectrum of a PRBS signal generated
ficient approximate system information is using n = 7 and Tel = 3 (with ~ = 1 s) is
usually available to allow informative dy- shown in Figure 7-1 (plot generated numeri-
namic tests to be designed. cally using the spa function in MATLAB,
The perturbation signal most often used 1989).
to satisfy the preceding identifiability condi- An analytical expression for the power
tions (as well as closed-loop identifiability spectrum of a PRBS signal generated using
conditions; see Section 7-3-5) is the shift registers is given by (Davies, 1970)
pseudo-random binary sequence (PRBS)
signal. The design of PRBS signals for _ a2(N + 1) Tel [ sine wTel/2) 12
perturbation of single-input-single-out- Sew) - N wTel/2
put (SISO) and multi-input-multi-output (7-5)
(MIMO) systems for control-relevant identi-
fication is discussed in the following sec- where a = signal amplitude. At low fre-
tions. quencies the power spectrum has the ap-
Discrete sampling of a continuous system proximate value a 2(N + l)Tel /N. At w =
leads to inevitable information loss, and 2.8/Tel , the spectrum is reduced by half,
therefore selection of sampling interval is defining the upper limit of its useful range.
an important topic in identification for dis- Due to the periodicity of the autocovariance
crete-time control models. Sampling too function of a PRBS signal, its power spec-
slowly can distort the spectrum of a sam- trum is discrete (Eykhoff, 1974) with points
pled signal in the frequency range of inter- in the spectrum separated by 2'lT /T and the
est as a result of aliasing (or foldover), lead- first point at 2'lT/T. Therefore, the fre-
ing to faulty identification. The selection of quency range of a PRBS signal useful for
sampling interval and design of anti-aliasing system excitation is
filters to avoid this problem is discussed in 2'lT 2.8
most texts dealing with system identification - s w< - rad/s (7-6)
(e.g., Ljung, 1987, Section 14.5). T - Tel
Selection of n and Tel allows a user to
7-2-2 Pseudo-random Binary tailor this frequency range for excitation of
Sequence Signals a particular system. If a priori estimates of
the system dominant time constant T dom and
7-2-2-1 PRBS Signals for S1S0 Systems deadtime T d are available, then the follow-
As described by Davies (1970), a PRBS sig- ing can be used as a guideline for designing
nal can be generated using shift registers, a PRBS signal:
and such signals can be characterized by
two parameters: n, the number of shift reg- 1 1
------ s w S ------
isters used to generate the sequence, and fJ( Tdom + Td/ 2 ) (Tdom + Td/ 2 )/a
Tel' the clock period (i.e., the minimum time (7-7)
between changes in level of the signal, as an
integer multiple of the sampling period Ts). where a is specified to ensure that suffi-
Identification of Distillation Systems 101
SPECfRUM
loo
10-1
frequency
FIGURE ,.1. Power spectrum of a PRBS signal generated using n = 7 and Tel = 3
(Ts = 1 s).
ciently high frequency content is available perturbation signal (Le., how low in fre-
in the perturbation signal, commensurate quency does one wish to model the plant).
with how much faster the closed·loop reo Choosing higher values of fJ results in
sponse is expected to be relative to the lower-frequency information: fJ = 3 will
open-loop response: provide information down to a frequency
roughly corresponding to the 95% settling
time of process, fJ = 4 for the 98% settling
time, and {3 = 5 for the 99% settling time.
For example, with a = 2, the designer ex- Comparing Equations (7-6) and (7-7)
pects the closed-loop time constant to be gives
one-half that of the open-loop response (i.e.,
twice as fast). As discussed in more detail in
Section 7-3, the accuracy of the frequency
response of a fitted model, relative to that and
of the real process, is greatest in those fre- 21T'afJ
quency ranges where the perturbation sig- N=-- (7·9)
nal has most energy. A more aggressively 2.8
tuned model-based controller (larger a) will Consider the design of a PRBS signal for
require an identified model that is accurate a first-order plus dead-time process where
of higher frequencies. Therefore, the per- 'T = 'Tdom = 100 sand 'Td = 20 s. Choosing
turbation signal should be designed (e.g. a = 1 (Le., only the open-loop dynamics are
using Equation 7-8) to ensure sufficient en- of interest and the controller is not required
ergy at higher frequencies. to be tightly tuned) gives Tel = 308 s. Thus,
{3 is specified to tailor how much low- the minimum time between changes in level
frequency information will be present in the of the PRBS signal is roughly 3 time con-
102 Practical Distillation Control
stants, that is, very nearly what can be con- in input levels varies widely with different
sidered conventional step tests where the combinations of input moves.
process is allowed to reach steady state af- As described by Eykhoff (1974) and
ter each step. In this particular example, Briggs and Godfrey (1966), PRBS signals
because the control system is expected to be applied to multiple inputs simultaneously
tuned so that the closed-loop bandwidth is must satisfy certain conditions to allow
not significantly wider than the open-loop identification of multivariable systems. Most
one (a situation that arises frequently in important, for a system with p inputs, the
industrial control problems), the perturba- cross-correlation functions cf>uu (7") for each
I }
tion signal need only emphasize low- pair of inputs uiu j ' i, j = 1,2, ... , p, must
frequency excitation. How few step changes be negligible for 7" less than the maximum
(Le., the duration of the test) will suffice system settling time (this can be taken as
depends on the presence of noise. In the the largest 99% settling time among the
absence of noise or corrupting disturbances, input-output pairs, T99%,max)' PRBS signals
a single step (followed by a step in the generated independently for each input
opposite direction to ensure that the system would satisfy this requirement. However, a
returns to its original level) in this example more practical approach is to generate mul-
would provide the required information. tiple PRBS signals as delayed versions of a
Higher-frequency information will be neces- single PRBS. To satisfy the requirement on
sary only when the system to be identified is the cross-correlation functions, the mini-
known to have faster modes (i.e., lower mum allowed delay between any two ap-
dominant time constant) or the control sys- plied sequences is given by D:
tem is expected to be tuned aggressively so
that frequency response in the higher ranges NTcI
is important.
D = --
p > T99% ,max (7-10)
7-2-2-2 PRBS Signals for MIMO Systems N can be increased as necessary to satisfy
When a dynamic test is performed by per- the inequality.
turbing only one input at a time, the estima- Another approach to conduct a multi-
tion problem is greatly simplified because variable test is to conduct one long test in
SISO models can be estimated one at a which PRBS signals are applied consecu-
time (assuming that a noise model is not tively to the inputs, one at a time. During a
estimated). This approach has advantages period in which a PRBS signal is applied to
and disadvantages from an operational point one particular input, the other inputs are
of view. By varying only one input, the be- held constant if possible, but allowed to
havior of process outputs is more pre- vary if necessary (e.g., due to automatic or
dictable, reducing the probability of enter- operator initiated control action). Data for
ing an undesirable operating region and/or the entire test are collected as a single batch,
violating constraints. The probability of op- from which multiple-input-single-output
erator intervention becomes more likely with (MISO) models of the form described by
a test specifically designed to perturb all Equation 7-4 are identified for each process
inputs simultaneously (for example, with in- output variable (that are then converted to
dependent PRBS signals applied to each SISO models if required for the control
input), because it is difficult to predict when system design). Thus, the effect of one or
certain combinations of input levels might more inputs wandering during a period in
drive a process into an undesirable operat- which another input is perturbed is ac-
ing region. This is a particular problem for counted for by the estimation procedure.
multivariable systems that exhibit strong Design of perturbation signals for multi-
gain directionality, that is, the gain of a input-multi-output systems that are ill-
particular output to simultaneous changes conditioned (e.g., high purity distillation
Identification of Distillation Systems 103
filtering of the input and output data, the power spectrum is equivalent to the
squared magnitude
YF(t) = L(q)y(t) (7-11)
is given by Ljung (1987): The input signal power spectral density fPJ w).
Because the input spectrum directly
weights the term IGo(e jW ) - G(e jw )1 2
<PeF(W) = (IGo(e jW ) - G(e jw )1 2 <1>uC w )
(i.e., the frequency-domain bias), error
between the true system and the esti-
1L( e jW ) 12
+<1>,,( w») . 2 (7-15) mated model is reduced at those fre-
IH( eJW ) I quencies at which the input power
spectrum is highest. Therefore, it is im-
<I>/W) and <I>,,(w) are the power spectral perative to define the frequency range of
densities of the input signal u and the dis- importance to the control problem and
turbance v, respectively, computed from the ensure that the input signal is rich in
expression frequency content in this range.
Prefilter L(q). The prefiltering also acts as a
frequency dependent weight, providing
<l>xCw) = L cx(k)e-jwTk (7-16) another way to selectively emphasize the
k= -00 frequency range of importance for con-
trol.
cx(k) is the covariance of a time series x at The structure of G(q) and H(q). The choice
lag k, evaluated using of model structure plays a significant role
in frequency-domain bias in the parame-
1 N-k ter estimation problem. Note that the
cx(k) = - L x(i)x(i + k), noise model H(q) weights the objective
N i=1
function in a manner similar to the pre-
k = 0, N - 1 (7-17) filter, and therefore has a direct effect on
frequency-domain bias in the estimated
When dealing with a deterministic signal, G(q).
Identification of Distillation Systems 105
The disturbance power spectral density 4>,lw). given by (Box and Jenkins, 1976)
The parameter estimation problem, when
examined in the frequency domain, is a
multiobjective minimization whenever a
disturbance model H{q) is present [be-
cause both G{q) and H{q) contain pa-
where
rameters to be estimated]. In this case
there is trade-off between reducing bias
1 N-k
between Go{q) and G{q), and fitting the
spectrum of the disturbance v{t). This
cuy(k) = N E u'(i)y'(i + k),
i=1
problem becomes particularly pro-
k = 0, N-1
nounced when autoregressive terms that
are common to both G{q) and H{q) are
included in the parameter estimation
problem.
N is the number of observations, whereas
Variance issues can be examined in the
the prime denotes that detrending and
frequency domain using an expression of
prewhitening operations were performed on
the form (Ljung, 1987)
the time series; under these conditions,
cuy{k), the estimated cross-covariance func-
.
vanance -
(n- -
<1>,,( w) )
- (7-20)
tion, can be used to obtain an estimate of
N <l>u( w) the kth impulse response of a system. Other
uses of correlation analysis include discern-
ing the deadtime of a system, the 95% set-
tling time, and also, most importantly,
where n = the number of parameters in determining if an input signal u has a sig-
the model nificant effect in describing the dynamics of
N = the length of the data set the process. This latter use involves evaluat-
ing the estimated impulse response coeffi-
Equation 7-20 clearly shows that reducing
cients against the hypothesis that the data
the number of model parameters, increasing reflect only white noise. The standard devi-
the length of the data set, and increasing ation of Uk is given by
the magnitude of the input signal all con-
tribute to variance reduction in system iden-
tification.
w( k) = (!(1 +
0,
cos( 'lTk/M» , lui :s; M
lul>M
C(q)
H(q) = A(q)D(q) (7-29)
(7-26)
Note that the A polynomial contains poles
that are common to both the disturbance
where M is the truncation parameter for and process transfer functions.
the window. Equation 7-27 is written for multi-input
Spectral analysis can also be applied to systems by adding terms on the right hand
the residual time series to obtain norm- side
bound uncertainty descriptions that are use-
ful for robust control. A more detailed anal-
ysis of this is explained in Section 7-4.
deliver superior performance over a con- capable of giving unbiased estimates for the
troller lacking a disturbance model. As parameters in G(q), even when structural
mentioned previously, Equation 7-29 is ca- inadequacies are present in H(q), including
pable of modelling stochastic autoregressive the case where no disturbance model is
integrated moving average (ARIMA) distur- present [Le. H(q) = 1], are particularly use-
bances [as described by Box and Jenkins ful in this situation. The prediction error
(1976)] or randomly occurring deterministic estimates implemented by Ljung (1988) fall
disturbances. However, when performing into this category.
off-line batch identification and off-line con- Another consideration is frequency do-
troller design, it is assumed that the distur- main bias in the estimated G(q) that is
bance process does not change with time introduced by including a disturbance model
(Le., its structure is fixed). This assumption in the estimation problem. From Equations
is very often not valid in typical processing 7-15 and 7-19 it is clear that H(q) acts as a
systems. Disturbances affecting a distillation frequency domain weight on the estimation
column, as well as most other process units, of G(q). Including the term D(q) [and
are typically random in structure. Process A(q)] emphasizes high frequencies in the
units within a plant are highly intercon- estimation of G(q); the model will more
nected and integrated through process and accurately predict fast dynamics and hinder
utility streams. A particular process unit the low-frequency and steady-state predic-
may be disturbed at any instant from an tive ability of the model. Including C(q) has
upset or change in operating condition from the opposite effect; high-frequency fit is de-
any number of other process units that in- emphasized. These effects are a current area
teract with it. These upsets arise from a of research in the literature.
multitude of different sources; for example, Various substructures arising from Equa-
equipment failure, scheduled or unsched- tion 7-27 and the resulting parameter esti-
uled maintenance operations, abrupt mation problems are now described.
changes in ambient conditions due to pas- ARMAX ARMAX identification uses
sage of fronts, changes in raw material type the structure
or composition, operator control moves mo-
tivated by any number of causes (e.g., shift A(q)y(t) = B(q)u(t - n k ) + C(q)e(t)
changes, optimization moves, response to (7-30)
disturbances), and so on. In addition, dis-
turbances from different sources propagate This structure has been popular in the con-
through a unit in different ways, again sug- trol literature, particularly for designing
gesting that each will have a different struc- adaptive controllers.
ture if modelled with a linear disturbance ARX The ARX structure is
model in the form of Equation 7-29. Con-
trol structures that employ an identified A(q)y(t) = B(q)u(t - n k ) + e(t) (7-31)
model for G(q) only, and ignore distur-
bance dynamics, have been found in prac- Using this structure, the problem of deter-
tice to deliver acceptable control perfor- mining the right order for A and B is often
mance. circumvented by overparametrization, that
Again, we do not consider model-based is, the orders of the A and B polynomials
adaptive schemes (particularly for multivari- (n = na = n b ) are selected to be high. The
able constrained systems) to be viable at theoretical justification for this choice of
this time. Thus the identification problem as variables is that for an infinite number of
discussed in this chapter is to obtain the observations, u white noise, and e a zero-
best possible estimate for G(q) from the mean stationary sequence, the estimate of
given batch of data. Estimators that are the first n impulse response coefficients of
108 Practical Distillation Control
G(q) will be unbiased. A high-order ARX it separately parametrizes the input and dis-
model is therefore capable of approximating turbances, avoiding transfer functions that
any linear system arbitrarily well. have common poles. The estimation prob-
Although high-order ARX modelling lem is nonlinear, but reliable and fast esti-
possesses a number of attractive theoretical mation methods are available. Nonstation-
properties, the conditions under which they ary disturbances are handled by forcing
apply are seldom seen in practice. Further- D(q) to have one or more roots on the unit
more the presence of the autoregressive circle [Le., write as (1 - q-l)dD(q), where
term A(q) results in an emphasis on the d is typically equal to 1].
high-frequency fit, which is not necessarily Output E"or The output error model is
good for control system design. The bias a simplified form of the Box-Jenkins model
problem with autoregressive terms becomes structure
worse if low-order ARX models are used.
FIR The FIR (finite impulse response) B(q)
model structure y(t) = F(q) u(t - nk) + e(t) (7-34)
yet) = B(q)u(t - n k ) + e(t) (7-32) The output error structure retains the ad-
vantage of separate parametrizations for the
is convenient because some predictive con- input and disturbance, and does not require
trollers are based on this model structure a choice of the structure for the disturbance
[and it is easily converted to step-response model.
form for others that require it, e.g., dynamic
matrix control (DMC)]. Although the ap- 7-3-3-2 Prediction-Error Parameter
propriate order of the finite impulse re- Estimation
sponse depends on the selected sampling The objective minimized in prediction-error
time and the settling time of the process, methods is the sum of the squared predic-
the result is usually high (30 to 100). When tion errors
u is persistently exciting, e is stationary, and
u and e are uncorrelated, the estimated 1 N 1
impulse response coefficients will be unbi- V= - E -(yet) - .9(t)2
N t=1 2
ased. Again, as with ARX, these conditions
are seldom fully observed in practice. As 1 N 1
mentioned in the introduction, direct esti- = - E -eJ,(t) (7-35)
N t=1 2
mation of the impulse response model is
inefficient, resulting in parameter estimates In the case of FIR and ARX models, the
with high variance. Direct identification of predictor structure is
the FIR model requires special test and
estimation procedures, as described in sub- .9(t) =B(q)u(t) + (1-A(q»y(t)
sequent text. (7-36)
Box-Jenkins Box and Jenkins (1976)
proposed a model of the form
which can be written in regression form as
B(q) C(q) (7-37)
yet) = F(q) u(t - nk) + D(q) e(t)
(7-33) lP is the regression vector
8= [ -
iN
1: (fJ(t)(fJT(t)
]-1 -1 1: (fJ(t)y(t)
N
able in commercial identification packages.
tainty) of the fitted model around this fre- specified structure is fit to these data, sub-
quency is an important objective when the ject to the constraint that the Nyquist curve
model is to be used for control system de- of the fitted model passes exactly through
sign. the known point(s) on the Nyquist diagram,
From previous discussions, one way to thus reducing the bias in the fitted model at
improve the accuracy of the fitted process this frequency to zero. As shown in the
model G(q) in a particular frequency range preceding references, the constraint can be
is to ensure that the perturbation signal is expressed as a set of linear equations, re-
rich in these frequencies. Designing a per- sulting in a quadratic optimization problem
turbation signal to emphasize system excita- when the objective function for fitting the
tion around the frequency where the Nyquist model is quadratic (as is the case with pre-
curve most closely approaches the point diction-error methods). For certain model
( -1,0) poses a problem because a priori structures, analytical solutions to the
knowledge of the closed-loop frequency re- quadratic optimization are available; other-
sponse for a given controller tuning would wise a numerical procedure must be used.
be required. Although an iterative design To perform a relay test to obtain auxil-
procedure could be adopted, an alternative iary information for fitting a discrete time
is the use of a simple relay test (Astrom and model, a relay is connected in a sampled
Hagglund, 1984) that determines the point feedback loop, as shown in Figure 7-2. The
of intersection of the Nyquist curve with the sampling frequency for the relay test must
negative real axis of the open-loop system be the same as the sampling period used to
Go(q) (i.e., the critical point, or the ultimate collect the data from the perturbation test.
gain and frequency). In an industrial setting, If the process has a phase lag of at least - 7T'
a typically tuned PID controller (or other radians (which will always be the case in
controller design) provides only a small practice because all real processes exhibit
amount of phase lead, and the frequency at some deadtime), under relay feedback it
which the Nyquist curve most closely ap- will enter a sustained oscillation with period
proaches the point (-1,0) (as well as the Pu (the critical period), giving the ultimate
crossover frequency of the compensated sys- frequency Wu = 27T'/Pu • The ultimate gain is
tem, of equal importance) is not often not given approximately by Ku = 4h/(a7T'),
far from the region of the critical point of where h is the height of relay and a is the
the open-loop process. In these situations, amplitude of the principal harmonic of the
knowledge of the open-loop critical point is output. The relay test is easier and safer to
of practical value in system identification apply than the procedure suggested by
for control purposes. Ziegler and Nichols because it does not
Astrom and Hagglund used the relay test require adjustment of controller gain in or-
to provide information for PID controller der to generate sustained oscillation. It only
tuning. The use of such auxiliary. informa- requires approximate knowledge of the pro-
tion (i.e., knowledge of one or more points
on the system Nyquist curve, obtained from
experiment or from other process knowl-
edge) to improve the quality of models iden- +
tified for control purposes has been de- SeIpaInt
scribed by Wei, Eskinat, and Luyben (1991)
and Eskinat, Johnson, and Luyben (1991).
The basic idea is as follows. Data are col-
lected from the process to be identified FIGURE 7·2. Configuration for a relay test to esti-
using a conventional perturbation proce- mate crossover frequency for use as auxiliary informa-
dure (e.g., step or PRBS tests). A model of tion in identifying a discrete-time model.
Identification of Distillation Systems 111
cess gain (ideally at the critical frequency, The literature remains vague about pre-
but because this is not likely available, filter design; questions regarding the sys-
steady-state gain can be used) in order to tematic choice of structure (low pass, high
specify an appropriate relay amplitude. pass, bandpass) and the choice of filter pa-
Because the relay is a nonlinear device, rameters remain unanswered issues. How-
when applied to a linear process the mea- ever, recent work in this area has begun to
sured critical point is an approximation to provide some answers. In Rivera et aI., 1990
the real one. However, as shown by Wei, and Rivera, Pollard, and Garcia, 1990, the
Eskinat, and Luyben (1991), the informa- derivation of a control-relevant prefilter for
tion is sufficiently accurate to be of value in SISO feedback control is described that ex-
system identification for control purposes. plicitly incorporates the model structure, the
The presence of noise may require the desired closed-loop response, and the set-
relay test to be implemented with a dead point-disturbance characteristics of the
band within which the response variable can control problem into the statement of the
move without causing a jump in level of the prefilter:
input. In a very noisy or highly disturbed
process, it may be difficult to obtain useful L(q) = H(q)G- 1(q)(1 - ij(q»
results from a relay test.
Xij(q)(r(q) - d(q» (7-40)
the plant (Balhoff and Lau, 1985). In in- Tel = the desired closed-loop time
ternal model control (IMC) (Morari and constant, specified by the engi-
Zafiriou, 1989), fi is specified directly via neer
the factorization of nonminimum phase ele-
ments and the choice of a filter. In accor- Assuming step setpoint and disturbance
dance with this philosophy for the model in changes for the control system, the resulting
Equation 7-41 we can use the structure prefilter for FIR and DE estimation is fourth
order,
increasingly more jagged. One way to over- the method described in Section 7-3-4-1) is
come this variance problem is to estimate violated in this case.
the FIR parameters using a biased least-
squares estimator, as discussed in Section 7·3·4·3 Biased Least·SqUllTeS Estimators
7-3-4-3. Direct estimation of a truncated impulse
Work is currently being carried out response model results in an estimation in
(Rivera, 1991b) that unifies the control-rele- which the bias-variance trade-off is shifted
vant prefiltering approach with the crossover toward reduction in bias at the expense of
frequency approach described in Section variance, due to the large number of param-
7-3-4-1. A fundamental analysis of mod- eters (and their lack of independence).
elling requirements for control shows that Therefore, an estimated impulse model
there are four factors that determine the might give an excellent fit to the data that
adequacy of a model for feedback control were used in the estimation, but its use as a
purposes: prediction tool on a different set of data
may be extremely limited because of the
1. The choice of plant and noise model resulting high variance of the parameter
structure. estimates (and thus predictions). There are
2. The desired closed-loop speed-of- a number of ways to reduce the variance of
response. estimates to improve the predictive ability
3. The shape of the closed-loop response of models:
(Le., underdamped or overdamped).
4. The nature of the setpoints and distur- 1. Estimate a parsimonious transfer func-
bances that will be encountered by the tion model (low number of parameters)
closed-loop system. and convert the nominal estimated model
to impulse or step form (with no loss of
Under certain circumstances (the presence information).
of an integrator in the model and control 2. Perform a longer dynamic test to in-
requirements demanding very fast, under- crease the size of the data set, (which, as
damped responses, for example), the power mentioned previously, may be quite ac-
spectrum of the prefilter shows a very strong ceptable for some processes).
amplification of the crossover frequency, in 3. Use a biased estimator that shifts
which case both approaches yield equivalent bias-variance trade-off toward allowing
results. However, when considering the case some bias in order to reduce variance.
of a first-order system with substantial
deadtime (ratio of time delay to time con- Biased estimators, as the name implies,
stant equal to 1), the requirement of an produce estimates that are biased and not
overdamped closed-loop response with a least squares, but are more desirable from a
correspondingly smooth manipulated vari- prediction point of view because the reduc-
able response results in modelling require- tion in prediction mean square error (MSE)
ments that emphasize the intermediate and they achieve more than compensates for the
final segments of the plant's step response minimal bias that is introduced. We briefly
without demanding a close fit of the plant's describe a few approaches that have been
ultimate frequency. This occurs because for used in the literature for biased estimation
the specified control requirements, the fre- of truncated impulse response models and
quency responses of the plant and the loop we provide references for further informa-
transfer function Go{s)K(s) differ substan- tion.
tially at the crossover point; the assumption Ridge Regression Consider the finite im-
that the controller does not introduce sub- pulse response (FIR) model Equation 7-32.
stantial phase shift (a key assumption for The prediction equation for this model
114 Practical Distillation Control
structure is given by Equation 7-36 [with tween the parameters in an FIR model, not
A(q) = 1], and the least-squares parameter all of the ti and Pi are required to explain
estimates are given by Equation 7-38, which the total variation that exists in the data
can be written in matrix notation as matrix X. For the same reason, not all the
ti and Pi are required to develop a satisfac-
8 = [X TXr 1X T y (7-50) tory predictive relationship between X and
the response vector y. Therefore, in PLS, X
With a typical order of 30 to 100 for is written as
B(q), the parameters will not be indepen-
dent, resulting in poor conditioning of XTX X= t 1 P[ + t 2 pr + ... +tqP~ + E
and an inversion that is sensitive to small
= TpT +E
variations in the data. In the simplest form
of ridge regression, Equation 7-50 is re- where q is referred to as the number of
placed by latent variables, and is usually very much
less than r. E is a residual matrix that
contains the random errors in X as well as
the discarded components of X that have
no significant predictable effect on y. In
where I = an identity matrix of appropri- PLS, the regression is performed between y
ate dimension and and the ti in Equation 7-53 and is an inte-
c = a specified positive number gral part of the numerical procedure that
performs the decomposition (in the general
For more information, the reader is re- PLS problem where there are multiple re-
ferred to Draper and Smith (1966) and sponse variables that may be correlated, the
Hoerl and Kennard (1970). numerical procedure includes decomposi-
Partial Least-Squares Estimation An- tion of the response matrix Y; see for exam-
other biased estimator that has been used ple, Kresta, MacGregor, and Marlin, 1991).
to estimate FIR models (Ricker, 1988) is PLS requires the user to select the num-
partial least squares (PLS), a general method ber of latent variables used in the decompo-
for solving poorly conditioned least-squares sition. The object in PLS is to find the
problems. A brief explanation of the method minimum number of latent variables that
follows. For more information, the reader is results in a satisfactory predictor of the re-
referred to Ricker (1988), Lorber, Wangen, sponse variable. This involves evaluating the
and Kowalski (1987), Hoskuldsson (1988), predictive ability of fitted models with in-
Geladi (1988), and a tutorial paper by Geladi creasing number of latent variables, usually
and Kowalski (1986). using cross-validation (the model is fit on
A data matrix X (dim m x n), can al- one segment of data, and evaluated for pre-
ways be decomposed into two matrices, t dictive ability on another segment). The
(m x r) (the "scores") and P (n x r) (the number of latent variables is increased until
"loadings"): no further significant decrease in the sum of
squares of residuals between the predictions
X -- . 1';''P~T
.,
- 1
- 1+
t pT 2 pT
t2 + ... + t rpT
r from the fitted model and the validation
(7-52) data set is observed. The number of latent
variables will be far less than the number of
where r = rank(X) FIR parameters in a typical FIR model.
t.,I p.I = are
....
the ith columns of t•
and With fewer parameters to estimate, a more
P, respectively acceptable balance between bias and vari-
ance in the estimated model is obtained. As
Because of the correlation that exists be- pointed out by Rivera et aI., 1990 and
Identification of Distillation Systems 115
Rivera, Pollard, and Garcia (1990), pre- collection of closed-loop data, two levels of
filtering prior to PLS estimation gives the identifiability can be achieved:
engineer a further handle on the bias-vari-
ance trade-off in a frequency-dependent System identifiability (SI). Given that certain
sense. strict conditions are met with regard to
the model structure and the order of the
compensator, model parameter estimates
7-3-5 Identifiability Conditions for that correctly represent the plant can po-
Closed-Loop Systems tentially be obtained from the data.
When data are collected with a process in Strong system identifiability (SSI). This means
closed-loop operation, certain identifiability that no special restrictions need be im-
conditions must be satisfied to ensure that posed on the model structure or the
proper identification can be performed. order of the compensator in order to
Historical plant data collected under closed- obtain reasonable estimates. Frequency-
loop conditions do not generally satisfy these domain bias remains as an issue, how-
conditions and cannot be used for dynamic ever.
model identification. The pitfalls associated
It is clearly desirable to obtain experimental
with ignoring these requirements have been
well documented in the literature (e.g., Box conditions that are SSI because regression
and MacGregor, 1974). Two issues are in- techniques can be applied in a direct fash-
volved: ion with no other special considerations.
Conditions that lead to SSI are
the system was really identifiable. One is It is possible to avoid detuning by intro-
better off carrying out closed-loop identifi- ducing the signal at an alternate point in
cation under circumstances that generate the loop, the controlled variable setpoint.
SSI, which implies that a carefully designed The expression for the input power spec-
external signal should be used. trum in this case is
It should be noted that nonparametric
identification techniques (such as correla-
tion analysis and spectral analysis) cannot
be applied directly to closed-loop data from
normal plant operation. This has been pre-
viously pointed out by Box and MacGregor
(1974), among others. The standard applica-
tion of these techniques results in model where 11 = G oK(1 + GOK)-l is the com-
estimates that are the inverse of the con- plementary sensitivity function. Note that
troller. the effect of an external setpoint change on
the estimation of G(q) is weighted by
7-3-5-2 Closed-Loop Frequency Domain Bias o
IG l 1112, as opposed to 1131 2, which does not
attenuate the low frequencies when the
An expression similar to Equation 7-15 can
controller is tightly tuned. Furthermore, a
be derived for closed-loop systems to ana-
tightly tuned controller may also lead to
lyze frequency-domain bias when a regula-
amplification of the higher frequencies, as
tor is present. Such an analysis is more
Example 7-1 shows.
complicated than its open-loop counterpart.
We will thus focus on the effect of the
feedback controller C on the power spec- Example 7-1
trum of the input signal, in order to provide Consider the first-order plant given by:
some insight into the problem of
frequency-domain bias generated from
closed-loop dynamic tests. If an external k 0.096
G o(q) -- -
q _-
a -- q - 0.904 (755)
-
signal U d is added onto the manipulated
variable U, the power spectrum of the input
is given by
This plant corresponds to a first-order
transfer function with an open-loop time
constant of 10 min. Assume that the desired
closed-loop response is also a first-order
transfer function, represented by the ex-
where 13 = (1 + GOK)-l is the sensitivity pression
function of the closed-loop system that acts
as a weight to the parameter estimation (1 - 8)
problem in a manner similar to prefiltering. 11(q) = 1 - e(q) = -q---8- (7-56)
The effect of the controller is to attenuate
low-frequency information, as shown in Ex-
ample 7-1. For controllers with integral ac- where
tion, the sensitivity function is 0 at w = 0,
which means that the controller attenuates
the low-frequency portion of the external
signal. Significant detuning of K, the feed-
back controller, may be required in order to 1'. is the sampling time (and control inter-
obtain an appropriate steady-state fit. val) and 'Tel is the closed-loop time constant.
Identification of Distillation Systems 117
A PI controller tuned using the rules de- creasing 'Tel' and when 'Tel < 'Tal' the high
scribed by Prett and Garcia (1988) would frequencies are amplified. In this example,
result in such a closed-loop system. G011/ is an allpass when Tel = 'Tal' which
In this example we set r.
= 1.0 min and represents an ideal situation. We thus rec-
examine the amplitude ratios of G 0 11/ and ommend that when injecting an external
8 for the three cases: signal at the controlled variable setpoint,
the controller should be tuned such that the
closed-loop speed of response equals the
Tel = 1 min (closed-loop speed faster than
open-loop speed of response.
open-loop).
Tel = 10 min (closed-loop speed equivalent
to open-loop).
Tel = 100 min (closed-loop speed slower than 7-3-6 Treatment of Nonlinearity
open-loop).
Distillation columns are inherently nonlin-
ear. Severe nonlinearity is most often seen
Figures 7-3 and 7-4 compare the normalized in high purity columns, but a number of
amplitude ratios of Go 11/ and 8. 8 resem- other factors can contribute as well, for
bles a high-pass filter with bandwidth (the example, non ideal vapor-liquid equilib-
frequency at which the amplitude ratio first rium. Although columns operated above ap-
reaches 1/ {i) defined by l/Tel' G 011/ is a proximately 98% purity often exhibit severe
low-pass filter for slow closed-loop speed- nonlinearity, behavior varies from column
of-response; however, its ability to attenuate to column and it is difficult to define a
the high-frequency range decreases with in- specific purity above which severe nonlin-
_ _ _ _ _ _""'
__:::c
_____________________________________ _
100
.g
~
..
"0
.e
:=
Q.
~
10-1
Frequency (Radians/Minute)
FIGURE 7-3. Amplitude ratio for G Ol 1/. - : 7'el = 1 min; - - -: Tel = 10 min;
100 min.
- ' - : 7'e1 =
118 Practical Distillation Control
Frequency (RadiaDI/MiDute)
FIGURE 7-4. Amplitude ratio for E. - : Tel = 1 min; - - -: Tel = 10 min; - ' - :
100 min.
Tel =
earity can be expected. As with other non- composition measurements are given by
linear systems, if operated over a sufficiently
small region, linear models often provide an _ 1 - XD
adequate system description. Although large X D = In t (7-57)
l-x~
setpoint changes are uncommon in distilla-
tion systems, disturbances often drive re- - XB
X B = In ---;et (7-58)
sponse variables far enough from steady- XB
state conditions that the performance of
linear model-based control systems can de-
teriorate significantly. This section presents where x = composition measurement
a number of tools for adapting the basic set = setpoint
linear identification approach when signifi-
cant nonlinearity is present in distillation Georgiou, Georgakis, and Luyben (1988)
systems. applied dynamic matrix control (DMC)
The most widely applied tool for identify- based on transformed composition mea-
ing nonlinear distillation systems is the use surements to three simulated distillation
of linearizing transformations on composi- systems: a moderate purity methanol-water
tion measurements (Alsop and Edgar, 1987; system (99% purity), the same system at
Koung and Harris, 1987). The Koung and higher purity (99.9%), and a very high pu-
Harris transformation is based on an analy- rity system (10 ppm impurity). As purity
sis of the fundamental performance equa- increased, significant performance improve-
tions described by Eduljee (1975), which are ments were obtained by using nonlinear
particularly accurate for columns operated DMC (i.e., variables transformed) over lin-
near minimum reflux. The transformed ear DMC and a conventional LV control
Identification of Distillation Systems 119
strategy using multiple PI loops. However, Further information on the use of lin-
with the very high purity column, the per- earizing transformations can be found in
formance of the conventional PI-based con- Shinskey (1988).
trol system was superior to that of the non-
linear DMC, suggesting that the simple
Koung and Harris transformation was in- 7·4 MODEL VALIDATION
adequate at this purity level. The more
7-4-1 Classical Techniques
complex Alsop and Edgar transformations
were not investigated in the Georgiou, Having performed parameter estimation,
Georgakis, and Luyben study. the remaining problem in identification is to
A number of different identification assess the validity of the model. As noted
approaches were used in the Georgiou, previously, the issue of importance here is
Georgakis, and Luyben study. Linear trans- whether the model meets the intended pur-
fer function models were readily obtained pose, which is control system design. We
for the moderate purity column using con- briefly describe some of the commonly used
ventional pulse testing techniques (Luyben, validation techniques.
1990). However, at higher purities, nonlin- Simulation Simulation of the measured
earities make it difficult to obtain a suitable (y) versus predicted (9) output is the most
linear transfer function model. Estimated common validation tool in identification.
transfer function parameters can vary Adequately determining how "close" the
widely, depending on the size and direction predicted output is to the actual output,
of the perturbations to the inputs. Luyben however, is a more challenging task than it
(1987) proposed an identification procedure appears.
to overcome this difficulty in high purity Impulse or Step Responses Examining
distillation systems. In the proposed proce- the model's response to a standard input
dure, steady-state gains are obtained from such as a step or a ramp is a commonly used
detailed and highly accurate steady-state technique that allows an engineer to com-
simulation (rating) programs. Relay tests (as pare the results of identification versus per-
described in Section 7-3-4-2) are then ap- sonal process understanding. A model with
plied to the actual system to obtain the wrong gain sign or one with dynamic behav-
critical gains and frequencies for the diago- ior that contradicts the engineer's own intu-
nal elements of the plant transfer function itive understanding of the process will result
matrix. Individual transfer functions are ob- in repeating some step(s) of the identifica-
tained by finding the best transfer function tion procedure.
fit to the zero frequency and ultimate fre- Cross- Validation Generally speaking, if
quency data. The utility of this identifica- a model is validated using the same data
tion procedure for very high purity columns that were used for parameter estimation,
when combined with linearizing transforma- the fit can be expected to improve with
tions was demonstrated by Georgiou, increasing number of parameters in the
Georgakis, and Luyben (1988). model. To judge the quality of a fitted model
As discussed in Chapter 12, it is difficult in this circumstance requires a measure that
to predict in advance whether a linearizing accounts for both the decrease in the fit's
transformation will be helpful for a given loss function and the potential loss in pre-
column. A useful procedure for testing the dictive power of a model with increasing
value of a linearizing transformation for a number of parameters. One commonly used
particular column is to put competing mod- measure is the Akaike information theoretic
els on line and compare the variance and criterion (AIC) (see Ljung, 1987, Section
autocorrelation structure of their prediction 16.4). An alternate approach is to perform
errors over an extended period. cross-validation, in which a portion of the
120 Practical Distillation Control
experimental data set is retained for valida- section is that they are open-loop measures
tion purposes and not used in parameter for the goodness-of-fit. An increased under-
estimation. Simulating the model output standing of modelling for control design has
over the cross-validation data set allows an led to the conclusion that a model that may
engineer to assess the true predictive ability appear to be a good fit in the open-loop can
of the model. result in very bad closed-loop performance
Residual Analysis Residual analysis con- and vice versa (see ;\'str6m and Witten-
sists of testing the residual time series mark, 1989, Chapter 2). Clearly, there is a
against the hypotheses that the noise series need for control-relevant validation mea-
is Gaussian and uncorrelated with the input sures that determine the adequacy of the
series. If a nonunity disturbance model is model for control purposes. To this end,
*
chosen, that is, H(q) 1, then residual current research draws from robust control
analysis is performed on the calculated pre- theory and the structured singular value
diction errors. Residual analysis is discussed paradigm to develop these validation mea-
in detail by Box and Jenkins (1976). The sures (see Rivera, Webb, and Morari, 1987
measures that are used in residual analysis and Smith and Doyle, 1989). The key chal-
are the calculated autocorrelation lenge here, however, is obtaining appropri-
ate estimates of the plant uncertainty from
k) = cep(k) the residual time series.
Pep( 2 (7-59) The most conventional means for obtain-
(Fep
ing model uncertainty is to use 100(1 - a)%
and for the cross-correlation between resid- confidence intervals generated from the so-
ual and the plant input lution of the linear least-squares problem
Equation 7-38. These confidence intervals
can be computed from the identification
data set and an arbitrary plant input a(such
as a step or an impulse) as follows:
residual time series, as shown by Kosut where <1>" = an estimate of the distur-
(1987). Kosut uses spectral analysis to ob- bance power spectral density
tain an estimate of unmodelled dynamics.
This result is applied by Rivera et al. (1990) <1>,,( w) = <l>v( w) - 1!I2<1>i w)
and Rivera, Pollard, and Garcia (1990) to <l>v(w) = computed power spectrum
generate uncertainty bounds that capture for v
both bias and variance effects in the identi- 111 =degrees of freedom of the
fication. A brief summary follows. spectral estimator (which is
Consider plant residuals represented by proportional to the length of
the symbol v, which reflect both unmod- the data set)
elled dynamics and the effect of noise 1I(t): 12117-2 = the two-way Fisher statistic
, for a user-specified confi-
v(t) = (Go(q) - G(q»)u(t) + lI(t) dence level.
design, model structure selection, and pa- behaved to allow one input to be perturbed
rameter estimation that are pertinent when while holding the others constant, then this
the end use of the identified models is for greatly reduces the complexity of the esti-
process control. Ideally, one would consider mation problem, because SISO models can
the model requirements and trade-offs pre- be estimated one at a time (assuming the
sented by different choices of these design multivariable model Equation 7-4 is deemed
parameters and then proceed with the dy- adequate and no disturbance model is to be
namic process test and identification based identified). Determining the structure,
on the particular design parameters se- deadtime, and polynomial orders for a
lected. In practice, however, constraints transfer function model Equation 7-27 can
(process as well as management) sometimes be performed readily. The resulting transfer
make it impossible to conduct the desired function forms can be converted numeri-
process test. Prejudice for selection of a cally to truncated impulse or step response
certain model structure also affects the pro- form without any loss of information.
cess test, as described in the following text. Regardless of whether inputs are per-
The presence of disturbances and mea- turbed one at a time or simultaneously, it
surement noise directly affects the variance may become necessary to allow an operator
of estimated parameters. In general, the to make moves in other inputs in order to
size of perturbations applied to the inputs keep a process within an acceptable operat-
should be as large as permitted (to maxi- ing region. In either case the structural and
mize the signal-to-noise ratio) in order to parameter estimation problem becomes
reduce the required duration of the experi- more complex if a multi-input model rather
ment. In situations where measurement than a single-input model must be identi-
noise is low and a process operates rela- fied, particularly if the number of inputs
tively smoothly between disturbances that becomes large. The trade-offs that occur
occur only infrequently, a shorter duration between duration of experiment, signal-to-
experiment can be designed. In these situa- noise ratio (input amplitudes), model struc-
tions, a small number of step changes (per- ture (impulse vs. low-order transfer func-
haps combined with a relay test) will pro- tion) must now also include the experience
vide sufficient information to obtain an level of the person performing the identifi-
acceptable control model. However, this is cation. As dimensionality increases, the
only true if (low-order) transfer functions problem of satisfactorily determining dead-
are to be identified. If an a priori decision is times, structures, and polynomial orders for
made to statistically estimate truncated im- transfer function models becomes more
pulse (or step response) models, then this complex (particularly if the system exhibits
dictates that a long duration experiment be dynamic behavior that is more complex than
performed, to improve the variance of the simple first or second order with deadtime,
resulting estimates. As mentioned previ- as very typically happens). In these situa-
ously, a biased least-squares estimator might tions the probability of obtaining an incor-
also be performed with this model structure rect structure becomes higher. Depending
to obtain a more favorable bias-variance on the training and experience of the user,
trade-off. direct identification of impulse or step mod-
Other practical considerations may also els (that require only one parameter to be
influence the decision to identify truncated specified-the process settling time) may be
impulse models directly over transfer func- the best alternative. However, it must be
tion forms: dimensionality (number of in- recognized that as dimensionality increases,
puts and outputs) and the ability to conduct experiments of increasing duration are re-
tests on a multivariable process one input at quired. At some point, even the most ac-
a time. If the process is sufficiently well commodating operations superintendent
Identification of Distillation Systems 123
may not permit such a long test, and a more points with a time stamp so that segments
efficient approach will be required. of data corresponding to periods of upsets
The following are some suggestions for or unscheduled events can be identified
increasing the probability of performing a later.
successful dynamic test. The first suggestion
is to perform a "pre-test" on a system some
7·6 EXAMPLE
time before the actual dynamic test. The
pre-test allows one to identify potential As an example, an identification is per-
problems that may hamper (or postpone) formed on a moderate purity refinery de-
the actual test. Typical things to look for are propanizer. Feed to the column is the dis-
equipment problems (valves, transmitters, tillate flow from a debutanizer, which is
analyzers, etc.), inadequate tuning of regu- manipulated by the debutanizer pressure
latory controllers, equipment in "nonnor- controller. Flow rate of this stream is mea-
mal" operating mode or range (e.g., bypass sured, but its composition is not. Analyzers
valves), and functioning of the data-logging on the overhead and bottom streams of the
program(s), to name a few. The pre-test depropanizer provide composition measure-
also allows one to explain to the operations ments every 3 min. A multivariable control
staff what the dynamic test needs to accom- design resulted in three inputs: feed flow
plish (and to agree on how it should be rate (a feedforward variable), reflux, and
performed to ensure that quality and oper- reboiler steam flow rates (manipulated vari-
ating constraints are not violated). In addi- ables) and two controlled outputs: overhead
tion to information obtained from process and bottoms compositions. Discrete impulse
instrument diagrams, the pre-test can help models were required for the control pack-
to identify points that should be logged dur- age, one for each input-output pair.
ing the test. In addition to the points for All data analysis, model structure deter-
which models will be required, any point mination, parameter estimation, and model
that may potentially be used to verify nor- validation was performed using the System
mal operation (or conversely to later explain Identification Toolbox (Ljung, 1988) in
an unusual operating occurrence) should be MATLAB (1989).
logged. Points on the unit under test as well The data set for the identification con-
as points on upstream (or interconnected) sisted of 3 min samples taken over approxi-
units that have a direct effect on the test mately 60 h. Discrete impulse response
unit sh01;lld be logged. Redundant or use- models relating reflux flow, reboiler steam
less data can always be discarded, but miss- flow, and feed flow to overhead and bot-
ing data can, in the worst scenario, invali- toms composition were obtained as follows.
date a test. Data from the pre-test can also The entire data set was used initially to
provide information on the system dynam- obtain least-squares estimates of truncated
ics, which allows one to design the pertur- impulse models relating each input -output
bation signals for the subsequent dynamic pair, which were converted to step response
test. models for examination. From plots of the
During the actual test it is important to step response models, approximate dead-
have an operations engineer monitor the times and model orders were determined. A
system to record any events (e.g., opening of smaller portion of the data (the portion
a bypass valve, upset or shutdown of an considered to be most reliable) was then
upstream unit, flushing of a heat exchanger, used to estimate low-order transfer function
tuning of a controller, putting a controller models [of MISO output error (OE) struc-
in manual, unscheduled maintenance oper- ture], using trial and error search of values
ation, etc.) that might invalidate a segment of deadtime and model polynomial orders
of test data. The data logger should record around those determined from the esti-
124 Practical Distillation Control
mated impulse models. The best model was tain level of skepticism due to the abnormal
selected as the one that minimized the operating condition).
Akaike information criterion (AlC) and that Due to exceptional operating problems
satisfied model adequacy criteria based on and disturbances in the debutanizer, feed
cross-correlation analysis (between the flow to the depropanizer experienced
residuals of the fitted OE model and the greater fluctuation than normal during the
corresponding inputs), cross-validation us- dynamic test period. This was not entirely
ing a segment of the original data set that undesirable because it excited the column
was held back from the estimation, and for identification of models relating feed
examining the poles and zeros of the fitted flow rate and the response variables. It
model (if the 95% confidence region of a sometimes can be difficult to arrange inten-
pole of the transfer function for a particular tional perturbation of feedforward variables
input-output pair overlaps with a zero of (because this usually means upsetting an
the same transfer function, it is over- upstream unit). Because of the wide swings
parametrized). Disturbance models were not in feed flow rate during the test, operators
identified because the control package used did occasionally have to manipulate reflux
assumed disturbance models (randomly oc- and reboiler steam rates to maintain opera-
curring deterministic steps), and it was not tion within quality constraints as much as
considered necessary to modify it for this possible. As described earlier in the chap-
particular application. Nevertheless, simul- ter, when data for identification are ob-
taneous identification of process and distur- tained when feedback is present (inad-
bance transfer function models can be read- vertently, as may have occurred in this case,
ily accomplished using the available fitting or due to a controller that is on-line), it is
routines in the toolbox. recommended that external perturbations be
All data were normalized by subtracting applied somewhere in the feedback loop.
off sample means (see function detrend in During the depropanizer test, random per-
the toolbox). Variance normalization (to unit turbations were applied to reflux and reboil
standard deviation) can also be performed steam flow rates to as great a degree as
(for an example, see Section 5.4 of the possible, considering the unusual operating
MATLAB manual). conditions. However, these conditions did
Only the identification of MISO models preclude applying preprogrammed PRBS
for bottoms composition is shown here. Sim- signals to the inputs. As well, it was impos-
ilar procedures were followed to obtain sible to perform a dynamic test in which
overhead composition models. only one input at a time was perturbed
Normalized test data are shown in Figure (while holding the others constant), which
7-5 (a: bottoms composition (% C3); b: would have simplified the model estimation
reflux flow rate; c: reboiler steam flow rate; stage of the identification.
d: feed flow rate). Several additional operating problems
Early in the test it was realized that the arose during the dynamic test that were not
column was overrefluxed (and reboiled). apparent in the pre-test, which had been
Over roughly one shift (sample number 125 conducted several weeks earlier. As shown
to sample number 275-approximately 7.5 in Figure 7-5b, a problem in the reflux con-
h), both reflux and reboiler steam flow rates trol loop caused high-frequency oscillations
were gradually reduced to bring the column in the reflux flow. Tuning of the PID flow
back to a more normal operating region. controller did not resolve the problem.
These data were retained for preliminary Troubleshooting for mechanical problems
estimation of the impulse models (to obtain in the loop hardware also did not immedi-
rough estimates of dead times and model ately solve the problem. Because the oscilla-
orders) and for cross-validation (with a cer- tions were high frequency, and would sim-
6
3 f\
h
r
~
2
I \
)
0
-1, v~
-2
l(I II
V V\
·3
V
V
~ VV \j
sample number
(a)
800
400
200
·200
-400
-6000
200 400 600 800 1000 1200
sample number
(b)
FIGURE 7·5. Depropanizer test data. (a) bottoms composition; (b) reflux flow rate;
(c) reboiler steam flow rate; (d) feed flow rate.
6~--~----~----~--~r---~-----r----'-----r---~-----,
3
•g
.=•
1>0
2
••
1 1
-1
-20
5 10 15 20 2S 30 35 40 45 SO
. (a)
1.5
0.5
••
1 0
-0.5
5 10 15 20 2S 30 35 40 4S so
(b)
DGURE 7·7. Estimated impulse models. (a) bottoms composition/reflux; (b) bot-
toms composition/reboiler; (c) bottoms composition/feed.
Identification of Distillation Systems 127
ply be filtered by the process, and it was still turbances, models identified for use in pro-
possible to make setpoint changes to the cess control need not be accurate at low
reflux flow controller, it was decided to pro- frequencies (Le., near steady state). The
ceed with the test while troubleshooting need to obtain a high quality estimate of
continued. A mechanical problem was fi- steady-state gain for control models is
nally resolved near the end of the test overemphasized in the identification folk-
(around sample number 975). lore. Differencing data will also remove
Power spectra were generated for all in- nonstationarity (low-frequency drift), but
puts using the spa function in the toolbox to will result in fitted models in which bias at
determine the useful frequency range of the higher frequencies is reduced (at the ex-
input perturbations. The spectrum for re- pense of accuracy at lower frequencies that
flux flow rate is shown in Figure 7-6. Exami- are of interest for process control) because
nation of the frequency functions (Le. Bode taking the derivative of a signal distorts its
plots) obtained from spectral analysis (gen- spectrum (Le., amplifies power) at higher
erated by the spa function) and the Bode frequencies.
plots of the final estimated models indicated For systems that do not exhibit strong
that the input perturbations did not contain gain directionality (as is the case with this
sufficiently high frequency information. The moderate purity tower), input perturbations
resulting control models would be consid- should be independent (noncorrelated).
ered adequate only if relatively sluggish Cross-correlation functions for the input se-
controller tuning proved to be adequate. quences were generated and plotted (not
Power spectra for the other inputs were shown) using the routine xcov in the Signal
similar. Processing Toolbox in MATLAB. No sig-
Low-frequency drift in the data should nificant cross-correlation between the in-
also be removed by prefiltering (for exam- puts was found.
ple, using the filter function idfilt in the Preliminary least-squares estimates of the
toolbox as a high-pass filter). Because any impulse responses for each input-output
controller with integral action will be able pair were obtained by fitting a three-
to regulate effectively against very slow dis- input-one-output ARX model [Equation
SPECI'RUM
frequency
I8mple number
( c)
IS00 r----------r---------,----------,---------~----------r_--------,
-500
-1000
1.S .----.----.----.----.----.-----,.------',.-----,.-----,.-----,
..
O.S I
".'"
....
0
...
co
~
0
H
~
( c)
7-31, with na = 0, i.e., A(q) = 1 and nbi = will be low]. The response of bottoms com-
50, the order of Bj(q) for input i, i = 1,2,3] position to reboiler steam flow rate rate
using the toolbox routine arx. Alternately exhibits inverse behavior, suggesting that the
the toolbox routine era, which obtains im- order of B/q), i = 2 (reboil), for a transfer
pulse models by a cross-correlation tech- function model should be at least 2.
nique, could have been used. The estimated Low-order transfer function models were
impulse models are shown in Figure 7-7. identified using sample numbers 630 to 1193
The toolbox routine idsim was used to ob- (approximately 28 h). Because no distur-
tain the corresponding step response mod- bance models were to be estimated, an out-
els (by simulating the response of the esti- put error (OE) structure was selected for
mated three-input-one-output impulse identification. An ARX structure could also
model to a unit step change applied to each have been used (Equation 7-31), but as
input, one at a time). shown by Equation 7-29 this structure im-
The unit step response models are shown plicitly assumes a disturbance model of the
in Figure 7-8. From these plots it was deter- form H(q) = 1/A(q). This is undesirable
mined that the responses of bottoms com- mainly because H(q) = 1/A(q) results in a
position to reflux, reboiler, and feed flow fit in the frequency domain that emphasizes
rate each has a deadtime of approximately reducing bias at high frequencies, as shown
five sampling periods (15 min). Each of the by Equations 7-15 and 7-19 [with H(q) =
responses appears to be low order [the or- I/A(q), the quantity 1/IH(e jwT )1 2 becomes
der of F(q) for a transfer function model larger at higher frequencies, thereby result-
Xlo-'
8
.
~
....
<=
0 4
:l
.'"...
.....
~
3
;§
2
-1 0
5 10 15 20 2S 30 35 40 45 SO
(a)
-1
-2
..
...<=
0
-3
:
II
...
.
.
~
-4
'"
....
g
-5
-6
-7
-80
5 10 15 20 2S 30 3S 40 45 SO
(b)
FIGURE 7-8. Unit step responses obtained from estimated impulse models. (a)
bottoms composition/reflux; (b) bottoms composition/reboiler; (c) bottoms composi-
tion/feed.
Identification of Distillation Systems 131
nO"'
8
.u
"
'"..
0
4
:!
..'"
.....
3
'"
:§
2
-1 0
5 10 15 20 2S 30 35 40 45 50
(c)
B2(q)
+ F 2( q) u 2(t - nk2)
nbt = 1 bu = 0.57 X 10- 3 (± 0.11 X 10- 3)
B3(q)
+ F3(q) u 3(t - n k3 ) + e(t) (7-66)
nfl = 1 111 = - 0.92 (±0.014)
132 Practical Distillation Control
Bottoms composition/reboil steam flow calculated using the toolbox routine th2zp
rate: and plotted (along with their 95% confi-
dence regions, which are not shown) using
n k2 = 4 the toolbox routine zpplot. Possible redun-
dancy in poles and zeros was only a concern
nb2 =2 b21 = 1.03 (±0.64) for the transfer function relating bottoms
b22 = - 1.34 (±0.74) composition and reboil steam rate because
this input-output pair had two poles. The
n/ 2 = 2 121 = - 1.59 (±0.15) pole-zero plot obtained from zpp/ot for
this input-output pair showed no overlap-
122 = 0.634 (±0.14) ping of the 95% confidence regions for the
poles and the zero, and therefore over-
Bottoms composition/feed flow rate: parametrization was not suspected.
A plot of the residuals for the fitted
nk3 = 5 three-input-one output OE model is shown
nb3 = 1 b31 = 0.98 X 10- 3 (±0.07 X 10- 3 ) in Figure 7-9. The autocorrelation function
for these residuals (Figure 7-10a) shows sig-
nf3 = 1 131 = - 0.86 (±0.015) nificant values (Le., outside the 95% confi-
dence limits given by the dotted lines), which
None of the confidence intervals encom- is expected because a disturbance model
passes zero, indicating that all the estimates was not estimated. At this point a distur-
are statistically significant. bance model could be identified by adding
The poles and zeroes of the transfer the polynomials C(q) and D(q) (Equation
function for each input -output pair were 7-27) in the toolbox routine pem and deter-
1.5
0.5
'iii
... ~
'tI
III
0
...
II
-0.5
-1
-1.5
0 100 200 300 '00 500 600
sample number
FIGURE 7·9. Residual plot for estimated OE models.
Identification of Distillation Systems 133
1~~------~---------r--------~--~~--~r-------~
Correlation function of residuals. Out ut * 1
0.5
-0.5~--------~--------~--------~----------~------~
o 10 15 20 25
lac
(a)
=
. . Cross
.=. . .=."".~. . . =. . ."=corr. . . . .=
. . =. . .~. . =
. . =.between
. . . .=
. . ~". "=. . .=
. . "=.input
. . 7....".=. . .=....1"=. . .and residuals from out ut 1
0.4r=
...... . . ""=. . .r.""="function
...= . . .= . . .= . . .~. . . =_.=
=
=."
... . ". =. . . .=..".~. _=."_.=. =. _~."_=._.=. __=. ~.~m=". _=. . .~. "_-.. , ="
...
0.2
o ---------------------~-----------------
-0.2
- 0 . 4 · ' · " · ' · ' ...........................
-25 -20 -15 -10 -5 o 5 10 15 20 25
lac
(b)
0.4r=~~~~~~~.~ =. . . . . . . . . . . . . . . . . . . . . . . . . . .
Cross corr. function between input 2 and residuals from out ut 1
.. .. ..... ...... ...... ~-~~.~~~~-~.~~~.=.*-~~~=.~====~==~~--~~~
0.2
o
-0.2
..=
. . .=. .=. . =
. .=
... =
.. =
... =
... =
. .=
.. =
----
Cross corr. function between input 3 and residuals from out ut 1
0.4 ~=:::;;:.::.= .....=.. . . . =.. . . =.;:::
:r.: ...=.::. . . . =_.:::
. . . :r.:.=
. . . .=.::
....... :=.!.: . . .:=. . . . =_..:.:._=. . _=.::r.::::... _=_.::::::;:.=.._=_...=. ._~..!:::%..•
.....:r..:::. . =_=. . .=
. . .:.:.:-,
----
0.2
o ---------~--------------------
-0.2
-0.4 L...::;;===".;;;;; . . .;;;;;. .;;;;;. . .;;;;;...'.::;L.'••;;;;;••••••;;;;;.•••;;;;;•••;;;;;•••••;;;:l.... ".....";::......;::......".•...."".....•""••••";;;;;..... ;;;;;;•••.;;;;;;.•_.;::•.•..•;;;0•• ""••.•""•••...""_.•""......;;;;;..•••""-;::...• =;.;.;.....0
.. ;;;;;. .;;;;;. . ."". . .;;;;;. . . ;;;;;. . .;;;;;. . . ;;;;;. . . ::;L•••.•;;;;;•••• =="".;;;;;
-25 -20 -15 -10 -5 0 5 10 15 20 25
lac
(d)
FIGURE 7-10. Auto- and cross-correlation functions for estimated OE models. (a)
Residual autocorrelation function; cross-correlation functions between residuals and
inputs; (b) input 1 (reflux); (c) input 2 (reboiler); (d) input 3 (feed).
134 Practical Distillation Control
6.--------r----.--,------"r--------r------~r_----__.
-2
-4
-6~-------L--------~------~--------~------~--------~
o 200 400 600 800 1000 1200
sample Dumber
:dO-I
7
~ 6
Ii:
CD
'"
" 5
~
0
...
Col
0 4
.0
CD
III
~
0
3
Qo
III
CD
Iliil
2
.s
Qo
rIl
~
;::l
1
0
0 10 20 30 40 50 60 70 80
parameter number
(a)
2
1
==0
.0
CD
'" 0
"a
Qo
-1
0
...
Col
0 -2
.0
CD
III
-3
"
0
It
CD
l1li -4
.s
Qo
rIl -5
=
~ -6
-7
0 10 20 30 40 50 80 70 80
parameter number
(b)
FIGURE 7·12. Unit step responses obtained from estimated output error model. (a)
bottoms composition/reflux; (b) bottoms composition/reboiler; (c) bottoms composi·
tion/feed.
136 Practical Distillation Control
dO-S
8~------~------~---------'-------~---------~-------r---------~------~
'tI 7
qj
~
........
a
6
..
0
.D
0
CI
5
4
a
qj
0
~
III 3
Gl
c:
.s
.
~
2
fI)
·a
p 1
0
0 10 20 30 40 50 60 70 80
parameter number
( c)
ing it more difficult to detect model inade- one-output OE model using the toolbox
quacy. routine idsim. For brevity, only the unit step
Recall that the model was estimated us- response models are shown here, in Figures
ing sample numbers 630 to 1193. For cross- 7-12 (a: bottoms composition/reflux; b:
validation, we are interested in the predic- bottoms composition/reboiler; c: bottoms
tive ability of the model on a segment of composition/feed). These model responses
data that was not involved in the estimation agree with our physical intuition of the pro-
(in this case, sample numbers 1 to 629, cess (settling times, sign of the steady-state
Figure 7-11). The model appears to predict gains, and dynamics, particularly the inverse
dynamic response very well for these data, response present in the transfer function
despite the abnormal operating condition, relating bottoms composition to reboiler
over a range of frequencies (fast dynamic steam flow rate, which was attributed to the
changes are predicted as well as slower dynamic response of liquid level at the bot-
changes), even during the initial period when tom of the tower).
reflux and reboiler steam rates were stepped
down to a more normal operating region.
As mentioned previously, due to the lack of
7-7 NOMENCLATURE
higher frequency excitation in the input per-
turbations, these simulation results do not II . 112 two-norm objective func-
provide useful information on the predictive tion' lIell2 = (Ek~oeD1/2
ability of the model at higher frequency. a Speed-of-response increase,
Discrete impulse and step response mod- PRBS design;
els for each input-output pair were ob- e-T,/Tdom, used in prefilter
tained from the estimated three-input- design;
Identification of Distillation Systems 137
140
(a) Summary of Column Control Sensor column may come from sensors not located in the
Issues column. In designing a column control system it is
important to take a more complete plantwide per-
Sensor Location: Distillation columns are distributed spective. The problems with the operation and con-
processes. Temperature, pressure, and composition trol of the column may best be corrected by better
vary throughout the column. Choosing the proper control elsewhere in the plant. Care should also be
location of control sensors is critical to the perfor- given that the local column control strategy does
mance of the column control system. The locations not inadvertently cause serious problems to other
that best facilitate the control system are highly unit operations downstream.
specific to the column and to the separation it is
designed to maintain.
Analyzer Type: Information that reflects the state of
the separation is necessary for all column control (b) Summary of Column Control Valve Issues
strategies. Compositions are sometimes measured
directly using process analyzers but they are most Range: The valve range is the maximum flow attainable
frequently measured indirectly using temperature. with the valve fully opened. It represents a limita-
The effectiveness of temperature-based sensors in tion on how large an adjustment a control system
controlling composition varies greatly from column can make. Such physical limitations need to be
to column. It is important for the control system considered in both the column design and in the
designer to realize that controlling temperature design of its control systems.
does not always solve the underlying composition
control problem. Resolution: The valve resolution represents a physical
limitation on how small an adjustment a control
Analyzer Focus: The degrees of freedom of most multi- system can make. Typically, valve resolutions are
component distillation columns suggest that it is not 2% or greater. If more subtle adjustments are nec-
possible to control the composition of all the com- essary the control system will struggle to find a
ponents. The choice of analyzer locations and which steady state.
component composition to determine at each is
important to the success of the column control Linearity: Typically, installed valves are not linear with
strategy. respect to the relationship between flow and signal
from controller. Highly nonlinear valves can cause
Sensor Sensitivity: The zero and span of the sensor are problems in control loops. The effect of the nonlin-
important to the performance of the control sys- earity is especially difficult in control loops that
tems. Sensors with too broad a span lack the resolu- have slow dynamics.
tion and sensitivity to meet the control objectives.
Sensors with too narrow a span tend to operate off Size: The size of a control valve is set by its CV.
scale and function in the control system more like Typically, control valves are sized too large for the
two position relays. flow service. In such cases the problems of linearity
and valve resolution are greatly exaggerated.
Sensor Consistency: One of the main purposes of a
column control system is to stabilize the operation Dynamics: The dynamics of the distillation column is
of the column. If the control sensor is not stable in an important issue. The nature and speed of re-
its reading such local variation can be propagated sponse is different for each valve and is a function
by the control system to other parts of the column of the role of the valve in the column control
operation. This problem is particularly severe in strategy.
complex multivariable control strategies in which
errors in a single sensor measure can affect the
Role: The local role of each control valve is a major
issue. Is it to be set by the inventory control sys-
movement of a number of manipulated variables.
tem? Is it to be set by the separation control
Sensor Reliability: Sensors that fail can cause very se- system? Is it to be set manually by the control room
vere problems in a control system. It is important operator?
for the control system designer to know the failure
probability of key sensors. It may be critical to
Plantwide Implications: A distillation column is typi-
cally only one of many unit operations that make up
design a control structure that degrades gracefully.
a processing plant. Manipulation of column control
It may also be important to design fault detection
valves affects more than the local operation of the
and sensor validation schemes that alert the control
column. The plantwide implications of changes in
system to the occurrence of such failures.
each control valve need to be considered thor-
Non-Column Sensors: It should be realized that some oughly. Such implications should strongly influence
of the most useful information in controlling the the design of the local control system.
141
142 Practical Distillation Control
the separation in the column and they help Loop sensitivity is a measure of how the
establish how the column interacts with control sensor responds at steady state to
other unit operations on a plantwide basis. changes in the manipulated variable. Math-
ematically, the loop sensitivity can be de-
fined as the partial derivative of the sensor
8-2-2 Separation Control Concerns signal with respect to changes in the signal
to the valve:
The second level of the column control ad-
dresses controlling the separation taking
place in the column. At this level the main as·
valve issue is deciding the role of the re- loop sensitivity = - '
aMj
maining valves (Le., valves not used in the
first level strategy). The control engineer
must decide which of the valves available where Si = signal from control sensor i,
for separation control are to be manipu- expressed as a percentage of
lated by the control system and which are to the maximum signal
be set manually by the control room opera- Mj = signal to control valve j, ex-
tor. pressed as a percentage of the
Perhaps the most confusing aspect of the maximum signal
second-level control concerns the sensors
used to monitor the state of the separation Both valve and sensor have a range be-
taking place in the column. There are three yond which they cannot operate. The units
fundamental sensor questions to be ad- of the loop sensitivity should be chosen to
dressed early in the design of the separation clearly reflect these physical limitations.
control strategy: Such limitations are included if the loop
sensitivity is expressed in percentage units
What kind of composition sensor? indicated in the preceding text. The numer-
How many measurements? ator is the change in the sensor reading
Where in the column should they be lo- expressed in percent of sensor span. The
cated? denominator is the change in manipulated
variable expressed in percent of the range
of the signal to the valve.
These sensor questions are highly specific to Measured in such percentage units, the
the column and to the separation taking ideal loop sensitivity is 1% per percent. In
place in the column. They must be an- such cases, if the control system is well
swered before the control strategy can be balanced a 1% change in the signal to the
properly designed.
control valve will result in a 1% change in
the sensor. If the sensitivity is either signif-
icantly larger or smaller than 1, control
8-2-3 Loop Sensitivity Issues problems are likely to develop. These prob-
Loop sensitivity issues are important consid- lems can seriously diminish the effectiveness
erations in selecting valves and locating sen- of a feedback control loop and should be
sors. Although such issues are fundamental understood and considered in the selection
to both levels of column control, they are and location of all column control sensors.
frequently overlooked (or ignored) in the
design of control strategies. Failing to prop- High Sensitivity
erly understand the importance of sensitiv- High loop sensitivity can be a problem in
ity issues can lead to serious operational distillation control. Control systems highly
and control problems. sensitive to changes in the manipulated
Selection of Controlled and Manipulated Variables 143
variables tend to overcompensate. The con- of the flow range of the valve. Also, the
trol system is inclined to "hunt" (without effective valve resolution is greatly reduced.
much success) for a valve position to satisfy If the flow for a linear valve can be posi-
the setpoint. In such cases, steady-state op- tioned ± 1.0% of the maximum, the over-
eration is seldom achieved. sized valve can be positioned to only
There are two aspects of the high sensi- ± 10.0% of the same maximum. The solu-
tivity problem: one concerns the sensor and tion to the problem of an oversized valve
the other concerns the actuator. The sensor can be easily found by changing the valve
problem occurs when the control sensor has trim to a more appropriate size.
a narrow span compared to the range over The problem of high sensitivity may well
which the process variable can be manipu- be the result of the separation. High purity
lated. In such cases, even small changes in separations typically have temperature pro-
the manipulated variable will cause the sen- files that are flat at both ends but break
sor to saturate. During disturbances, most sharply in the middle of the column. The
of the time the sensor will present to the sharp break indicates where the primary
control system either the maximum or the separation of the components takes place;
minimum value of its output. In effect, it the flat portions represent where the prod-
will function in the control loop much like a ucts are being purified. The sensitivity is
two position relay. very low except on the few trays where the
The problems of high sensitivity also can separation is taking place. In such cases the
be associated with the manipulated vari- problem can be reduced by using multiple
able. Valves and other actuators all have a sensors. Temperature sensors, well placed
minimum resolution with respect to posi- throughout the sensitive section of the col-
tioning. These limitations restrict the fine umn, can be averaged to provide a net feed-
adjustments often necessary for high gain back signal that has lower gain and better
processes to reach a steady operation. If the linearity.
fine adjustment necessary for steady state is
less than the resolution of the valve, sus- Low Sensitivity
tained oscillation will likely occur. Consider, Low loop sensitivity is a more common
for example, a steam valve with a resolution problem in distillation control. In such cases,
of ± 1.0%. If a valve position of 53.45% is even large changes in the signal to the con-
necessary to meet the target temperature, trol valve will have little effect on the signal
then the valve will, at best, settle to a limit from the control sensor. For example, a
cycle that hunts over a range from about 55 loop sensitivity of 0.05% per percent indi-
to 53%. If the process gain is 10, the hunt- cates a change in the valve position of 10%
ing of the valve will cause a limit cycle in will result in only a 0.5% change in the
the control temperature of 20%. sensor. The most obvious problem caused
There are many considerations that by such low loop sensitivity is leverage. The
should be balanced in trying to improve a control system is severely limited with re-
poor loop sensitivity. The problem can be spect to the magnitude of errors that can be
addressed to some extent by resizing the corrected by the valve. The magnitude of
sensors and/or valves. For example, a com- change necessary to correct for an error
mon problem in distillation control is over- may well exceed the capability of the con-
sized control valves. The actuation charac- trol valve. When valves saturate (fully open
teristics of a typical oversized valve are or fully closed), the control objective cannot
shown in Figure 8-2. Note that 90% of the be met until the error returns to a more
maximum flow is achieved when the signal manageable magnitude. For processes with
to the valve is only 10% of its maximum. low sensitivities, the range of manageable
The result is high sensitivity through most errors may be very small.
144 Practical Distillation Control
100
so.o -++-++f++++f++++I-+++-H++t+tffiH-ttt1-tt-t11~r'l'tt11H"ttt1
Oversized Valve
20.0 -+f-+++-H-H-,*,,++If++++-I-++t+t++-I+tt+1H-tt+t++t-t+~I-+++Tt
Positioner
Resolution
0.0
0.0 20.0 40.0 60.0 80.0 100
Signal to Valve
(%of Max)
The second problem caused by low pro- its saturation limits. From an operations
cess sensitivity concerns the effect sensor perspective, such valve action is universally
noise has on loop performance. In order to unacceptable.
compensate for the low process sensitivity, It should be noted that the problem
the feedback controller gains are typically caused by noise for processes with low sen-
set high. Such high controller gains will sitivity is even more pronounced in control
cause problems if the control system is con- systems using multivariable control algo-
nected to a sensor containing noise in its rithms. In such systems, the noise is ampli-
signal. (The only sensors the author can fied and reflected to more than one manipu-
think of that do not contain noise are the lated variable. Problems that were only of
ones used in typical academic simulations.) local concern in a conventional control sys-
A high gain controller amplifies and reflects tem can be inadvertently propagated
sensor noise to its manipulated variables. throughout the entire system by the multi-
Consider, for example, a process sensitivity variable algorithm.
of 0.01. For such a process, a reasonable The problem caused by the noise can be
value of controller gain might be 100.0 (this reduced by either lowering the controller
value provides a closed loop gain of 1.0). In gain or adding filters. Lowering the con-
such cases, sensor noise as small as ±0.5% troller gain reduces the sensitivity of the
will cause the valve to swing wildly between control loop to noise; but, it also reduces
Selection of Controlled and Manipulated Variables 145
~ STEAM
8~--------------------------------~
6 ~
...>-
~
.....
en
4 •• dT/dO
dT/dQ
c
C1>
(/) ......
....
C1>Qe
L.. ......
::JU
tV ......
L.. 2
C1>
Q.
E
C1>
~
~.
~
0
i
I
I
I
I
I
I
-2
0 10 20 30 40 SO
Tray Number
FIGURE 8-5. Ethanol-water temperature gain matrix plot.
these locations would not be good choices. rithm called principal component analysis
Temperatures on adjacent (or nearly adja- are discussed in many text books (Forsythe
cent) trays will be highly correlated and and Moler, 1967; Golub and Reinsch, 1970;
therefore inadequate choices of control sen- Lawson and Hanson, 1978), and software is
sors. With respect to the problem of inter- readily available through a number of stan-
action, a better sensor location is near the dard numerical packages (for example,
column extremes. For example, locating MATLAB, 1985).
sensors on tray 1 and tray 50 would yield Singular-value decomposition was devel-
much less interaction. The problem with oped to minimize computational errors in-
this choice is sensitivity. As can be seen in volving large matrix operations. When
Figure 8-5, the sensitivities of all trays near applied to a steady-state gain matrix the
the top and the bottom of the column are component matrices provide insight into the
extremely low. multivariable control problem. To under-
Selecting sensor locations for a multivari- stand the use of SVD in addressing the
able control strategy requires balancing the sensor placement problem, it is helpful to
problem of sensor interaction with the prob- consider the physical interpretation of these
lem of loop sensitivity. The gain matrix can component matrices presented in Figure 8-7.
provide insight into both problems; how-
ever, further analysis is required. Singular- Sensor Placement Based on U-Vector Plots
value decomposition (SVD) provides a basis The U and K matrices are both measures
for further analysis. The SVD analysis and of sensor sensitivity. The geometric proper-
an approach to select multiple sensor loca- ties of the U matrix provides subtle insights.
tions are described in the following sections. They are not obvious in the gain matrix
alone. The U vectors present tray sensitivi-
ties in a orthogonal coordinate system.
SVD Analysis of the Gain Matrix These are useful in considerations of both
Singular-value decomposition (SVD) is a sensor interaction and sensor sensitivity. The
numerical algorithm useful in analyzing the analysis and interpretation of the U vector
multivariable aspect of the gain matrix. It is as follows:
decomposes a matrix into three unique
component matrices as UI indicates the most sensitive combination
of tray temperatures in the column. The
principal component of the U1 vector is
the most sensitive location for a tempera-
where K = an n X m matrix ture sensor in the column.
U = an n X n orthonormal matrix, U2 indicates the next most sensitive combi-
the columns of which are nation of tray temperatures, orthogonal
called the left singular vectors to the tray temperatures represented in
V = an m X m orthonormal ma- U1• The principal component of the U2
trix, the columns of which are vector represents location sensitivity but
called the right singular vec- exhibits the least possible interactions
tors with the primary sensor.
I = an n X m diagonal matrix of
scalars called the singular val- It should be noted that although the U
ues and organized in descend- vectors are noninteracting (by definition),
ing order such that 0"1 ~ 0"2 ~ the locations indicated by the principal
••• ~ O"rn ~ 0 components of each of the U vectors may
exhibit considerable interaction. Unfortu-
SVD and a closely related statistical algo- nately, it is seldom possible to select two
150 Practical Distillation Control
U. U2
- 0.0001 -0.0009
K .. BT;/BM.: BT;/BM2: ..• :B1j/BMm -0.0002 -0.0011
The gain matrix is composed of a set of physically -0.0002 -0.0014
scaled vectoR that describe the steady-state tem- -0.0003 -0.0016
perature sensitivity on each of the n trays to changes -0.0003 -0.0019
in each of the m manipulated variables. -0.0004 -0.0023
-0.0004 -0.0027
U = U.: U2 : ••• :Un -0.0005 -0.0031
The left singular VectOR are the most important -0.0006 -0.0036
component matrix. It provides a sensor coordinate -0.0008 -0.0041
system for viewing the sensitivity of the column to -0.0009 -0.0047
changes in the manipulated variables. This coordi- -0.0011 -0.0054
nate system is such that U. indicates the vector -0.0013 -0.0061
direction that is most responsive to changes in the -0.0016 -0.0069
manipulated variables. U2 indicates the next most -0.0020 -0.0079
responsive vector direction. The third vector is the -0.0025 -0.0090
next most responsive, and so on. The properties of -0.0031 -0.0102
U provide that each vector direction is orthogonal -0.0038 -0.0115
to the other vector directions. -0.0049 -0.0131
-0.0063 -0.0149
v = V.:V2 : ••• :Vm -0.0082 -0.0170
The right singular vectOR are not directly used in
-0.0110 -0.0195
the analysis of the sensor placement problem; how-
-0.0150 -0.0223
ever, its interpretation is helpful in completing the
-0.0209 -0.0255
physical picture provided by SVD. The right singu-
-0.0298 -0.0291
lar vector provides a manipulated variable coordi-
-0.0434 -0.0332
nate system for viewing the sensitivity of the col-
-0.0647 -0.0375
umn. This coordinate system is such that V. indi-
cates tbe combination of control actions that most
-0.0984 -0.0419
-0.1530 -0.0458
affect the column, V2 indicates the next most effec-
-0.2436 -0.0492
tive combination of control actions, etc. The prop-
-0.3982 -0.0568
erties of V are such that the vector directions are
-0.6121 -0.0754
orthogonal.
-0.5583 -0.0023
I = diag(O"., 0"2' •.. ,O"m> -0.2237 0.1733
The scalar singular value provides an indication of -0.0338 0.2104
the multivariable gains in a decoupled process. This -0.0323 0.3240
information is useful in evaluating the difficulty of a -0.0292 0.3909
proposed dual-ended control system. -0.0250 0.4101
-0.0204 0.3904
FIGURE B-7.
-0.0160 0.3456
-0.0121 0.2892
sensors totally free of interaction. This pro- -0.0089 0.2317
cedure does identify two sensitive locations -0.0064 0.1792
'with little interaction relative to other -0.0046 0.1346
choices. -0.0032 0.0987
-0.0022 0.0706
Consider the results of the singular-value
-0.0015 0.0495
decomposition of the ethanol-water gain -0.0010 0.0339
matrix as shown in Figure 8-8. The U matrix - 0.0007 0.0226
can be considered directly, but a graphic -0.0003 0.0100
analysis presents a clearer picture of the 0"2 V. v2
sensor selection problem. In Figure 8-9 the ... ] [-0.99996 0.00891]
elements of the U vectors are plotted versus 0.12738 . 0.00891 0.99996
tray locations. The maximum absolute val- 1Condition number = 84.81
ues of the U vectors indicate the locations
of the principal components. These are gen- FIGURE 8-8. SVD analysis of the ethanol-water
erally good locations to consider in placing temperature gain matrix.
Selection of Controlled and Manipulated Variables 151
0.6.,..-------------------.,
0.4
• UI
• U2
0.2
en
L-
....u
0
0.0
Q)
>
L-
IV
::J
01-0.2
c
r.J')
-0.4
2nd Sensor
!I
-0.6
Primary Sensor
Location
-0.8 +r,.,.,r'T"I"''I''''I''''IPT''I'',.,.,r'T''I''''I''''I''''I"....,.,.,,......,.,.,~'''",......,.,.,~'''"~,.,.,~
o 10 20 JO 40 50
Tray Location
FIGURE 8·9. U vector plots for the ethanol-water column.
the control sensors. Likewise, in this exam- considers the individual sensor interactions
ple trays 19 and 13 can be chosen for plac- in more detail.
ing temperature sensors. Note that in this example the results of
A modified version of the principal com- the two methods are similar except for the
ponent analysis is shown in Figure 8-10. It location of the primary sensor. In the modi·
consists of only one function defined by the fied analysis the primary sensor was se-
difference between the absolute values of lected to be at tray 18 to reduce the interac-
the elements of the U vectors: tion that exists between the two vector com-
ponents at tray 19.
applied to the control problem, the answer multivariable gains are well balanced and
is seldom clear. Nearly all columns have 2 should result in a system with sufficient de-
degrees of freedom, but they may be too grees of freedom to meet the dual-ended
narrow for a control system to successfully control objectives. On the other hand, high
regulate. Singular-value decomposition pro- condition numbers evidence an imbalance
vides insight into the pragmatic side of the in the multivariable gains. Imbalance signi-
degrees-of-freedom ,question through the fies the column to be much more sensitive
condition number. in one vector direction than it is in the
The condition number (CN), in general, other. Consequently it may have difficulty
is defined as the ratio of the largest singular meeting the dual-ended control objectives.
value to the smallest singular value. For the At the extreme values of the condition
dual-ended control problem, it is defined as number the results are unambiguous: A
the first singular value divided by the sec- condition number of 1 (very unlikely) clearly
ond singular value: indicates two full degrees of freedom,
whereas a condition number of infinity indi-
eN = O"largest = 0"1 cates only 1 degree of freedom. What about
O"smallest 0"2 the condition numbers between 1 and infin-
ity? Consider, for example, the ethanol-
It provides a numerical indication of the water column. The SVD analysis yielded a
balance of sensitivities in the multivariable condition number of 84.8. This indicates
system. Low condition numbers indicate the that the column is 84.8 times more respon-
0.6
0.5
0.4
0.3
N
::>
..... 0.2
(/)
10
<
,....I 0,1
::>
.....
(/) 0.0
10
<
-0.1
-0.2
-0,3
-0.4
0 10 20 30 40 50
Tray Location
FlGURE 8·10. Modified principal component analysis.
Selection of Controlled and Manipulated Variables 153
suIts from these analyses lead to a single weaknesses of the various control specific
control specific subsystem to be studied fur- subsystems.
ther. The global approaches that will be Minimum of the Condition Numbers
presented systematically consider all possi- Consider first the control specific condition
ble control specific subsystems. number as a basis of comparing sensor pair
There are two problems with a global subsystems. Using the subsystem condition
approach to sensor placement. The first numbers, the sensor placement problem can
concerns the large number of control spe- be formulated as
cific subsystems. (For the ethanol-water
column there are 1200 unique combinations
of dual sensor locations.) This can be a min
-.-. CNi,i max
[ or - - - -1 - 1
large but manageable computational prob- l, ] i,j CNi,i
lem. The second problem is the selection of
a suitable basis for comparing the subsys-
tems. Three different formulations of the where CNi,i = control specific condition
global optimization problem are described number
in the following text along with a rationale i, j = tray location of the control
for using SVD to measure the strengths and sensors
FIGURE 8·12. Control specific condition number versus sensor locations for the
ethanol-water column.
Selection of Controlled and Manipulated Variables 155
For the sake of scale and clarity the In general, the problem with basing sen-
surface is plotted as the reciprocal of the sor locations solely on control specific con-
condition number. In the reciprocal condi- dition numbers is the neglected issue of
tion number plots, the global optimum ap- sensor sensitivity. Balanced gains are impor-
pears as a maximum. The surface defined by tant to the dual-ended control problem, but
the control specific condition numbers for even a perfectly balanced system that has
the ethanol-water column is shown in Fig- low sensitivities cannot be controlled.
ure 8-12. Note, that the "best" sensor loca- 2. Maximum of the Smallest Singular Val-
tions suggested by this optimization strategy ues Another perspective of the global
are curiously at tray 1 and tray 44. The problem to consider is the multivariable
problem with these two locations is loop sensitivity of each of the control specific
sensitivity. The individual sensitivities shown subsystems. The singular values of the sub-
in Figure 8-4 indicate a 10% change in system are a measure of the multivariable
either the distillate or the steam flow affects sensitivity. In terms of placing sensors for
the two temperatures by less than 0.03° C. maximum sensitivity the focus typically
This is far too weak a response for either a should be on the weakness of the subsys-
dual- or single-ended control system to be tem. The weakness of the system is re-
effective. flected in the smaller of the singular values.
FIGURE 8-13. Minimum singular value versus sensor location for the ethanol-water
column.
156 Practical Distillation Control
FIGURE 8-14. Intersivity versus sensor location for the ethanol-water column.
Selection of Controlled and Manipulated Variables 157
The surface defined by the intersivity in- The results are similar to those found
dex is the product of the two surfaces using the principal component analysis of
discussed earlier. The result for the the overall gain matrix with one important
ethanol-water column is shown in Figure difference. The position of the weak sensor
8-14. The maximum intersivity occurs when is essentially the same; yet the position of
the sensor pairs are located at trays 25 and the strong sensor is adjusted (moved up the
12. This choice results in a control specific column several trays) for the sake of a mul-
subsystem with both good sensitivity and tivariable balance.
balance. 4. Choosing a Method for Locating Sen-
sors for Dual-Ended Control Five methods
have been described for selecting sensor
Method 1: Principal Component Analysis. Requires only locations in dual-ended temperature con-
a single SVD analysis of the full column gain ma- trol. A brief summary of the characteristic
trile Results in sensors that represent a good bal- strengths and weakness of each method is
ance between interaction and sensitivity. Can be
presented in Figure 8-15. Generally, Meth-
applied to a control system of any reasonable di-
mension.
ods 1 and 2 are the most commonly used
because they involve a single SVD analysis.
Method 2: Modified Principal Component Analysis. Re-
They also yield similar results to the more
quires only a single SVD analysis of the full column
gain matrix. Weighs more heavily the interaction extensive procedures of Methods 4 and 5.
between sensors than does Method 1. Can only be
applied to the analysis of a 2 x 2 multivariable
control problem. 8·4 SELECTING SENSOR TYPE:
Method 3: Global Minimum Condition Number. Re- TEMPERATURE VERSUS
quires an SVD analysis of all possible sensor combi- COMPOSITION
nations. Considers only the problem of sensor inter-
action. Because it does not consider the serious Historically, temperature has been the sen-
problem of sensor sensitivity it can yield very mis- sor of choice in monitoring and controlling
leading results. the separation. Temperature measurement
Method 4: Global Maximum of the Minimum Singular is simple, fast, inexpensive, reliable, and
Value. Requires an SVD analysis of all possible does not need the complex sampling sys-
sensor combinations. Directly considers sensor sen- tems that many analyzer-based schemes
sitivity but forces the sensor interaction also to be
considered by looking at the multivariable weak-
require. In many columns, however, control-
ness. Yields good results that are similar to those ling temperature does not meet the separa-
obtained by Methods 1 and 2. Can be applied to tion control objectives. In such cases more
any dimension control problem; however, for sys- direct measurements of composition should
tems greater than 2 x 2 the number of sensor com- be considered.
binations that should be considered becomes very
large.
Method 5: Global Maximum of the "1ntersivity." Inter- 8-4·1 Limitation of Temperature
sivity is a function defined to explicitly include both
Sensors
sensor sensitivity and sensor interaction. Requires
an SVD analysis of all possible sensor combina- Before considering the use of process ana-
tions. Weighs more heavily the interaction between
lyzers, perhaps it is appropriate to consider
sensors than do Methods 1, 2, and 4. As a result the
secondary sensor is the same but the primary sensor the limitations and failings of conventional
is moved to a weaker location for the sake of temperature-based control. Temperature is
interaction. Can with some difficulty be applied to useful in controlling column separation only
control problems with dimensions greater than to the extent that it is a reliable indicator of
2 x 2. composition. The use of temperature sen-
FIGURE 8-15. Summary and comparison of methods sors to indirectly measure composition is
for locating sensors for dual-ended control. loosely based on the physics of the
158 Practical Distillation Control
liquid-vapor phase transition of mixtures. When a temperature control has been im-
In general, the composition, temperature, plemented to insure temperature remains
and pressure of a boiling mixture are ther- constant, the following problem develops:
modynamically related as indicated by
dT= 0
T = !(X I , X 2 , ••• , Xi' ... ' X n ' P) (8T/13P)
dx = (13T/8x) dP
where T =temperature of the boiling
mixture
In such cases, variations in pressure are
Xi = composition of component i in
forced by the temperature control system to
a n component mixture
propagate to the composition. The magni-
P = vapor pressure of the boiling
tude of the problem is defined by the mag-
mixture
nitude of pressure variation and by the
physical properties of the mixture.
The effectiveness of a temperature-based
It is relatively easy to estimate the pres-
composition control scheme is largely re-
sure problem by considering the thermody-
lated to the complexity of the relationships
namics of the mixture at the control tray. If
defined by the thermodynamics of the mix-
the ratio of the two partial derivatives is
ture. Such schemes are the most effective
even modestly large, temperature control
for binary mixtures if the column pressure is
may actually cause more problems than it
constant. In such cases composition is only
solves. It should be noted that it is relatively
a function of temperature. Therefore, if the
easy to develop a strategy that compensates
control system is able to hold temperature
for pressure variations. Such strategies do
constant, then the composition will also be
need, however, a pressure measurement for
held constant. For more complex (and more
each of the temperature control points.
realistic) systems, holding temperature con-
stant does not imply that composition will
Nonideal and Muiticomponent Mixtures
also be constant.
The nonideal and multicomponent nature
of mixtures can cause problems in tempera-
Pressure Variations
One problem of temperature-based column ture-based control systems more difficult to
control is pressure variations. It is impor- solve than pressure variations. In thermody-
tant to realize all separation control systems namically complex systems, temperature can
be a very poor indicator of composition.
manipulate (either directly or indirectly) the
Controlling temperature in such systems
liquid and vapor rates inside of the column.
As these internal flows change, so will the may be counterproductive to the fundamen-
tal compensation control objectives of most
pressure distribution in the column. Even in
columns where pressure is formally con- systems. Consider, for example, the
trolled, variations are likely to occur at the azeotropic column shown in Figure 8-16. An
extensive simulation-based study was made
tray locations where temperature measure-
for the Department of Energy (Roberts et
ments are made.
aI., 1991) in which three different sensor-
The potential problems caused by such
pressure variation are demonstrated by the based control systems were evaluated with
following relationship for a simple binary respect to the effect various levels of varia-
tion in feed composition have on product
mixture:
purity. The study results (summarized in
Figure 8-17) indicate a control system using
8TI dx
dT=- 8TI dP
+- a well placed internal composition measure-
13x 13P ment would be expected to yield results
Selection of Controlled and Manipulated Variables 159
ChHIed Water
orders of magnitude better than the conven-
tional temperature-based schemes.
The acetic acid study also showed an
interesting result with respect to single-
ended versus dual-ended temperature con-
trol. The conventional wisdom is dual-ended
control, if possible, should give better per-
formance than single-ended contro!' The ra-
tionale is one of degrees of freedom. If a
column exhibits two controllable degrees of
freedom, tying down those degrees of free-
dom with a dual-ended control system
should reduce variation throughout the col-
umn. The study indicates, however, fixing
the degrees of freedom of the temperature
profile actually results in higher variation in
composition.
Temperature-based control systems can
yield good (perhaps even exciting) results in
terms of temperature performance. Never-
theless, the underlying control objective is
composition and not temperature, The per-
formance of all column control systems (be
they single-ended, dual-ended, or even more
FIGURE 8·16. Azeotropic distillation column. complex) should always be evaluated with
respect to product compositions. Unfortu-
nately, this is seldom applied in practice.
Costs: On-line process analyzers are typically expensive as well as to stages inside the column. (Re-
both in terms of the cost of the device and in terms call, the temperature sensor can be used to
of the cost of maintenance.
infer composition only when a liquid and
Auxiliary Equipment: On-line analyzers typically re- vapor are in thermal equilibrium. This con-
quire considerable auxiliary equipment such as dition seldom exists in product and feed
samplers, calibration standards, environmental en-
closures, etc., which must also be designed, pur-
streams.)
chased, and maintained. In discussing the role of analyzers for
distillation control it is important to distin-
Temperamental Nature: Process analyzers and the nec-
essary auxiliary equipment tend to demand a lot of
guish between off-line and on-line analyz-
attention to insure reliable results. ers. Off-line analyzers are typically labora-
tory-based equipment and procedures. On-
Speed of Response: Most analyzers are relatively slow,
particularly if they are multiplexed to a number of
line analyzers are directly connected to the
sample points. automatic control system. The role of these
on-line analyzer will be discussed first.
Application Specific: Unlike temperature sensors in
which only the span needs to be specified, on-line
Figure 8-19 illustrates three basic ana-
analyzers have to be selected and tailored in con- lyzer control schemes using reflux as a ma-
siderable detail for each application. nipulated variable. The first scheme (Figure
8-19a) uses the analyzer to measure the
FIGURE 8-18. Headaches and operational concerns
in using on-line process analyzers for control.
product composition and a composition
controller to directly adjust reflux flow. The
second scheme (Figure 8-19b) also uses
product composition, but the role of the
the use of process analyzers will undoubt- composition controller is to make feedback
edly increase. adjustments to a temperature controller. In
the third scheme (Figure 8-19c) the ana-
lyzer monitors an internal tray instead of
8-4-3 Schemes for USing Analyzers In
the product. The composition controller di-
Distillation Control
rectly manipulates the reflux flow based on
Analyzers can have a broader role in distil- the internal composition in much the same
lation control than can temperature. They way as the temperature schemes discussed
can be applied to feed and product streams in earlier sections.
FIGURE 8-19. Analyzer control schemes. (a) Scheme 1: product composition con-
trol. (b) Scheme 2: temperature composition cascade control. (c) Scheme 3: internal
composition control.
Selection of Controlled and Manipulated Variables 161
The first two schemes historically repre- sensitive location to measure composition is
sent the most common use of analyzers in typically at the product streams. Inside the
the control of column separation. The third column is where most of the action is taking
scheme, however, can offer advantage if loop place and it should be considered in select-
sensitivity is a problem. The rationale for ing a location for the control analyzer.
placing analyzers directly in the product
stream is based on a desire to directly con-
trol the product purity. This rationale may
8-4-4 Analyzer Resolution
not always result in the best control system.
Requirements versus Location
The sensitivity issues discussed for tempera-
ture-based control also can be a problem in Locating separation control analyzers inside
control systems using analyzers. The least the column as indicated in Figure 8-19c can
1.uwl,---------------------....,
BASE CASE
--- ....
;.:-:--:.:...::..:.. .....•..........•..._..............
0.999
- >:;._...._ ........_..__._...... __e.
0.998 • •••• __••••••••••••••••_._•••_ ••••••
0.997
0.996
o.
o.
1.00,.....-------------------.....
0.00
........
~
"- -1.00
~
II dX(acetic)/dQ
r::
• dX(Acetic)/dR
'0
"-
X
'0
-2.00
-3.00
o 10 20 30 40 50
Tray Locatlon
(a)
FIGURE 8-21. Sensitivities for an azeotropic column. (a) Acetic acid; (b) water; (c)
solvent.
on selecting a first-level configuration see in the decanter. Note, however, that such
Section 8-6.} schemes force the inevitable variation in
Azeotropic columns are different from solvent inventory to occur inside the col-
conventional columns in that a decanter is umn. The control configuration considered
used to separate and return the solvent to in this example (LQ) uses reflux to control
the column as reflux. Except for losses at the inventory in the column. The inventory
the decanter, the solvent is trapped in a in the decanter is allowed to float uncon-
recycle loop between the column and the trolled.
decanter. For such systems the role of the Although the solvent losses at the de-
reflux valve is very important. It can be used canter are not great, they are continuous
to control either the inventory of solvent in and the solvent inventory in the system will
the decanter or the inventory of the solvent gradually decrease. Eventually, an operator
in the column. Conventional schemes typi- will need to add fresh solvent to make up
cally use the reflux valve to control the level for the loss of inventory in the system. It is
164 Practical Distillation Control
0.4"7"""-------------------,
-0.2
o 10 20 30 40 50
Tray Locat i on
(b)
better if such variations (both the gradual sponds to both the reflux and the steam but
loss and the typically sudden correction) are the response is relatively weak. Of the three
forced to occur at the decanter instead of components the solvent composition is the
inside the column. If the decanter is large most responsive to changes in both the
enough to maintain the short term variation steam and the reflux.
in inventory, a control room operator can This example clearly shows the problem
easily maintain the overall solvent inventory that often exists with analyzers located in
without disturbing the operation of the col- the product streams. In this case, composi-
umn. Control adjustment can be made on tion at the top and the bottom of the col-
simple high and low alarms on the decanter umn essentially are not responsive to
level. changes in either of the two manipulated
In the azeotropic example only the LQ variables. Control loops based on either the
control configuration is considered. For this bottom acetic acid product or on the solvent
configuration the acetic acid composition is water overhead would be very weak. Conse-
responsive to the reflux but shows little sen- quently, performance of such loops would
sitivity to changes in the steam. Water re- be poor, at best.
Selection of Controlled and Manipulated Variables 165
4~--~~------------------------------~
0
......
Qe
......
Qe
-2 • dX(Solvent)/dO
0
"0
......
• dX(Solvent)/dR
X
"0
-4
-6
o 10 20 30 40 50
Tray Location
(c)
Poor analyzer sensitivity in the product aration control strategy that manipulates
streams is not always the case. Consider the only steam. Disturbances in either the feed
analysis of the ethanol-water column. The rate or the feed composition cannot be
composition gain matrix and corresponding compensated for by manipulating the steam
sensitivity plots are shown in Figures 8-22 rate in any DQ control configuration. All
and 8-23. Note that ethanol is the more strategies that require the distillate flow to
sensitive component and its peak sensitivity be set manually are very sensitive to mate-
is different for each manipulated variable. rial balance disturbances in the feed.
The gain analysis indicates the strongest
composition control loop would be one con-
8-4-7 Selecting Analyzer Locations and
trolling the ethanol composition on tray 18
Focus for Dual-Ended Control
by manipulating the distillate rate. If steam
is to be used also as a manipulated variable, The use of analyzers can present intriguing
the composition of ethanol on tray 1 (the possibilities for multivariable column con-
reboiler) would result in the strongest sec- trol systems. Analyzers better represent
ondary loop. product purity than do temperature sensors.
Note that for this example it would not They also can provide a much better condi-
be advisable to consider a single-ended sep- tioned multivariable problem for the control
166 Practical Distillation Control
8Xil 8Xil 8Xil 8Xil system to solve. In order for these potential
8D Q 8Q D 8D Q 8Q D advantages to be realized, it is important to
determine the proper location and focus for
the analyzer system.
- 0.222 0.019 1.361 -0.119
- 0.249 0.022 1.500 -0.131
Properly locating analyzers is somewhat
- 0.278 0.024 1.647 - 0.143 more complicated than locating tempera-
-0.310 0.027 1.801 -0.155 ture sensors. There are two design ques-
- 0.344 0.029 1.964 -0.166 tions that need to be addressed concerning
- 0.382 0.032 2.138 -0.176 composition-based control: Where should
- 0.425 0.034 2.323 - 0.186
- 0.471 0.037 2.521 -0.196
the analyzers be located and what compo-
- 0.524 0.039 2.735 -0.205 nent composition should be presented to
- 0.583 0.042 2.967 -0.214 the control system? The composition gain
-0.651 0.045 3.220 -0.223 matrices and their singular value decompo-
- 0.728 0.048 3.496 -0.231
sitions can provide insight into answering
-0.817 0.051 3.801 - 0.239
-0.922 0.055 4.138 -0.247 these critical questions.
-1.044 0.059 4.514 -0.254
-1.190 0.063 4.935 - 0.261
-1.366 0.067 5.411 -0.267 Identifying Locations for Similiu
-1.581 0.072 5.950 - 0.274 Composition Measurements
-1.847 0.078 6.566 - 0.280
-2.182 0.085 7.274 -0.286
Consider an analysis of the ethanol-water
- 2.612 0.093 8.091 -0.292 column. Figure 8-24 shows the U-vector
-3.175 0.103 9.039 -0.298 plots that result for the SVD analysis of the
-3.928 0.115 10.142 -0.304 two composition gain matrices shown in
-4.962 0.131 11.427 -0.311 Figure 8-22. Each plot indicates. two ana-
K ••hanol = -6.423 0.152 K wa•er = 12.921 -0.318
-8.555 0.180 14.641 -0.326
lyzer locations that might yield a. promising
- 11.785 0.222 16.586 -0.335 balance between sensitivity and interaction.
- 16.901 0.286 18.694 -0.343 The results suggest two possible measure-
- 25.442 0.389 20.755 -0.350 ment schemes: one that measures the
-40.548 0.568 22.156 -0.348 ethanol composition on trays 1 and 18 and
- 67.802 0.887 21.168 -0.317
-105.170 1.283 14.578 -0.213
another that measures the water composi-
-110.770 1.073 5.543 -0.075 tion on trays 21 and 37.
- 63.662 - 0.048 1.367 -0.009 A control specific SVD analysis of each
-10.864 -0.718 0.177 0.009 scheme provides further insight and a basis
- 12.297 -1.373 0.158 0.015 for comparison. This will be discussed after
-13.611 - 2.071 0.135 0.018
-14.752 -2.787 0.110 0.018
the rationale for identifying measurement
-15.696 - 3.499 0.087 0.017 systems for dual-ended composition has
-16.446 - 4.190 0.066 0.014 been more fully developed.
-17.020 -4.852 0.049 0.012
-17.449 -5.481 0.036 0.009
-17.763 -6.077 0.026 0.007
-17.989 - 6.642 0.018 0.005 Identifying Locations for Different
-18.149 -7.177 0.013 0.003 Composition Measurements
-18.263 -7.685 0.009 0.002 Identifying locations using the U vectors
-18.342 -8.168 0.006 0.002 assumes the same component composition
-18.398 - 8.627 0.005 0.001
-18.437 - 9.067 0.003 0.001
determination at each measurement point.
-18.465 - 9.604 0.001 0.000 Measuring different components at each lo-
cation should also be considered. The com-
FIGURE 8-22. Composition gain matrices for the position sensitivity plots illustrated in Fig-
ethanol-water column. ure 8-23 provide guidance in this direction.
Selection of Controlled and Manipulated Variables 167
25~------------~-------------------,
-25
L ......
'OQe
................ -50
•a d(water)/dO
d(water)/dO
xQe
'0 ........
• d( ethano 1)1 dO
0 d(ethano1)/dO
-75
-100
-125~~~~~~~~~~~~~~~~~~~
o 10 20 30 40 50
Tray
FIGURE 8-23. Composition sensitivity plots for the ethanol-water column.
0.4
0.3
c • Ul-water
0
0.2 • U2-water
(f) ~
~
-
(f)
0
0 0.1
Q. a.
'- E 0.0
0 0
~
u u
Q) '- -0.1
> Q)
~
:::) ro -0.2
3 2nd Sensor
-0.3
,Ilocat1on
-0.4
0 10 20 30 40 50
Tray Locat1on
FIGURE 8·24. U vector plots for the water composition of the ethanol-water
column.
168 Practical Distillation Control
The most sensitive location for each com- a control scheme needs to be considered.
ponent is one obvious combination to con- Figure 8-26 reports the singular values and
sider. In this example, the maximum ethanol condition numbers resulting from the con-
sensitivity is at tray 18 and the maximum trol specific SVD analysis of each proposed
water sensitivity occurs at tray 21. The close measurement system. (Recall, the control
proximity of these locations, however, hints specific SVD analysis is of the 2 X 2 gain
of interaction problems. Figure 8-25 sug- matrix defined by the location and type of
gests the next most sensitive location to sensors selected.)
measure ethanol is at the bottom. Locating In this example two composition-based
an ethanol sensor at the bottom and the schemes look acceptable. Scheme 2 (ethanol
water sensor at tray 21 is another measure- at trays 1 and 18) has a condition number of
ment scheme that should be considered. 11. Scheme 5 (ethanol at tray 1 and water at
tray 21) with a condition number of 4 is
even better. The singular values of both
schemes also indicate strong sensitivity.
Comparing the Different Measurement It is important to note that composition-
Schemes based measurement schemes 2 and 5 are far
The rationale just presented identified four better from a control point of view than the
possible measurement schemes: one based temperature-based measurement scheme.
on ethanol measurements, one based on The temperature-based system has a condi-
water measurements, and two schemes tion number of 126. Such large condition
based on both ethanol and water measure- numbers indicate a very narrow degree of
ments. To properly evaluate these schemes freedom. In this case, a dual-ended control
the multivariable problem each presents to system would face a problem 126 times more
0.2
0.1
c
0
0.0
(f) ~
~
(f)
0 0 -0.1
a.. 0.
'- 0E -0.2 • UI-ethanol
0
~ U • U2-ethano I
u
Q)
0 -0.3
> c
ro
::::> ~
-0.4
~
lLJ
I
I
I
2nd Sensor
-0.5 l/ocatlon
-0.6
0 10 20 30 40 50
Tray Number
FIGURE 8-25. U vector plots for the ethanol composition of the ethanol-water
column.
Selection of Controlled and Manipulated Variables 169
Water-21
Scheme 3 22.54 0.1789 126.01
Water-37
Ethanol-18
Scheme 4 112.9 0.130 864.1
Water-21
SchemeS Ethanol-1
294.7 7.436 3.963
Water-21
FIGURE 8-26. Comparison of various sensor schemes for the ethanol-water col-
umn.
lyzer in the feed of the column should ing can often lead to rather simple changes
always be considered if the plantwide envi- in the process. Those changes will reduce or
ronment results in significant composition eliminate the need for complex column con-
variations in the column. trol systems.
In the role of monitoring and document-
ing variation, it is not essential to have an
8·5-2 Recycle Inventory Control
on-line measurement. In such applications
Distillation columns are frequently part of versatility and accuracy are far more impor-
plantwide recycle systems. Their purpose is tant than speed. These studies are typically
often to separate solvents and intermediates conducted off-line by collecting and submit-
from salable products so they can be recy- ting samples to an analytical laboratory.
cled for further processing. From a
plantwide control perspective it is very im-
8·6 SELECTING MANIPULATED
portant to monitor and control the compo-
VARIABLES
nent inventories in these recycle streams
(see Chapter 20). Such sensor information is The foregoing sections primarily concern
necessary for the control of these plantwide selecting and locating sensors to be used by
inventories. the column control systems. An equally im-
An analyzer placed in the column feed portant concern is the selection of the vari-
can often serve a dual role: It can be used ables to be manipulated by those control
by the control room operator in the man- systems. The selection of the controlled and
agement of the plantwide inventories; it also manipulated variables fundamentally de-
can provide information necessary for the fines the structure of the control system.
feedforward control discussed previously. The selection of an appropriate control
structure is critical to the success of the
control system. It establishes the difficulty
8-5·3 Measuring and Documenting
of the local column control problems and
Variation
defines how the operation of the column
Another important role of the analyzer is to interacts with the more global plantwide
measure and document variation. Part of concerns.
this role is related to quality assurance. To- The first step in selecting manipulated
day, customer contracts typically require variables is to consider which valves will be
statistical control charts be maintained on used for inventory control and which valves
the product quality as well as other key will be used for controlling the separation.
process variables. The purpose of such Designating the role of each valve and es-
charts is to document the state of control tablishing the manipulated variables for the
(in the statistical sense) of the process as separation control system defines the first-
well as the product. level structure of the column control sys-
The role of the analyzer in measuring tem.
and documenting variation should be much Consider the 73 tray acid recovery col-
broader than providing information for umn shown in Figure 8-27. In this case two
quality assurance control charts. In a more valves must be used to control the invento-
general sense, such information can be used ries in the accumulator and reboiler, and
to improve process operations. Control sys- the remaining two are available for the con-
tems that respond to disturbances are im- trol of the column separation. A list of
portant; however, it is also important to choices commonly considered are given in
understand where composition disturbances Figure 8-28. Note that the inventory control
originate and how they propagate through a system manipulates the valves directly. The
complex process system. Such understand- separation control system also includes
Selection of Controlled and Manipulated Variables 171
numbers and singular values for each con- 1 DO 46.9 53.8 7.63 0.142 0.595
trol structure. These numbers can provide 2 LO 249.0 287.0 18.37 0.064 19.36
3 LB 46.9 53.9 3.87 0.072 0.607
insight into the difficulty of the separation 4 LIDO 123.0 142.0 8.95 0.063 9.486
control problem and can be used to screen 5 L/DB 48.7 56.3 3.95 0.072 0.621
the control structures under consideration. 6 DQ/B 42.1 48.0 7.33 0.153 0.569
7 LO/B 82.7 95.0 6.18 0.065 6.79
The condition number provides a mea- 8 L/DQ/B 62.9 72.4 4.62 0.064 5.12
sure of the degrees of freedom of the sepa-
ration control problem. High condition "The control specific analysis is based on temperature
sensors on trays 16 and 44. These locations were cho-
numbers are an indication that dual-ended sen using the principal component analysis of the full
control may not be a practical considera- U vectors for each control structure outlined in Sec-
tion. Singular values also can be used to tion 8-3-3. The result of the study indicated identical
sensor locations for all eight first-level control struc-
identify another practical aspect of the tures.
dual-ended control. Low singular values in-
dicate low sensitivity. To compensate for FIGURE 8-29. Condition numbers and singular val-
the low sensitivity the controller gain must ues for various control structures in an acid recovery
column.
be set high. High controller gains typically
result in saturated valves. Also, high con-
troller gains can make the entire multivari-
able system sensitive to even low levels of principal component method outlined in the
sensor noise. sensor selection section.
High singular values are less common; Typically, the choice of sensor locations
however, they suggest another practical is not a strong function of control structure.
problem that needs to be considered in the For this example, sensor location analysis
question of dual- versus single-ended con- for all eight control structures indicated tray
trol. High singular values indicate multivari- locations 16 and 44 to be the best choice.
able directions of high sensitivity. Such high In this example, the control structures
sensitivity may place demands on the con- using either distillate and steam or reflux
trol system that cannot be met within the and bottom as valves in the separation con-
resolution of the valve. In a dual-ended trol strategy are better conditioned than the
control the hunting typical of such loops can schemes using reflux and steam. The two
affect the entire multivariable system. schemes with condition numbers greater
Consider the SVD analysis of each tem- than 100 (LQ and L/DQ) can be dis-
perature control structure for the acid re- counted as candidates for dual-ended con-
covery column. Figure 8-29 indicates the trol.
results of both overall SVD and control Considering the singular values can fur-
specific analyses. The overall analysis is of ther narrow the list of structures to be con-
the full 73 X 2 temperature gain matrix for sidered for dual-ended control. In this ex-
each control structure. (The gain matrices ample, six of the eight structures have
were determined for each control structure singular values less than 0.10. (For most
using a steady-state simulation of the col- systems, singular values less than 0.1% per
umn operating at its base case.) The control percent indicate that the multivariable sys-
specific analysis for each structure is of the tem will be prone to saturate the control
2 X 2 gain matrix defined by the location of valves and will be sensitive to noise.) The
the two control sensors. The location of the two structures with the largest second singu-
control sensors was determined using the lar value are DQ and DQ/B. (It is interest-
Selection of Controlled and Manipulated Variables 173
ing to note that these are the only struc- SVD Analysis
tures that use the steam and distillate valves DQ-StructunI La-Structure
in the separation control strategy.) D Q L Q
0.8923
-0.9991 -0.0418
the separation control strategy. The columns -0.0418 0.9991 0.8923 0.4515
h
the DQ structure (T - 20 min) than in the
LQ structure (T - 60 min).
This example illustrates some of the sub-
'<a 1'" tle but important complexities of distillation
dynamics. Note that manipulating Q results
(a) , in open loop overshoot. The initial effect of
changing the steam is caused by the change
CONTRJl VALVES in vapor rate, which occurs relatively fast. A
6L -5% 60 -5% secondary effect results from changes in re-
flux and distillate at the top of the column.
IljI~
~ These secondary effects are slower and in
the opposite direction. Such effects are not
extremely serious; however, they can cause
trouble in tuning the feedback control.
I~ !
I-- !-Ptb
It is important to consider the nature of the ~ ax,,- :tlO% 0.00431 0.25636 0.02092 0.00813
disturbances in the selection of a control
structure. What are the anticipated column AXel
disturbances and how will the more promis- Change in Overhead Purity
ing control systems react to each? Such • Includes a feedforward controller that adjust
lha distillate ftow for changes in feed ftow
information is necessary in order to fully
interpret the results of the steady-state and FIGURE 8·32. Change in overhead purity for vari-
dynamic analyses. ous acid recovery column disturbances.
Again, consider the acid recovery col-
umn. The steady-state analysis indicates,
from the point of view of condition number, were considered. Because of the dynamic
that DQ should be a good control structure advantage offered in manipulating steam,
to consider. The dynamic analysis further only DQ, LQ, and RQ control structures
indicates that although the dynamics for were considered (R = LID). For all three
changes in Q are good, the dynamics for structures a sensor-valve pairing of T(16)
changes in D are extremely slow. One logi- with Q was considered. For the DQ scheme
cal conclusion might be to implement a a sensor-valve pairing of T(44) with D was
single-ended control system within the DQ considered also. The temperature control
structure that does not use D as a manipu- loop in all four schemes also included pres-
lated variable. Considering only the results sure compensation.
of the steady-state and dynamics analyses, Of the four schemes considered, the two
the most logical single-ended control struc- using the DQ structure showed the poorest
ture is one that controls the temperature on performance. The DQ strategy using D as a
tray 16 using the steam flow. Although this manipulated variable is sensitive to all three
rationale is clever, it would obviously lead disturbances. The DQ strategy using Q is
to a problem if feed disturbances occur. extremely sensitive to the feed composition.
Feed flow or composition changes require (This strategy would also be very sensitive
adjustments in the distillate flow that are to feed flow disturbances if the feedforward
simply not provided for in this design. controller were not present to adjust the
A disturbance rejection study of the acid distillate flow for changes in the feed flow
recovery column illustrates the problem rate.) The LQ and RQ strategies demon-
more clearly. The study considered three strate about the same disturbance rejection
common disturbances: overhead pressure, capabilities for feed flow disturbances.
feed flow rate, and feed composition. Each However, RQ shows a marked advantage
disturbance was varied from its base operat- with respect to changes in feed composition
ing condition. The results are presented as and changes in column pressure.
steady-state variations in purity of the acetic
acid product stream as seen in Figure 8-32. Propagation of Variation to the Plant:
The schemes selected for this study were The preceding discussion considers how dis-
based on the results of the steady-state and turbances affect product purity. It is impor-
dynamic analyses. Because of the high con- tant to be aware of other issues concerning
dition numbers, only single-ended schemes how the column and its control system re-
176 Practical Distillation Control
act to these disturbances. These other is- Plant and Column Design
sues concern the relationship between the The easiest column to control is one that
column and other unit operations in the has no disturbances. The calmest column, in
plant. A column is typically connected to the plantwide sense, is one that has no
other unit operations through its feed and disturbances. During the control system de-
product streams. In highly heat-integrated sign (and after) the control engineer should
plants it may also be coupled through heat become aware of where and how distur-
exchangers and economizers. All these con- bances originate and propagate through the
nections suggest that variation in the col- plant to the column. Understanding the dis-
umn will propagate to other parts of the turbance paths may suggest simple process
plant. changes to eliminate or reduce the magni-
There are several related questions to be tude of such disturbances. For example, well
considered during the selection of a control placed inventory tanks can help reduce
structure: What plantwide disturbances are propagation of many disturbances from unit
caused by the actions of the column control to unit. Such inventories can often be justi-
system. What units are disturbed and how fied on the basis of improved quality and/or
does that affect the plantwide objectives? A productivity.
larger perspective in some cases may sug- The control engineer also should con-
gest using a different column control strat- sider the control effectiveness of the local
egy. One that propagates its variation to column design. For example, internal con-
more benign locations in the plant should densers are often used in column design.
be considered. Such designs cost less to fabricate; however,
they limit the control possibilities. An inter-
Component Inventories in Recycle Systems nal condenser-based design allows only con-
Another question concerns the problem of trol structures that vary the distillate to be
plantwide inventories. Specifically, this con- implemented. If the column analysis indi-
cerns component inventories in recycle sys- cates that a reflux-based control structure is
tems. A distillation column designed to appropriate, the internal condenser may not
consider only local performance may not be cost effective after all.
behave well in such systems. Closely con- It is obvious that design changes neces-
trolling the purity of the product streams sary to improve the control system are best
may inadvertently serve as a dam to one or identified before the process design phase is
more of the components. If the component completed. Once fixed in concrete and steel,
flow into the system is greater than the rate even simple modifications may be difficult
the column control system allows it to leave, unless the control engineer builds a very
accumulation occurs. The component convincing case. Nevertheless, if such
buildup will continue until the column con- changes are to occur, the control system
trol system cannot handle it anymore or designer must also be involved in the col-
until an operator intervenes. In such cases umn design as well as the overall plant
the best solution is a less ambitious column design.
control strategy. Some situations may not
suggest using any separation control strat-
8-7 SUMMARY AND
egy. It is important to remember that the
CONCLUSIONS
local column control system should conform
to the plantwide concerns and goals. This chapter attempts to discuss a number
A more extensive detailed discussion of of technical as well as philosophical issues
plantwide concerns that should be consid- related to the preliminary design of a col-
ered in the design of the local column con- umn control system. These issues relate pri-
trol system is included in Chapter 20. marily to the selection of control structures,
Selection of Controlled and Manipulated Variables 177
placement of sensors, and pairing those sen- Forsythe, G. E. and Moler, C. B. (1967). Com-
sors with available manipulated variables puter Solution of Linear Algebra. Englewood
within that structure. Proper choices at this Cliffs, NJ: Prentice-Hall.
level typically have a far greater impact on Golub, G. H. and Reinsh, C. (1970) Singular
value decomposition and least squares solu-
the overall performance of the column con-
tions. Numer. Math. 14,403-420.
trol systems than does the choice of control
Roberts, 1. J., Muly, E. c., Garrison, A A, and
algorithms. Moore, C. F. (1991). On-line chemical compo-
The more philosophical issues concern sition analyzer development, phase II, Final
the role and responsibility of the control Report DOE DE FC07-88ID 12691, Depart-
strategy specialists. As distillation control ment of Energy.
strategies become more complex there is a Lawson, C. L. and Hanson, R. J. (1978). Solving
natural tendency for the control engineer to Least Squares Problems. Englewood Cliffs, NJ:
become more specialized. There is also a Prentice-Hall.
tendency for the role of the control engi- MATLAB User's Guide (1985). The Mathworks,
neer to be limited to solving control prob- Inc., South Natick, MA.
Moore, C. F. (1985). A reliable distillation con-
lems. A larger role should include working
trol analysis procedure for use during initial
with other specialists to minimize the dif-
column design. In 1985 Annual Meeting of
ficulty of the control problem that must be the AIChE.
solved. This is particularly true with respect Moore, C. F. (1991). A new role for engineering
to process design. Developing a clearer un- process control focused on improving quality.
derstanding of the fundamental cause of In Competing Globally Through Customer
column disturbances can lead to process Value- The Management of Strategic Supra-
improvements that reduce the need for systems, M. 1. Stahl and G. M. Bounds, eds.
complex control systems. New York: Quorum Books.
Moore, C. F. (1987). Determining analyzer loca-
tion and type for distillation column control.
In 14th Annual Meeting of the Federation of
Analytical Chemistry and Spectroscopy Soci-
References eties.
Cantor, D. L. (1987). A detailed steady-state Moore, C. F., Hackney, J. E., and Canter, D. L.
control analysis of an ethanol-water distilla- (1986). Selecting sensor location and type for
tion column. M.S. thesis, The University of multivariable processes. In The Shell Process
Tennessee. Control Workshop. Boston: Butterworths.
Deshpande, P. B., ed. (1988). Multivariable con- Roat, S. D., Downs, J. J., Vogel, E. F., and Doss,
trol methods. Research Triangle Park, NC: J. E. (1986). The integration of rigorous dy-
Instrument Society of America. namic modeling and control system synthesis
Downs, J. J. (1982). The control of azeotropic for distillation columns: An industrial ap-
distillation columns. Ph.D. thesis, The Uni- proach. In Chemical Process Contro[-CPCIII,
versity of Tennessee. 99-138. Elsevier, Amsterdam, 1986.
9
Selection of Controller
Structure
B. D. Tyreus
E. I. Du Pont de Nemours & Co.
new one well into the design phase. The can achieve and thus make the objectives
process equipment including buffer tanks less demanding. It is crucial to define what
and control valves are then already desig- the control system should do as well as to
nated and the task is to select the appropri- understand what disturbances it has to deal
ate variables to be controlled and design with.
controllers that will tie the controlled vari- Process understanding is another key ac-
ables to the control valves in such a way tivity for successful control system design.
that the resulting controller structure meets Nobody seems to disagree with this notion,
the desired objectives. but in practice more time is usually spent on
I will also assume that the final controller designing and implementing algorithms and
structure will be built up around conven- complex controllers than on analyzing pro-
tional control elements such as PID con- cess data and figuring out how the process
trollers, ratio, feedforward, and override really works. Modelling and simulation are
control blocks as found in all commercial key parts in the process understanding step.
distributed control systems. I will not ad- Rigorous dynamic simulation is the third
dress the design of multivariable controllers important activity in control system design.
such as DMC, which is treated in Chap- A flexible dynamic simulation tool allows
ter 12. rapid evaluation of different control struc-
tures and their response to various dis-
turbances. The attributes of a dynamic sim-
9-1-1 Control Design Principles
ulator ideal for control system work are
The process control field is characterized by outlined in Chapter 5.
algorithms and mathematical procedures for The methodology I advocate for con-
analysis, design, and control; but if you are troller structure selection is the following:
looking for another algorithm for selection
• Define the objectives of the control sys-
of controller structures you will not find it
tem and the nature of the disturbances.
here. Not that I am against algorithms
• Understand the principles of the process
per se, I just happen to believe that they
in terms of its dynamic behavior.
compete poorly against a combination of
• Propose a control structure consistent with
other activities in process control design,
the objectives and the process characteris-
namely, definition of the control system ob-
tics.
jectives, process understanding and rigorous
• Assign controllers and evaluate through
dynamic simulation. Others have expressed
simulation the proposed control structure
a similar view (Downs and Doss, 1991). Be-
with the anticipated disturbances.
cause these three activities are central to
the design principles expressed in this chap- Before we embark on the details of this
ter, I will say a few more words about each methodology it will be useful to first exam-
of them. ine the nature and roles of the manipulative
Defining and understanding the control variables that will be used by the controllers
system objectives should be a collaborative in the proposed control structures.
effort between process engineers, control
engineers, and plant personnel. Left to any
9-2 MANIPULATIVE VARIABLES
one of these contributors alone, the objec-
tives can be severely biased. For example, a When a process engineer works with a
control engineer might be tempted to make flowsheet or a detailed steady-state simula-
the control system too complex in order for tion of a distillation column a certain num-
it to do more than is justified based on ber of variables have to be specified or
existing disturbances and possible yield and "controlled" in order to get a solution. For
energy savings, whereas a process engineer example, for a two-product column two
might underestimate what process control specifications have to be given (e.g., product
180 Practical Distillation Control
broadens considerably for a simple two- are only left with five dynamic degrees of
product column. However, the steady-state freedom. This means that only five levels
and dynamic degrees of freedom remain and pressures can be controlled with the
unchanged as 2 and 3, respectively. control valves provided. A close look at
Figure 9-3, however, shows that there are
six integrating variables that we might at-
9-2-1 Manipulative Variables and
tempt to regulate; two column pressures
Degrees of Freedom
and four levels. How do we resolve this
The relationship between control valves in dilemma? The answer is that we should not
the process and degrees of freedom for con- attempt to control all levels and pressures
trol is important for the proposed design at the same time with the control valves
methodology, and I will therefore illustrate provided. Qualitatively it is not hard to see
the key ideas with a few more examples. that the level in the base of the second
Consider first the extractive distillation column primarily depends on the total
system in Figure 9-3. It has 10 control valves amount of extractive agent in the system
as we might expect from combining two and not on the bottoms flow rate or boilup
columns. However, the process has 5 for that column. An attempt to control the
steady-state degrees of freedom and not 4 level in the base of the second column with
as we initially would assume (four composi- say bottoms flow, while all other levels are
tions and the recycle rate of extractive also controlled, would lead to an unstable
agent). Figure 9-4 shows one possible com- system.
bination of values or manipulators that could For another example of the relationship
be used to reach a solution on a steady-state between degrees of freedom and control,
simulator. Because the total degrees of free- consider Figure 9-5, which shows an
dom for actual plant operation equals the azeotropic column with an entrainer. The
number of control valves in the process, we entrainer separates from the light compo-
nent in the overhead decanter and is then should be allowed to float. This way any loss
typically refluxed back into the column. To of entrainer will show up in the reflux drum
simulate such a column at steady state we instead of in the column where it affects the
need three manipulators. Two of these could separation. In some installations I have seen
be the condenser and reboiler duties, but this loss of a dynamic degree of freedom
the third has to be the entrainer reflux rate was not recognized and the entrainer was
because this would be the only way of intro- level controlled with reflux, which caused
ducing that component to the balance equa- the amount of entrainer to vary indetermi-
tions. Of the six control valves in Figure 9-5 nately in the column. This resulted in col-
we would thus need three for steady-state umn instabilities and confusion for the op-
control, leaving only 3 dynamic degrees of erators. If there is a need to make up en-
freedom. However, we have four integrating trainer on a continuous basis, the entrainer
variables we might want to control; column level in the drum can be controlled with the
pressure, the base level, the decanter inter- entrainer makeup valve.
face level, and the decanter total level. We To summarize this section, we can say
have thus lost 1 dynamic degree of freedom. that the total degrees of freedom for actual
In practice it means that the level of en- plant operation equals the number of valves
trainer in the decanter should not be con- available for control in that section of the
trolled by the refl:ux of entrainer. Instead plant. To find out how many integrating
the entrainer reflux should be flow con- variables (pressures and levels) we can con-
trolled or set by some other variable, and trol with the available valves, we first have
the amount of entrainer in the reflux drum to subtract the degrees of freedom required
Selection of Controller Structure 183
t=====:J. Entrainer
for steady-state control from the total de- ery, or yield and not on how well it holds
grees of freedom. Recycle streams will re- integrating variables such as levels and pres-
duce the dynamic degrees of freedom if the sure. The integrating variables must obvi-
recycle contains a component that neither ously be controlled too, but their control
enters nor leaves the process. If it is impor- performances do not directly translate into
tant to control the inventory of such profits. However, as all practitioners know,
"trapped" components, it is necessary to inventory controls can be the most trouble-
provide extra control valves to account for some of all loops and can preoccupy the
the loss in degrees of freedom. operators to the point where the economi-
cally important composition and recovery
9-3 A METHODOLOGY loops are almost neglected. This problem
was recognized early by industrial practi-
FOR SELECTION OF
tioners (Buckley, 1964) and has been re-
CONTROLLER STRUCTURE
solved by designing the level and pressure
The economic performance of a distillation controls before dealing with the composi-
system is linked to its steady-state degrees tions controls. I too have found this to be a
of freedom. In other words the economic very workable approach. However, I have
benefits of a column control scheme depend also found that the choice of manipulators
on how well it controls compositions, recov- for inventory control significantly impacts
184 Practical Distillation Control
the control performance of the composition denser cooling will be good candidates for
loops (Tyreus, 1987). Others have made the pressure control. On a column with a partial
same observation (WaIler et al., 1988a). condenser we need to determine whether
I am now ready to summarize a method- removing the vapor stream affects pressure
ology for selection of controller structure: more than condensing the reflux. Some-
times the dominating effect is not immedi-
1. Count the control valves in the process ately obvious. If the vent stream is smaIl we
to determine the overall degrees of free- might assume that condenser cooling should
dom for control. be manipulated for pressure control. How-
2. Determine from a steady-state analysis ever, if the vent stream contains noncon-
the steady-state degrees of freedom. densables, these will blanket the condenser
3. Subtract the steady-state degrees of free- and affect condensation significantly. In this
dom from the overall degrees of freedom situation the vent flow, although smaIl, is
to determine how many inventory loops the best choice for pressure control.
can be closed with available control There is only one caution in applying the
valves. principle of using the stream with the most
4. Design pressure and level controls and impact for inventory control and it has to do
test for disturbance rejection. with the loss of dynamic degrees of freedom
S. Design composition controls based on discussed in Section 9-2-1. If the largest
product stream requirements. stream is a recycle stream, we need to ex-
6. Design optimizing controls with remain- amine if it contains components that either
ing manipulative variables. should stay completely within the process
(e.g., an extractive agent) or enter and leave
at a relatively small rate. I have already
9-3·1 Level and Pressure Controls
discussed the case of a component that re-
The assignment of manipulative variables mains in the process and the process loses a
for level and pressure control is, in princi- dynamic degree of freedom. The other case
ple, quite simple: choose the stream with where a component enters and leaves at a
the most direct impact. For example, in a small rate compared to what is recycled in
column with a reflux ratio of 100, we have the process presents a similar problem. If
101 units of vapor entering and 100 units we use the larger recycle stream for inven-
leaving the reflux drum for every unit of tory control without satisfying the overall
distillate leaving. We should therefore use material balance for the component, we also
reflux flow or vapor boilup to control the lose a dynamic degree of freedom. I will
drum level. If we were to violate this assign- illustrate this point further in the examples.
ment principle and select distillate flow for
level control, it only takes a change of
9-3·2 Composition Controls
slightly more than 1% in either vapor boilup
or reflux flow to overwhelm the drum level The composition control loops on a column
controller and saturate the distillate valve. are the most important steady-state con-
Finding manipulative variables for pres- trols. The purpose of composition control is
sure control obeys the same general princi- to satisfy the constraints defined by product
ple as for level control. Column pressure is quality specifications. The constraints must
generated by boilup and is relieved by con- be satisfied at all times, particularly in the
densation and venting. To find an effective face of disturbances.
variable for pressure control we need to The objective of a composition control
determine what affects pressure the most. loop is usually clear: Hold the controlled
For example, on a column with a total con- composition as close as possible to the im-
denser, either the reboiler heat or the con- posed constraint without violating the con-
Selection of Controller Structure 185
straint. This objective translates to on-aim, variables can be used for process optimiza-
minimum variance control. tion. Because process optimization should
To achieve good composition control we be done on a plantwide scale (Fisher,
need to examine two things: process dynam- Doherty, and Douglas, 1988), it is difficult
ics and disturbance characteristics. Process to provide guidelines for how the manipula-
dynamics includes the measurement, the tive variables should be used on a local, unit
process itself, and the control valve. Tight operation basis. In many cases the best
process control is possible if the equivalent strategy might be to operate the manipula-
deadtime in the loop is small compared to tive variables at their maximum values (e.g.,
the shortest time constant of a disturbance full cooling load on a condenser) or to set
with significant amplitude. To ensure a small them to a near optimum value such as con-
overall dead time in the loop, we always seek stant reflux flow (Tyreus, 1987).
a rapid measurement along with a manipu-
lative variable that gives an immediate and
9-4 EXAMPLES
appreciable response. In distillation such a
measurement often translates to a tray tem- I will illustrate the methodology for selec-
perature, and a good manipulative variable tion of controller structure with two ex-
is vapor flow, which usually has a significant amples. Because most discussions of dis-
gain on tray temperatures and compositions tillation control in the literature use
and also travels quickly through the column. conventional, two-product columns I will ex-
In situations where the apparent dead- pand the scope and use multiproduct,
time in the composition loop cannot be kept sidestream columns in my examples.
small compared to significant disturbances
we need to focus our attention on these
9-4-1 A Column with a Stripping
disturbances. Sometimes the important dis-
Section Sidestream
turbances can be anticipated or measured,
in which case direct feedforward control is Figure 9-6 shows the flow sheet configura-
appropriate. In other situations we might be tion of a column with a stripping section
able to affect the way disturbances affect vapor sidestream. The multicomponent feed
our composition variable by rearranging comes from an upstream unit in the process
other control loops, particularly the pres- and the sidestream goes downstream for
sure and inventory loops (Shinskey, 1985; further processing. The distillate is the
Tyreus, 1987; Waller et aI., 1988a). Al- product stream. It has a certain purity spec-
though several researchers have proposed ification in terms of sidestream material.
algorithms for determining the disturbance The feed contains small quantities of heavy
sensitivity of different control structures boilers that are removed with a purge from
(Waller et aI., 1988b; Skogestad, Jacobsen, the reboiler.
and Lundstrom, 1989), I have found that The control scheme objective of this col-
direct dynamic simulation of the strategies umn is to operate close to the quality con-
resulting from assignment of manipulative straint of the distillate product while pre-
variables for pressure and level control gives venting high boilers from entering the
the best insight to the viability of a pro- sidestream. The major disturbances will be
posed composition control scheme. feed rate (due to overall changes in produc-
tion rates) combined with feed composition
increases of sidestream material.
9-3-3 Optimizing Controls
At this stage, we consider all prior in-
After the pressure, level, and composition sights to the process and its dynamic behav-
loops have been assigned, there typically ior. Particularly relevant here is the control
remain a few manipulative variables. These of base level. We know that manipulation of
186 Practical Distillation Control
I+--DI(}-- CW
-+--+-
D =12 Ib-molhl
1 ppm impurity
~~------4=~--~
Vapor
SS .. 106.51b-moIIh
0.07 wt% impurity
the vapor sidestream has a very indirect of six control valves giving 6 degrees of
effect on the base level. For example, if the freedom. Three of these are needed to con-
level increases and we attempt to reverse trol the process at steady-state, leaving three
this trend by removing more sidestream va- manipulative variables for pressure and level
por, our actions will not be felt by the level controls. The column has a total condenser
until the reduction in vapor flow up the free from noncondensables. We therefore
column has resulted in a decrease in reflux select condenser cooling as the best choice
flow and the liquid has worked itself down for pressure control. Control of the reflux
the column. Furthermore, this scheme as- drum level is also straightforward. Because
sumes that reflux flow controls drum level the reflux ratio is 100:1, reflux flow is the
or that there is a tray temperature con- only reasonable manipulator for drum level.
troller manipulating reflux. Because the col- However, we need to worry about the po-
umn has many trays, it will take appreciable tential loss of a dynamic degree of freedom
time for reflux changes to be felt in the unless we ensure that the material balance
base. For this same reason reflux flow is for the distillate product is satisfied. One
also a poor manipulator for base level con- simple way of forcing distillate component
trol. to leave the column when it accumulates is
We next assign manipulative variables for to ratio the distillate flow to the reflux flow.
pressure and level controls. We have a total The effective manipulator for reflux drum
Selection of Controller Structure 187
Tme(O·3tn)
(a)
!iD..---------------:T18
17.5
15.5
(b)
FIGURE 9-7. Inventory controls for the example col-
umn with stripping section sidestream. FIGURE 9-8. Inventory controller response to a 20%
feed rate increase.
effective in removing the noncondensables At this point it might seem that we are
and thereby quickly increasing heat transfer done with the inventory controls and ready
whenever needed. to tackle the composition controllers. How-
Control of base level is pretty much re- ever, we need to take a closer look at how
stricted to the use of reboiler steam due to sidestream material leaves the system. The
the large vapor boilup to tails flow ratio. If partial condenser separates the noncon-
we used sidestream flow for base level con- densables from sidestream material and re-
trol we would introduce deadtime in the cycles essentially all sidestream component
level loop from reflux moving down the col- back into the column. The base level con-
umn. If we used tails flow we would have a troller will also send most of this compo-
range ability problem with the tails flow con- nent back into the column. To prevent
trol valve. To ensure that bottoms flow ma- sidestream component material from get-
terial is provided a way out from the column ting trapped in the column and cause a loss
we could ratio tails flow to steam. Alterna- of a dynamic degree of freedom we need to
tively, we could manipulate bottoms flow provide a route for it to escape. This can be
directly by way of a tray temperature con- accomplished by ratioing the sidestream to
troller. the partial condenser reflux.
t ~ BcItIrns~-;- -
7
6
I Feed rate ,
(lb mol/h)
5
" Sidestream rete (lb mol/h)
4
t
3 I ')
~r-----------------------_.~
I
52
51.5 _
51 ~ c
50.5~ ~
50 ! ~
~
49.5 E ~
I
• 49
!l
'" 46'!
48.5~
46
... 47 ~
46
47.5
47 ~~~~~~--~~--~~~~,~
BOOr--------------------------,
7110
~~~~~~~~~~~~~~~
Tllne (0-3 h)
o 0.5 1.0 Time (h)
(e)
(e)
FIGURE 9·13. Inventory controller response (a) to a
FIGURE 9·12. Inventory controller response to a 10% increase in feed rate and a 7% feed composition
20% feed rate increase. increase of high boiler; (b) to a simultaneous feed rate
and feed composition increase; (c) to feed rate and
feed composition increase.
best recovery. As before the optimum policy any remaining manipulative variables can be
for these variables should come from a used for steady-state optimization. This
plantwide optimization but for this column could amount to something as simple as
chances are that maximum cooling in the holding a reflux rate constant or maximizing
condenser could be justified to minimize condenser cooling, but it could also be the
vent losses of product material. The result of running a plantwide optimizer.
sidestream to reflux ratio will then deter-
mine the recovery of product from the bot-
References
toms stream. Buckley, P. S. (1964). Overall process control.
Techniques of Process Control. New York:
9-5 CONCLUSION Wiley, Chapter 13.
Downs, J. J. and Doss, J. E. (1991). Present
An empirical method for selection of distil- status and future needs-A view from North
lation column controller structures has been American industry. Paper presented at the
presented. We first determine how many Fourth International Conference on Chemi-
degrees of freedom exist for closing inven- cal Process Control, CPC IV, South Padre
tory loops on the process. These inventory Island, 1991.
loops are then assigned by using manipula- Fisher, W. R., Doherty, M. F., and Douglas,
J. M. (1988). The interface between design
tive variables with the most impact on the
and control. 3. Selecting a set of controlled
controlled variables. Ratio controls are pro- variables. Ind. Eng. Chern. Res. 27, 611.
vided at this step to prevent a near loss of Shinskey, F. G. (1985). Disturbance-rejection ca-
dynamic degree of freedom from internal pabilities of distillation control systems. Pro-
recycles. The resulting inventory control ceedings of the American Control Confer-
scheme with appropriate ratio controls is ence, Boston, MA, June 19-21.
then tested through dynamic simulation to Skogestad, S., Jacobsen, E. W., and Lundstrom,
verify its disturbance rejection capability. P. (1990). Selecting the best distillation con-
The ratio controllers usually provide excel- trol structure. AIChE I. 36, 753-764.
lent disturbance rejection toward feed rate Tyreus, B. D. (1987). Optimization and multi-
changes, eliminating the need for separate variable control of distillation columns. Pro-
feedforward controllers. ceedings of the ISA-87, ATUlheim, California.
Composition controllers are assigned Waller, K. V., Finnerman, D. H., Sandelin,
P. M., Haggblom, K. E., and Gustafsson,
next. Here the objective is to find fast loops
S. E. (1988a). An experimental comparison of
that measure composition or a surrogate for four control structures for two-point control
composition. This usually amounts to mea- of distillation. Ind. Eng. Chern. Res. 27, 624.
suring a tray temperature and controlling it Waller, K. V., Haggblom, K. E., Sandelin, P. M.,
with a manipulator that directly or indi- and Finnerman, D. H. (1988b). Disturbance
rectly affects the vapor rate in the column. sensitivity of distillation control structures.
After the composition loops are assigned, AIChE I. 34, 853-858.
10
Control Structures,
Consistency, and
Transformations
Kurt E. Haggblom and Kurt V. Waller
Abo Akademi
192
Control Structures, Consistency, and Transformations 193
In principle, the basic variables can be the assumption of "constant flows" in the
transformed in an unlimited number of ways column sections. As this assumption re-
resulting in an unlimited number of possible quires that the heat of vaporization of
control structures with different sets of in- the mixture distilled is independent of its
put and output variables. Even if the composition, it may be a very inaccurate
above-mentioned suggestions only are con- assumption for real columns. Another draw-
sidered, the number of different control back with the methods based on this as-
structures for a distillation column is huge. sumption is that the boilup rate usually is
Because the properties of the control struc- used as a manipulated variable. In reality,
tures and distillation columns vary consider- however, the boilup rate is always a depen-
ably, the selection of an appropriate control dent variable that cannot be manipulated
structure is undoubtedly one of the most directly.
important tasks-perhaps the most impor- Haggblom (1986) and Haggblom and
tant one-in the design of distillation con- Waller (1986, 1988a) have developed rigor-
trol systems. ous analytical methods, which do not re-
A prerequisite for a rational selection of quire any limiting assumptions, for the mod-
control structure is that one can assess the elling of control structures based on
feasibility of different structures. In prac- steady-state models. Also Skogestad and
tice, this assessment requires that all control Morari (1987b) have presented similar
structures under consideration be described methods. By these methods, the gains of
by simple analytical models. It is possible to any control structure that can be described
determine such models from simulations or by a steady-state model can be determined
experiments with the control structures in from the gains of any other adequately
question, but obviously this is not an attrac- modelled control structure. It has also been
tive approach. Furthermore, it is not a triv- shown that the process gains of a control
ial task even to determine an adequate lin- structure have to satisfy certain consistency
ear steady-state model from simulated data relations, which follow from the external
for a nonlinear process (Haggblom, 1988b). material balances for the column. These
Therefore, it would be desirable to have relations can, for example, be used to show
methods by which models for different con- that the assumption of "constant flows" in
trol structures could be derived analytically the column sections generally is incorrect
from a single process model. Such an ap- and even unnecessary.
proach would be time saving and it would Recently the methods for modelling of
also make the whole procedure of control control structures have been extended to
structure selection suitable for computer dynamic models (Yang et aI., 1990;
implementation. Haggblom and Waller, 1991). The general
Actually, methods for modelling control dynamic case is much more complicated
structures have been proposed and used in than the steady-state case. There are two
the literature. McAvoy (1983) and Shinskey situations, however, when dynamic models
(1984), for example, use approximate analyt- can be treated essentially as steady-state
ical models in the derivation of linear mod- models (Haggblom and Waller, 1988a, 1991).
els for different distillation control struc- One situation is when the reflux drum and
tures. Other attempts at deriving control reboiler holdups can be considered per-
structure models have been made, for ex- fectly controlled. The other situation is when
ample, by Tsogas and McAvoy (1981), Krid- the same inventory control system is used in
iotis and Georgakis (1986), Skogestad and different control structures. This is possible,
Morari (1987a), Takamatsu, Hashimoto, and for example, if the reflux drum holdup is
Hashimoto (1987), and Koung and MacGre- controlled by L + D and the reboiler holdup
gor (1991). All these methods are based on by V' + E, where V' is an estimate of the
194 Practical Distillation Control
treatment, the pressure control loop is ne- control structure is to be derived from the
glected.) model for the (L, V) structure. See Sections
A steady-state model (Le., the steady- 10-4 and 10-5.
state part of a dynamic model) for the (L, V) The flow gains are also important in cer-
structure can be written tain "generalized" inventory control
schemes (Haggblom and Waller, 1991). For
example, it would be optimal from a
steady-state point of view to control the
reflux drum holdup by aD - Kf;[ aL -
Kf;~ av if F and z are not measured. If liD
is the controller output, D would then be
determined from aD = allD + Kf;[ aL +
Kf;r av. This means that changes in Land
V are fed forward to D before they affect
the holdup in the reflux drum. A suboptimal
approach would be to use feedforward only
from L, that is, to calculate D from aD =
allD + Kf;[ aL. Note that D is still a de-
pendent variable in these generalized inven-
where y and x are the primary outputs. tory control schemes. This means that
Two disturbance variables, the feed flow Equations 10-1 may be used to describe the
rate F and the feed composition z, are also (L, V) structure irrespectively of how the
included. a denotes a deviation from a inventory control system is implemented.
nominal steady-state value (e.g., ay = y - Most works on control structure mod-
y, where y is the nominal steady-state value elling reported to date are based on the
for y). The superscript LV is used to distin- assumption of "constant flows" in the col-
guish the gain between two variables (de- umn sections. Basically, this assumption re-
noted by subscripts on K) in the (L, V) quires that the heat of vaporization (ex-
structure from the gain between the same pressed in units consistent with the flow
two variables in another control structure rates) of the mixture distilled is indepen-
(see Equations 10-2 and 10-3). dent of its composition and that the reflux
This model of the (L, V) structure may stream is saturated liquid. According to this
not look familiar. When transfer function assumption, the flow gains in Equation 1O-1b
models are used in the literature, usually would have the following values: Kf;[ =
only the first part, Equation 10-la, is con- -KJif = -1, Kf;r= -KJi~ = 1 (if V de-
sidered or even recognized. However, when notes boilup), Kf;~ = 1 - KJi]: = qF (where
the inventory control loops are closed, the qF expresses the degree of vaporization of
inventory control manipulators become de- the feed such that qF = 0 for saturated liq-
pendent variables, and Equation 1O-1b uid and qF = 1 for saturated vapor), and
shows how they depend on L, V, F, and z. Kf;~ = KJi~ = O. These flow gains mean that
The numerical values of the flow gains in the relationship aD = - aL + av +
this equation are far from irrelevant. They qF aF, or D = - L + V + qFF (+ a con-
mainly depend on the properties of the col- stant) would hold.
umn itself, not on the inventory control In Section 10-3 we will show that the flow
system. Therefore, both the primary and gains of the (L, V) structure are related to
secondary gains in any control structure the composition gains in Equation 1O-1a
where D or B, or some function of them, is through so-called consistency relations that
used as a primary manipulator will depend can be rigorously derived from the external
on these flow gains. This means that the material balances for the column. This
flow gains are needed if a model for such a means that the preceding assumptions can
196 Practical Distillation Control
'Wood and Berry, 1973; Hiiggblom and Waller, 1988b (mass units, V is steam).
bWaller et aI., 1988a; Hiiggblom, 1989 (mass units, V is steam).
CHiiggblom, 1988b; Hiiggblom and Waller, 1989 (molar units, V is heat).
dMcAvoy and Weischedel, 1981; Hiiggblom and Waller, 1988a (molar units, V is boilup).
·Yang et aI., 1990 {molar units, V is boilup, z is n-butanoO
be checked if the composition gains and crease, even if the heat input to the column
some steady-state data are known. Table (or the vapor boilup in the bottom of the
10-1 shows calculated flow gains for some column) remains unchanged. As a result of
distillation columns taken from the litera- this, the overhead vapor flow rate will in-
ture. As can be seen, the flow gain K fjL for crease, which in tum means that the distil-
example, is often in the range - 0.6 to - 0.8, late flow rate D does not decrease as much
but a value as low as - 0.25 has also been as L is increased. Hence, Kfjr > - 1, and
found. These values are even more remark- in practice, -1 < Kfjr < o. A more rigor-
able considering that the reflux in experi- ous treatment of the flow dynamics is given
mental columns with a total condenser (as in Haggblom (1991).
the column for which Kfjr = - 0.25) usu-
ally is subcooled, in which case one would 10-2-2 Material Balance Structures
expect K fjr < - 1 according to the assump- (0, V) and (L, B)
tion of constant flows. This means that the
"standard" assumption of constant flows and Two other frequently used control struc-
the corresponding values of the flow gains tures are the material balance structures
given previously generally are incorrect. (D, V) and (L, B). In the (D, V) structure,
Moreover, the assumption is unnecessary D and V are used as primary manipulators
when the flow gains can be determined from whereas Land B usually are used as inven-
the composition gains. tory control manipulators. (Figure 15-4
Because of the practical importance of shows an implementation of this control
flow gains like K fjL an explanation as to scheme.) The (D, V) structure may thus be
why K fjr often is significantly different from described by the steady-state model
-1 is appropriate. In practice, all mixtures
are nonideal and the heat of vaporization, K~:][fl.Dl
for example, is dependent on the composi- KDV fl.V
xV
tion of the mixture. This is especially true if
the composition is not expressed in molar
units. When the reflux flow rate L is in-
K~tl[fl.Fl
K DV fl.z (1O-2a)
xz
creased, the concentration of light compo-
nent in the column will also increase (if
other inputs are constant). This usually [fl.L] [Kfb Kft] [fl.D]
=
fl.B K3b K3t fl.V
means that the heat of vaporization of the
KDV KDV][fl.F]
+ [ DV K ~v fl.z
mixture in the column will decrease, which LF
has the effect that the amount of vaporized (l0-2b)
KBF Bz
liquid in all parts of the column will in-
Control Structures, Consistency, and Transformations 197
In the (L, B) structure, the opposite decom- 1984, pages 128-129) used an inventory
position into subsystems is made, that is, the control scheme for the reflux drum where a
primary outputs are controlled by Land B, change in D was fed forward to a corre-
and the column inventory is usually con- sponding change in L. Van Kampen found
trolled by D and V. The steady-state model it beneficial to increase the factor k by
for the (L, B) structure is thus which variations in the distillate was ampli-
fied to above unity. The reflux flow rate was
thus determined from ilL = IlllD - k IlD,
where liD is the output from the inventory
controller. From Equation 10-2b it follows
that k should ideally be chosen as k =
- KEK. Because KEK = 1/KJ;r (see Sec-
tion 10-4), van Kampen's suggestion that
k > 1 is in agreement with -1 < KJ;r < 0,
which is the normal situation in "real"
columns (see Table 10-1 and Section 10-3).
An interesting property of the two mate-
rial balance structures is that the primary
KLB gains in each control structure are not inde-
+ [KLB KLB][IlF]
DF
:: (1O-3b) pendent of each other if the primary out-
VF Kvz Ilz
puts are product compositions. For exam-
ple, for the (D, V) structure it can be shown
As for the (L, V) structure, Equations 10-2b (see Section 10-3) that i5K~J' + BK;'J' = O.
and 10-3b are valid irrespectively of how the Similar relations exist between other pri-
inventory control system actually is imple- mary gains both in the (D, V) and the (L, B)
mented (provided that the implementation structure. This has important practical con-
is compatible with the choice of primary sequences when models for these control
manipulators in the control structure). structures are determined through identifi-
These control structures may seem very cation from numerical (experimental or sim-
similar to the (L, V) structure. The only ulated) data: the gains should be deter-
difference between the (L, V) and the mined in such a way that these consistency
(D, V) structure is that the roles of Land relations are satisfied; otherwise the exter-
D are switched. However, this has a pro- nal material balances (i.e., the law of mass
found effect on the properties of the control conservation) will be violated.
structures; the properties are generally very For the material balance structures it is
different (Waller et aI., 1988a, 1988b). This not possible to determine all flow gains from
also applies to the disturbance gains, al- consistency relations. If the model is deter-
though they are gains between the same mined from numerical data, this means that
input and output variables. This has to do some of the flow gains have to be explicitly
with the fact that the assumption of con- determined from such data. It is then possi-
stant flows does not generally hold. All these ble to design, for example, a generalized
differences are disclosed by the analytical inventory control scheme. It is also possible
relations existing between the gains in the to derive models for other control struc-
different control structures (see Section tures, for example the (L, V) structure.
10-4). Without the flow gains, this is not possible.
A generalized inventory control scheme If a complete steady-state model, including
similar to the ones discussed for the (L, V) all flow gains, is known for a control struc-
structure has been used with the (D, V) ture, it is possible to derive the model for
structure for a long time in industrial opera- any other control structure (see Section
tion. Van Kampen (1965) (see Shinskey, 10-5).
198 Practical Distillation Control
vLV + y-KDV
--:~yV
D LV + BKxV
LV + XKLV
BV = 0 (10-7b)
vLV + y-KDz
LV + BKxz Bz = P
LV + XKLV
[ Kk:
KfJ~ KfJ~][IlF] --:~yz
D (10-7d)
+ (lO-lb)
Kk:' Ilz Equation 10-6a, for example, simply
states that a change in L affects D and B
where y, x, and z now specifically denote so that the sum of the resulting changes in
the (mole or weight) fraction of the same D and B is zero at steady state, which is
component in the distillate, bottoms, and completely in accordance with Equation
feed stream, respectively. This model has to 10-5a (IlF = 0 because L does not affect
satisfy the external material balances for the F). Similarly, the change in L will also, due
column, that is, to the component material balance in Equa-
D +B =F (to-4a) tion 10-5b, affect y, x, D, and B in such a
Dy + Bx = Fz (l0-4b)
way that Equation 10-7a holds.
If the primary outputs are product com-
Linearization of the material balances at positions expressed as fractions of the same
the nominal steady state (D, B, P, y, X, z) component, the gains in Equations 10-1 have
Control Structures, Consistency, and Transformations 199
to satisfy Equations 10-6 and 10-7. If the literature. The listed references are the
primary outputs are any other kinds of vari- original source of the model and the work
ables (e.g., temperatures), or if they are where the gains were determined. The unit
fractions of different components, only for the disturbance in the feed composition
Equations 10-6 apply. is weight or mole percent of the light com-
Equations 10-7 can be rearranged by ponent, except for the ternary butanol col-
means of Equations 10-6 to give umn, where it is mole percent of the least
volatile component. It should also be noted
DK yL
LV + BKxL LV that the boilup manipulator varies in the
Kf;r = -KkK = y -x
(lO-8a) five cases. For the methanol-water and the
ethanol-water column it is the flow rate of
DK yV
LV + BKxV LV steam to the reboiler, for the methanol-
KDVLV -- _KLV-
BV- (1O-8b) ethanol column it is the flow rate of heat
}i-x from the reboiler to the column, and for the
z - x - DK LV - BK LV benzene-toluene and the ternary butanol
Kf;~ = 1 - Kk~ = yF xF column it is the boilup rate. As can be seen,
}i-x the flow gains of "real" columns can drasti-
(1O-8c) cally differ from the flow gains following
from the assumption of constant flows.
LV = _KLV __ F - DK LV - BK LV
xz Note that although the flow gains in
KDz Bz
yz
Y_ X Equation 1O-1b often can be calculated from
consistency relations (Equations 10-6 and
(1O-8d)
10-7), it is advisable when a model is deter-
mined from numerical data first to estimate
These equations show that if nominal all gains, including the flow gains, then to
steady-state data and the gains in Equation reconcile them subject to the consistency
10-1a are known, the gains in Equation relations. This will reduce the uncertainty
10-1b can be calculated when the primary (variance) of the parameter estimates
outputs are product compositions expressed (Haggblom, 1988b).
as fractions of the same component. Unless
the primary gains and the nominal steady-
10-3-2 (D, V) Structure
state data are related in a very specific way,
the popular assumption of constant flows is Let us next consider the (D, V) structure,
obviously incorrect and, in fact, inconsistent which is described by the steady-state model
with the external material balances. [Recall
that this assumption gives the flow gains
Kf;r = -KkK = -1, Kf;~ = -Kk~ = 1 (if
l
ay] = [K~; K~:][aD]
DV KxV DV av
V denotes boilup), Kf;~ = 1 - Kk~ = qF' ax KxD
and Kg = Kk;' = 0.] Moreover, the as-
sumption is unnecessary if the primary out-
puts are fractions of the same component. +[ K~[
Klj[
K~t][aF]
Kljt az
(1O-2a)
If the primary outputs do not fulfill the
requirements (e.g., because they are tem-
peratures), Equations 10-8 do not apply and
[aLl = [KEK Kf!][aD]
DV KBV
KBD DV av
the assumption of constant flows cannot be ""AB
checked, but this does not make the as-
sumption correct, of course.
Table 10-1 shows values of the distillate + [KE:DV KE;][aF]
DV (1O-2b)
flow gains for five columns studied in the
KBF K Bz az
200 Practical Distillation Control
Consistency relations for this control struc- give an indication of the reliability of the
ture may be derived in a similar way as flow rate measurements.
those for the (L, V) structure, except that The consistency relations do not give any
now the independent variables are D, V, F, information about the secondary gains asso-
and z. When independent changes in these ciated with the inventory control manipula-
variables are considered, one at a time, tor L. Therefore, these gains always have to
Equation 10-2b and the total material bal- be included in a complete model for the
ance, Equation 10-5a, give the expressions (D,V) structure. However, in practice this
never seems to be done, although the gains
are needed, for example, if a model for the
DV - -1
K BD DV- 0
K BV (L, V) structure is to be derived from a
- , - ,
model for the (D, V) structure. Normal
DV -1
K BF - , K3i=0 (10-9) practice is to apply the generally incorrect
assumption of constant flows in the column
sections, in which case the gains would have
When these expressions are taken into ac- the values KfK = -1, Kf: = 1 (if V de-
count, Equations 10-2 and the component notes boilup), Kff = qF' and Kfi = O. The
material balance, Equation 1O-5b, give the fact that the assumption of constant flows
consistency relations cannot be checked from the consistency re-
lations for the (D, V) structure is not a
justification for using it.
DK DV
yD + BKxD
DV = X- Y- (10-lOa)
When independent changes in these vari- In this section we will treat only transfor-
ables are considered, Equations 1O-3b and mations of input variables. An application
10-5a give the consistency relations of output variable transformations is given
in Section 10-10. Furthermore, we will only
K DL
LB -- 0, K DB
LB -- -1 , consider transformations between the
LB -1
(L, V), (D, V), and (L, B) structures in this
K DF - , (10-11)
section. We will deal with the general case
in Section 10-5.
Similarly, Equations 10-3 and 10-5b to-
gether with Equation 10-11 give the consis-
tency relations
10-4-1 Transformation
from (L, V) to (D, V)
(1O-12a)
Recall the steady-state model for the (L, V)
DKLB
yB
+ iiKLB
xB
= y- - j (l0-12b) structure,
DKLB
yF
+ iiKLB
xF
= Z - y- (1O-12c)
(10-12d)
gives
KE~
ilL = -KEr
1
IlD - -KEr IlV
KE~ KE~
- ----rv IlF - K Ilz
LV (10-13)
KDL DL 1 K DV
LV
Elimination of ilL
from Equations 10-1 by K DL
LV -K LV
DL
means of Equation 10-13 now gives the
K BL
LV
model for the (D, V) structure:
K DL
LV
+ [K~r
K~:][IlF]
K K Ilz
(lO-2a)
DV DV
xF xz
nized: Published models are usually given as Elimination of av from Equations 10-1 by
Equation 10-la or 10-2a with no flow gains means of Equation 10-15 gives the model
included. If needed, the flow gains are usu- for the (L, B) structure:
ally chosen according to the assumption of
constant flows. As was shown in Section
10-3, this assumption is generally incorrect.
Equation 1O-14b clearly shows that dis-
turbances generally affect the outputs dif-
ferently in two structures, that is, generally
the disturbance sensitivity of different struc-
tures is different. This has also largely been
overlooked in the distillation control litera-
ture. One reason for this is the simplifying
assumptions commonly used in simulation
studies (e.g., the assumption of constant
flows), which, for example, result in the flow
gains Kf;~ = 0 and Kf;~ = O. According to KLB
Equation 1O-14b, the disturbance gains of [ DF KLB][4F]
Dz (10-3b)
+ LB KLB 4z
the (L, V) structure and the (D, V) struc- KVF Vz
ture would then be equal. However, from
practice it is known that generally this is not
true (Waller et aI., 1988a, 1988b). where
Equations 10-14 are convenient for calcu-
lation of the gains in the (D, V) structure if
a model for the (L, V) structure is known.
If the model for the (D, V) structure is
known, the same equations can be used to
calculate the gains in the (L, V) structure.
10-4-2 Transformation
from (L, V) to (L, B)
A model for the (L, B) structure can be
derived from a model for the (L, V) struc-
ture in the same way as the model for the
(D, V) structure. In the (L, B) structure, (10-16a)
the reboiler holdup is controlled by V in-
stead of by B as done in the (L, V) struc-
ture. For this situation, the lower part of
Equation 10-1b gives
1 LV
K BL
4V= - 4 B - - 4 L
LV
K BV LV
K BV
KkJ Kk;'
- K LV 4F - K LV 4z (10-15)
BV BL (10-16b)
204 Practical Distillation Control
[ ~~~ ~~~l
VF Vz
ture cannot be modelled by a steady-state
model, we will not consider it further in this
chapter.
K LV However, there are numerous other con-
K LV _KLV~
Dz DV KLV
BV trol structures, for example, control struc-
tures where flow ratios are used as mani-
K Bz LV
pulators. Because it gets very tedious and
- K LV
BV complicated to treat such control structures
(10-16d) as explicitly as we have done in the previous
sections, and the list of such control struc-
When the consistency relations in Equa- tures is endless, we will in this section in a
tion 10-6 between the flow gains in the unified manner treat all possible control
(L, V) structure are taken into account, the structures that can be described by steady-
gains associated with the output D in Equa- state models. Basically, we continue to con-
tions 1O-16c and 1O-16d become sider simple two-product distillation, but as
we show in Section 10-5-4, the results apply
Kf;2 = 0, Kf;: = -1, to distillation columns of any complexity.
In order to derive and state the general
KLB
DF -1
- , Kf;: = 0 (10-11)
results we have to use a compact matrix
In Section 10-3-2 we derived the same ex- notation. The matrix operations used are
pressions without using the consistency rela- very elementary, however; it is mainly the
tions for the (L, V) structure. compact notation we need.
Conclusions similar to those drawn for
the (D, V) structure from Equations 10-14 10-5-1 Compact Description
can be drawn for the (L, B) structure from of Control Structures
Equations 10-16. Equations 10-16 can be
used, for example, to calculate the gains for Consider the model for the (L, V) struc-
the (L, B) from the gains for the (L, V) ture, Equations 10-1. If we define y =
structure or vice versa. [y X]T, U = [L V]T, v = [D B]T, and
w = [F Z]T, the model can be compactly
expressed as
10-5 CONTROL STRUCTURE
MODELLING - THE GENERAL CASE ay = K)'Uau + Kyw aw (10-17a)
In the last sections we have been dealing
with the most basic distillation control
av = Kvu au + Kvw aw (10-17b)
Control Structures, Consistency, and Transformations 205
where Kyo, K vu , Kyw, and K~ denote. the tion of deviation variables give the relation-
gain matrices explicitly shown 10 Equations ships
10-1.
With suitable definitions of y, u, v, and w, a", = B",u au + B",y av + B",.. aw
this model could be used to describe any
(IO-20a)
control structure, but in the following we
will assume that y, u, v, and ware defined as all = B .. u au + Bn av + B .... aw (lO-20b)
before, that is, they are vectors of the basic
"physical" variables of the process. We will
where the B matrices contain partial deriva-
often in general terms refer to this control
tives of the new variables with respect to
structure as the "base" structure and to the
the base variables. The columns of B ... and
variables as follows: y is a vector of primary
B that correspond to unmeasured distur-
outputs, U is a vector of primary manipula- "" in w should contain only zeros.
bances
tors, v is a vector of dependent (or sec-
Because u and v (and w) are the variables
ondary) manipulators, and w is a vector of
that physically affect the process, it must be
disturbance variables.
possible to determine u and v from "', 11,
Consider now a control structure where
and measurable disturbances in w. This
'" is the vector of primary manipulators and
means that also the relationships
11 the vector of dependent manipulators (due
to inventory control). This control structure
can be described by the model au = Mu. a", + M u" all + Mu.. aw
(10-21a)
11 = lI(U, v, w) (10-19b)
c(y,v,u,w) =0 (10-22)
The variables '" and 11 are only allowed to Linearization and introduction of deviation
be functions of variables included in the variables give the relationships
base structure. If the functions contain dis-
turbance variables, the latter have to be Fey ay + Fcv av + Fc:a au + Few aw =0
measured (or estimated) disturbances. Lin-
(10-23)
earization of Equations 10-19 and introduc-
206 Practical Distillation Control
Elimination of av from Equations 10-20 Kvu and Kvw are needed if H",v * 0, that is,
by Equation 1O-17b gives if the primary manipulators of the new con-
trol structure are functions of the secondary
atll = (H",u + H",vKvu) au manipulators in the base structure. As we
+ (H",w + H",vKvw) aw (1O-31a)
have seen, in some cases the secondary gains
can be determined by means of consistency
all = (Hvu + HvvKvu) au relations; in other cases they have to be
+ (Hvw + HvvKvw) aw (l0-31b) provided. Note also that it is unnecessary to
specify the secondary manipulators by
The primary manipulators til have to be defining H vu ' H vv , and Hvw if only the
independent of each other even when aw primary gains are to be determined. The
= O. This means that the transformations primary gains are, in other words, unaf-
have to be such that the matrix H"ID + fected by the choice of secondary manipula-
H",vKvu is nonsingular. Also, if the manipu- tors.
lators of the base structure are the variables
that in reality affect the column, it must be
10-5-4 Complex Distillation Columns
possible to determine u (and v) from til, 11,
and measurable disturbances in w. Because The expressions derived in the preceding
11 is a dependent variable, u must in princi- text are based on the model given in Equa-
ple be determinable from til and w. This tions 10-17. The results thus apply to any
implies that u can be solved from Equation process that can be described by a model of
10-31a, giving the same general form as the model in
Equations 10-17. We will show that a distil-
au = (H",u + H",v K vur 1 lation column with any number of feed
streams, sidestreams, and chemical compo-
x(atll - (H",w + H",vKvw) aw) nents can be described by such a model.
(10-32) Consider a distillation column having n F
feed streams with the flow rates F; (i =
Elimination of au from Equations 1O-17a 1, ... , n F ), ns side streams with the flow
and 1O-31b by Equation 10-32 then gives the rates Si (i = 1, ... , n s ), and a total of nc + 1
model in Equations 10-18, where components in the different streams. Let us
-1 define
K y", = Kyu(H",u + H",vKvu) (10-33a)
where XD,j' XB,j' X Fbi , and XSi,j are mole Routines for identification of consistent
or weight fractions of component j (j = models from numerical data are also in-
1, ... , n) in the distillate stream D, the cluded in the package (see Section 10-9).
bottoms stream B, the feed stream F;, and
the sidestream S;, respectively.
10-6 APPLICATION 1:
Note that the composition of a stream is
fully determined by nc component fractions NUMERICAL EXAMPLES
when the total number of components is OF CONTROL STRUCTURE
nc + 1. In spite of this, the fractions of any TRANSFORMATIONS
number of components, even nc + 1, could In this section we will illustrate by means of
be included in the composition vectors of numerical examples how models for arbi-
the product streams. However, the composi- trary control structure can be derived from
tion vectors of the feed streams should con- a model for a given base structure. As base
tain at most nc component fractions be- structure we will use the (L, V) structure.
cause we want the input variables to be The examples are control structures for a
independent of each other (mathematically). 15-plate pilot-scale distillation column sepa-
Because a change of any of the nc compo- rating a mixture of ethanol and water. The
nent fractions will affect the outputs, the primary outputs are the temperatures on
number of component fractions should gen- plates 4 and 14, counting from the top of
erally be exactly nco The component ex- the column. Table 10-3 (see Section 10-9)
cluded can, of course, be arbitrarily chosen. shows nominal steady-state data for the col-
Let us further define umn. The steady-state model for the (L, V)
structure, which was experimentally deter-
y~ u~ [H
mined and reconciled with the models for a
[:;]. v= [~], few other control structures (see Section
10-9), is
[ll.AL]
way.
= [-1.6250 2.1976] [ll.D ]
B - 1.0000 0.0000 ll.V
10-6-1 (D, V) Structure
0.0961 1.5304] [ ll.F]
In the (D, V) structure, the primary output [
+ 1.0000 0.0000 ll.z
variables are controlled by D and V, and
the secondary outputs (the reflux drum and (10-39b)
[!i] [~ ~ ][ !~] + [~ ~ ][ !~ ]
reflux drum holdup is usually controlled by
= L + D and the reboiler holdup by B, but we
(10-37b) will use Land B as inventory control ma-
nipulators. The primary and secondary vari-
ables are thus '" = [LID V]T and v =
which give the transformation matrices [L B]T. Note that the primary steady-state
gains are not dependent on the choice of
inventory control manipulators (see Section
H",u = [~ ~], H",v = [~ ~ ], H",w = 0 10-5-3).
In the transformation from the (L, V)
structure to the (D, V) structure, the vari-
Hpu = [~ ~], Hn = [~ ~], H.,w= 0 able transformations only resulted in inter-
(10-38) changes between Land D. In the (LID, V)
structure, the manipulator LID is a nonlin-
ear function of two base variables. Lin-
Use of the transformations in Equations earization of the variable transformations at
210 Practical Distillation Control
the nominal steady state (I, 15) gives lator includes a measurable disturbance.
The inclusion of F in the manipulator V IF
[!~] [~ ~][ !~ ]
similar to those of the (LID, V) structure.
= The primary manipulators for this con-
trol structure are'" = [LID VIF]T. They
+[ ~ ~][ !~] (lO-40b)
give the primary transformation matrices
_[1/D
The transformation matrices are thus H",u -
0 l~F ].
H = [1/15
0
0]
l' H = [ -LI15 2 0]
\flu \fIv 0 o'
[-LI152 0]
~]
H =
\fIv 0 0 '
(10-41)
H",w = [ -;IFZ (10-43)
H\fIw = 0
If the column inventory is controlled by L
Hvu = [~ ~], and B, the secondary transformation matri-
ces H"u' H"v' and H"w are the same as in
H"w = 0 the two previous cases. The resulting model
is
Numerically, the following model is ob-
tained:
[ tlT4] [-1.6167 1.8639] [tl(LID)]
tlT14 = - 8.3967 70.7935 tl(VIF)
~T4 ] [ -1.6167 0.0093][~(LID)]
[
~T14 = - 8.3967 0.3540 ~V 0.0008 -0.0212] [tlF]
[
+ -0.0414 -0.7545 tlz
-0.0026 -0.0212] [~F]
[
+ -0.1688 -0.7545 ~z
(lO-44a)
(1O-42a)
[ ~L]
tlB
= [37.1432
22.8568
167.4339] [~(LID)]
-167.4339 tl(VIF)
[ ~L 1 = [37.1432
0.8372] [~( LID) ]
22.8568 -0.8372 0.3380
~B ~V [ 0.5830] [tlF]
+ 0.6620 -0.5830 ~z
[
0.0366
0.583°1[~Fl (lO-44b)
+ 0.9634 -0.5830 ~z
(1O-42b)
10-6-4 (L I D, V I B) Structure
The control structure considered by many
10-6-3 (L I D, V I F) Structure
to be the best general control structure is
The (LID, VIF) structure is an example of the (LID, VIB) structure (Takamatsu,
a control structure where a primary manipu- Hashimoto, and Hashimoto, 1982; Shinskey,
Control Structures, Consistency, and Transformations 211
1984; Skogestad, Lundstrom, and Jacobsen, TABLE 10·2 Steady·State Data for the Wood
1990). Due to the flow ratios used as pri- and Berry Column
mary manipulators, this control structure Flow rate Composition
has a certain built-in disturbance rejection Flow Ob/min) (wt% methanol)
capability. The interaction between the pri-
Feed, F 2.45 46.5
mary control loops as measured by relative
Distillate, D 1.18 96.0
gains is also considerably lower than, for Reflux, L 1.95 96.0
example, in the (L, V) structure (Shinskey, Bottoms, B 1.27 0.5
1984; Haggblom, 1988a). Steam to reboiler, V 1.71
The primary transformation matrices for
this control structure are
most studied model in the distillation con·
liD
,. =
H..,.u [ 0
trol literature over the years. The process is
a methanol-water distillation in an eight·
plate pilot-plant column. Nominal steady·
[ Ll y] = [12.8 -18.9][t:.L]
[ t:.L]
t:.B =
[44.0223
15.9777
81.9295] [t:.(L/D)]
-81.9295 t:.(V/B)
t:.x 6.6 -19.4 t:.V
0.3265 (10-47)
[ 0.4075] [t:.F]
+ 0.6735 -0.4075 t:.z
(l0-46b)
The process model is given in the usual
form, showing the gains associated with the
primary outputs, but not the flow gains. As
10-7 APPLICATION 2: USE OF we have already pointed out, the flow gains
are needed if a model for another control
CONSISTENCY RELATIONS IN
structure is to be derived from the model
TRANSFORMATIONS
for the (L, V) structure.
In this section we will demonstrate the use- However, because the outputs in this case
fulness of consistency relations in control are product compositions, and appropriate
structure transformations. As an example steady-state data are known, the consistency
we will use the well known Wood and Berry relations following from the external mate-
(1973) column model, which is probably the rial balances can be used to calculate the
212 Practical Distillation Control
flow gains. Equations 10-8 give the flow the feed flow rate in the Wood and Berry
gains column. As shown by Equations 10-47 and
10-51, Ryskamp's control structure has a
KEr = -Kkt = -0.25 (10-48a) significantly larger steady-state sensitivity to
disturbances in F than the (L, V) structure.
KE~ = -Kk~ = 0.49 (lO-48b) However, the difference is smaller than for
the other structures.
KE~ = 1 - KkJ = 0.37 (1O-48c)
We would once more like to stress the
importance of using the correct flow gains
These values differ considerably from the in control structure transformations. For ex-
values generally used in the literature, that ample, if the standard assumptions (KEr =
is, KEr = -1, KE~ = 1, and KE~ = o. -1, KE~ = 1, and KE~ = 0) were used in
Now all necessary data for arbitrary con- the calculation of the model for the (D, V)
trol structure transformations are available. structure from the model for the (L, V)
For the (D, V) structure, Equations 1O-14a structure, the result would be
and 1O-14b give the model
I
[GLY GLY ]
One situation is when the reflux drum and
yL yV
reboiler holdups can be assumed to be per-
fectly controlled (Haggblom, 1986). In real- GxL GxV
LV LV
ity, this assumption is always an approxima-
tion, but because inventory control is a much [ 2454e-' -3493e-'
easier task than composition (or tempera- 20s + 1 22s + 1
ture) control, it can usually be made to work - 3.270e- s _ 5 .51ge -s (10-54)
I
drum holdup is controlled by L + D and
the reboiler holdup by V' + B, where V'is [GDY GDV ]
yD yV
~[
tem has many advantages (Shinskey, 1990; 0.3986e-'
Haggblom and Waller, 1991). In Ryskamp's 20s + 1 5s + 1
control structure, for example, the reflux -1.917e- s -0.332ge- s
drum holdup is usually controlled by L + D. 20s + 1 4s + 1
By means of an example taken from the
(10-56)
literature, we will next illustrate that the
transformations derived for steady-state
models can be applied to dynamic models The primary outputs are the mole fractions
when the reflux drum and reboiler holdups of the more volatile component on plates 3
(levels) are perfectly controlled. and 17. Time is expressed in minutes.
In a study of noninteracting control sys- Because the outputs are not product
tem design methods, Bequette and Edgar compositions, the consistency relations be-
(1989) used a distillation model, which they tween the primary gains and the flow gains
obtained using "standard" assumptions, in- (which were not given by Bequette and
cluding perfect level control and constant Edgar) do not apply. Furthermore, the con-
molar flow rates. From the nonlinear state- sistency relations never apply to complete
space model they first developed for the transfer functions, only to their steady-state
17-plate benzene-toluene distillation, they parts, because they are derived from
determined two transfer function models: steady-state material balances. However, the
one for the (L, V) structure, one for the transformations between models for differ-
(D,V) structure. The models are ent control structures are valid for transfer
functions when the reflux drum and reboiler
G~:l[flL
G;: flV 1 (10-53)
holdups are perfectly controlled. According
to Equation 1O-14a, the relationship be-
tween the (D, V) structure and the (L, V)
214 Practical Distillation Control
Numerically, we obtain
10-9 APPLICATION 4:
Gf;r = -1.71 (lO-60a) IDENTIFICATION OF CONSISTENT
20s +1 MODELS
LV -- -1.71
GDL (lO-60b)
22s + 1 In the previous sections we have shown that
the process gains of a control structure have
and to satisfy certain consistency relations, which
can be derived from basic physical laws such
(20s + 1)(6.74s + 1) as material balances. We have also shown
G LV = 2.71 (lO-61a)
DV (22s + 1)(4s + 1) that the process gains of different control
structures are related through exact analyti-
(2.97s+ 1) cal relationships. However, gains that are
Gf;~ = 2.71 ( ) (lO-6lb)
4s + 1 determined from numerical data by stan-
dard methods do not generally satisfy these
As seen from Equations 10-60 and 10-61, relationships (Haggblom, 1988b; Haggblom
the two expressions for both Gf;r and Gf;~ and Waller, 1988a). The reason is that the
are identical with respect to steady-state process usually is nonlinear with respect to
gains, and as similar as can be expected, the variables used and that experimental
Control Structures, Consistency, and Transformations 215
data are inaccurate. In order to obtain The natural way to reconcile the gains is
"consistent" models, one should estimate to use the method of least-squares, that is,
the gains subject to the known relationships. to modify the gains subject to the known
Even if a model for a control structure relationships in such a way that a quadratic
has been determined subject to known con- objective function of the adjustments is min-
sistency relations, it is never an exact de- imized. If also the variances of the gain
scription of the process. If models for other estimates have been estimated, and the
control structures are to be determined from choice of weighting matrices used in the
this model, the inaccuracy of the model may objective function is based on these vari-
be magnified in the calculations. In order to ances, there is no loss of generality in the
get a model that is accurate enough for proposed two-step procedure (Haggblom,
arbitrary control structure transformations, 1988b). However, the method can be used
it may therefore be necessary to estimate irrespectively of how the initial estimates of
models for a few control structures and to the process gains have been obtained and
reconcile these models subject to the analyt- whether estimates of their variances are
ical relationships existing between the mod- available or not.
els. More explicitly, the objective function to
In this section we will describe a proce- be minimized subject to the known relation-
dure for estimating consistent steady-state ships is a sum of weighted squares of the
control structure models by means of an adjustments of all gains in all control struc-
example. According to this procedure, mod- tures considered. Even if only a few models
els for an arbitrary number of control struc- are reconciled, the dimension of this esti-
tures are first determined by some standard mation problem is huge. Moreover, the
method. The next step is to reconcile the known analytical relationships between the
models subject to the analytical relation- gains in different control structures are non-
ships existing between the models and the linear with respect to the gains. Therefore,
consistency relations derived from external a numerical solution according to this for-
material balances. The theory and several mulation is not practical.
applications are described in more detail in It is possible to reduce the dimension of
Haggblom (1988b, 1989). A program pack- the problem by elimination of most of the
age for solving this kind of estimation and gains in the objective function by means of
reconciliation problem is also described in the known relationships, but the number of
Haggblom (1988b). remaining gains will still be considerable
and the estimation problem will be nonlin-
ear. For example, if two disturbance vari-
ables are included in the model and the
10·9·1 Reconciliation of Control
primary outputs are not product composi-
Structure Models
tions, the number of gains to be simultane-
In principle, consistency relations could be ously adjusted is 12. In practice, this means
directly taken into account when process that there will be local minima (at least
gains are estimated from steady-state data. numerically) and that a unique solution is
However, if the process gains of several difficult to obtain.
control structures are estimated, it is incon- However, it is also possible to formulate
venient simultaneously to estimate all gains the reconciliation problem as an ordinary
subject to all known consistency relations. It linear estimation problem, which has an an-
is more convenient first to estimate the gains alytical solution. Suppose that initial esti-
by standard (unconstrained) methods, then mates of the gains for a number of control
to reconcile the estimates subject to the structures are known. From the set of gains
relevant relationships. for each control structure, it is possible to
216 Practical Distillation Control
calculate the corresponding gains for the TABLE 10-3 Nominal Steady-State Data
base structure [the (L, V) structure] by for the Abo Akademi Column
means of an "inverse" transformation, which Feed flow rate, F 200kg/h
can be derived from Equations 10-33. Each Distillate flow rate, D 6Okg/h
control structure thus gives a set of gains for Bottoms flow rate, B 140kg/h
the base structure, which can be regarded Feed composition, Z 30wt%
as estimates obtained from different experi- Distillate composition 87wt%
Bottoms composition 5wt%
ments. The task is then to estimate one set Reflux flow rate, L 6Okg/h
of gains for the base structure from these Rate of steam flow to reboiler, V 72kg/h
"experimental" gains. Feed temperature 65°C
If the objective function is expressed as a Reflux temperature 62°C
sum of weighted squares of the adjustments
of each of these gains (i.e., of the "errors"
of these gains), the resulting estimation
problem is linear. The main problem is the The nominal steady state at which the
choice of weights for the gains calculated models were determined is given in Table
from the gains for different control struc- 10-3. The actual operating point varied
tures. The simplest solution is to choose the somewhat from experiment to experiment,
same weight for all calculated values of a but the given steady state is used in the
given gain. The optimal value of the gain is reconciliation of the models.
then the average of the calculated values. Table 10-4 shows the experimentally de-
However, it is also possible to use weights termined process gains for the four control
that are directly related to the accuracy of
the gains determined for each individual
control structure (Hiiggblom, 1988b). TABLE 10-4 Experimentally Determined Process
When the .gains for the base structure Gains lor the Abo Akademi Columna
have been estimated as outlined, the next
step is to reconcile them subject to the (L, V) Structure
structures. The gains in the gain matrices In the study previously mentioned, dy-
are ordered in the same way as in Equa- namic models were determined from step
tions 10-36, 10-39, etcetera (see Section responses. Because it was desired to approx-
10-6). The fact that complete models for the imate the initial responses fairly well with
control structures were not determined does first order transfer functions, the estimates
not preclude a reconciliation of the models of the steady-state gains, which are deter-
as previously described. The unknown gains mined by the ultimate responses, are not
are simply given the weight zero. As a result necessarily as accurate as possible. Another
of the reconciliation, consistent estimates of practical difficulty in the estimation of
the unknown gains are also obtained. steady-state gains from step responses is
Before the process gains can be recon- that it is not easy to determine when a new
ciled, the weighting matrices have to be steady state has been reached. Based on
selected. Obviously, the weights should re- these facts and on a detailed examination of
flect the reliability of the estimates so that the original experiments, the weights for the
more reliable estimates are given higher reconciliation were chosen according to
weights than less reliable ones. In accor- Table 10-5, where each entry is the weight
dance with this, estimates that are very un- for the gain in the corresponding position in
reliable, or perhaps not even available, can Table 10-4.
be given the weight zero. According to sta- Table 10-6 shows the result of the linear
tistical considerations, the choice of weights reconciliation procedure using these
as the inverse of the variances of the initial weights. [Note that the weights for the gains
estimates is optimal under certain condi- in each control structure had to be trans-
tions (Haggblom, 1988b). formed to weights for the gains in the (L, V)
TABLE 10-5 Weights for Reconciliation of Gains TABLE 10-6 Reconciled Gains for the Abo
for the Abo Akademi Column Akademi Column
TABLE 10-7 Consistent Control Stnldure Models for the AIIo Abdemi Colnmn
structure as described in Haggblom (1988b).] loops. Basically, relative gain analysis is used
Most of the reconciled gains are close to the to pair inputs and outputs in a control
experimentally determined gains, but for structure, but based on the calculated inter-
some gains the differences are considerable. action indices it is often used for choosing
These differences were to be expected from the control structure (i.e., the inputs and
the examination of the experiments. For outputs to be used) from a number of alter-
example, the estimate of IK;[I was too natives.
small in the original work because a good fit In its usual form, relative gain analysis is
with the initial response was considered a steady-state analysis and thus based on
more important. Further, the gains in the process gains only. This means that the
(D, V) structure associated with a distur- transformations and consistency relations
bance in the feed composition, and all previously derived are readily applicable. If
primary gains in the (D I(L + D), VIB) the process gains of a system are given for
structure except those associated with this one control structure, such as the conven-
disturbance, were determined at an operat- tional (L, V) structure, the process gains for
ing point slightly (but significantly enough) any other feasible control structure can be
different from the nominal steady state given calculated. The relative gains for any distil-
in Table 10-3. Finally, it can be mentioned lation control structure can thus be calcu-
that a sign error (already corrected in Table lated from known process gains for one
10-4) in the recording of one of the gains structure.
was detected through the reconciliation. When the reflux drum and reboiler
In the reconciliation, only the gains of holdups can be assumed perfectly con-
the base structure were determined when trolled, or when the same inventory control
the objective function was minimized. The system is used in the control structures con-
gains of the other control structures given in sidered, the transformations also apply to
Table 10-6 were calculated by means of dynamic models. In this case the relative
Equations 10-33. In a similar way the gains gains for the control structures can be calcu-
for any control structure can be calculated lated as functions of frequency. Note, how-
once the transformation matrices are deter- ever, that the transformations require that
mined. As a numerical illustration, Table the complete model, including the flow
10-7 shows the gains for 15 control struc- transfer functions, be known because the
tures, which have been calculated from the flow transfer functions cannot be obtained
reconciled gains for the (L, V) structure from steady-state consistency relations.
(which also are included in Table 10-7). An In this section we will not deal with the
extended control structure notation is used calculation of relative gains from known
in the table; (L, V; D, B), for example, de- models for different control structures. In-
notes the (L, V) structure with D and B as stead, we will present some analytical rela-
inventory control manipulators. tions between the relative gains for a num-
ber of control structures. These relations
can be used, for example, to calculate rela-
tive gains without first deriving models for
the various control structures. Some of the
10-10 APPLICATION 5: RELATIVE expressions also apply to frequency depen-
GAIN ANALYSIS dent relative gains, but in general they have
to be modified for the dynamic case because
The relative gain concept, introduced by the consistency relations do not apply to
Bristol (1966) and developed by McAvoy transfer functions. The results given in this
(1983) and Shinskey (1984) among others, is section are from Haggblom (1988a), where
a measure of interaction between control the subject is treated more extensively.
220 Practical Distillation Control
10-10-1 Some Analytical Relations gains A;r, A~b' and A;f are related by the
between Relative Gains expression
In the (L, V) structure the relative gain for
the y - L pairing is given by 1 ) = (1 _ _
( 1 _A_ 1 ) (1 _ _
1 )
LV A DV ALB
yL yD yL
LV
K LVKLV)-1
xL yV
(10-67)
AYL = (1- KLVK LV (10-62)
yL xV
Equation 10-67 has previously been de-
Analogously, the relative gain for the y - D rived by Jafarey, McAvoy, and Douglas
pairing in the (D, V) structure is given by (1979) with use of the constant flow assump-
tion aD = aV - aL. As no limiting as-
sumptions were made in the preceding
(10-63)
derivation, the relationship is exact for any
separation process, whether the assumption
of constant flows holds or not and regard-
and the relative gain for the y - L pairing less of, for example, the type of primary
in the (L, B) structure by output variables (compositions or tempera-
tures).
A control structure often used and stud-
(10-64) ied is Ryskamp's (1980) control structure
(D I(L + D), V), which is essentially equiv-
alent to the (LID, V) structure. This struc-
The models for these control structures are ture can be interpreted as a hybrid between
given in Equations 10-1, 10-2, and 10-3, the (L, V) structure and the (D, V) struc-
respectively. ture. The relative gain A~~, where R =
The transformations from the (L, V) LID, can be expressed as a weighted aver-
structure to the (D, V) structure and the age of the corresponding relative gains for
(L, B) structure are given by the (L, V) structure and the (D, V) struc-
ture. The expression is
KDV
yD KDV]
yV ,LV ,DV
[
DV DV
RV AyL + PAYD
K xD K xV AYR = 1 +0
(10-68)
1 yL
[KLV
= K LV
DL
K LV
xL
(10-69)
functions of L and V IB, respectively, form TABLE 10·8 Data for Calculation of Relative Gains
another class of control structures. For these for the Abo Akademi Column
control structures, the relative gain A~f,
LID = 1.0000 VIB = 0.5143
where S = VI B, can be expressed as K;l = -0.0435 K;r = 0.0458
KH> -0.2260 K;r = 0.5432
KiX = -0.6154 K i;~ = 1.3523
(10-70)
a~ = Hfy ay + H fx ax (10-75b)
and
Consider now the standard (D, V) struc-
ture, described by the model (here we only K DV
consider the gains between the primary ma- DV ax
a..,. . = ay - -K yV (1O-81a)
nipulators and the outputs) xV
K~X
ay] = [K~X K~:][aD] a~ =- K DV ay + ax (10-81b)
[ax K~X K~: aV
(10-76) yD
(D, V) structure:
is controlled by B.
Art = Ay + 1.1 b.x (10-9Oa)
It can further be shown that one-way
steady-state decoupling is obtained in any At = -S.2 b.y + b.x (lO-90b)
Note that primary output variables are primary outputs are fed both to L and V, as
temperatures in the Abo Akademi column, illustrated by the two numerical examples.
whereas they are product compositions in More recently, Bequette and Edgar
the Wood and Berry column. (1989) used a number of methods to calcu-
late noninteracting control schemes for a
specific column using D and V as manipula-
tors. They ended up with Rosenbrock's
10-11-3 Discusalon of Output
scheme, connecting the composition sum to
Decoupllng Structures Suggested
D and the difference to V.
In the literature
Weber and Gaitonde (1982) used the ma-
A number of output transformations and nipulators L and V and connected the vari-
noninteracting control structures have been able k(1 - y) + x to L and (1 - y)/x to V.
suggested in the literature. These can be The constant k. was calculated from the
examined with respect to noninteraction at nominal operating point as x/(1 - ji).
steady state in the light of the results and McAvoy (1983) discussed Weber's and
numerical examples in the two previous sec- Gaitonde's scheme and stated that a
tions. "better," that is, less interacting, choice for
As a way to eliminate the often harmful k would be i/(1 - i). Further, he sug-
interaction between the primary control gested a scheme where
loops in dual-composition control of distilla-
tion, Rosenbrock (1962) as early as 1962
suggested a scheme, where "the 'top loop' 7j = x + ( i _) Y (10-93)
l-z
now manipulates the flow of top product to
control the sum of the two measured com- is connected to D, stating that the scheme
positions." The bottom loop manipulates V is almost one-way steady-state decoupled
to control the difference between the mea- (McDonald and McAvoy, 1983). The other
sured compositions. Rosenbrock stated that output was the separation factor S con-
the scheme often can be shown to be nonin- nected to the boilup V. The separation fac-
teracting. tor is defined as
As shown in Section 10-11-1, a one-way
decoupled structure is obtained if D con-
trols the sum of the product compositions
when 15 = ii. The two numerical examples
S= L~ )C :y X) (10-94)
of the disturbances. Note, however, that this and B, thus in a way combining all struc-
only applies to disturbances that have been tures previously discussed (except those
taken into account in the design. where output variable transformations were
Assuming that the (L, V) structure is used). We note a certain resemblance with
used as base structure, u are the variables a recently suggested "combined balance
that physically affect the primary outputs. structure" (Yang, Seborg, and Mellichamp,
Therefore, these variables have to be calcu- 1991), where the energy balance structure
lated when the control structure is imple- (L, V) is combined with the two material
mented. From Equation 10-98 it follows that balance structures (D, V) and (L, B). In the
they can be determined from approach of Yang, Seborg, and Mellichamp,
the manipulators for the primary loops are
chosen as
[41/1,] [ - 0.0408 0.0398] [4L] Even the recently discovered "super mate-
41/1x = - 0.6330 1.4375 4V rial balance" structure (D, B) (Finco,
Luyben, and Polleck, 1989) is covered by
[ 0.0032 -0.1215
+ -0.7828
-0.OOI3][4D]
4B
Equation 10-103 (a = b = 0). The com-
bined balance manipulators can also be
viewed as a special case of the ORD manip-
(10-101)
ulators.
The static ORD structure was first stud-
When the control structure is implemented, ied both experimentally and through simu-
L and V are calculated from lations in Haggblom and Waller (1990). Its
self-regulating properties have further been
[4L]
4V
= [-43
-19
1.2] [41/1,]
1.2 41/1x
studied through one-point control in
Sandelin, Haggblom, and Waller (1991),
where the structure was compared with a
1.1 0.09] [4D] number of other control structures for the
+ [ 1.0 0.12 4B
(10-102) o
Abo Akademi column. The incorporation of
"dynamics" into the structure is studied in
where 1/1, and I/Ix are the outputs from the Waller and Waller (1991) and Haggblom
primary controllers in the ORD structure. (1992).
Experimental results for this control struc-
ture are given in Chapter 15. [Note that the
ACKNOWLEDGMENT
numerical values given here differ some-
what from those given in Chapter 15 and in The results reported have been obtained
the papers cited in the following text be- during a long-range project on multivariable
cause slightly different models were used process control supported by the Neste
for the (L, V) structure.] Foundation, the Academy of Finland,
It can be noted that both primary manip- Nordisk Industrifond, and Tekes. This sup-
ulators 1/1, and I/Ix are functions of L, V, D, port is gratefully acknowledged.
Control Structures, Consistency, and Transformations 227
of Continuous Distillation Units. Amsterdam: trol of distillation columns. Int. Chem. Eng.
Elsevier. 27, 669-677.
Rosenbrock, H. H. (1962). The control of distil- Tsogas, A. and McAvoy, T. J. (1981). Dynamic
lation columns. Trans. Inst. Chem. Eng. 40, simulation of a non-linear dual composition
35-53. control scheme. Proceedings of the 2nd World
Ryskamp, C. J. (1980). New strategy improves Congress of Chemical Engineering, Montreal,
dual composition column control. Hydrocar- pp. 365-369.
bon Processing 59(6),51-59. Van Kampen, J. A. (1965). Automatic control by
Ryskamp, C. J. (1982). Explicit versus implicit chromatographs of the product quality of a
decoupling in distillation control. In Chemical distillation column. Convention in Advances
Process Control 2 (D. E. Seborg, and T. F. in Automatic Control, Nottingham, England.
Edgar, eds.). New York: Engineering Founda- Waller, K. V. and Finnerman, D. H. (1987). On
tion/ AIChE, pp. 361-375. using sums and differences to control distilla-
Sandelin, P. M., Hiiggblom, K. E., and Waller, tion. Chem. Eng. Commun. 56, 253-258.
K. V. (1991). Disturbance rejection properties Waller, K. V., Finnerman, D. H., Sandelin,
of control structures at one-point control of a P. M., Hiiggblom, K. E., and Gustafsson, S. E.
two-product distillation column. Ind. Eng. (1988a). An experimental comparison of four
Chem. Res. 30, 1187-1193. control structures for two-point control of dis-
Shinskey, F. G. (1984). Distillation Control, 2nd tillation. Ind. Eng. Chem. Res. 27, 624-630.
ed. New York: McGraw-Hill. Waller, K. V., Hiiggblom, K. E., Sandelin, P. M.,
and Finnerman, D. H. (1988b). Disturbance
Shinskey, F. G. (1990). Personal communication.
sensitivity of distillation control structures.
Skogestad, S., Lundstrom, P., and Jacobsen,
AIChE J. 34, 853-858.
E. W. (990). Selecting the best distillation
Waller, J. B. and Waller, K. V. (1991).
cuntrol configuration. AIChE J. 36, 753-764.
Parametrization of a disturbance rejecting and
Skogestad, S. and Morari, M. (1987a). Control decoupling control structure. Report 91-13,
configuration selection for distillation Process Control Laboratory, Abo Akademi,
columns. AIChE J. 33,1620-1635.
Abo.
Skogestad, S. and Morari, M. (1987b). A system- Weber, Rand Gaitonde, N. Y. (1982). Non-
atic approach to distillation column control. interactive distillation tower control. Proceed-
Inst. Chem. Eng. Symp. Ser. 104, A71-A86. ings of the American Control Conference,
Skogestad, S. and Morari, M. (1988). Under- Arlington, VA, pp. 87-90.
standing the dynamic behavior of distillation Wood, R K. and Berry, M. W. (1973). Terminal
columns. Ind. Eng. Chem. Res. 27, 1848-1862. composition control of a binary distillation
Smart, A. M. (1985). Recent advances in distilla- column. Chem. Eng. Sci. 28, 1707-1717.
tion control. Adv. Instrum. 40, 493-501. Yang, D. R., Waller, K. V., Seborg, D. E., and
Takamatsu, T., Hashimoto, I., and Hashimoto, Mellichamp, D. A. (1990). Dynamic structural
Y. (1982). Multivariable control system design transformations for distillation control con-
of distillation columns system. Proceedings of figurations. AIChE 1. 36, 1391-1402.
the PSE, Kyoto, pp. 243-252. Yang, D. R, Seborg, D. E., and Mellichamp,
Takamatsu, T., Hashimoto, I., and Hashimoto, D. A. (1991). Combined balance control
y. (1987). Selection of manipulated variables structure for distillation columns. Ind. Eng.
to minimize interaction in multivariable con- Chem. Res. 30, 2159-2168.
11
Diagonal Controller Tuning
William L. Luyben
Lehigh University
229
230 Practical Distillation Control
trol structure, as long as the interaction use because we include the possibility of
among the loops is explicitly taken into ac- using ratios, sums, differences, and so
count and not ignored. Constraints can be forth, of various manipulated variable
handled by using override controls to sup- flow rates.
plement the conventional control structure. 3. How should the controlled and manipu-
The purpose of this chapter is to present lated variables be paired?
a fairly simple approach to designing an 4. How do we tune the SISO controllers in
effective diagonal control system for a mul- this multivariable environment?
tivariable plant.
(b)
F'-----......
Steam
FIGURE 11·2. Alternatives: (a) Constant retNx; (b) constant vapor boilup; (c)
constant reflux ratio.
232 Practical Distillation Control
load disturbance
L(s} = load disturbance -
B
We use only one load variable in this devel- FIGURE 11-4. Block diagram of closed-loop multi-
opment in order to keep things as simple as variable system.
possible. Clearly, there could be several load
disturbances, and we would just add addi- where B l , B 2 , ••• ,BN are the individual
tional terms to Equations 11-3. Then the controllers in each of the N loops.
L(s} in Equation 11-4 becomes a vector and
In the closed-loop system the manipu-
Qc. becomes a matrix with N rows and as lated variables are set by the controllers:
many columns as there are load distur-
bances. Because the effects of each of the
load disturbances can be considered one at m=
-
B E
dS6
= B [xset - -x]
ds} -
(11-6)
a time, we will do it that way to simplify the
mathematics. Note that the effects of each Substituting Equation 11-6 into Equation
of the manipulated variables could also be 11-4 and solving for ~ gives
considered one at a time if we were looking
only at the open-loop system or if we were
considering controlling only one variable.
However, when we go to a multivariable
closed-loop system, the effects of all manipu-
lated variables must be considered simulta-
neously.
Equation 11-7 gives the effects of setpoint
Figure 11-4 gives the matrix block dia-
and load changes on the controlled vari-
gram description of the open-loop system
ables in the closed-loop multivariable envi-
with a feedback control system added. The !
matrix is the identity matrix. The ;s}matrii
ronment. The matrix (of order N X N) mul-
tiplying the vector of setpoints is the
contains the feedback controllers-:-With the closed-loop servo transfer function matrix.
multiloop SISO diagonal controller struc~ The matrix (N X 1) multiplying the load
ture the !!<s} matrix has only diagonal ele- disturbance is the closed-loop regulator
I
ments; all the off-diagonal elements are transfer function vector.
zero: Remember that the inverse of a matrix
Bt has the determinant of the matrix in
the denominator of each element. There-
o o
B - [ o (11-5) fore the denominators of all of the transfer
ds> 0 functions in Equation 11-7 will contain
o BN Det[I + ~ ;).
= - ('j-
Diagonal Controller Tuning 235
Now we know that the characteristic in Chapter 1, and additional aspects are
equation of any system is the denominator treated in several other chapters.
of its transfer function set equal to zero.
Therefore the closed-loop characteristic
equation of the multivariable system with
11-3 SELECTION OF
feedback controllers is the simple scalar MANIPULATED VARIABLES
equation Once the controlled variables have been
specified, the next step is to choose the
manipulated variables. We do not worry at
this stage about how to pair the variables in
a diagonal multiloop SISO controller struc-
ture but only about what variables to ma-
We can use this closed-loop characteristic nipulate. Should we manipulate reflux and
equation to tune controllers in multivariable vapor boilup (R-V) to control distillate and
processes. bottoms compositions, or should we manip-
ulate distillate and vapor boilup (D-V)?
11-2 SELECTION OF
CONTROLLED VARIABLES 11·3·1 Morarl Resiliency Index
Engineering judgment is the principal tool The Morari resiliency index (MRI) can be
for deciding what variables to control. A used to give some guidance in choosing the
good understanding of the process leads in best set of manipulated variables. The set of
most cases to a logical choice of what needs manipulated variables that gives the largest
to be controlled. Considerations of eco- minimum singular value over the frequency
nomics, safety, constraints, availability and range of interest is the best.
reliability of sensors, and so forth must be The MRI gives an indication of the in-
factored into this decision. herent controllability of a process. It de-
It should be remembered that controlled pends on the controlled and manipulated
variables need not be simple directly mea- variables, but it does not depend on the
sured variables. They can also be computed pairing of these variables or on the tuning
from a number of sensor inputs. Common of the controllers. Thus it is a useful tool for
distillation examples are reflux ratio, re- comparing alternative processes and alter-
flux-to-feed ratio, reboiler heat input calcu- native choices of manipulated variables.
lated from the flow rate and temperature The MRI is the minimum singular value
change of a hot oil stream, and so forth. of the process open-loop transfer function
Sometimes we are unable to measure matrix ~(j,,)' It can be evaluated over a
directly the variable that we would really range of frequencies w or just at zero fre-
like to control. This is frequently true in quency. If the latter case, only the steady-
distillation because it is often difficult and state gain matrix is needed:
expensive to measure product compositions
directly with sensors such as gas chro-
matographs. Instead, temperatures on vari-
ous trays are controlled. The selection of
the best control tray to use requires a con- The singular values of a matrix are a
siderable amount of knowledge about the measure of how close the matrix is to being
column, its operation, and its performance. "singular," that is, having a determinant
Varying amounts of nonkey components in that is zero. A matrix that is N X N has N
the feed can affect very importantly the best singular values. We use the symbol U'j for
choice of control trays. Some aspects of this singular value. The U'j that is the biggest in
inferential control problem were discussed magnitude is called the maximum singular
236 Practical Distillation Control
value, and the notation u max is used. The Ui solving the equation
that is the smallest in magnitude is called
the minimum singular value and the nota- Oet
[
(A - 207.4)
369.96
369.96 1
(A - 733.57)
tion u min is used.
The N singular values of a real N X N = A2 - 940.97A + 15,272 = 0
matrix are defined as the square root of the A = 916.19,16.52
eigenvalues of the matrix formed by multi-
plying the transpose of the original matrix Therefore the minimum singular value is
by itself. v'16.52 = 4.06 = MRI.
As another example, consider the follow-
ing steady-state gain matrices for three al-
i = 1,2, ... ,N (11-10)
ternative choices of manipulated variables:
reflux and vapor boilup (R-V), distillate and
vapor boilup (0-V), and reflux ratio and
The eigenvalues of a square N X N matrix
vapor boilup (RR-V).
are the N roots of the scalar equation
R-V O-V RR-V
Oet[ A! -~] = 0 (11-11)
GMil 16.3 -2.3 1.5
GM12 -18.0 1.04 -1.12
GM21 -26.2 3.1 -2.06
where A is a scalar quantity. Because there GM22 28.63 1.8 4.73
are N eigenvalues, it is convenient to define MRI 0.11 1.9 0.89
the vector A., of length N, that consists of
the eigenvalues: AI' A2, A3,· .. , AN: Based on these results, the 0-V system
should be easier to control (more resilient)
because it has the largest minimum singular
Al value.
A2 One important aspect of the MRI calcu-
A= A3 (11-12) lations should be emphasized at this point.
- The singular values depend on the scaling
use in the steady-state gains of the transfer
AN functions. If different engineering units are
used for the gains, different singular values
We will use the following notation. The will be calculated. Thus the comparison of
expression ~[Al means the vector of eigen- alternative control structures and processes
can be obscured by this effect. The practical
values of the matrix A.
solution to the problem is to always use
For example, the MRI for the Wood and dimensionless gains in the transfer func-
Berry column at zero frequency would be tions. The gains with engineering units
calculated should be divided by the appropriate trans-
mitter span and multiplied by the appropri-
ate valve gain. This gives the gains that the
control system will see.
The differences in MRI values should be
GT G = [ 12.8 6.6] [ 12.8 -18.9]
,;:;M(O),;:;M(O) -18.9 -19.4 6.6 -19.4 fairly large to be meaningful. If one set of
manipulated variables gives an MRI of 10
[ 207.4 -369.96] and another set gives an MRI of 1, you can
= -369.96 733.57 conclude that the first set is better. How-
ever, if the two sets have MRIs that are 10
The eigenvalues of this matrix are found by and 8, there is probably little difference
Diagonal Controller Tuning 237
from a resiliency standpoint and either set 11-4-1 Review of Nyquist Stability
of manipulated variables could be equally Criterion for 5150 Systems
effective.
As you should recall from your undergradu-
ate control course, the Nyquist stability cri-
11·3·2 Condition Number terion is a procedure for determining if the
closed-loop characteristic equation has any
Another popular index for choosing the best
roots (zeros) in the right half of the s plane.
set of manipulated variables is the condition
number of the steady-state gain matrix. The
If it does, the system is closed-loop unsta-
ble.
condition number is the ratio of the maxi-
The method is based on a complex vari-
mum singular value to the minimum singu-
able theorem, which says that the difference
lar value,
between the number of zeros and poles that
a function has inside a closed contour can
eN = u max /U min (11-14) be found by plotting the function and look-
ing at the number of times it encircles the
A large condition number is undesirable origin. The number of encirclements of the
because it indicates that the system will be origin (N) is equal to the difference be-
sensitive to changes in process parameters tween the number of zeros (Z) and the
(lack robustness). Therefore, sets of manip- number of poles (P) that the function has
ulated variables that give small condition inside the closed contour.
numbers are favored. We use this theorem in SISO systems to
find out if the closed-loop characteristic
equation has any roots or zeros in the right
11-4 TUNING DIAGONAL half of the s plane. The s variable follows a
CONTROLLERS IN A closed contour that completely surrounds
MULTIVARIABLE ENVIRONMENT the entire right half of the s plane as shown
in Figure 11-5a.
Tuning is a vital component of any con-
troller design procedure once the structure The closed-loop characteristic equation
for an SISO system is
has been determined. For multivariable sys-
tems, we need a procedure that simultane-
ously tunes all of the multiloop SISO con-
trollers in the diagonal structure, taking into
account the interaction that exists among all
the loops.
The biggest log-modulus tuning (BLT) If the process is open-loop stable, the func-
procedure outlined in the following text is tion F(S) will have no poles in the right half
one way to accomplish this job. It provides of the s plane. Therefore, P = 0 for open-
settings that work reasonably well in many loop stable systems and the number of en-
processes. These settings may not be opti- circlements of the origin (N) is equal to the
mum because they tend to be somewhat number of zeros (Z).
conservative. However, they guarantee sta- Now it is convenient to plot the total
bility and serve as a benchmark for compar- open-loop transfer function GM B(S) in-
isons with other types of controllers and stead of plotting F(s)' This displa~~s every-
other tuning procedures. thing to the left by one unit; so instead of
The procedure is easy to use and easy to looking at encirclements of the origin, we
program, and it is easy to understand be- look at encirclements of the ( -1,0) point in
cause it is based on a simple extension of the GMB-plane. These curves are illustrated
the SISO Nyquist stability criterion. in Figure 11-5b and c.
238 Practical Distillation Control
Then the number of encirclements of the load responses. The method yields set-
(-1,0) point made by U(iw) as W varies tings that give a reasonable compro-
from 0 to 00 gives the number of zeros of the mise between stability and performance
closed-loop characteristic equation in the in multivariable systems.
right half of the s plane. 3. Make a plot of U(iw)' Using the guessed
value of F and the resulting controller
settings, a multivariable Nyquist plot
11-4·3 BLT Tuning Procedure
of the scalar function U(iw) = -1 +
BLT tuning involves the following five steps: Ded! + ~(j ..!(iW)] is made.
4. Make a plot of Lcm' The closer the W
1. Calculate the Ziegler-Nichols settings for contour is to the (-1,0) point, the
each individual loop. The ultimate gain closer the system is to instability.
and ultimate frequency Wu of each di- Therefore the quantity W/(1 + W) will
agonal transfer function Gjj(s) are cal- be similar to the closed-loop servo
culated in the classical SISO way. To transfer function for a SISO loop
do this numerically, a value of fre- GMB/(l + GMB). We define a multi-
quency W is guessed. The phase angle variable closed-loop log modulus
is calculated, and the frequency is var-
ied to find the point where the Nyquist
plot of Gjj(iw) crosses the negative real
Lcm = 20 log 101 1 : wi (11-21)
axis (phase angle is -180°). The fre- The peak in the plot of Lcm over the
quency where this occurs is Wu' The entire frequency range is the biggest
reciprocal of the real part of Gjj(iw) is log modulus L~:-.
the ultimate gain. 5. The F factor is varied until L':: is equal
2. A detuning factor F is assumed. F should to 2N, where N is the order of the
always be greater than 1. Typical values system. For N = 1, the SISO case, we
are between 1.5 and 4. The gains of all get the familiar + 2 dB maximum
feedback controllers Kci are calculated closed-loop log modulus criterion. For
by dividing the Ziegler-Nichols gains a 2 x 2 system, a + 4 dB value of L~:
KZNi by the factor F, is used; for a 3 X 3, + 6 dB; and so
forth. This empirically determined cri-
KZNi terion has been tested on a large num-
Kci = F (11-18)
ber of cases and gives reasonable per-
where KZNi = Ku;/2.2. Then all feed- formance, which is a little on the con-
back controller reset times T/i servative side.
are calculated by multiplying the The method weighs each loop equally,
Ziegler-Nichols reset times TZNi by the that is, each loop is equally detuned. If it is
same factor F, important to keep tighter control of some
T/j = TZNiF (11-19)
variables than others, the method can be
easily modified by using different weighting
where factors for different controlled variables.
2'lT . The less-important loop could be detuned
TZNi = -2
.1 .
w", (11-20)
more than the more-lD1portant
. 1oop.
The F factor can be considered as a
11-4-4 Examples
detuning factor that is applied to all
loops. The larger the value of F, the Two examples are given in the following
more stable the system will be but the text. The first is a 2 x 2 system and the
more sluggish will be the setpoint and second is a 3 X 3 system.
240 Practical Distillation Control
1m
milt. dB =16.71
<a> EaqJiric:Il
F1GURE 11-6. W plots for Wood and Berry column. (a) Empirical settings; (b)
Ziegler-Nichols settings; (c) BLT settings.
max. dB = 30.83
(b)ZN
1m
max. dB =3.99
-2
(c)BLT
FIGURE 11·6. Continued.
242 Practical Distillation Control
step change in xrt, verifying that the multi- troller tuning parameters are compared with
variable system is indeed closed-loop stable. those using the empirical settings in Figure
Note that the W plane curve using the 11-7.
ZN settings does encircle the ( - 1, 0) point,
showing that the system is closed-loop un-
stable with these settings. OgunlUlilce and RIly Column
Figure 11-6c gives the W plots for the W plots and time responses for this 3 X 3
system when the BLT settings are used system are given in Figures 11-8 and 11-9
(K cl = 0.375, Kc2 = - 0.075, Til = 8.29, for two sets of controller tuning parameters:
and T/2 = 23.6 with a detuning factor F = empirical and BLT. The open-loop transfer
2.55). Time responses using the BLT con- function matrix is
0.5
-,\
\ Bmpirical
\
\
\
~_----.... BLT
01---0.6
20
1ime(mia)
FIGURE 11·7. Time responses for Wood and Berry column for unit step change in
Xlsetpoint: empirical and BLT settings.
Diagonal Controller Tuning 243
Empirical controller settings are Kc1 = 1.2, by using two detuning factors: F detunes
Kc2 = - 0.15, Kc3 = 0.6, T/1 = 5, T/2 = 10, the ZN reset and gain values and FD de-
and Tl3 = 4. BLT controller settings are tunes the ZN derivative value. The opti-
Kc1 = 1.51, Kc2 = - 0.295, Kc3 = 2.63, mum value of FD is that which gives the
T/1 = 16.4, T/2 = 18, and Tl3 = 6.61 with a minimum value of F and still satisfies the
detuning factor F = 2.15. + 2N maximum closed-loop log modulus
Both of these cases illustrate that the criterion.
BLT procedure gives reasonable controller
settings.
If the process transfer functions have
greatly differing time constants, the BLT
procedure does not work well. It tends to
11-5 PAIRING
give a response that is too oscillatory. The
problem can be handled by breaking up the Now that the controlled and manipulated
system into fast and slow sections and ap- variables have been chosen and we have a
plying BLT to each smaller subsection. procedure to tune controllers once the pair-
The BLT procedure was applied with PI ing has been specified, the final question
controllers. The method can be extended to that must be answered is "How do we pair
include derivative action (PID controllers) the controlled and manipulated variables?"
1m
max. dB :: 4.37
-2
(0) BmpiricalleUioaa
1m
DIlL dB =6.01
OR
------------ -----------
", .... - -. ...... Fmpirical
/~ ...... ,....
~
,...-,
I ,
,
,,
, Ilmpirical
"-
" -r=::::::o.-.;;:;..--.--..::.:--=:-=:--:=:--=-
0 .....- ....
BLT
0 ....._ _ _...
-25
60
Time (min)
FIGURE 11-9. Time responses for Ogunnaike and Ray column for unit step change
in xl setpoint: empirical and BLT settings.
246 Practical Distillation Control
the Wood and Berry column is "integrally unstable" for any controller
tuning.
[XD]
xB
= [-18.9
-19.4 12.8][
6.6 R
V] (11-29)
and the NI for this pairing would be For a 3 x 3 system there are three elements
in the x vector, so three curves are plotted
of each-of the three elements in the vector
[[! + Qv(i..!ifl»]-I§(io) for the specified
load variable L. Figure 11-10 gives the TLC
curves for the 3 x 3 Ogunnaike and Ray
(-18.9)(6.6) - (12.8)( -19.4) column.
( -18.9)(6.6)
-40
-m~-----------;------------;-----------~
0.01 0.1 1.0 10.0
Rational procedures for selecting manipu- Another key concept is a procedure for
late variables and pairing variables have tuning diagonal multiloop SISO controllers
been discussed. in a multivariable environment. The tuning
One of the key concepts is how one de- is done considering all controllers simulta-
termines stability in a multivariable system. neously. The application of these methods
The closed-loop characteristic equation of a will be illustrated in later chapters (Chapters
MIMO system has been developed. 18, 22, 23, and 24).
12
Dynamic Matrix Control
Multivariable Controllers
David A. Hokanson
Exxon Chemical HoI/and
James G. Gerstle
Exxon Chemical Company
248
Dynamic Matrix Control Multivariable Controllers 249
I
~(. x~:~ (-:}~
"1'1
u, -I Process t- y,
v,
I ;~':'" wI &~
~II~II',~ W:,'if. h ''w'&
It!1~i '''};W@lw~;~n.@''
fw. I m.,i<i;' , Pi Vii
I ~¥ iiI," 'IY- I;'
~I~ I tw M~ I·~
: "
II"I'"~ I~~ 'I'f~"~
~i.~ { ~; f '" Mi'·, :' :I
'.' ·,1 ~W.'I"~~<'~%"
. _= 1 »>,, 1 ' ·
I,':'I' ~ m. I ;0 ~ '.,
9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
=o
o 1 2 3 4 5 6 7 8
Input Output
o 2 3 4 5 6 7 8 9 10 11 12 13 14 15 o 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Input Output
FIGURE 12·2. Step response convolution model: discrete form.
~
.....
252 Practical Distillation Control
The sequence aj (j = 0, 1,2,3 ... ) is the This is essentially the definition of the step
step response convolution model, which is response convolution model described be-
the basis for DMC. Remembering that ev- fore. What is interesting about this is that
erything so far is in deviation variables, at is we can calculate what y will be at each
no more than a series of numbers based on interval for successive moves in u. For ex-
a set length (time to steady state) and a set ample purposes only, let us examine a series
sampling interval. There is no set form of of three moves "into the future" -one now,
the model, other than that steady state is one at the next time interval, and one two
reached by the last interval. In fact, the last time intervals into the future:
value at is the process steady-state gain.
Because DMC is based on the step re- l1u o = U o - u_ 1
sponse .convolution model, let us look at it
in more detail. To keep things relatively I1U1 = U1 - Uo (12-8)
simple, we will examine a SISO problem
l1u 2 = u 2 - u1
with independent variable u and dependent
variable y. Furthermore, we will use a model
length of 30 (a popular model length) to then
minimize the number of indefinite parame-
ters (that is, the n's and m's). Y1 - Yo = a 111u O
For a lined out process with a single step
in u at time 0, Y2 - Yo = a 2 11u o + a 1 11u 1
r~uo 1
bances, nonlinearity in the models, and Y4 - Yo a4 a3 a2
modelling errors. One way to correct for
aUl
these is by calculating a prediction error b
at the current time and then adjusting the Y29 - Yo a 29 a 28 a 27 aU 2
predictions. In equation form, Y30 - Yo a 30 a 29 a 28
Y3l - Yo a 30 a 30 a 29
(12-11) Y32 - Yo a 30 a 30 a 30
Yi
where ya is the actual, measured value of yf
y. Then the vector yP (from Equation 12-10)
is corrected by b for predictions 1 to 29 as yf
yr
yr = yf + b (12-12) + (12-14)
yfg
y! - Yo = a! auo + yF
12-2-3 Control Solution
Y2 - Yo = a2 auo + a! au! + yr From the preceding text, we have a set of
Y3 - Yo = a3 auo + a2 au! + a! aU2 + yj future predictions of y based on past moves
and a set of future values of y with a set of
Y4 - Yo = a4 auo + a 3 au! + a2 aU2 + Y4 three moves. Let us now assume that we
have a desired target for y called ySP. (We,
(12-13)
of course, need this for any control action!)
Y29 - Yo = a29 auo + a28 au! + a27 aU2 + y~ One criterion we can use is to minimize the
error squared over the time between now
Y30 - Yo = a 30 au o + a29 au! + a28 aU2 + y~ and the future (that is, to use the least-
squares criterion). In vector equation form,
Y3! - Yo = a 30 auo + a 30 au! + a 29 aU2 + y~
(12-20)
In the preceding example, the number of where M is a diagonal matrix of move sup-
moves calculated was set to three for exam- pressionfactors and the vector ~ is a vector
ple purposes. Cutler (1983) states that the of zeros. These factors make the matrix well
time covered by the calculated moves should conditioned and reduce the aggressive
cover the time it takes for the variable to moves the controller can impart upon the
reach 60% of steady state. In actual prac- plant. Figures 12-3 and 12-4 show an exam-
tice, this number is usually fixed to one-half ple of a setpoint change and its effect on a
of the time horizon covered by the overall simulated controlled and manipulated vari-
controller. able with varying move suppression.
25
c: 20
o
---------
'fil
.~
1J 15
CD
~as
> 10
~
~ 5
-1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Time .....
Ught
-------
Medium Heavy
12
fO
~ 8
~ 6
o~
-g 4
12
°li
.-._.-._-
~ O~------~--~~~-----------------------------------
(2) -5 -3 -1 3 5 7 9
Time~
Ught Medium Heavy
words, with the moves in the process known, If we let L be the fractional drum level,
find the dynamic matrix A using least-
squares criterion. Least-squires smoothing V
can again be used to reduce the resulting L=- (12-25)
ill-conditioned matrix and reduce the prob- ~ot
lems with a noisy signal. This technique has
then
been developed into a commercially avail-
able package and is described by Cutler
(1991). (12-26)
II 12 13 14 1"_1
where fi = inlet flow VOl V0 2 V0 3 V04 VO,,_I
Fo = outlet flow (12-24)
(12-27)
V = drum volume
o Time-"
FIGURE 12·5. Integrating process with deadtime: response to step input.
Dynamic Matrix Control Multivariable Controllers 257
times as long as if simply one variable was (PRBS) inputs will be used based on a
moved while holding the others constant. switching time of 0.5 hour (see MacGregor,
• The independent variable moves must not 1989, for discussion of PRBS signals and
be correlated. For example, the indepen- switching times). For our four independent
dent variables reflux and reboiler heat variable case, we now need to conduct the
should not be moved at the same exact test for over 96 hours (because the number
time' when doing a test for an overhead of parameters for each independent vari-
composition model. able is 4 x 30, or 120).
Before executing the test directly, we
To illustrate the preceding points, let us must again make sure that the PRBS
examine the test of a rather simple energy moves are randomized. Specifically, corre-
balance (LV) controlled distillation tower lated moves must be avoided. For example,
shown in Figure 12-6. reflux and reboiler heat should not be moved
Assuming a relatively constant feed com- at the same time even though it might be
position, there are four variables that affect tempting to do this. Instead, the moves
the overhead composition: should be made at significant intervals apart
(i.e., at least 30 minutes in the preceding
1. feed rate example).
2. feed temperature One other caution on multivariable tests:
3. reflux Because multivariable tests tend to be long,
4. heat input they have the tendency to include slow
rolling effects (i.e., nonstationary load dis-
turbances) that are unmodelled such as
After identifying these independent vari- day-night shifts in ambient temperature.
ables, the next problem is to select the right These disturbances can cause problems with
feed temperature. Although one could model identification techniques. To mini-
imagine that T2 would give better control, mize the influence of these disturbances,
selecting this can lead both to an incorrect data differencing should be used. (Again see
overhead composition model on all four MacGregor, 1989, for details')
variables and very poor control. The reason
is that T2 is not independent of the other
12·3-4 Nonlinear Transformations
variables. Moving feed rate, reflux, or heat
input will affect T2 • Instead, the true inde- Because distillation control problems tend
pendent variable is the temperature of the to be nonlinear, there are two ways of deal-
feed upstream of the bottoms-feed pre- ing with the' nonlinearities with a linear
heater, Tt . controller such as DMC:
With the four true independent variables
• Operate in a region that can be assumed
selected, the next problem becomes to de-
to be linear.
sign a test. Let us assume for this example
that a rough "time-to-steady-state" is 4 • Transform the variables so that they be-
come more linear.
hours. The test length and sampling interval
are determined by the number of parame- Three examples of linear transformations
ters that are needed by the model. that can be used in DMC distillation control
If we are doing an analysis that will lead follow.
directly to a convolution model (e.g., a 30
parameter DMC model), let us assume that Logtuithms for High Purity Distillation Control
a plant test would require a 1 min sampling The compositions of high purity distillation
time and a 24 hour test (six times the time- towers are definitely nonlinear. In many
to-steady-state) for SISO identification. cases, using the logarithm (natural or base
Manual pseudo-random binary sequence 10) will linearize these variables over their
Dynamic Matrix Control Multivariable Controllers 259
normal operating range so that they can be pipe resistances in series with the control
fitted with a linear identification package. valve is
Shinskey (1988) describes the use of loga-
rithms for PID tower composition control, L
and others (Georgiou, Georgakis, and Q = -;=====:=- (12-29)
Va + (1 - a)L2
Luyben, 1988; see others in Section 12-6)
have applied it to DMC control. A word of
warning, however: The logarithm method
does not always work for every high purity where Q = fraction maximum flow (lin-
tower. The use of logarithms must be evalu- ear)
ated on a case by case basis. Methods of L = fraction valve position
doing this include using a detailed steady- a = valve characteristic
state simulation to find regions where the
logarithm transformation can be used, com- Different a's yield the curves shown in Fig-
paring the logarithm and linear results from ure 12-7.
a model identification package, or closely
studying the prediction errors from an on- Differential Pressure Transformation
line DMC controller. As long as a tower remains below the
flooding point, a differential pressure mea-
Valve Transformations surement can be assumed linear over the
Valves that are highly nonlinear should be normal operating range of a tower. How-
linearized unless they can be assumed linear ever, if a tower floods easily, the differential
over a small operating range. Here, knowl- measurement will become very nonlinear.
edge of the actual valve characteristics can A transformation using hyperbolic functions
be a big help. Otherwise, doing a fit with to linearize differential pressure in a case
one of the standard valve transformations where tower flooding is a problem was
from Perry's Handbook (Perry and Chilton, developed and used successfully in one ap-
1973) can lead to a good model. The lin- plication (Hokanson, Honk, and Johnston,
earization equation that takes into account 1989).
~ 0.4 ! / ,/'
~ 0.3 ! /
! / /1
I
0.2 / 1
./
0.1 . ~I
O~ ______-L________L -______-L________L -______ ~
,
I
I
I
~cv 300 ppm
' ( O.75bar
I
I
I
oJ
2500 ppm
nGURE 12-8. Energy balance tower with bottoms-feed preheater: DMC example.
CV - controlled variable; FF = feedforward variable; MY = manipulated variable.
Dynamic Matrix Control Multivariable Controllers 261
the top as a 200 ppm error in the bottoms the least-squares solution. Prior to imple-
and a 0.02 bar error in the differential pres- mentation, the least-squares controller is
sure. These values are then used by the checked for any manipulated variable con-
package to normalize each variable before straint violation in any of the moves. If
entering the least-squares control solution. there are, the moves that are in violation
are set at the constraint limits and the
least-squares solution is re-solved.
12-4·2 Constraints
Here is an example of how the LP works
A DMC controller must deal with two dif- within the DMC Corporation controller.
ferent types of constraints: Suppose everything in the example tower is
fine, and the controller is operating within
• Manipulated variable constraints.
the range of all variables. Then, the LP
• Controlled variable constraints.
directs the DMC controller to target the
Manipulated variable constraints can be a overhead composition setpoint, bottoms
hard limit (e.g., a valve at 100% open), an composition setpoint, and passes on the re-
operator entered limit (e.g., a reboiler maxi- sulting steady-state differential pressure as
mum heat input), or a maximum desired the differential pressure setpoint to the
rate of change to a manipulated variable. least-squares solver. The least-squares solver
All are essentially equivalent, and must be then plans the moves of the reboiler heat
dealt within a DMC control package. and reflux accordingly, taking into account
Controlled variable constraints arise what effects the feed rate has on the con-
when the set of controlled variables targets troller.
cannot be reached. For example, if the dif- Suppose, however, the feed increases
ferential pressure in the preceding example dramatically. Then, the LP sees that the
becomes close to the target of 0.75 bar, differential pressure may have to be vio-
somehow information must reach the least- lated and redirects the bottoms composition
squares solution that the bottoms com- to a new steady-state value (violating the
position specification should be violated. setpoint) while at the same time directing
Likewise, if the differential pressure is well the differential pressure to hold at its high
below 0.75 bar, the least-squares DMC solu- setpoint. Similar events occur if the heat or
tion should not be solved to hold the 0.75 reflux is moved to its maximum limit.
bar target.
The handling of these constraints in
ODMC is described in Prett and Garcia 12·5 IMPLEMENTATION STEPS
(1988). Here, the entire DMC controller FOR A DMC CONTROLLER
with manipulated variable constraints is The actual implementation of a DMC con-
converted into quadratic program (OP). The troller consists of a number of steps:
controlled variable constraints are placed as
inequalities that are part of the quadratic
objective function. If no constraints exist, 1. Initial design
ODMC results in the same solution as 2. Pretest
least-squares DMC. 3. Plant test
In the DMC Corporation controller, both 4. Model identification
manipulated and controlled variable con- 5. Off-line simulation
straints are handled by a linear program 6. Operator interface and on-line controller
(LP), which "directs" the least-squares solu- building
tion. The LP runs prior to the DMC con- 7. Controller commissioning
troller and calculates the optimum steady- 8. Measuring results
state solution. Then it passes its results via
the setpoints of the controlled variables to In the following discussion, it is assumed
262 Practical Distillation Control
that the models are generated from plant There are two manipulated variables that
test data. directly affect these controlled variables: re-
flux flow and reboiler heating medium flow.
Depending on the circumstances, these may
be modified to internal reflux (especially in
12-5-1 Initial Design
this case because the airfin condenser can
The initial controller design phase of the induce significant changes in subcooling) and
DMC controller involves the selection of reboiler heat input.
the manipulated, feedfolWard, and con- The feedfolWard variables for this case
trolled variables. Although the design may could be the feed rate and T1• However, in
change over the course of the controller many cases feed temperature changes with
development and implementation, it will al- feed rate. If there is a significant correla-
ways be strongly affected by the decisions tion, Tl may not be a good feedfolWard
taken in the initial design. Thus, the initial variable either. Any correlation will nor-
design will have a strong impact on the mally show up in the test data, and if it is
success of the application. found to be significant, the feed tempera-
The selection of independent variables ture can be dropped as a feedfolWard vari-
(manipulated and feedfolWard variables) for able during the model identification and
the DMC application is very important at controller building step.
this stage. Of prime importance is that each All of this logic results in an initial DMC
independent variable be truly independent. controller design:
In other words, independent variables must
not have an effect on one another. Further- Manipulated variables
more, manipulated variables (which can be • Internal reflux
regulatory controllers or control valves) must • Reboiler heat input
have a direct and independent path to each Feedforward variables
valve. They must not be impeded or aided • Feed rate
by overrides, decouplers, and so forth. • Feed temperature
Selection of the controlled variables is Controlled variables
fairly straightfolWard at this stage because • Overhead composition
they can usually be modified once the test is • Bottoms composition
concluded. Tower operating targets and • Tower differential pressure
tower constraints are usually the main con-
trolled variables. However, all possible vari- However, we are not finished. The bottoms
ables are collected during the test so that level may cause some problems that need to
they can be added as required during the be examined during the pretest. The exist-
model identification and controller building ing bottoms level controller impacts the
steps. tower through the heat it provides in the
To examine this phase, let us look again bottoms to feed the preheater. If it has
at the energy balance example in Figure been a problem controller and/or has a big
12-8. The objectives of the controller are to impact on tower performance, the bottoms
control both the top and bottoms composi- level should be included as a controlled
tions tightly but to give up on the bottoms variable. This will also add bottoms flow as
composition if the tower reaches its flooding a manipulated variable to the controller.
limit as determined by the tower differential If the bottoms level pedorms acceptably
pressure. These three-overhead composi- and does not swing the tower around (per-
tion, bottoms composition, and tower dif- haps because of the small duty from the
ferential pressure-become our prime con- preheat exchanger), it can be left out. In
trolled variables. this case, it still must be checked for any
Dynamic Matrix Control Multivariable Controllers 263
nonlinear or gap action. Either will ruin the model will cover) should be made. This
overall quality of the resulting DMC model. will give all involved a rough idea of how
Finally, one more important note about long the full plant test must last.
this level. The decision to include the level
in the DMC controller must be done prior To conduct this test, open loop moves
to the test. Once the test commences, it will are made to the reflux, reboiler, feed rate,
be almost impossible to change that deci- and feed temperature (if possible). The
sion without a second test. pretest is conducted over one or two shifts,
The decision on whether to use tray tem- and moves are heavily influenced (of course)
peratures as controlled variables in this by the console operator.
tower can be done after the test is com-
pleted. Normally, all tray temperatures are
collected during the test. If the analyzer
12-5-3 Plant Test
response is found to be sluggish when ana-
lyzing the test data, a tray temperature can With the pretest complete, the actual test
be used in conjunction with the analyzer as can now be conducted. As previously dis-
a DMC controlled variable. The decision on cussed, the length of the test is set by the
this is usually made after the test is com- number of independent variables and the
pleted. time-to-steady-state. Using the a rough rule
of six times the time-to-steady-state, and
assuming the pretest yielded a 4 h time-to-
12-5-2 Pretest
steady-state, the test should then be con-
With the rough design set, the next step is ducted for four full days.
to perform a "pretest." Here, all the inde- During the test, each variable should be
pendent variables are checked to make sure moved in a PRBS (pseudo-random binary
that each can be moved. For feedforward sequence) fashion based on the moves de-
variables, changes to the process may make termined during the pretest. Although an
enough moves in the variables. If not, a automatic PRBS signal can be generated for
method must be devised to move them. moving the controller, the authors recom-
Again, in the tower depicted in Figure mend against using an automatic PRBS sig-
12-8, there are four independent variables: nal for plants with limited experience in
T 1 , feed rate, reboiler heating, and reflux. these types of tests. Controlling the test
During the pretest and actual test, the re- manually will begin the process of console
boiler and reflux setpoints will be collected operator involvement and acceptance and
and there should be no problems moving will provide the operators with an excellent
them. However, during the pretest, Tl and learning opportunity. Operators are used to
the feed rate are checked for the type of keeping the controlled variables at a con-
"natural" process moves that occur and, if stant value. The opportunity to observe the
possible, how they can be moved. responses to independent variable changes
Also during the pretest, three other things will expand their existing knowledge and
must be accomplished: dispel any untrue legacies that exist.
In theory, a PRBS test implies that the
1. The data collection package must be decision to make a move or stay at the
tested. In the full plant test the package current value is entirely random. In prac-
must operate flawlessly. tice, this is not always possible due to the
2. Move sizes for each independent vari- operating objectives of the column. For ex-
able should be roughly determined. ample, again consider the initial controller
3. An estimation of the time-to-steady-state design for Figure 12-8. A switching fre-
(i.e., the time horizon that the DMC quency of 0.5 hour has been selected and
264 Practical Distillation Control
the reboiler heat input is currently being independent variable at a time. Concentrat-
moved. At the switch time, the process is ing on one variable also minimizes "corre-
first examined with the console operator to lated" independent variable moves. During
see if a move "has to be made" in order to the test, around 20 moves should be made
keep the compositions within specification, in each independent variable. These moves
the column out of flood, or any other pro- are not made consecutively, but rather split
cess event that requires a move. between two more different time periods.
If no move needs to be made, the process Further, the order in which the indepen-
should then be examined to see if a PRBS dent variables are moved is changed after a
move should be avoided. This can occur if cycle through the variables. For example, in
there are no methods of forcing a change in Figure 12-8, the first cycle in the plant test
one of the feedforward variables (e.g., feed could be to move the reboiler first, reflux
temperature) and that variable has had a second, then feed third. (There is no "easy"
substantial change. In this case, making no way of moving the feed temperature so we
move will allow an uncorrelated response of must wait for normal process changes.) The
that variable. Finally, if none of these cases second cycle would be slightly different: first
exist, then the PRBS move is made by a reflux, then feed rate, and finally reboiler.
random decision (i.e., a coin flip). Figures 12-9 and 12-10 illustrate the
From a test management standpoint, it is moves made during a plant test on a
usually easier to concentrate on moving one tower similar to that depicted in Figure 12-8
Independent Variables
'Reftux
Time -
Dependent Variables
Bottoms
/ composition
Time -
FIGURE 12·10. Dependent variable test data.
but without the feed temperature. It is cess flow should be thoroughly examined to
actually the tower composition controller make sure that the correlation between feed
described by Gerstle, Hokanson, and rate and feed temperature makes process
Andersson (1990). sense. Then, and only then, should the feed
temperature be "thrown out" and the iden-
tification run with only three independent
12·5·4 Model Identification
variables.
In this stage, knowledge of both the model Another example: Suppose the feed rate
identification package and the process is versus differential pressure showed very lit-
very important. The decision on which mod- tle correlation. Before totally dropping this
els to use and what models to discard de- correlation, the fact that there is very little
pends on both a thorough knowledge of the correlation must be verified through process
process plus the goodness of fit information knowledge. (For example, the feed is a satu-
from the identification package. In the ex- rated liquid and enters below the lower
ample, suppose the feed temperature moved differential pressure tap.) Otherwise, the
quite frequently during the test, but each data must be further analyzed for a better
move was a function of feed rate. This can correlation or the test must be run again.
be verified by running a cross-correlation Besides keeping models in and out of
check. If such a correlation exists, the pro- the final problem, additional controlled
266 Practical Distillation Control
Kp = 0.3005 Kp = 0.0072
--------
Kp =0.1003
---------
=0.0016
Kp
---------
= -2.6132 Kp Kp =1.3249
ftGURE 12-11. OveraU tower model. Eacb model represents tbe response over 240
min.
variables can be added (with care). In the shows a model developed from the data
example Figure 12-8, the tray temperatures shown in Figures 12-9 and 12-10.
should be examined to see if they might
improve overall response by including them
in the model. In the controller discussed in 12-5-5 Controller Building
Gerstle, Hokanson, and Andersson (1990), and Simulation
such a tray temperature was included. After picking the models, the controller fre-
Also during the model identification quency and DMC model size are selected.
phase, the issue of nonlinearity is also exam- These parameters are determined by the
ined. like the decision on which variables DMC package used, the dynamics of the
to use in the final model, both process process, and a comparison between the
knowledge and identification package good- longest and shortest dynamics. For example,
ness of fit are used to determine where the DMC Corporation package allows a
linearization should be used. In the example model size in multiples of 30. If the longest
problem, the compositions may be better time to steady state is found to be 240 min,
modelled by using logarithms instead of us- the fastest a 30 parameter controller can
ing the actual data. run is once every 8 min. In the example
At the end of this stage, a multivariable problem, this may not be fast enough to
DMC model is produced. Figure 12-11 control the differential pressure, so a larger
Dynamic Matrix Control Multivariable Controllers 267
model (e.g., a 60 parameter model) could be mode." During this time, controller predic-
used. tion errors (the b's from Equation 12-11)
With these parameters selected, the final are checked to make sure that the DMC
controller model is built and then simulated models are reasonable. Also, calculated ma-
off-line. Here the two "tuning constants" nipulated variable moves are checked to
-the equal concern error and move sup- ensure that they are reasonable.
pression factors-are tuned. Finally, the controller is turned on and
observed. Move suppression factors, equal
concern error weightings, and occasionally
12-5-6 Controller and Operator
the models themselves are modified to im-
Interface Installation
prove overall controller effectiveness. Al-
With the simulation complete, the con- though this phase should be straightfor-
troller is moved to the on-line control sys- ward, the controller commissioning can take
tem and supporting software is developed. as long as the plant test phase.
A key item during this phase is building the
operator interface for the controller. If this
is the first multivariable controller in the
12-5-8 Measuring Results
plant, the DMC controller operation will be
substantially different than other single loop The control performance of a tower DMC
controllers. Turning the DMC controller on controller is usually hard to measure with
and off alone will be different, so the sup- typical control performance measurement
porting interface software must be written tools [e.g., integrated absolute error (lAB)].
so that the console operators can easily This is because DMC controllers are usually
operate the controller. Besides turning the installed to handle tower constraints that
controller on and off, the interface must aid sometimes keep the tower away from all of
the console operators in understanding the its composition setpoints. In the face of no
basics of the DMC multivariable controller. constraints, the authors have found that a
Besides potentially being faced with the well tuned DMC controller on a conven-
new concepts of multivariable control, the tional tower (for example, the one depicted
console operators may also be faced with in Figure 12-8) usually equals and some-
another new concept-that of a "setpoint times exceeds the tightness of control found
range." With the controlled variable con- from a well designed, conventional dual-
straint handling capability comes the ability composition controller. Of course, if little
to set a high and low setpoint for each control is applied on a tower before the
controlled variable. Again, the interface implementation of DMC, the improvements
software should be developed to adequately in control around a given composition set-
illustrate this concept. point will be very noticeable.
All these concepts mean both operator However, in the case where DMC re-
training and considerable work on the oper- places a well designed, conventional con-
ator interface. Some of the DMC appli- troller, reduced composition variation is
cation papers referenced (e.g., Cutler usually not the driving force behind its in-
and Hawkins, 1987; Hokanson, Houk, and stallation. Instead, constraint handling is
Johnston, 1989) contain example displays what is sought from DMC. Here, the results
and discuss this aspect in more detail. of DMC are usually measured in improve-
ments in plant performance. As the tower is
held more tightly against its constraints by
12-5-7 Controller Commissioning
the DMC controller, higher unit throughput
After building all the necessary software, or less energy consumption are usually easy
the controller is turned on in "prediction to measure.
268 Practical Distillation Control
272
Distillation Expert System 273
logical model. The column model also char- signing columns are of little use for control
acterizes the column on the basis of its studies because their output is the recom-
various flow ratios, sending this information mended number of equilibrium stages.
to the logical model to assist in configuring Models used for control purposes must start
the control loops. with a fixed set of stages and report the
The logical model also requires informa- results of flow changes. They need not be as
tion on the economic objective of the distil- accurate as models used for column design
lation and any constraints placed on it such because their results are principally used to
as the inability to condense all of the over- compare one control system against another
head vapor. Equipment characteristics also for purposes of selecting the best. Addition-
enter into the logic because they limit ally, they can be fitted to an accurately
choices for some control loops. Finally a designed column, or one already operating.
decision is made on the best control system Modelling begins with the overall mate-
for this particular column and a drawing is rial balance
created illustrating the recommended con-
figuration. F=D+B (13-1 )
where components I and h represent the for a constant distillate-to-feed flow ratio.
light and heavy key, respectively, and S is The hyperbolic curve labelled S is the solu-
the separation factor between them. tion of Equation 13-4 for a constant separa-
If the mixture is binary, then Yh = 1 - Yl tion factor. The point where they cross is
and a solution can be found. If that is not the current operating point for the column.
the case, a solution can still be found for The separation factor is a function of
four components in the feed, providing that reflux-to-distillate flow ratio L / D, relative
the lightest does not leave with the bottom volatility, the number of theoretical equilib-
product (lighter-than-light key) and the rium stages in the column, and the enthalpy
heaviest does not leave with the distillate of the feed (Shinskey, 1984, page 79). If
(heavier-than-heavy key). The complete D / F is increased, its line will move upward,
derivation for this four-component model is pivoting about the point where feed compo-
given by Shinskey (1988, pages 411-414). sition falls on each axis. But if reflux ratio is
Figure 13-2 plots the preceding relation- held constant as D / F is changed (by chang-
ships as Yh against Xl for a four-component ing reflux L proportionately), then both
separation. It was generated by the Distilla- product compositions will change along the
tion Consultant for a column having a feed S curve, one becoming less pure while the
consisting of 1, 20, 40, and 39 mol% other becomes more pure. If instead, reflux
lighter-than-light, light, heavy, and heavier- ratio is increased by increasing L while D is
than-heavy keys, respectively. Distillate con- held constant, then compositions will move
tained 2 mol% heavy key and the bottom downward along the D / F line, both prod-
product cuntained 0.5 mol% light key. The ucts becoming purer. This scenario gives a
reflux ratio was 3.0 and there were 25 theo- qualitative indication of the interactions that
retical stages in the column. can be observed between product composi-
The line labelled D / F represents the tions when one flow rate or flow ratio is
solution of material-balance Equation 13-3 adjusted.
5
s\ '\UI8
\ \ \\
\ \\\ /
~~ V
V
~~ ./'
I'\. /v
f
V ~ ~
V ~ ~ r-- to--
/'
1/ ~
8
/
8 1
FIGURE 13·2. Operating curves for a typical column plotted by the Distilla-
tion Consultant.
Distillation Expert System 277
These two relationships are sufficient to as the VI B ratio is then held constant,
define a column, but there are other operat- compositions must remain on its curve re-
ing curves that may be more meaningful in gardless of variations in the other flow ra-
a given situation. The remaining curves in tios. The verticality of the curve would then
Figure 13-2 are all derived from the first make it possible to adjust Yh over a wide
two. Consider, for example, the case where range by adjusting reflux, for example, with
the reflux-to-feed flow ratio LIF is held little influence on Xl; in effect, Xl would be
constant. Increasing D IF under this condi- regulated simply by holding VIB constant.
tion will cause LID to fall and thereby, Then if bottom composition were to be con-
separation will decrease; hence the curve trolled by manipulating VIB, that controller
representing constant L IF lies to the right would have little work to do-a favorable
of the S curve above the operating point, situation.
and underneath it (higher separation) below The material-balance line may be nearly
the operating point. If overhead vapor flow vertical or nearly horizontal, or anywhere in
V is held constant while D IF is increased, between, depending on the BID ratio. If it
reflux falls, causing LID and hence separa- is nearly vertical, the material balance will
tion to change more than in the case of easily regulate bottom composition; if it is
constant reflux flow. As a result, the curve nearly horizontal, top composition will be
for constant VIF is still steeper than the easily regulated.
L IF curve. The last curve in Figure 13-2
represents a condition of constant vapor-
13·2·2 Relative Gain Analysis
to-bottom flow ratio VIB, analogous to re-
flux ratio but at the other end of the col- Relative gain analysis is a simplified method
umn. Notice that it is the steepest of all the for determining the interaction among con-
curves. trol loops in a multivariable process such as
There is one other significant operating distillation (Shinskey, 1984, pages 131-154).
curve, corresponding to a constant ratio Each viable configuration for the two com-
LIB. It falls between the VIF and VIB position loops can be evaluated, and the
curves, but it represents a combination of combination having the least interaction can
flows not easily coordinated, being at oppo- be selected as a good candidate for control-
site ends of the column, and so is not given ling the products. The Distillation Consul-
further consideration. tant uses relative gain as one measure of
In general, columns having relatively im- control-loop performance, and as a basis for
pure products, low reflux ratios, and few estimating another (disturbance rejection)
trays, tend to have their operating curves as well.
well spread as in Fig. 13-2. But those mak- Relative gain is defined as the ratio of
ing high-purity cuts from low-volatility mix- the open-loop gain for a selected loop when
tures have curves grouped tightly together; all loops on the process are open, to its
in the extreme they are almost superim- open-loop gain when all the other loops are
posed. Yet the curves always fall in the closed. In the first case, all other manipu-
same order. The material-balance line, on lated variables m are held constant, whereas
the other hand, is quite independent of the in the second, all other controlled variables
curves. Its slope is dependent on feed and c are held constant by their controllers:
product compositions because it is the bot-
tom-to-distillate flow ratio BID. (oc.;om.)
The slopes of the various curves at the A.. = I J m (13-5)
TABLE 13-1 Properties of Relative Gains cause the variables are not all independent.
A Property There are fewer degrees of freedom than
there are control loops: A relative gain of
<0 Conditional stability-do not close loop infinity is indicated.
o Control depends on other loops Negative relative gains reveal conditional
0-1 Interaction increases gain
and period of oscillation
stability because numerator and denomina-
1.0 Interaction is incomplete, if any tor have opposite signs. A controller given
>1 Interaction reduces control effectiveness the correct action for negative feedback
00 Loops are completely dependent when other loops are open will produce
positive feedback when they are closed. If
the action is reversed to handle the multi-
gain expression, it is a dimensionless num- loop situation, however, it will be incorrect
ber and is unaffected by the choice of scales when other controllers are placed in manual
for the variables. Relative gains are dis- or encounter constraints. The dynamic be-
played in an n X n array with the manipu- havior of the loop is also poor, tending to
lated variables listed across the top and the exhibit inverse response and a very long
controlled variables down the side. A rela- period of oscillation. A relative gain greater
tive gain array describing the interactions than 1 is always accompanied by a negative
between the two composition loops in a number in the same row and column.
distillation column would appear as Interaction between loops having relative
gains in the 0-1 range forms only negative
feedback loops, but loop gain increases and
(13-6)
so does the period of oscillation as A de-
creases from 1 to O. Interaction in the case
Top and bottom compositions are desig- of relative gains greater than 1 forms weak
nated y and x, and the manipulated vari- positive feedback loops having little effect
ables m 1 and m 2 could be any selected flow on loop gain or period, but forming a very
rates or flow ratios. An important feature of stiff system. The controllers tend to share
a relative gain array is that the numbers in disturbances, but recovery is very slow, as is
each column and each row sum to 1.0. In setpoint response, as A increases.
Equation 13-6 then, AY2 is 1 - Ayl , and so Controllers in interacting loops may re-
is Ax!; as a result, Ax2 = Ayl • quire detuning when all are placed in auto-
Table 13-1 summarizes the properties as- matic at once, depending on the relative
sociated with certain groups of relative gain gain values and on individual loop dynam-
numbers. A relative gain of 1.0 indicates ics. In general, fast loops can upset slow
that m j will form only one closed loop with loops, but not vice versa. The slow con-
ci ' because the same open-loop gain is ob- troller in an interaction between a slow and
tained whether other loops are open or not. a fast loop effectively moves both manipu-
There could be one-way interaction-rela- lated variables because of the response of
tive gain will not reveal this. the fast controller to its output. As a conse-
In a 2 X 2 array, a value of 1.0 must be quence, its gain must be reduced in direct
accompanied by a value of 0 in the same proportion to its relative gain (for relative
row and column. A relatiye gain of 0 is gains less than 1). For loops having similar
caused by the numerator in Equation 13-5 dynamics, the controller time constants also
being o. Control could still be possible in have to be increased as well; the detuning
this case (if partial interaction exists), but may be shared by the interacting con-
would depend on both loops being closed trollers, either equally or unequally as de-
all the time. The opposite situation, the sired. For cases of relative gains greater
denominator of 0, indicates that the loops than 1, less detuning is required; a halving
cannot all be closed at the same time be- of the controller gain is the largest adjust-
Distillation Expert System 279
ment necessary, with little or no changes to TABLE 13·2 Relative Gains for the Column
the time constants. of Figure 13·2
Relative gains for the two composition mj
loops on a distillation column can be calcu- mz D L S
lated from the slopes of the operating
curves: B 00 0.412 0.487
V 0.613 10.09 2.982
VIB 0.640 5.098 2.457
A = 1/(1 __(O_YI_O_X):..-.l) (13-7)
yl (iJYliJxh
The two remaining numbers in each array
The calculated relative gain would ap- are of no interest because they are either
pear in the upper left and lower right posi- negative or represent dynamically unfavor-
tions in the 2 X 2 array, with its comple- able pairings, and are therefore not re-
ment appearing in the other two positions. ported by the Distillation Consultant. The
If the two slopes have the same sign, then pairings appearing in the table therefore
the resulting relative gain will be greater are selected as being the most viable candi-
1 O.
than or less than This would be the case dates for controlling top and bottom com-
for any combination of the curves in Fig. positions simultaneously.
13-2. Additionally, the closer together the Those combinations using one of the
slopes are, the higher is the degree of inter- product flows (B or D) as a manipulated
action. However, the combination of the variable have relative gains falling in the
material-balance line with any of the curves 0-1 range, whereas the other combinations
will cancel the negative sign in Equation give numbers greater than 1. Although the
13-7 and produce a set of relative gains in highest of the low numbers (0.640) is closer
the 0-1 region. Formulas for the slopes of to 1.0 than the lowest of the high numbers
all the curves are derived in Shinskey (1984, (2.457), the latter relative gain is more fa-
pages 147-153) for a saturated-liquid feed; vorable. Recognize that the range of the
the Distillation Consultant includes a cor- lower numbers is severely restricted in com-
rection for feed enthalpy in its calculations, parison to that of the others, and that their
as well. control loops are also slower. As a conse-
As mentioned previously, columns mak- quence, the best combinations for this col-
ing high-purity products will have their op- umn would have top composition controlled
erating curves tightly grouped together. by reflux ratio (S) and bottom composition
Equation 13-7 indicates that curves with controlled by either boilup (V) or boilup
very similar slopes will produce very high ratio (VI B). (Although the curves in Figure
relative gains. This will limit the viable con- 13-2 are labelled LIF, VIF, etc., the F can
trol-loop configurations for high-purity be dropped from the relative gain table
columns to those manipulating the material because as an independent variable it has
balance because its line is independent of no effect on relative gain.)
the grouped curves. Relatively low-purity
separations will have their operating curves
13-2-3 Establishing a Performance
spread apart, as those in Figure 13-2. The
Index
relative gains for those curves were calcu-
lated by the Distillation Consultant and are Relative gain only reveals one aspect of
presented in Table 13-2. control-loop performance-that of loop in-
Each of the nine relative gains reported teraction and detuning requirements. Of
in Table 13-2 represents a 2 X 2 relative perhaps more importance, is the response
gain array, being the number in the upper- of the product-quality loops to changes in
left and lower-right corners of each array. load: feed rate and composition, and heat
280 Practical Distillation Control
input and removal. Consider a control loop Subscript n represents the conditions at the
in a steady state with controlled variable c new steady state, at which time e and its
at setpoint r, at the moment a step change derivative are both 0, so that only the inte-
in load q is imposed. This would corre- gral term is meaningful. The change in m
spond to the condition at time 0 in Figure between the steady states is due entirely to
13-3. The change in load then drives c away the integration of e between time 0 and
from r, causing the controller to move ma- time n:
nipulated variable m in an effort to recover.
Eventually, a new steady state is reached (13-9)
with m settling at a new value correspond-
ing to the current load. The output of an
ideal PID feedback controller would re- Solving for integrated error, we have
spond to the deviation e between c and r
as f edt = Am-
PT[
100
(13-10)
mn = -
100 (
P
e + 1 -f "edt +
T[
t
0
de)
TD -
dt + mo Integrated error should be minimized be-
(13-8) cause it represents excess operating cost.
Suppose the controlled variable in Figure
where P = the proportional band 13-3 were to represent octane number of a
expressed in percent gasoline blend, with the setpoint corre-
= integral time constant
T[ sponding to the specification limit. The load
TD = derivative time constant change raises octane number above specifi-
t = time cations, so that temporarily the system is
IOO~--~--~--~--~--~--~--~--~--~~--~--~--~--~
60 .......
20 ............... .
10 ............... ..
giving away excess octane. This higher- TABLE 13·3 Load Facton for Composition Control
octane product costs more to make, but its Load
selling price remains the same because it is m Qj Qo Zj
based on the specification. Meanwhile,
blending continues at some liter per minute D/F 0 0 1/(1 - x;/z;)
rate. Integrating the octane deviation against B/F 0 0 I/O - Y;/Zj)
L/F 0 -1 /LDz - 1
time and multiplying the result by the flow
V/F -1 0 /LDz - I/O + D/L)
rate gives the number of octane-liters given L/D 0 -1 -1
away. This can be converted into a loss in D/(L + D) 0 -1 I/O +D/L)
revenue by multiplying by the incremental V/B -1 0 /LDz - /LBz - 1/(1 + D/L)
price difference of an octane number per
liter of gasoline. This is the cost of poor
control and the incentive to minimize inte- the process model. For example, the load
grated error. factor for manipulating D in the face of a
Observe that there are three terms af- change in feed composition is found by dif-
fecting integrated error: Il.m, P, and 'rl . Of ferentiating Equation 13-3:
these, Il.m is the only term that can be
driven to 0, essentially by uncoupling the
manipulated variable from the load. If a
particular manipulated variable does not
:~~~
Iy,x = -z,-. ~_i_X-i -1---~-;/-Z-i
= (13-12)
last three manipulated variables tend to be sider the relative gain from Table 13-2 for
in the vicinity of 1.0. Conclusions to be compositions controlled by reflux and
drawn from Table 13-3 are that product-flow boilup:
manipulation is insensitive to heat-balance
disturbances, and reflux and boilup manipu- (ox/oVh
lation are relatively insensitive to feed-com- Axv = 1I.YL = ( ) = 10.09 (13-13)
ox/oV x
position disturbances.
The remaining factors affecting inte- If the top composition loop were then re-
grated error are the proportional and inte- configured to manipulate reflux ratio, a new
gral settings of the controller. They should relative gain would result:
be minimized to minimize integrated error.
However, a point is reached where further (ox/oV)s
reduction in the value of either will produce AxV -- AyS -- (ox/oV)x = 2.98 (13 14)
-
undamped oscillations: This is the limit of
stability. Tuning rules are available for con-
Interestingly, reconfiguring the top loop re-
trollers (Shinskey, 1988, pages 141-150) that
sults in a reduction in process gain for the
provide stability while minimizing inte-
bottom loop:
grated absolute error (IAE). The latter is a
preferable criterion to integrated error be-
(ox/oV)s 2.98
cause minimizing it ensures stability as well ~-....:....- = - - = 0.295 (13-15)
as a minimal value of integrated error. The (ox/oVh 10.09
controller producing the response curve in
Figure 13-3 was tuned for minimum lAB, Similarly, reconfiguring the bottom loop can
and the controlled variable scarcely crosses change the gain of the top loop. Consider
setpoint at all, so that lAB is only slightly changing the manipulated variable for bot-
higher than integrated error. tom-composition control in the previous ex-
Although the derivative time constant ample from boilup V to boilup ratio V/B.
does not appear in the integrated-error ex- This changes the process gain of the top
pression, it contributes by allowing a reduc- loop as follows:
tion in the integral time constant, typically
to half the value that produced minimum (oY/OL)V/B 5.098
--:------:- = - - = 0.505 (13-16)
lAB without derivative action. Integral time (oy/oL)V 10.09
should be set relative to the delay or dead-
time in response of the controlled variable Because the proportional band needed by a
to the manipulated variable. It is a function controller for stability varies directly with
of loop interaction only insofar as it should the process gain, it follows that selecting the
be increased for relative gains less than 1 manipulated variables that present the low-
when the interacting loops have similar dy- est relative gain (of the numbers above 1.0)
namics. will allow the lowest setting. This has been
The proportional band setting varies di- demonstrated both in simulation and on
rectly with the ratio of deadtime to domi- real columns. Therefore, the relative gain
nant time constant, and with the steady-state has as much impact on integrated error as
process gain. The dominant composition does the load factor.
time constant for a column is usually similar The case of relative gains less than 1 is
for all manipulated variables, but process roughly the reciprocal of those greater than
gain is not. Process gain is included in the 1. Controllers in this situation must be de-
relative gain expression Equation 13-5, and tuned for muitiloop stability, both in pro-
therefore can be inferred from calculated portional and integral settings, inversely with
values of relative gain. For example, con- relative gain. As a result, loops having rela-
Distillation Expert System 283
tive gains less than 0.8 tend to show poor balance system. These columns, because of
recovery from feed-composition distur- high reflux ratios, also tend to be very sensi-
bances, and are not recommended unless tive to heat-balance upsets for all but mate-
there are no selections having relative gains rial-balance systems.
less than 5. The Distillation Consultant selects that
The Distillation Consultant estimates in- combination that gives the best overall per-
tegrated errors for each of the combinations formance estimate (in this example the re-
in Table 13-2, relative to that of the flux-ratio-boilup-ratio configuration) with
reflux-boilup system at 100 in response to second place given to the reflux-ratio-boilup
heat-balance disturbances. For simplicity, system. The user is, of course, free to over-
the integrated errors for the two com- ride the selection based on preference or
position loops are combined into a single conditions not considered by the Consul-
number, and the effects of heat-input and tant.
heat-removal disturbances are similarly
combined. The results appear in Table 13-4
13-2-4 ApplIcations and Objectives
for the column whose relative gains are
given in Table 13-2. One aspect of distillation that makes con-
Observe that the reflux-boilup system trol-system selection so difficult is the wide
gives the highest integrated error for heat- variety of separations that can be encoun-
balance disturbances and the lowest error tered from industry to industry, and even
for feed-composition disturbances, where- within a single industry. Petroleum refining
as material-balance configurations (manip- is the most extensive user of distillation,
ulating either B or D) have the opposite followed closely by the chemical industry.
behavior. In this situation, the preferred But distillation is also used in beverage pro-
configuration is usually a compromise, ma- duction, tall-oil separation from wood-pulp-
nipulating reflux ratio (S) at the top and ing wastes, fatty-acid separation for soap
boilup or boilup ratio at the bottom. These manufacture, and even in the separation of
systems combine good response to both dis- the isotopes of boron for the nuclear indus-
turbances with favorable relative gains. A try.
high-purity column like an ethylene frac- In petroleum refining, we encounter the
tionator or butane splitter, however, may multiproduct distillation of crude oil, lubri-
have no relative gains between 0.7 and 20, cants, and cracked oils, each of which is a
so that the best choice will be a material- specialized operation. Light-hydrocarbon
separation is the most common distillation
operation, where a mixture of gases or liq-
uids will pass through a series of columns
TABLE 13-4 Integrated Error Estimates
with progressively heavier fractions distilled
m1 in each. These columns are named for the
m2 D L S light-key component removed: demetha-
nizer, deethanizer, depropanizer, debu-
Heat-Balance Disturbances tanizer, deisobutanizer, and so forth.
B 12 10 Even identically named columns may not
V 8 100 15 have identical functions, however. For ex-
VIB 8 51 12 ample, depropanizers appear in many dif-
ferent parts of a refinery, from crude-oil
Feed-Composition Disturbances
fractionation to the fluid-catalytic cracking
B 51 72 unit to the alkylation unit and the ethylene
V 42 15 15 plant, and each may have a different objec-
VIB 49 15 16 tive. Feedstocks differ, and so do product
284 Practical Distillation Control
. r-...............~..!
~~ .••••=............................... s ~
changes in distillate flow are imposed on late composition can also be controlled by
reflux to avoid the lag that would otherwise manipulating bottom-product flow, if base
result if the level controller had to do it level can be controlled by heat input and
alone. the temperature gradient is low.
Other features of the figure include a
submerged water-cooled condenser with
pressure controlled by a hot-vapor bypass
13-3-3 Controlling Both Product
valve, and a kettle reboiler whose heat input
Compositions
controls column differential pressure. Note
that the analyzer samples the overhead va- When economics dictate that both composi-
por rather than the liquid product because tions must be controlled, the Distillation
its dynamic response avoids the lag of the Consultant selects the best combination of
accumulator; with a total condenser, there manipulated variables as described in Sec-
will be no steady-state difference between tion 13-2-3. The selection is then converted
the two compositions. to control loops consistent with configura-
If bottom composition is to be controlled tions proven in the field. For example,
by manipulating its own flow, base level Ryskamp developed a reflux-ratio configu-
must be controlled by manipulating heat ration (Ryskamp, 1980) that proved to be
input. The latter loop may not be successful responsive over a wide range of operating
owing to the possibility of inverse response conditions and is included in the recom-
of level to heat input; this is particularly mendations of the Distillation Consultant.
common when the reboiler is fired. Then an The author later applied the same concept
alternative structure must be found. Bottom to boilup-ratio manipulation and included it
composition may be controlled by manipu- in the Consultant. A system illustrating both
lating distillate flow, if heat input can be of those configurations is shown in Figure
used to control accumulator level; this loop 13-5.
is successful where there is a narrow tem- The reflux ratio is actually manipulated
perature gradient across the column, so that as Dj(L + D) rather than LjD, and boilup
most of the heat goes into changing vapor ratio as Bj(V + B) rather than VjB. Again,
flow rather than raising temperature. Distil- these have been found to give improved
Distillation Expert System 287
e.
y............ ··············1
[...~..! ~
. , , •........... ···..··..·····1..·..·..·········:1.
. ·'t'········~·rn··-e··i
! LlJI: .............! :
: :
L............................................................................................;
FIGURE 13-5. Product compositions are controlled by reflux ratio and boilup ratio
to minimize interaction,
dynamic response and also decouple the given. Having controlled pressure, it is often
level and composition loops. Notice that desirable to minimize its value because in
the lower composition loop sets column most separations relative volatility is im-
temperature in cascade. This configuration proved with falling pressure. This enables
improves response to feed composition energy to be saved per unit of product dur-
changes. No feedforward is required to this ing periods when plenty of cooling is avail-
system because both composition con- able. Capacity increases are also possible
trollers already manipulate flow ratios. for light-hydrocarbon separations.
(Feedforward is only indicated when a Floating-pressure control applies to those
product flow, reflux, or boilup is selected as separations that are conducted above atmo-
a manipulated variable.) An overhead air- spheric pressure and ambient temperature.
cooled condenser is shown, with its drain Whether energy savings will be significant
valve manipulated for pressure control. depends primarily on the variation likely to
The Distillation Consultant can also ac- be experienced in the cooling medium. Air-
commodate a partial condenser, a flooded cooled condensers experience more temper-
accumulator, several reboiler types as well ature variation than those that are water-
as direct steam injection, and so forth. With cooled, and temperate installations show
all these possibilities, it can produce more more diurnal, weather-related, and seasonal
than 30,000 different diagrams, each repre- variations than tropical installations.
senting a solution to a potential problem. To implement a floating-pressure system,
a controller sensing the condenser con-
straint needs to adjust the pressure set-
13·3-4 Floating Pressure Control
point. In the case of the hot-vapor bypass
The manipulated variable for column pres- valve in Figure 13-4, the constraint is the
sure control is selected as a function of the closed position of the valve. A valve-posi-
type and location of the condenser because tion controller (VPC) is connected to the
it is the means for adjusting heat removal. output of the pressure controller and set to
Two different examples were previously hold the bypass valve 10% open in the steady
288 Practical Distillation Control
state by adjusting the pressure setpoint. The sidestream is normally removed as a liquid
VPC should have integral-only action and above the feed point, and its light impurity
be tuned to be slow enough to avoid upset- is controlled by the flow of the distillate
ting composition loops when cooling changes product. A typical example of a Class I
take place, as in a storm. Speed of response column is the ethylene fractionator, whose
may be improved by a feedforward loop, sidestream is a high-purity ethylene prod-
which adjusts the ratio of heat input to feed uct, separated from the ethane bottom
as a function of pressure (Shinskey, 1988, product at high recovery; methane is re-
pages 431-433). moved from the condenser to limit its con-
In the case where the valve position does tent in the ethylene. Controls over the key
not indicate condenser loading, as when re- components in the side stream and bottoms
flux is manipulated to control pressure, an- should be selected on the basis of relative
other measure must be used, for example, gains because this tends to be a very inter-
liquid level in the condenser. Then the con- acting system. However, control over
denser level controller sets the pressure set- methane in the ethylene product is achieved
point. If column temperature is used to by manipulating the flow of distillate.
control composition, pressure compensation Class II columns have the opposite ar-
is needed for that temperature measure- rangement: bottom product is the smallest
ment to continue to infer and control com- stream and the sidestream is taken as a
position. The Distillation Consultant in- vapor below the feed. Again, control over
cludes several means for implementing the light and heavy keys should be assigned
floating-pressure control, including the as for a simple column, with bottom flow
placement of feedforward loops and appli- manipulated to limit the heavier-than-heavy
cation of compensation to temperature key in the sidestream.
measurements. Class III columns have the sidestream as
their smallest flow, which may be withdrawn
either above or below the feed, depending
13-4 RULE BASE FOR on the boiling point of the middle key. If
SIDESTREAM COLUMNS not withdrawn, this middle-boiling compo-
nent can accumulate until it interferes with
To apply the methods developed for simple
separation of the major products or, as in
columns to sidestream columns requires
the case of fusel oil, reaches a solubility
their classification into functional groups. A
limit and disrupts vapor pressure by creat-
sidestream column will have most of its
ing a second liquid phase. Again, the small-
stages dedicated to separating its light- and
est stream should be manipulated to control
heavy-key components; this much of it re-
the content of its principal constituent in
sembles a simple column. The third stream
one or the other product.
is intended to remove a troubling third com-
ponent: either a lighter-than-light or heav-
ier-than-heavy or a middle key. That third 13-4-2 Configuration Rules
stream is usually the smallest of the three
The configuration rules begin with the same
products, and should be treated as a special
consideration as for simple columns-the
purge stream apart from the two major
economics of the separation. They turn out
products.
to be essentially the same as for simple
columns, the principal distinction being that
the smallest stream has no significant eco-
13-4-1 Classifications
nomic contribution. However, where the
A Class I column is defined here as one Distillation Consultant posed but one eco-
whose smallest product is the distillate. The nomic decision for the simple column, it
Distillation Expert System 289
must formulate three for the side stream col- small. Base level control is assigned to heat
umn-one for each class. input and accumulator level control is as-
The smallest flow is then arbitrarily as- signed to reflux. There must be a tie, how-
signed to control composition, either of it- ever, between side stream and reflux flow to
self or the adjacent major product as de- close the material balance, or control of
sired. Relative gain analysis is used to select both levels will be lost. Figure 13-6 shows
the other two composition loops, if eco- an ethylene fractionator having bottom-
nomics dictate that they both require con- product composition controlled by bottom
trol; otherwise, material-balance control is flow, which fits this case. Reflux ratio is
recommended, with feedforward from feed manipulated to control the ethane content
rate. of the ethylene sidestream. The output of
Sidestream columns present special prob- the dynamic compensator acting on the re-
lems relative to inventory controls. Assign- flux flow signal represents what that flow
ment of a sidestream to control its own would be approaching the withdrawal point
composition usually works quite well, but it several trays from the top of the column.
cannot be assigned to level control without Note that a single analyzer on the sidestream
some help. Consider, for example, a Class II reports impurity levels of both the heavy
column whose largest flow is its vapor (H) impurity ethane and the light (L) im-
sidestream. The largest flow in any process purity methane, which are both controlled.
should always be under level control, but Methane is withdrawn overhead as a vapor
the flow of vapor has no direct influence on product, so that pressure must be controlled
liquid level. The assignment of base level by manipulating the flow of refrigerant to
controlled by sidestream vapor will work, the condenser.
however, if column differential pressure is The rule base for sidestream-column
controlled by heat input (Shinskey, 1984, configuration is about three times as com-
page 252). plex as that for simple columns. The con-
A similar problem appears with a Class I figuration program is capable of drawing
column whose bottom flow is relatively about 300,000 different diagrams for
;... r~·.·.·.·.~~·.·.·.·~::::::::::::::::::::::::::::::::::::::::~t"::"::"r::::::::::::::::::::::::::::::::::::::::1········
:i ~ ~qlfi . . · ·. · · · · · · ·. · · . . ····+
ii ["'9 ["'~'T+- 1
1J,·····:···~
: ~.~IJ1.- . "!
··....·········::::::::::::::::di..·····.. ·······~
re··: !.......................................
r·~··i
·+::::::::::::::!:~~~~~~;;;;;;;;;;;;;;;;;;;;;;;;;;Ll.... . . . . . . . . . . . . ..1
FIGURE 13-6. Both the ethane and methane content are controlled in the
sidestream from this ethylene fractionator. (Note: Crossing lines do not intersect.)
290 Practical Distillation Control
Rejerenus
13-5 NOMENCLATURE Ayral, T. E. (1989). On-line expert system for
B Bottom-product flow process control. Hydrocarbon Proc. 68, 61-63.
c Controlled variable Babb, M. (1990). New expert system guides con-
D Distillate-product flow trol operators. Control Eng. 37, 80-81.
Finco, M. V., Luyben, W. L., and Polleck, R. E.
d Derivative
(1989). Control of distillation columns with
e Deviation from setpoint
low relative volatilities. I &EC Res. 28, 75-83.
F Feed flow Kraus, T. W. and Myron, T. J. (1984). Self-tuning
L Reflux flow PID controller uses pattern recognition ap-
m Manipulated variable proach. Control Eng. 31.
p Proportional band Leung, K. S. and Lam, W. (1988). Fuzzy concepts
Qi Heat inflow in expert systems. Computer 21, 43-56.
Qo Heat outflow Ryskamp, C. J. (1980). New strategy improves
q Load dual-composition control. Hydrocarbon Pro-
r Setpoint cessing 59.
S Separation Shinskey, F. G. (1984). Distillation Control, 2nd
ed. New York: McGraw-Hill.
t Time
Shinskey, F. G. (1988). Process Control Systems,
V Vapor boilup
3rd ed. New York: McGraw-Hill.
x Bottom composition Shirley, R. S. (1987). Status report 3 with lessons:
y Distillate composition An expert system to aid process control. Pa-
z Feed composition per given at IEEE Annual Pulp and Paper
a Partial derivative Industrial Technical Conference, 1-5 June
4 Difference 1987.
14
Robust Control
Sigurd Skogestad
University of Trondheim
The objective of this chapter is to give the powerful tool to analyze the robust stability
reader a basic knowledge of how robustness and performance of multivariable control
problems arise and what tools are available systems.
to identify and avoid them.
We first discuss possible sources of model
uncertainty and look at the traditional 14-1 ROBUSTNESS AND
methods for obtaining robust designs, such UNCERTAINTY
as gain margin, phase margin, and maxi-
mum peak criterion (M circles). However, A control system is robust if it is insensitive
these measures are difficult to generalize to to differences between the actual system
multivariable systems. and the model of the system that was used
As an introductory example to robustness to design the controller. Robustness prob-
problems in multivariable systems we then lems are usually attributed to differences
discuss two-point control of distillation between the plant model and the actual
columns using the LV configuration. Be- plant (usually called model-plant mismatch
cause of strong interactions in the plant, a or simply model uncertainty). Uncertainty
decoupler is extremely sensitive to input in the plant model may have several origins:
gain uncertainty (caused by actuator uncer-
tainty). These interactions are analyzed us- 1. There are always parameters in the lin-
ing singular value decomposition (SVD) and ear model that are only known approxi-
RGA analysis. We show that plants with mately or are simply in error.
large RGA elements are fundamentally 2. Measurement devices have imperfec-
difficult to control and that decouplers tions. This may give rise to uncertainty of
should not be used for such plants. It is the manipulated inputs in a distillation
shown that other configurations may be less column because they are usually mea-
sensitive to model uncertainty. sured and adjusted in a cascade manner.
At the end of the chapter we look at In other cases, limited valve resolution
uncertainty modelling in terms of norm may cause input uncertainty.
bounded perturbations (as). It is shown that 3. At high frequencies even the structure
the structured singular value f.L is a very and the model order is unknown, and the
291
292 Practical Distillation Control
uncertainty will exceed 100% at some Robust performance (RP). The system also
frequency. satisfies the performance specifications
4. The parameters in the linear model may for the worst-case model uncertainty.
vary due to nonlinearities or changes in
the operating conditions.
14·2 TRADITIONAL METHODS
FOR DEALING WITH MODEL
Other considerations for robustness in- UNCERTAINTY
clude measurement and actuator failures,
constraints, changes in control objectives, 14-2-1 Slngle-lnput-Slngle-Output
opening or closing other loops, and so on. Systems
Furthermore, if a control design is based on For single-input-single-output (SISO) sys-
an optimization, then robustness problems tems one has traditionally used gain margin
may also be caused by the mathematical (GM) and phase margin (PM) to avoid
objective function, that is, how well this problems with model uncertainty. Consider
function describes the real control problem. a system with open-loop transfer function
In the somewhat narrow use of the term g(s)c(s), and let gc(jw) denote the fre-
used in this chapter, we shall consider ro- quency response. The GM tells by what
bustness with respect to model uncertainty factor the loop gain Igc(jw)1 may be in-
and shall assume that a fixed (linear) con- creased before the system becomes unsta-
troller is used. Intuitively, to be able to cope ble. The GM is thus a direct safeguard
with large changes in the process, this con- against steady-state gain uncertainty (error).
troller has to be detuned with respect to the Typically we require GM > 1.5. The phase
best response we might achieve when the margin tells how much negative phase we
process model is exact. can add to gc(s) before the system becomes
To consider the effect of model uncer- unstable. The PM is a direct safeguard
tainty, it needs to be quantified. There are against time delay uncertainty: If the system
several ways of doing this. One powerful has a crossover frequency equal to we [de-
method is the frequency domain (so-called fined as Igc(jwe)1 = 1], then the system be-
H-infinity uncertainty description) in terms comes unstable if we add a time delay of
of norm-bounded perturbations (As). With (J = PM/We' For example, if PM = 30° and
this approach one also can take into ac- We = 1 rad/min, then the allowed time d~
count unknown or neglected high-frequency lay error is (J = (30/57.3) rad/l rad/mm
dynamics. This approach is discussed to- = 0.52 min. It is important to note that
ward the end of this chapter. Readers who decreasing the value of We (lower closed-
want to learn more about these methods loop bandwidth, slower response) means
than we can cover may consult the books by that we can tolerate larger deadtime errors.
Maciejowski (1989) and Morari and Zafiriou For example, if we design the controller
(1989). such that PM = 30° and expect a deadtime
The following terms will be useful: error up to 2 min, then we must design the
control system such that We < PM/(J =
Nominal stability (NS). The system is stable (30/57.3)/2 = 0.26 rad/min, that is, the
with no model uncertainty. closed-loop time constant should be larger
Nominal performance (NP). The system sat- than 1/0.26 = 3.8 min.
isfies the performance specifications with
no model uncertainty. Maximum Peale Criterion
Robust stability (RS). The system also is In practice, we do not have pure gain and
stable for the worst-case model uncer- phase errors. For example, in a distillation
tainty. column the time constant will usually in-
Robust Control 293
crease when the steady-state gain increases. bances). One particular approach is the loop
A more general way to specify stability mar- transfer recovery (LTR) method where un-
gins is to require the Nyquist locus of gc(jw) realistic noise is added specifically to obtain
to stay outside some region of the - 1 point a robust controller design. One may say that
(the "critical point") in the complex plane. model uncertainty generates some sort of
Usually this is done by considering the max- disturbance. However, this disturbance de-
imum peak M t of the closed-loop transfer pends on the other signals in the systems,
function T: and thus introduces an element of feed-
back. Therefore, there is a fundamental
M t = max IT(jw) I, T=gc(l +gC)-1 difference between these sources of uncer-
w tainty (at least for linear systems): Model
(14-1)
uncertainty may introduce instability,
whereas signal uncertainty may not.
The reader may be familiar with M circles For SISO systems the main tool for ro-
drawn in the Nyquist plot or in the Nichols bustness analysis has been GM and PM,
chart. Typically, we require M t = 2. There and as previously noted, these measures are
is a close relationship between M t and PM related to specific sources of model uncer-
and GM. Specifically, for a given M t we are tainty. However, it is difficult to generalize
guaranteed GM and PM to MIMO systems. On the
other hand, the maximum peak criterion
1 may be generalized easily. The most com-
GM ~ 1 +-,
Mt mon generalization is to replace the abso-
lute value by the maximum singular value,
PM ~ 2 arcsin ( _1_)
2M
>
-
~
M
rad (14-2) for example, by considering
t t
M t = maxii( T(jw)) ,
For example, with M t = 2 we have GM > w
1.5 and PM > 29.0° > IIMt rad = 0.5 rad.
T = GC(I + GC) -I (14-3)
Comment The peak value Ms of the
sensitivity function S = (1 + gc) - 1 may be
used as an alternative robustness measure. The largest singular value is a scalar posi-
1I Ms is simply the minimum distance be- tive number, which at each frequency mea-
tween gc(jw) and the -1 point. In most sures the magnitude of the matrix T. As
cases the values of M t and Ms are closely shown later, this approach has a direct rela-
related, but for some "strange" systems tionship to important model uncertainty de-
it may be safer to specify M t rather than scriptions and is used in this chapter.
Ms. For a given value of Ms we are guar- Comment In Chapter 11, a different
anteed GM > MsI(Ms - 1) and PM > generalization of M t to multivariable sys-
2arcsin(1/2Ms) > IIMs· tems is used: First introduce the scalar
function W(jw) = detU + GC(jw» - 1
[for SISO systems W(jw) = GC(jw)] and
14-2-2 Multi-Input-Multi-output
then define Lc = IW1(1 + W)I. The maxi-
Systems
mum peak of ILcl {in decibels) is denoted
The traditional method of dealing with L r:ax and is used as part of the biggest
robustness for multi-input-multi-output log-modulus tuning (BLT) method. For
(MIMO) systems (e.g., within the frame- SISO systems Lr:ax = Mt •
work of "optimal control," linear quadratic Even though we may easily generalize
Gaussian (LQG), etc.) has been to intro- the maximum peak criterion (e.g., M t ~ 2)
duce uncertain signals (noise and distur- to multivariable systems, it is often not use-
294 Practical Distillation Control
LV DV Vr ~..
p(Mv)
L D. Yo L D. }b
F.z F. ZF
v
........~
B. Xe
the model is simple and displays important split D jF (with V constant the increase in
features of the distillation column behavior. L yields a corresponding decrease in D);
The RGA matrix for this model is at all that is, the column is very sensitive to
frequencies: changes in the external flows D and B.
Similarly, if we increase V by only 0.01
RGA(G LV ) = ( 35.1 - 36.1) (14-5) (with L constant), we see that the predicted
-36.1 35.1 steady-state change in Yv is - 0.00864.
Again, this is a very large change, but in the
The large elements in this matrix indicate direction opposite that for the increase
that this process is fundamentally difficult to in L.
control. We therefore see that changes in Land
V counteract each other, and if we increase
Interactions and JU-Conditionedness L and V simultaneously by 0.01, then the
Consider the case with no composition con- overall steady-state change in Yv is only
trol. The effect on top composition of a 0.00878 - 0.00864 = 0.00014. The reason
small change in reflux L with V constant is for this small change is that the composi-
tions in the column are only weakly depen-
0.878 dent on changes in the internal flows (i.e.,
ayv(s) = 5 1 aLes) changes in the internal flows L and V with
7s+
the external flows D and B constant).
If we increase L by only 0.01 (that is, LjF Summary Because both L and V affect
is increased 0.4% from 2.7 to 2.701), then both compositions Yv and X B , we say that
we see that the steady-state increase in yv the process is interactive. Furthermore, the
predicted from this linear model is 0.00878 process is "ill-conditioned," that is, some
(that is, Yv increases from 0.99 to 0.99878). combinations of aL and aV (correspond-
This is a rather drastic change, and the ing to changing external flows) have a strong
reason is that the column operation is very effect on the compositions, whereas other
dependent on keeping the correct product combinations of aL and aV (correspond·
296 Practical Distillation Control
ing to changing internal flows) have a weak sition (YD + x B )/2. All columns with both
effect on the compositions. The condition products of high purity are sensitive to
number, which is the ratio between the gains changes in the external flows because the
in the strong and weak directions, is there- product rate D has to be about equal to the
fore large for this process (as seen in the amount of light component in the feed. Any
following text, it is 141.7). . imbalance leads to large changes in product
compositions (Shinskey, 1984).
Singular Value Analysis of the Model The low plant gain, !l( G) = 0.0139, is
The preceding discussion shows that this obtained for inputs in the direction (~~) =
column is an ill-conditioned plant, where
the effect (the gain) of the inputs on the f:!. = (~::) which corresponds to changing
outputs depends strongly on the direction of . the internal flows only (t1B = t1L - t1V::= 0).
the inputs. To see this better, consider the From Jhe output vector ~ = (_~:~~) we
SVD of the steady-state gain matrix
see that the effect is to move the outputs in
(14-6) different directions, that is, to change YD -
x B. Thus, it takes a large control action to
or equivalently, because V T = V-I, move the compositions in different direc-
tions, that is, to make both products purer
GD = u(G)u, Gf:!. = !l(G)~ simultaneously. The condition number of
the plant, which is the ratio of the high and
where u denotes the maximum singular low plant gain, is then
value and !l the minimum singular value.
The singular values are y( G) = u( G)/!l( G) = 141.7 (14-7)
x 10-3 ,----,-----.------r----r---.,--.-----,-----.----r----,
.-
f/. . ,
I, "
·
I, I,
S \\
·
"
"
2 : \\
···
""
"
""
· "
:,·
""
"
, "
"
\' \~ ,
:t
"
~
.... ,
..... ,_
~YD
lV-I \,~,=------ - - - - - I
:" \\
~B
o '",----
o 10 20 30 40 so 60 70 80 90 100
min
FIGURE 14-2. Response for decoupling controller using LV configuration.
Setpoint change in YD = 1O- 3/(5s + 1). Solid line: No uncertainty; dotted line:
20% input gain uncertainty as defined in Equation 14-9.
the direction of the large inputs may not the simulated response to a setpoint change
correspond exactly to the low plant gain in top composition.
direction, and the amplification of these
large input signals may be much larger than 14-3-3 Use of Decoupler When There
expected. As shown in succeeding simula- Is Model Uncertainty
tions, this will result in large values of the
controlled variables y, leading to poor per- In practice, the plant is different from the
formance or even instability. Consider the model, and the dotted lines in Figure 14-2
following decoupling controller (a steady- show the response when there is 20% error
state decoupler combined with a PI con- (uncertainty) in the gain in each input chan-
troller): nel ("diagonal input uncertainty"):
~L = 1.2 ~Lc' ~V = 0.8 ~~ (14-9)
kl
C1(s) = -:;-GLV-1(S) ~ L and ~ V are the actual changes in the
manipulated flow rates, whereas ~Lc and
k l (1 + 75s) (39.942 ~~ are the desired values (what we believe
-31.487)
s 39.432 -31.997 ' the inputs are) as specified by the con-
troller. It is important to stress that this
kl = 0.7 min- l (14-8) diagonal input uncertainty, which stems
from our inability to know the exact values
We have GC I = 0.7 lsI. In theory, this con- of the manipulated inputs, is always pre-
troller should counteract all the directions sent. Note that the uncertainty is on the
of the plant and give rise to two decoupled change in the inputs (flow rates), not on
first-order responses with time constant their absolute values. A 20% error is rea-
1/0.7 = 1.43 min. This is indeed confirmed sonable for process control applications
by the solid line in Figure 14-2, which shows (some reduction may be possible, for exam-
298 Practical Distillation Control
a
pie, by use of cascade control using flow spect to the values of L and av makes it
measurements, but there will still be uncer- impossible to make them both large while at
tainty because of measurement errors). Re- the same time keeping their difference small.
gardless, the main objective of this chapter The result is a undesired large change in
is to demonstrate the effect of uncertainty, the external flows, which subsequently re-
artd its exact magnitude is of less impor- sults in large changes in the product compo-
tance. sitions because of the large plant gain in
The dotted lines in Figure 14-2 show the this direction. As we shall discuss in the
response with this model uncertainty. It following text, this sensitivity to input un-
differs drastically from the one predicted by certainty may be avoided by controlling D
the model, and the response is clearly not or B directly, for example, using the DV
acceptable; the response is no longer decou- configuration.
pled, and aYD and aXB reach a value of
about 2.5 before settling at their desired
14-3-4 Alternative Controllers:
values of 1 and O. In practice, for example,
Slngl.Loop PID
with a small time delay added at the out-
puts, this controller would give an unstable Unless special care is taken, most multivari-
response. able design methods (MPC, DMC, QDMC,
There is a simple physical reason for the LQG, LQG/LTR, DNA/INA, IMC, etc.)
observed poor response to the setpoint yield similar inverse-based controllers, and
change in YD. To accomplish this change, do generally not yield acceptable designs for
which occurs mostly in the "bad" direction ill-conditioned plants. This follows because
corresponding to the low plant gains, the they do not explicitly take uncertainty into
inverse-based controller generates a large account, and the optimal solution is then to
change in the internal flows (aL and aV), use a controller that tries to remove the
while trying to keep the changes in the interactions by inverting the plant model.
external flows (aB = -aD = aL - aV) The simplest way to make the closed-loop
very small. However, uncertainty with re- system insensitive to input uncertainty is to
x 10-3
min
FIGURE 14-3. Response for PID controller using LV configuration.
Robust Control 299
use a simple controller (for example, two Lundstrom, Skogestad, and Wang, 1991),
single-loop PID controllers) that does not which is designed to optimize the worst-case
try to make use of the details of the direc- response (robust performance) as discussed
tions in the plant model. The problem with toward the end of this chapter. Although
such a controller is that little or no correc- this is a multivariable controller, we note
tion is made for the strong interactions in that the response is not too different from
the plant, and then even the nominal re- that with the simple PID controllers, al-
sponse (with no uncertainty) is relatively though the response settles faster to the
poor. This is shown in Figure 14-3 where we new steady state.
have used the following PID controllers
(Lundstrom, Skogestad, and Wang, 1991):
14-3-5 Alternative Configurations:
YD -L: DV Control
Kc = 162, T[ = 41 min, The process model considered in the pre-
TD = 0.38 min (14-10) ceding text, which uses the LV configura-
tion, is fundamentally difficult to control
XB - V: irrespective of the controller. In such cases
Kc = -39, T[ = 0.83 min, one should consider design changes that
TD = 0.29 min (14-11)
make the process simpler to control. One
such change is to consider the DV configu-
ration where L rather than D is used for
The controller tunings yield a relatively fast
condenser level control (Figure 14-1b). The
response for x B and a slower response for
independent variables left for composition
YD. As seen from the dotted line in Figure
control are then D and V:
14-3, the response is not very much changed
by introducing the model error in Equation
14-9.
In Figure 14-4 we show response for (14-12)
a so-called mu-optimal controller (see
x 10-3
·2
I1YD
I1xB
o 10 20 30 40 SO 60 70 80 90 100
min
I1GURE 14-4. Response for Il--optimal controller using LV configuration.
300 Practical Distillation Control
= 75s
1
+1
(0.878
0.014
-0.864)
-0.014
tion for internal flows. We may then make
large changes in the internal flows V with-
out changing the external flows D. This was
(14-14)
not possible with the LV configuration;
x 10-3 r--.----...---,,.----r---.----.--~---.---.--...,
:r
DGURE 14-5. Response for decoupling controller using DV configuration.
Robust Control 301
where we had to increase both L and V in case a change in D from the composition
order to increase the internal flows. controller will immediately change L with-
The RGA behavior for various other out having to wait for the level loop.
configurations is treated in detail by
Shinskey (1984) for the static case and by
Skogestad, Lundstrom, and Jacobsen (1990)
14-4 RGA AS A SIMPLE TOOL TO
for the dynamic case. DETECT ROBUSTNESS PROBLEMS
14-4-1 RGA and Input Uncertainty
14-3·6 Limitations with the Example: We have seen that a decoupler pedormed
Real Columns very poorly for the LV model. To under-
It should be stressed again that the column stand this better consider the loop gain Gc.
model used in the preceding text is not The loop gain is an important quantity be-
representative of a real column. In a real cause it determines the feedback properties
column the liquid lag 8L (min), from the top of the system. For example, the transfer
to the bottom, makes the initial response function from setpoints ys to control error
for the L V configuration less interactive e = Ys - y is given by e = SYs = (I +
GC)-lys. We therefore see that large
and the column is easier to control than
found here. It turns out that the important changes in GC due to model uncertainty
parameter to consider for controllability is will lead to large changes in the feedback
not the RGA at steady state (with exception response. Consider the case with diagonal
of the sign), but rather the RGA at frequen- input uncertainty fl. l . Let fl.l and fl.2 repre-
cies corresponding to the closed-loop band- sent the relative uncertainty on the gain in
width. For the LV configuration the RGA each input channel. Then the actual ("per-
is large at low frequencies (steady state), turbed") plant is
but it drops at high frequencies and the
RGA matrix becomes close to the identity
matrix at frequencies greater than 1/8L •
Thus, control is simple, even with single- (14-16)
loop PI or PID controllers, if we are able tQ
achieve very tight control of the column.
However, if there are significant measure- In the simulation example we had fl.1 = 0.2
ment delays (these are typically 5 min or and fl.2 = -0.2. The perturbed loop gain
larger for GC analysis), then we are forced with model uncertainty becomes
to operate at low bandwidths and the re-
sponses in Figures 14-2 through 14-4 are GpC = G(I + fl.l)C = GC + Gfl.lC
more representative. Furthermore, it holds (14-17)
in general that one should not use a
steady-state decoupler if the steady-state If a diagonal controller C(s) (e.g., two PIs)
RGA elements are large (typically larger is used, then we simply get (because fl.l is
than 5). also diagonal) GpC = GC(I + fl. l ) and
In a real column one must pay attention there is no particular sensitivity to this un-
to the level control for the DV configura- certainty. On the other hand, with a pedect
tion. This is because D does not directly decoupler (inverse-based controller) we have
affect composition; only indirectly, through
its effect on reflux L through the level loop, C(s) = k(s)G- 1(s) (14-18)
does it have influence. In practice, it may be
a good idea to let the condenser level con- where k(s) is a scalar transfer function, for
troller set L + D rather than L. In this example, k(s) = 0.7/s, and we have GC =
302 Practical Distillation Control
k(s)I, where I is the identity matrix, and closed-loop time constant. A diagonal con-
the perturbed loop gain becomes troller (e.g., single-loop PIs) is robust (in-
sensitive) with respect to input uncertainty,
GpC = Gp(/ + d[)C but will be unable to compensate for the
= k(s)(I + Gd[G- 1 ) (14·19) strong couplings (as expressed by the large
RGA elements) and will yield poor perfor-
For the aforementioned LV configuration, mance (even nominally). On the other hand,
the error term becomes an inverse-based controller that corrects for
G LVI1[(G LV ) -1 the interactions may yield excellent nominal
performance, but will be very sensitive to
= (35.111 1 - 34.1112 -27.711 1 + 27.711 2 ) input uncertainty and will not yield robust
43.2111 - 43.2112 -34.1111 + 35.111 2 performance. In summary, plants with large
(14·20) RGA elements around the crossover fre-
quency are fundamentally difficult to con-
This error term is worse (largest) when d 1
trol, and decouplers or other inverse-based
and d 2 have opposite signs. With d 1 = 0.2
controllers should never be used for such
and d 2 = -0.2 as used in the simulations
plants (the rule is never to use a controller
(Equation 14-9), we find
with large RGA elements). However, one-
(13 way decouplers may work satisfactorily.
GLvd[(G LV )- 1 = 17:28 -11.1)
-13.8 (14·21)
The elements in this matrix are much larger 14-4-2 RGA and Element Uncertainty /
than 1, and the observed poor response with Identification
uncertainty is not surprising. Similarly, for
Previously we introduced the RGA as a
the DV configuration we get
sensitivity measure with respect to input
GDVd (G DV )-1 = (-0.02 0.18) (14.22) gain uncertainty. In fact, the RGA is an
I 0.22 0.02 even better sensitivity measure with respect
to element-by-element uncertainty in the
The elements in this matrix are much less
matrix.
than 1, and good performance is maintained
Consider any complex matrix G and let
even in the presence of uncertainty on each
Ajj denote the ijth element in its RGA
input.
matrix. The following result holds (Yu and
The observant reader may have noted
Luyben, 1987):
that the RGA elements appear on the diag-
onal in the matrix G Lvd/(G LV )-1 in Equa- The (complex) matrix G becomes
tion 14-20. This turns out to be true in singular if we make a relative change
general because diagonal elements of the -l/A jj in its ijth element, that is, if
error term prove to be a direct function of a single element in G is perturbed
the RGA (Skogestad and Morari, 1987): from gjj to gpjj = gil - l/Ajj).
n
(GdG- 1 )u = E Ajj(G)d j (14·23) Thus, the RGA matrix is a direct measure
j=l of sensitivity to element-by-element uncer-
tainty and matrices with large RGA values
Thus, if the plant has large RGA elements
become singular for small relative errors in
and an inverse-based controller is used, the
the elements.
overall system will be extremely sensitive to
input uncertainty.
Example 14·1
Control Implications The matrix G LV in Equation 14-4 is non-
Consider a plant with large RGA elements singular. The 1,2 element of the RGA
in the frequency range corresponding to the is A12(G) = -36.1. Thus the matrix G
Robust Control 303
certainty. Skogestad, Morari, and Doyle 14-5·2 Condhlons for Robust Stability
(1988) and Lundstrom, Skogestad, and
By robust stability (RS) we mean that the
Wang (1991) use an approach where one
system is stable for all possible plants as
considers numerically the relative uncer-
defined by the uncertainty set (using the
tainty generated by the time delay. This
a/s as previously discussed). This is a "worst
results in a complex perturbation .1 2, but
case" approach, and for this reason one
otherwise in the same w2 , except that it
must be careful to not include unrealistic or
levels off at 2 at high frequencies. This
impossible parameter variations. This is why
approach is used in the computations that
it is recommended not to include large indi-
follow. (However, for other reasons we
vidual variations in the gain k and the time
would probably have preferred to not let w2
constant 7' for a distillation column model.
level off at 2 if we had redone this work
Now, consider the distillation column ex-
today.)
ample with combined gain and time delay
We now have two sources of relative
uncertainty. For multivariable plants it
uncertainty. Combining them gives an over-
makes a difference whether the uncertainty
all 2 x 2 (real) perturbation block a, with
is at the input or the output of the plant.
.11 and .12 on its diagonal. To simplify, we
We will here consider input uncertainty,
may include the combined effect of the gain
and the weight WI then represents, for ex-
and time delay uncertainty using a single
ample, variations in the input gain and ne-
(complex) perturbation by adding their mag-
glected valve dynamics. We assume the same
nitudes together, that is, W = Ir11 + Iw 2 1, or
magnitude of the uncertainty for each input.
approximately
The set of possible plants is given by
(14-29)
get a simple analytic expression for JL for may be used as a robustness criterion. How-
robust stability: ever, here we are restricting the peak of S
at high frequencies primarily to get a good
JLt,.{M) = O=(w/T/) response (without too much oscillation and
overshoot). The robustness issues are taken
The peak value (with respect to frequency) 14-5-4 Conditions for Robust
of the weighted sensitivity wpS is also called Performance
the Roo norm. w/s) is the performance
weight. Typically, we use the weight Robust performance (RP) means that the
performance specification is satisfied for the
worst-case uncertainty. The most efficient
way to test for RP is to compute JL for RP.
If this JL value is less than 1 at all frequen-
cies, then the performance objective is satis-
This requires (a) integral action, (b) that the fied for the worst case. Although our system
peak value of lSI should be less than Ms has good robustness margins and excellent
(typically Ms = 2), and (c) that the closed- nominal performance, we know from the
loop response time should be less than 'Tel simulations in Figure 14-2 that the perfor-
(Le., the bandwidth should at least be (J)B = mance with uncertainty (RP) may be ex-
lj'TcI )· tremely poor. This is indeed confirmed by
For multivariable systems, the largest sin- the JL curve for RP in Figure 14-8, which
gular value of S, u(S), is used instead of the has a peak value of about 6. This means
absolute value lSI. In the introduction we that even with 6 times less uncertainty, the
mentioned that the maximum peak on S performance will be about 6 times poorer
308 Practical Distillation Control
313
314 Practical Distillation Control
This chapter discusses an experimental control structures was also sparked by the
comparison of different control structures findings mentioned in the preceding text
(mainly) for two-point control of a binary that in two-point control, different struc-
separation in a 15-plate pilot plant column tures do not only show different loop inter-
separating ethanol and water. actions but they also reject disturbances to
various degrees.
15-2 MANIPULATOR CHOICE FOR An unusual choice of manipulators was
recently proposed by Finco, Luyben, and
DECENTRALIZED CONTROL
Polleck (1989), who investigated, both
In multiloop SISO control of distillation, through simulation and industrial experi-
the manipulator choice for the top composi- ments, the (D, B) structure previously re-
tion loop has traditionally been reflux flow jected by researchers in the field as repre-
rate L or distillate flow rate D or a ratio of senting an infeasible structure.
the two. For the bottoms composition loop, A manipulator choice like L / D means
vapor boilup V' (manipulated through the that both Land D are simultaneously used
heat input to the reboiler, here denoted V) as manipulators. In linearized form the
or bottoms flow B or a ratio of the two has manipulator can be expressed as al/1 =
analogously been used. The structure that a a L + baD. Indeed, Yang, Seborg, and
uses L and V has been called energy bal- Mellichamp (1991) suggest something they
ance control; structures using D or B have call "combined balance control," where the
been called material balance control. (It manipulators are
could be mentioned that the manipulators
affecting the compositions in the column al/11 = a aL + (1 - a)( -aD) (15-1)
are Land V. Using D or B is actually only al/12 = baY + (1 - b)( -ilB) (15-2)
an indirect way of manipulating L or V
through the level control loops.) Equations 15-1 and 15-2 contain the energy
In addition to the use of simple flows as balance control (a = b = 1), the two tradi-
composition manipulators, as in the (L, V) tional material balance schemes (D, V) and
and (D, V) structures, ratios of flows often (L, B), and the new super material balance
have been suggested for use. An early ex- structure (D, B) (for a = b = 0), as well as
ample is Rijnsdorp's suggestion (Rijnsdorp, many of the ratio control structures (in lin-
1965) to use the ratio of reflux flow and earized form). By choosing a and b =1= 0 and
overhead vapor flow L/(L + D) as a ma- =1= 1 various combinations of these control
nipulator for the top composition control structures are obtained, which explains the
loop. The suggestion was experimentally name "combined balance control."
studied by Wood and Berry (1973), who The combination can be further devel-
compared the «L + D)/L, V) structure oped. Actually, by using transformation the-
with the conventional (L, V) structure. ory (Haggblom and Waller, 1988) based on
The book by Rademaker, Rijnsdorp, and consistency relations like total and partial
Maarleveld (1975) lists a large number of material balances, Haggblom and Waller
flow ratio manipulators discussed in the lit- (1990) derived manipulators that, in the
erature. steady state, not only decoupled the two
Renewed interest in ratio control awoke composition control loops, but also elimi-
in the 1980s. Ryskamp (1980) proposed nated disturbances in feed composition and
the scheme (D /(L + D), V). An extension feed flow rate. The manipulators obtained
of this scheme, (D/(L + D), V/B) was can be written
suggested and studied by Takamatsu,
Hashimoto, and Hashimoto (1982, 1984) and ill/11 = a aD + bilL + c ilB + d ilV (15-3)
by Shinskey (1984). Increased interest in ill/12 = e ilD + filL + g ilB + h ilV (15-4)
Experimental Comparison of Control Structures 315
The acronym DRD (disturbance rejection TABLE 15-1 Technical Data for the Experimental
and decoupling) has been used for the Distillation System
structure. Column
Diameter O.30m
Plate spacing O.25m
15·3 EXPERIMENTAL APPARATUS Number of plates 15
Number of bubble-caps/plate 18
A flowsheet for the experimental equipment Holdup area/plate 455 cm 2
is shown in Figure 15-1, which also shows Weir height 2.54cm
the instrumentation. The column is a IS-tray Holdup of bottom section 6.2dm3
bubble-cap column. From the two feed
tanks, each with a volume of 1000 I, feed is Thermosyphon
Heat transfer area 6.32m 2
pumped to the column through a feed pre- Holdup (including piping) 9.2dm3
heater. The reboiler is of a thermosyphon
type, which (like the feed preheater) uses Condenser
fresh steam of 3 bar pressure. The con- Cooling capacity 300MJ/h
denser is a water-cooled total condenser.
Reflux drum
The reflux drum is vented to the atmo- Cross area 212cm 2
sphere. Technical data for the system are Volume 14dm3
given in Table 15-1.
H H A Column
B Reboiler
C Condenser
D Reflux drum
E Distillate cooler
F Bottoms cooler
G Feed preheater
H Product tanks
J Feed tanks
~
Feed flow rate F 200kg/h
88 Distillate flow rate D 60 kg/h
Bottoms flow rate B 140kg/h
Feed composition z 30wt%
14
86 Distillate composition x D 87wt%
Bottoms composition x B 5wt%
Reflux flow rate L 60 kg/h
T 84 12 Steam flow to reboiler V 72kg/h
""-.
65°C
t Feed temperature
Reflux temperature 62°C
82
r-
IO
3
80 vary significantly in the column sections has
6
\ highlighted several interesting properties of
"", ~ the system that would have passed unno-
78 0
ticed if a separation with constant (molar)
-30 o 30 60 90 120 min flows had been investigated. One example is
~t
the generally incorrect (steady-state) rela-
tion, often used in the literature, IlD =
FIGURE 15-2. Illustration of recorded column tem-
- ilL (for F and V constant); that is, that a
peratures after step change of heat input to reboiler.
change in reflux flow L eventually brings
about an equally large (negative) change in
distillate flow D. How incorrect the relation
In the column temperatures are mea- may be is discussed, for example, by
sured on every second plate. An illustration Hliggblom and Waller (1988, 1989). For the
of a recorded temperature profile is shown experimental ethanol-water system dis-
in Figure 15-2. cussed here, IlD = - 0.6 ilL at the steady
state considered. This large steady-state
difference between IlD and - ilL has, in
15-4 MIXTURE DISTILLED tum, revealed the large difference in distur-
bance sensitivity that may occur between
Ethanol-water is the separation studied.
different control structures.
The system is treated on a mass basis, which
Data for the nominal steady state studied
means that the flows are far from being
in this chapter are given in Table 15-2.
constant in the column, which they often
are assumed to be in simulations. This is
thus more in agreement with how columns
15-5 CONTROL STRUCTURES
are operated in practice. Actually, keeping
a flow like reflux constant means keeping STUDIED
the flow constant on a volumetric or mass Six control structures are experimentally
basis or a mixture of the two. This is exten- studied. In all structures the controlled out-
sively discussed in Jacobsen and Skogestad puts are the same, that is, the temperatures
(1991). Flows cannot usually be manipu- on plates 4 and 14 (counting from the top).
lated on a molar basis, which is why it It was not possible to move the enriching
generally seems unsound to make such as- section measurement closer to the top be-
sumptions in simulations. cause the temperature is quite insensitive to
This study of a mixture that has such composition changes at high ethanol con-
properties as to make the internal flows centrations. Even at plate 4 a specially de-
Experimental Comparison of Control Structures 317
O/(L+O}
r-----------------------------------,,
,
,,
f,';::\L+O .d;,
,,, -~-~-~
a
(9--
bT4 L D Xo ,r--
-<)+' :
,
10
,10 , L I I
,,
F ~ ________________ J I
D Xo
z F
.--~-
d
14 z
v
:v v'
0,
I
t-------l>- -
't':::F : B v
B Xs
'----.p!q--
B Xs
FIGURE 15-3. Implementation of the (L, V) struc-
ture. FIGURE 15-5. Implementation of the (D I(L + D),
V) structure.
r---------------------------------
,,, O/(L+O)
r- - ----- --- - - -------------- ---------,
!,, :,,
,, ,,
o , f,';;\ L+O ~
,,,
: -\!:Srj-$
: _+1 :
, , r __ A '0
L D Xo
I I V.
al!.T4 6~T4 1D :L
I!::r
I
(!S,r- :c ________________ J I
D Xo
F F
z z
r--~0~~~@~·
1 IV v'
v l, v I
,,
,,
,,
L. _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ •
FIGURE 15-4. Implementation of the (D, V) struc- FIGURE 15-6. Implementation of the (DI(L + D),
ture. V IB) structure.
318 Practical Distillation Control
Gu G12 G 1F G 1z
Structure G 21 G 22 G 2F G 2z
o 50 min 100
-+t
FlGURE 15·9. Open-loop responses to a step change
in feed flow rate from 180 to 200 kg/h for four of the
strategies studied. Simulations with the experimentally
obtained transfer functions. [Reprinted by permission
from Waller et at. (l988b). Copyright 1988 American
Chemical Society.)
_::r: ~
0.4 ,
L, V ~T4
°C D, V T
D.T4 D/ (L+D) , V
a
1 0.1 D/(L+D), V/8
0 .. ~;;:::~-~=.~~-~-.-.-.-.-.-.-.-.-.-.-.-.-.-.
,
-------------
:~K
-0.1 •
TI4
T
-0.3
85 ' , :
k::~b!=4
-0.5
87 L
°C 1
T14 _.-._._._._._._._._._._._._._._._-
1 85
100
-------------------~~~~
kg/h
v
1
80 LL..-...>.--"----'--~===:=::;::::::.::=::
o 50 min 100
-d 50 L-~~ __-,--~---,__-,--~---,__-,--~
o 50 min 100
FlGURE 15·8. Open-loop responses to a step in- -+ t
crease jn feed composition of + 3 wt% for four of the
strategies studied. Simulations with the experimentally FlGURE 15-10. System response to setpoint changes
obtained transfer functions. [Reprinted by permission in T ,4 . Two-point control with (L, V) structure.
from Waller et al. (1988b). Copyright 1988 American [Reprinted by permission from Waller et al. (I 988b).
Chemical Society.) Copyright 1988 American Chemical Society.)
320 Practical Distillation Control
transfer functions (Haggblom, 1989; Waller setpoint changes in T14 is shown. The con-
et aI., 1988a). trol system is (L, V) with fixed controller
The feedback controllers for the temper- parameters. As seen, the system stability
ature loops were tuned on-line to give rea- considerably decreases when the setpoint
sonable responses. The controller settings for T14 is increased.
for two-point control are given in Table In addition to having different distur·
15-4. bance sensitivity and control loop inter-
Traditionally, when discussing differ- action, different structures could also be
ences between control structures, the inter- expected to show different degrees of non-
action between the primary control loops linearity. This aspect is highlighted in Fig-
has been the main concern. Table 15-3 re- ures 15-11 and 15-10. In both cases fixed
veals another important difference: the controller settings were used (Table 15-4).
(open-loop) disturbance sensitivity may be No difference concerning nonlinearity be-
very different. (Here "open-loop" means tween (L, V) and (D/{L + D), V) can be
that the temperature loops are open; only obsetved.
the level loops are closed.) Figures 15-8 and
15-9 also show that this degree of self-regu- 15-7 (0, V), (V, 0), AND (L, B)
lation may be very different for different
disturbances.
STRUCTURES
Many times it has been stated in the In the experimental identification of the
literature that high-purity separations may dynamics for the (D, V) structure, G 22 = 0
exhibit a strongly nonlinear behavior. In was obtained (Table 15-3). This means that
such cases it has been suggested that a a change in V results in such a change in L
suitable logarithmic transformation of the that the effects on T14 cancel. A low value
outputs would make the system significantly of K22 in the (D, V) scheme was also ob-
more linear (Skogestad and Morari, 1988). tained by McAvoy and Weischedel (1981)
Although the separation studied in this and by Tsogas and McAvoy (1981) for
chapter does not result in very pure prod- columns with high product purities and high
ucts, the system is quite nonlinear at the reflux ratio. In the study reported in this
operating point studied, as illustrated in paper the result was obtained for a low-pur-
Figure 15-10, where the system response to ity column with a low reflux ratio.
G 22 = 0 means that T14 is not control-
lable by V if the (D +- T4 ) loop is not in
operation (a backward arrow denotes feed-
back). The relative gain for the scheme is
zero, which means that "control depends on
other loop(s)" (Shinskey, 1984).
Because (D, V) thus lacks integrity, it
would be natural to use another scheme.
One appealing alternative is the other
"material balance" scheme (L, B), which
has a relative gain of 1 when (D, V) has a
relative gain of 0 (Haggblom and Waller,
o 50 min 100 1988).
-+ t The (L, B) scheme was experimentally
FIGURE 15-11. System response to setpoint changes tried on the column. However, satisfactory
in T14 • Two-point control with (D/(L + D), V) struc· control of the temperatures was not ob-
ture. tained during the experiments. One prob-
Experimental Comparison of Control Structures 321
',-_ ....
i ,
, I
I
This scheme was originally suggested by
, I
, I a Ryskamp (1980), who motivated it in the
O~~ " following way.
~T4
·C
:'\
I
i'
/
\
The scheme "holds reflux ratio constant
if the top composition controller output is
i 0
___I ~ r\ "' ____________________ _
constant. An increase of heat input from
\ I -
\1
tj the bottom composition controller does not
b
3
0 t - - - -____'--_ _-r---(
o 50 min 100
-+t
L,D
lem was the inverse response between boilup T
and reboiler level, which made it very
0.6
difficult to make this level control loop D
function properly, even for very small dis- L+D
0.5 1. =""""" or
1
A ..... ' O .. Ar"'eV~
make top product as impure as would occur the (Dj(L + D), VjB) scheme is always
with reflux constant (conventional control) lower than the relative gain for Ryskamp's
nor as overpure as would occur with distil- scheme when the latter has a relative gain
late flow constant {material balance larger than 1. If the relative gain for
controI)." Thus, this property of the scheme Ryskamp's scheme is 1, it is 1 also for the
results in a certain decoupling effect. The two-ratio scheme. In this investigation, the
scheme is sometimes said to result in "im- operating conditions for the two-ratio
plicit decoupling," in contrast to "explicit scheme were slightly different from those
decoupling" accomplished by external de- for Ryskamp's scheme. This is one explana-
coupling elements. tion why the relative gain calculated from
For this scheme, G 12 = 0 in our column, the process gains given in Table 15-3 (see
which means that the relative gain is 1 and Table 15-5) do not satisfy the preceding
that changes in V do not affect T4 or, more conditions.
precisely, that a change in V causes such a The response of T14 to changes in the
change in L that the net effect on T4 is O. manipulated variables Dj{L + D) and VjB
The implementation of the scheme did not follow a simple first-order deadtime
caused no practical problems. The two loops response, as is illustrated in Figure 15-14. In
were separately tuned and worked well to- spite of this, the transfer functions have
gether. Figure 15-13 is an experimental il- been identified as first-order deadtime
lustration of how the scheme (D j{L + D), transfer functions (see Table 15-3). For this
V) works. scheme, a change in feed flow rate did not
affect the uncontrolled temperatures T4 and
T14 much (Figure 15-15). At the same time,
15-9 (D I(L +D), V18) however, this change quite strongly affected
STRUCTURE the flows in the column, as shown in the
figure. It was not possible to eliminate the
This scheme is an extension of Ryskamp's
oscillations shown in the figure by detuning
scheme, suggested and studied by Taka-
matsu, Hashimoto, and Hashimoto (1982, the bottom level controller. Actually, it
turned out to be very difficult to make the
1984) and by Shinskey (1984), who have
(VjB +- T14 ) loop operate satisfactorily be-
shown that the steady-state interaction be-
tween the two primary loops often is small. cause of stability problems. An explanation
Shinskey (1984, page 156) has calculated
eight separation cases, and in all of them
he obtained a lower relative gain (less in-
teraction at steady state) for the scheme
(Dj{L + D), VjB) than for Ryskamp's
scheme. In fact, it can be theoretically shown
(Haggblom, 1988) that the relative gain for
o 1-......-----------------------
Tu
T
2::
ioo
E '
FIGURE 15-16. Illustration of the coupling between
the reboiler level control loop and the T14 control loop
for the scheme CD I(L + D), V IB). [Reprinted by per-
mission from Waller et al. (t988b), Copyright 1988
American Chemical Society.)
~ :~]
TI4 for the controlled variable in one-point con-
T trol using different control structures for a
t distillation column is obtained by the pa-
rameter IKw/Twl, that is, the absolute value
BO~
kg::~
of the ratio between the gain and the time
L
constant of the disturbance transfer func-
T
1 tion (when expressed by a first-order trans-
fer function as in Table 15-3). The ratio
V
T kg::~ I IKw/Twl has been found to be roughly pro-
portional to the integrated absolute error of
the controlled variable after a step distur-
k:;: !
bance (Sandelin, Hiiggblom, and Waller,
1991a).
B
T
100
'~l 0 50 min 100
The parameter IKw/Twl is shown in Table
15-6 for the two disturbances, the two loops,
and the four structures. The values for the
• •
bottoms loop in the (D, V) structure are in
-+ t parentheses because of the zero process
FIGURE 15-17. Experimental illustration of two- transfer function for this loop (which means
point control with the (D j(L + D), V jB) strategy for that this loop is not useful for one-point
a step change in feed composition from 27 to 33 wt%
at t ... 32 min. [Reprinted by permission from Waller
control). In the (D/(L + D), V) structure
et aL (1988b). Copyright 1988 American Chemical So- the prediction for the bottom loop is inaccu-
ciety.] . rate because of the longer deadtime in the
process transfer function for this loop in
Table 15-3.
Similar difficulties to get the bottom level Neglecting differences in implementation
control loop to function properly have difficulties for one-point control, Table 15-6
been experienced on industrial columns indicates that for disturbances in feed flow
(Shinskey, 1987). As a remedy Shinskey rate the best control of the top loop is
(1990) suggests using V' + B as the level obtained by (L, V) and (D/(L + D), V).
control manipulator. Because V'is not nor-
mally measurable, the manipulator is cho-
sen to be aV + B (see Figure 15-3 for nota- TABLE 15-6 Disturbance Sensitivity Parameter
tion), where a is estimated so that aV ::= V'. IK.. /
T.. I X 10 3
F z
Structure Top Bottom Top Bottom
15-10 COMPARISON OF FOUR Loop Loop Loop Loop
CONVENTIONAL CONTROL
STRUCTURES (L,V) 0.10 33 0.49 69
(D,V) 0.17 (24) 7.8 (150)
In this section the four conventional control (D/(L + D), V) 0.13 (36) 1.5 (76)
structures (L, V), (D, V), (D/(L + D), V), (D/(L + D), VjB) 0.25 1.3 1.8 130
Experimental Comparison of Control Structures 325
O.5r---~--~----~--~--~
!!t~l ·c
:: ~ \f'~L+D)'V/B I
__
::f=:y:
o
-+t
:~
50 min
o
....................................................................................
-0.3~---L----~ __-L~__~__~
o min 50
FIGURE 15·18. Experimental comparison of one·
point control of T14 for three structures. The distur· 90
bances are step changes in feed composition of the
following size: (D/(L + D), V): 27.1 wt% --> 33.0wt%
·c .......--~
Tu
--> 27.1 wt%; (D/(L + D), V/B): 32.1 wt% --> 27.1 ,---""-'
wt% --> 32.1 wt%; (L, V): 27.0 wt% --> 33.0 wt% --> 1 I
:c.::~:."":::::.""\~-::,.)j.."::::. <::::.-:::::::~~;.~:":.-.-.-.-.~.~
27.0wt%.
0, V
85
01 (L+Ol, V
01 (L+Ol, via
For the bottoms loop the best by far is o min 50
obtained by (DI(L + D), VIB).
--> t
For disturbances in feed composition, the
best top loop control correspondingly is FIGURE 15·19. Experimental comparison of four
achieved by (L, V). This structure is also studied conventional structures. At t = 10 min, a step
change from 27 to 33 wt% in feed composition is
the best structure for the bottom loop, introduced. (The measuring device on plate 4 measures
closely followed by (D I(L + D), V). (It temperature differences from a reference temperature;
should be remembered, however, that the thus no recording of the temperature level is obtained.)
transfer functions given in Table 15-3 are [Partly reprinted from Wallet et aI. (1988b), Copyright
quite inaccurate, which is also why the re- 1988 American Chemical Society.]
sults in Table 15-6 are quite inaccurate.)
A rough experimental illustration is given
in Figure 15-18. Considering that the distur- ence, affect the control qualities to such a
bance to the (DI(L + D), VIB) structure degree that the conclusions would be af-
was only 5 wt%, compared to 6 wt% for the fected. The relative gains as calculated from
other two structures shown, the agreement the experimentally obtained process gains
between Figure 15-18 and Table 15-6 is for the studied structures are given in Table
reasonable. 15-5. Of the four strategies, (D I(L + D),
The behavior of the uncontrolled output V) and (L, V) are the ones most suitable to
in one-point control of a two-output distilla- use when coping with disturbances in feed
tion process is extensively discussed in composition at the studied steady state in
Sandelin, Haggblom, and Waller (1991b). our column. The first one was also found
quite insensitive to controller tuning for the
temperature loops.
For disturbances in feed flow rate, on the
15·10-2 Two-Point Control
other hand, the scheme (Dj(L + D), VIB)
An experimental comparison of four con- is probably the one resulting in the best
ventional strategies is shown in Figure 15-19. control quality, as indicated by Figure 15-9.
Although the levels of operation are some- The conventional scheme (L, V), Ryskamp's
what different, this does not, in our experi- scheme (DI(L + D), V), and the two-ratio
326 Practical Distillation Control
92
pIes the two primary loops. The manipula-
tors in DRD are functions of D, L, B, and 90
V, as in Equations 15-3 and 15-4. DRD is
treated in detail by Haggblom and Waller T1488
(1990). t
86
As an illustration, consider a static DRD,
calculated on the basis of reconciled gains 84
of the (L, V) structure and designed so as
to (at steady state) eliminate disturbances in o 50 100 min 150
feed composition and feed flow rate. If the --+t
equations are solved with respect to L and
FIGURE 15-20. "Open-loop" (i.e., 41/1 = 0) re-
V, which are the manipulators with which sponses of DRD structure to a + 6 wt% step change in
the controlled temperatures are affected, feed composition. Structure given by Equations 15-5
the following expressions are obtained and 15-6 (-) and by Equations 15-7 and 15-8 (- - -).
Experimental Comparison of Control Structures 327
Figure 15-20 shows the experimental actual advantages are more indirect: a cer-
"open-loop" responses (i.e., ~1jJ = 0) with tain control is obtained over both products
the ORO structures for both controllers (for the disturbances considered in the de-
given by Equations 15-5, 15-6, and 15-7, sign at the ORO structure) even if one or
15-8 (Nurmi, 1990). The disturbance in Fig- both feedback controllers fail or are turned
ure 15-20 is a + 6 wt% change in feed to manual. Illustration of how indirect two-
composition, whereas it is only a + 3 wt% point control is obtained through one-point
change in Figure 15-8 for four conventional control using ORO is given in Figures 15-21
structures. Especially for the bottom loop, and 15-22, where also the corresponding
the ORO structure very well eliminates the experimental results with the (L, V) struc-
disturbance in the steady state. The devia- ture are shown (Sandelin, Haggblom, and
tion from the desired steady state during Waller, 1991b).
the first 20 min following the step distur- It should be noted that closing one of the
bance was expected. It is caused by the feedback loops may destroy the self-regulat-
process dynamics neglected in the static ing properties of the other output, as illus-
ORO controller used (see Haggblom and trated in Figures 15-21 and 15-22 for the
Waller, 1990, for a more detailed illustra- (L, V) structure (compare with Figure 15-8
tion). where both feedback loops are open). This
In two-point feedback control there property is extensively discussed and illus-
seems to be no direct advantage in using trated in Sande lin, Haggblom, and Waller
ORO compared to standard structures. The (1991b).
0,0
- - L,V
------ DRD --L,V
--------- DRD
-0.5
-0,5
°C °C
93 93
91 91
T14 T14
i 89 i 89
87
85 :---------o'-:-----:---~ 85
o 50 min 100 0 50 min 100
-t -t
FIGURE 15-21. Experimental one-point control of FIGURE 15-22. Experimental one-point control of
the bottom temperature at a step disturbance in the the bottom temperature at a step disturbance in the
feed flow rate from 190 to 210 kg/h. DRD (Equations feed composition from 33 to 27 wt%, DRD (Equations
15-5 and 15-6) and (L, V) structures used. [Reprinted 15-5 and 15-6) and (L, V) structures used, [Reprinted
by permission from Sandelin. Haggblom, and Waller by permission from Sandelin, Haggblom, and Waller
(I991b), Copyright 1988 American Chemical Society,] (I991b). Copyright 1988 American Chemical Society,)
328 Practical Distillation Control
I
used in Figures 15-23, 15-10, and 15-11 are
two SISO PI controllers with fixed parame-
~r' '~E
ters in all cases. More details are given in
Sandelin, et al. (1991).
:~~ I
15-13 SUMMARY AND
CONCLUSIONS
Various ways to control distillation have
1"
been investigated on a 15-plate pilot-plant
column separating ethanol and water. Em-
phasis has been put on two-point feedback
t~[~
control using multiloop SISO PI control.
Various structures have been studied
y:~ 40 0 _ t
1
100 min
respect to control quality and difficulties of
implementation: (L, V), (D, V), (L, B),
(DI(L + D), V), and (DI(L + D), VIB).
By far the most difficult structures to
implement were the two structures (L, B)
FIGURE 15-23. Experimental responses of the out- and (DI(L + D), VIB), and by far the
puts and the inputs to setpoint changes in T14" The worst control quality for the disturbances
control structure is (L, V) and two PI controllers are considered (feed composition and feed flow
used for feedback. The controller parameters have
been calculated by a min-max approach to achieve rate) was obtaine~ by (D, V), a structure
robustness against nonlinearities over operating range that for the studied system lacks integrity
considered. Note the different scaling for .1T4 com- because it has a relative gain of O. No re-
pared to previous figures. The controller settings are sults concerning control quality were ob-
(cf. Table 15-4): top loop: Kc = -125 kg h- 1/oC, tained for (L, B).
T; = 5.8 min; bottom loop: Kc = 5.4 kg h- 1/OC, T;-
2.7 min. [Reprinted with permission from Chemical Some properties of a special disturbance
Engineering Science, Sandelin, et al. (1991). Copyright rejection and decoupling strategy, called
1991 by Pergamon Press.] ORO, are also illustrated. One feature of
Experimental Comparison of Control Structures 329
this strategy is that it approximately gives 1. E. Rijnsdorp et aI., eds. Oxford: Pergamon
indirect two-point control through one-point Press, pp. 259-266.
control for the two disturbances considered. Hiiggblom, K. E. and Waller, K. V. (1986). Rela-
Finally, an illustration is given of a multi- tions between steady-state properties of distil-
lation control system structures. In Dynamics
loop PI controller with fixed parameters,
and Control of Chemical Reactors and Distilla-
tuned by a min-max linear quadratic theory,
tion Columns, C. McGreavy, ed. Oxford: Perg-
which copes with the nonlinearities of the amon Press, pp. 243-247.
system studied when operated over a larger Haggblom, K. E. and Waller, K. V. (1988).
operating range. Transformations and consistency relations of
distillation control structures. AIChE J. 34,
1634-1648.
ACKNOWLEDGMENT Haggblom, K. E. and Waller, K. V. (1989). Pre-
The results reported have been obtained dicting properties of distillation control struc-
tures. Proceedings of the American Control
during a long-range project on multivariable
Conference, Pittsburg, pp. 114-119.
process control supported by the Neste
Haggblom, K. E. and Waller, K. V. (1990). Con-
Foundation, the Academy of Finland, trol structures for disturbance rejection and
Nordisk Industrifond, and Tekes. A number decoupling of distillation. AIChE J. 36,
of people have been involved in this project 1107-1113.
over the years and the results reported stem Haggblom, K. E. and Waller, K. V. (1991). Mod-
from their endeavors: Dan Finnerman, Sten eling of distillation control structures. Report
Gustafsson, Kurt-Erik Hiiggblom, Piiivi 91-4, Process Control Lab., Abo Akademi.
Nurmi, Peter Sandelin, and Kai Wikman. Jacobsen, E. W. and Skogestad, S. (1991). Multi-
Their contributions are gratefully acknowl- ple steady states in ideal two-product distilla-
edged. Some of the results presented have tion. AIChE J. 37, 499-511.
been published previously by the author and McAvoy, T. J. and Weischedel, K. (1981). A
dynamic comparison of material balance ver-
co-workers in American Institute of Chemi-
sus conventional control of distillation
cal Engineers Journal, Industrial and Engi-
columns. Proceedings IFAC World Congress,
neering Chemistry Research and Chemical Kyoto, Japan, pp. 2773-2778.
Engineering Science. They are reprinted here Nurmi, P. (1990). M.Sc. Thesis, Abo Akademi (in
by permission. Swedish).
Rademaker, 0., Rijnsdorp, J. E., and Maar-
References leveld, A. (1975). Dynamics and Control of
Finco, M. V., Luyben, W. L., and Polleck, R. E. Continuous Distillation Units. Amsterdam: El-
(1989). Control of distillation columns with sevier.
low relative volatility. Ind. Eng. Chem. Res. Rijnsdorp, J. E. (1965). Interaction in two-vari-
28,75-83. able control systems for distillation control:
Finnerman, D. and Sandelin, P. (1986). M.Sc. II. Application of theory. Automatica 3,
Theses, Abo Akademi, Abo, Finland (in 29-52.
Swedish). Ryskamp, C. J. (1980). New strategy improves
Gustafsson, S. (1984). Unpublished work, Abo dual composition column control. Hydrocar-
Akademi. bon Process. 59, (6) 51-59.
Haggblom, K. E. (1986). Tech. Lic. Thesis, Abo Sandelin, P. M., Haggblom, K. E., and Waller,
Akademi (in Swedish). K. V. (1991a). A disturbance sensitivity pa-
Haggblom, K. E. (1988). Consistent control rameter and its application to distillation con-
structure modeling with application to distil- trol. Ind. Eng. Chem. Res. 29, 1182-1186.
lation control. Ph.D. Thesis, Abo Akademi. Sandelin, P. M., Haggblom, K. E., and Waller,
Haggblom, K. E. (1989). Reconciliation of pro- K. V. (1991b). Disturbance rejection proper-
cess gains for distillation control structures. ties of control structures at one-point control
In Dynamics and Control of Chemical Reac- of a two-product distillation column. Ind. Eng.
tors, Distillation Columns and Batch Processes, Chem. Res. 29, 1187-1193.
330 Practical Distillation Control
Sandelin, P. M., Toivonen, H. T., Osteras, M. Waller, K. V. (1986). Distillation column control
and Waller, K. V. (1991). Robust multiobjec- structures. In Dynamics and Control of Chem-
tive linear-quadratic control of distillation us- ical Reactors and Distillation Columns, C.
ing low-order controllers. Chem. Eng. Sci. 46, McGreavy, ed. Oxford: Pergamon Press, pp.
2815-2827. 1-10.
Shinskey, F. G. (1984). Distillation Control, 2nd Waller, K. V. (1990). Distillation: Control struc-
ed. New York: McGraw-Hill. tures. In Systems & Control Encyclopedia,
Shinskey, F. G. (1985). Disturbance-rejection ca- Supplementary Volume 1, M. G. Singh, ed.
pabilities of distillation control systems. In Oxford: Pergamon Press, pp. 174-181.
Proceedings of the American Control Confer- Waller, K. V. and Finnerman, D. H. (1987). On
ence Boston, pp. 1072-1077. using sums and differences to control distilla-
Shinskey, F. G. (1987). Personal communication. tion. Chem. Eng. Commun. 56, 253-258.
Shinskey, F. G. (1990). Personal communication. Waller, K. V., Finnerman, D. H., Sandelin,
Skogestad, S. (1988). Disturbance rejection in P. M., and Haggblom, K. E. (1986). On the
distillation columns. Proceedings of Chem- difference between distillation column control
data 88, Gothenberg, Sweden, June 13-15, structures. Report 86-2, Process Control Lab.,
pp. 365-371. Abo Akademi.
Skogestad, S. and Morari, M. (1987). Control Waller, K. V., Haggblom, K. E., Sandelin, P. M.,
configuration selection for distillation col- and Finnerman, D. H. (1988a). Disturbance
umns. AIChE J. 33, 1620-1635. sensitivity of distillation control structures.
Skogestad, S. and Morari, M. (1988). Under- AIChE J. 34, 853-858.
standing the behavior of distillation columns. . Waller, K. V., Finnerman, D. H., Sandelin,
Ind. Eng. Chem. Res. 27, 1848-1862. P. M., Haggblom, K. E., and Gustafsson,
Takamatsu, T., Hashimoto, I., and Hashimoto, S. E. (1988b). An experimental comparison of
Y. (1982). Multivariable control system design four control structures for two-point control
of distillation columns system. Proceedings of of distillation. Ind. Eng. Chem. Res. 27,
PSE, Kyoto, Japan, Technical Session, pp. 624-630.
243-252. Wood and Berry (1973). Chem. Eng. Sci. 28,
Takamatsu, T., Hashimoto, I., and Hashimoto, 1707-1717.
Y. (1984). Dynamic decoupler sensitivity anal- Yang, D. R., Seborg, D. E., and Mellichamp,
ysis and its application in distillation control. D. A. (1991). Combined balance control
Preprints IFAC World Congress, Budapest, structure for distillation columns. Ind. Eng.
Hungary, Vol. III, pp. 98-103. Chem. Res. 29, 2159-2168.
Tsogas, A. and McAvoy, T. J. (1981). Dynamic Yang, D. R., Waller, K. V., Seborg, D. E. and
simulation of a non-linear dual composition Mellichamp, D. A. (1990). Dynamic structural
control scheme. Proceedings of the 2nd World transformations for distillation control con-
Congress of Chemical Engineering, Vol. 5, figurations. AIChE J. 36, 1391-1402.
pp. 365-369.
16
Industrial Experience with
Double Quality Control
Henk Leegwater
DSM
Limit Limit
\V d \'\
() !\" I\.
A'YCrage
I V CTW Average
FIGURE 16·1. The better the control, the closer to maximum allowed impurity level
the unit can be operated, resulting in lower costs.
Edlyleoc
r---T"'""--~
amount of ethane in the ethylene product costs, product values, and the current plant
stream. First it reduces the energy con- operating environment. If, for instance, the
sumption of the column. Second the bottom refrigerating compressors are running with
stream will decrease and the more valuable open recycle valves to prevent surge, addi-
ethylene stream will increase. The value of tional cooling energy is free of charge. In
the top product is much higher: in fact that case the optimal bottom impurity will
ethane is sold as ethylene. decrease.
Bottom Impurity
16-4-2 Optimizing Throughput
Although the bottom stream, mainly the
ethane, is recycled via a furnace where it is In the last column of a butadiene plant,
cracked, the amount of impurity is also im- the product 1,3-butadiene leaves over the
portant for optimizing the column and even top and the C4-raffinate (l,2-butadiene,
the plant operation. Increasing impurities, butenes, etc.) leaves at the bottom. This
primarily ethylene, increases recycle costs column is the bottleneck of the whole plant,
(compression and separation) and reduces not only for the butadiene production itself,
capacity. Most significant perhaps, valuable but also for the C4-raffinate 1, which serves
ethylene is lost because it is partly cracked as the feedstock of the methyl-tertiary butyl
into less valuable products such as hydro- ether (MTBE) plant on the same site. The
gen, methane, propene, and so forth. column will be operated at the maximum
On the other hand, as the ethylene in the hydraulic load (reflux, vapor load). A lower
bottom stream decreases, separation costs reflux-feed ratio (R/F) results in higher
in the C2 splitter go up. So there is an feed: the lower the purities in both ends,
optimum impurity level, which can be calcu- the lower the reflux-feed ratio and the
lated if the separation costs, compression higher the throughput.
costs, and the behavior of ethylene in the Of course the product needs to meet its
ethane furnace are known. See Figure 16-3. specifications regarding impurity level: the
The optimal bottom impurity is not a top needs to be operated as close as possi-
constant; it varies as a function of energy ble to the maximum allowed impurity level.
FIGURE 16-3. Given a certain top impurity, putting more energy (reflux and
reboiler heat) in a column results in a purer bottom stream and a higher top product
recovery. Too much energy, however, is not worth the extra gain of top product; too
little allows the loss of valuable top product to the less valuable bottom stream.
334 Practical Distillation Control
1,3-hated_
C4 - nfl'iIIIfe 1
aoMTBB
So the two targets to optimize the operation afflicted with sample time and deadtime.
of the column are: throughput and bottom The sample time can vary from a few min-
impurity. utes to half an hour. Because of the costs of
To calculate the optimal bottom impurity installation and operation, one single gas
we need to know the values of 1,3-butadiene, chromatograph is sometimes shared to mea-
the bottom stream, C4-raffinate 1, and sure the impurity of several different
the C4 fraction. In the bottom stream the streams. In that case, the sample time can
1,3-butadiene is devalued to a less valuable be an hour or even several hours. The dead-
product. We do not have to consider the time of the measurement consists of two
costs of the reflux and heat input because components: the travel time through the
the column will be operated at its hydraulic sample line from the field to the place where
constraints: the amount of reflux and cool- the gas chromatograph is situated (1 to 5
ing will hardly vary as functions of through- min) plus the analysis time of the gas chro-
put and bottom impurity. See Figure 16-4. matograph itself; after the sample has been
In general, double quality control reduces taken, it takes time before the result of the
the internal streams in columns and there- measurement is available. See Figure 16-5.
fore tends to maximize throughput. In the case of the manually taken sam-
ples, which are hand processed in a labora-
tory, the deadtime can be an hour or more
and the sample time even a day or longer.
16-5 QUALITY MEASUREMENTS
Besides gas chromatographs, infrared-
Normally gas chromatographs are installed based measurement devices are available
to measure top and bottom qualities. Usu- for quality measurements. These are usually
ally the impurities are measured because continuous analyzers: however, a deadtime
their concentrations are lower and there- (travel time) is added to the control loop
fore the accuracy of the measurement is dynamics. Usually they are not sensitive
higher. Due to the nature of analyses by gas enough for low impurity measurement
chromatographs, these measurements are ranges and are used only as a final measure-
noncontinuous sampled signals: they are ment in separations that are not high purity.
Industrial Experience with Double Quality Control 335
.. t ..
'clood
'......
(a)
Deacltime
(trawl time)
(b)
D
xD
they tend to fight each other. The purer the work. Dependent on the disturbances, large
product streams, the higher the interaction deviations for the weak controlled quality
and the more severe the effects. If, for in- can be expected. Often this scheme will
stance, the top impurity is too high, the top need operator assistance if there are distur-
quality controller will increase the amount bances; at best it gives marginal perfor-
of reflux. However, more reflux also causes mance.
a more impure bottom, resulting in more
reboiler heat input via the bottom quality 16-7 HEAT AND MATERIAL
controller, which causes a less pure top,
BALANCE IN RELATION
which results again in more reflux, and so
TO SEPARATION
on. In many distillation systems this interac-
tion can cause flooding. See Figure 16-9 for In fact the separation of a column depends
the traditional double quality scheme. on the ratio between the down-going liquid
Only if one of the two loops is tuned slow stream and the up-going vapor stream. Rec-
compared to the other might this scheme tification or separation in the top section is
80 ton/h 76 ton/h
100 ton/h
80%light
20% heavy 180 ton/h 276 ton/h
20 ton/h 24 ton/h
FIGURE 16·10. An improperly balanced reflux and bottom heat causes a high
impurity in the bottom.
338 Practical Distillation Control
improved by a higher L /V ratio (more re- balance point of view it can be seen that the
flux), whereas stripping or separation in the separation will prejudice one of the ends as
bottom section is· improved by a lower L /V pointed out in Section 16-7. In fact, for a
ratio (more heat input). Both internal flows given feed, the net heat input needs to be
act on both flows leaving the column: the constant:
heat and mass balance are coupled. Thus Qnet = RllHR + FrebllHreb
increasing the amount of reflux to increase
the top purity must also be followed by an
where R = reflux (ton/h)
increase in heat input to compensate for the
llHR = difference of heat contents be-
higher amount of liquid to the bottom.
tween top flow and reflux
From the material balance one can con-
(kJ/ton)
clude also that an imbalance between these
Freb = reboiler flow (ton/h)
streams causes a unnecessarily high impu-
llHreb = difference of heat contents be-
rity in one of the two ends, as seen in
tween reboiler flow inlet and
Figure 16-10.
outlet (kJ /ton)
Figure 16-10 shows two situations for a
binary distillation column. In both cases the To compensate for feed flow variations the
feed, which is at dew point and thus just Qnet is normalized by the feed flow itself:
vaporized, contains 80 ton/h light compo- Qnet/F .
nent and 20 ton/h heavy component. On The background of this factor can also be
the left the reflux-feed ratio is R/F = 2; clarified by the following example. Consider
on the right R/F = 3. In the right-hand a vapor feed at its dew point. Normalizing
case there is an imbalance between the Qnet/F again by dividing it by the heat
amount of reflux and the heat input: the needed to condense a ton of feed, the factor
amount of flow out of the bottom is even Qnet/(F llHF) can be considered as the
more than the heavy component in the feed. fraction of the feed that is condensed and
The surplus of 4 ton/h must be light com- leaves the column at the bottom. That frac-
ponent, causing an impurity of at least 20%. tion is about the same as the fraction of
The effect will be less pronounced in a real heavy component in the feed. In case of
column, but the thoughts behind it remain subcooled feed, llHF is the heat needed to
true. vaporize a ton of feed, and the factor can be
We can conclude that after a major considered as the amount of feed that needs
change in reflux, the bottom heat input also to be vaporized to form the top product,
needs to be adapted, and vice versa. which is about the same as the fraction of
light component in the feed.
16-8 NET HEAT INPUT An operator gave the following metaphor.
Increasing R/F while holding Qnet/F con-
In the case of normal separations the mag- stant presses the concentration profile like
nitudes of the product streams are broadly the bellows of an accordion. Increasing
the amount of heavy component in the feed Qnet/F while holding R/F constant is mov-
for the bottom stream and the amount of ing the concentration profile upward like
light component for the top stream. There- moving the accordion as a whole without
fore the heat input and the amount of reflux changing the form of the bellow. See Figure
need to be in balance with each other: for 16-11.
example, the heat input must be adjusted
after a change in the reflux in such a way
16-9 SEPARATION INDICATORS
that the amount of product flow out of the
bottom (and thus also over the top) will To understand separation and Qnet better,
remain the same. Otherwise, from a mass two separation factors are introduced. The
Industrial Experience with Double Quality Control 339
,I
I
t t ,I
I
t t
XJiPt R/F > R/F al1llt/F < al1l t /F
a"", = CClIIItaDt RIF = cmstant
philosophy behind the control scheme was tom, top, or divided equally. The lower this
derived from the inverse Nyquist array number is, the more the concentration pro-
method: combinations of measured vari- file is pushed to the lower end (the bottom)
ables give new variables; the same is true of the column. This means that the top is
for manipulated variables. more in favor than the bottom, regarding
the impurities. The separation accent indi-
cator can be compared with the cutpoint:
16-9-1 Separation Performance
Indicator
SAl = X D - cXB
The separation performance indicator (SPI)
indicates the degree of total separation of
the column: the lower this number is, the The separation accent is related to QnetlF.
lower is the "total" impurity and the better
is the separation: 16-9-3 Interpretation
The constant c accounts for different scales
SPI = X D + cXB of top and bottom measurements and the
design of the column, especially the location
of the feed tray in relation to the feed
The separation performance is related to composition. By decreasing the reboiler heat
the amount of energy put into the column. input a little, a column has the following
steady-state conditions:
16-9-2 Separation Accent Indicator
XD = 400 ppm, XB = 0.5%, R/F = 2.5
The separation accent (SAl) indicates where
the accent of the separation is placed: bot- X D = 300 ppm, XB = 0.9%, R/F = 2.5
340 Practical Distillation Control
11
o~ ____Afo
~ , slopeo = dXoldt·
16-10-2 Column Simulation
t\;---'------
A dynamic mathematical model was espe-
cially made for developing a new control
t
Xs AX
LS slopes = dXsldt
strategy for both cases. Due to the high
purity separation, the behavior of the col-
umn is highly nonlinear in both a static and
AXo slopeo a dynamic sense. Therefore a rigorous non-
c = AXs or, c = --
slopes linear model is needed. It consists of dy-
namic tray-to-tray simulations based on dif-
DGURE 16-12. Two possibilities for the determina- ferential equations for conservation of mass,
tion of the factor c in the separation indicators, based
on concentration response to a change in reflux or mass of individual components, and energy.
reboiler heat. Algebraic equations are needed to describe
Industrial Experience with Double Quality Control 341
the equilibrium of ethylene and ethane, en- returned to manual mode. When the opera-
thalpies, hold-ups, and so forth. The model tors have a good understanding of the con-
was run interactively on a mainframe com- trol scheme, they will trust it, know its limi-
puter and was time-scaled faster than real tations, and restore it to automatic mode
time because of the slow responses of the after a major disturbance. If not, controllers
actual column. The reliability of this model will be put in manual and remain in manual.
was checked by process operators who nor- A lecture of several hours for every shift
mally operate this column in the plant. It was presented in the control room during
appeared to be a "hifi" model. the more quiet evening and night shifts,
sometimes together with the dynamic simu-
lation. The operators could then apply dis-
16-11 INTRODUCTION OF THE
turbances to the column simulator for bet-
NEW CONTROL SCHEMES INTO ter understanding of the new control scheme
INDUSTRIAL PRACTICE and gain confidence in it.
Much attention has been paid to the intro-
duction of the new control schemes to the
16-12 C2 SPLITTER WITH HEAT
operators. A good understanding of a con-
INTEGRATION
trol scheme is vital for its automatic opera-
tion for the long run. No control scheme The C2 splitter in DSM's naphtha cracker
can be in automatic for 100% of the number 3 has 120 trays. The column has a
time. Instruments can fail or need to be combined reboiler / condenser. The plant
calibrated. To cope with very severe dis- has an open-type ethylene refrigerating sys-
turbances, not necessarily related to the tem, which means that product ethylene
column itself, control must sometimes be itself flows through the refrigerating system
Fv
and acts at the same time as low tempera- tom concentration was controlled by the
ture (-110 to - 400 C) coolant. The eth- main reflux. After a change in the amount
ylene is condensed in the reboiler to pro- of feed, which is not measured directly but
vide the bottom heat input and selVes then can be derived from the upstream load, the
as reflux. One trim stream (Fz) is provided operators changed both flows manually.
to give more bottom heat than the condens- There appeared to be no consensus among
ing reflux alone, and one stream (Fl ) to operators about the influence of the main
provide more reflux than the main reflux reflux. Because the main reflux not only acts
itself. See Figure 16-13 for a visual repre- as reflux, but also as bottom heat, the final
sentation. Unfortunately the operators influence was not clear to all. Probable rea-
called the two trim streams bottom heat and sons for lack of consensus are the large time
auxiliary reflux. In fact, the two streams act constant in the order of duration of an
in the same way, except for sign: 1 ton/h operator shift (8 h) and the inverse re-
more bottom heat extracts 1 ton/h from the sponse: After an increase of the main re-
main reflux, while the amount of heat input flux, the bottom starts to become more pure
remains constant. The same is true for 1 because the influence of bottom heat domi-
ton/h less auxiliary reflux. For better un- nates first; at the end, the bottom will be
derstanding the bottom heat stream was less pure. Calculations and plant obselVa-
renamed "negative-auxiliary reflux." tions show that an extra 1 ton/h main re-
flux gives, after flashing, an extra internal
reflux of 0.96 ton/h and, due to the differ-
16-12·1 Basic Controls
ences in heat of condensation of the ethyl-
Pressure is controlled by the suction pres- ene reflux and the heat of evaporation of
sure controller of the third stage of the ethane in the bottom, an extra up-going
ethylene refrigerating compressor. The re- vapor flow of 0.91 ton/h. From the perspec-
flux/ condenser vessel is normally full due tive of a mass balance it can be seen that
to flooded condenser operation. The col- the bottom flow at the end will be 0.05
umn bottom level is controlled by regulating ton/h higher. If the bottom flow is 10 ton/h
the outgoing liquid stream. Two flow con- this means an increase of at least 0.5% in
trollers are provided: one main reflux flow impurity.
and one that is switchable to auxiliary reflux
and bottom heat (negative-auxiliary reflux). 16-12-4 Improved Control Scheme
Compared with other columns, through its
16-12·2 Quality Measurements
nature, this column is relatively stable: A
Two gas chromatograph are installed to change in main reflux automatically causes
measure top and bottom quality. The sam- more bottom heat, which is necessary for
ple time is 12.5 min and the total dead time mass balance reasons, as pointed out in a
is 16 min (2.5 min transport time through previous section. Although the feed flow is
the sample line from the field to the gas not measured, the top flow immediately
chromatograph building plus 12.5 min for shows the feed flow variations because the
analyzing). Two temperature measurements feed enters the column at its dew point:
provide the operator continuous early warn-
ing about changes in concentration.
~op = F + Vbottom
Vbottom = aR
16-12-3 Previous Control Scheme where
According to the instructions the top con- heat of vaporization 1 kg ethane
centration was controlled by the auxiliary
a=------------
heat of condensation 1 kg ethylene
reflux/bottom heat controller and the bot- =0.91
Industrial Experience with Double Quality Control 343
The factor a depends on the discharge pres- perature could be used for measuring the
sure of the compressor (the higher the pres- position of the concentration profile in the
sure, the lower the heat of condensation): column. The temperature at tray 14, which
is controlled by the auxiliary reflux, ap-
-R- 1 /( a+--
~op
1)
RIF
peared to be the most suitable. In fact it
acts as Qnet control. There was no need to
make the scheme more complicated by us-
ing internal reflux control, feed-auxiliary
By controlling the main reflux-top vapor
flow ratio, the main reflux-feed flow is in- reflux ratio, separator indicators, and so
herently controlled, which is necessary to forth.
cope with feed flow variations.
The flow from the lowest tray to the 16·12·5 Condenser - Reboller
bottom compartment is Level Control
Xo Top
sp + R/V,op
quality Main reflux R
control
mv control
-
X8 Bottom Temp.
sp +
quality ~ )- tray 14 Auxiliary reflux r
- control control
mv
T14
nGURE 16-14. Improved control scheme.
344 Practical Distillation Control
Fv xI'
FIGURE 16-16. Discharge pressure control of the third stage of the ethylene
compressor, preventing a too low pressure, which results in breakthrough of gas and
an excessive pressure, which spoils compressor energy.
Industrial Experience with Double Quality Control 345
l00c.
Reflux
- '-
hrVV ~
".
vu.
18
" ~vv
1\ A - .r "'" -
if
bar
~
AI'
17
~ V"\
- ~
R:
7S
-- . / '~ I'
~
tao/b
70
JJ 1 _\ LiL
/-
.-
/
~ / V \)
~
'~
/ h. /'\ I~ V
\J'-J
x.
0.,
/
J
vol" 6Marcb 1983
o
l00minutca
FIGURE 16-17. Influence after an imbalance between heat input and reflux, caused
by discharge pressure too low to condense the reflux. The fluctuating level in the
reflux-condenser vessel indicates this imbalance.
346 Practical Distillation Control
between reboiler heat and reflux (i.e., in compensated for pressure variations: a AP
Qnet). This causes a major upset in the measurement is in fact a volume flow mea-
column. Therefore, this setpoint needs to be surement. If the pressure increases, the
as close to 100% as possible. Severe upset same AP will be maintained by the reflux
occurred when the discharge pressure of the flow controller, resulting in a greater mass
third stage of the compressor became too flow. This increase was not detected by the
low to condense the requested reflux and flow controller, but it is seen by the ~oP
the setpoint of the level controller was 50%. measurement; therefore, the effect was
See Figure 16-17. magnified by the R/~op controller. The
increase in discharge pressure from 17.5 to
Reflux Measurement Not Properly 18.2 atm causes a direct increase of
Pressure CompellSllted ";(18.2 + 1)/(17.5 + 1) = 1.02 or 2%.
After an increase in discharge pressure the Thus ~op will increase. To keep an R/~op
reflux flow started to rise, which resulted in ratio of 0.69, the reflux will be increased by
a purer top stream. After the controller put the ratio controller:
the concentration back to its setpoint, the
reflux was 1.5 ton/h less than before. See Rnew 1 - 0.69
Figure 16-18. What occurred? Because the --- = -----r===============
ROld 1 - 0.69";(18.2 + 1)/(17.5 + 1)
reflux flow measurement takes place in the
vapor phase, this measurement needs to be = 1.05
P, discharge pressure
18,5 '
bar
,. v :r
P3 = 18.4 atm
17
~J P3 = 17,5 1atm L ........
auxiliary-reflux r
,~
3
ton/h
2
r = 2,0 Ton/h F = 31,0 ton/h r = 2,2 t~n/h
....
---- - - -1'-1-
Feed ton/h
35
30
75
R = 73,0 ton/h
Ii\. \ tonC-
Reflux
65
-
ton/h/
,.
r
R = 68/ ton/h
1\'- M ./~~ ~
'/
R - 67.0 105
r\
95
/ ~V
VtoP
300
'J
,5%
.......- I--""'\. XB
o
12 March 1983
FlGURE 16·18. Reflux measurement not properly pressure compensated,
After an increase in discharge pressure the same A.P of the flow transmitter
results in more tons per hour, which will be amplified by the Vtop/reflux ratio
controller, The quality controller ultimately corrects this increase,
Industrial Experience with Double Quality Control 347
Xo ml!
F
sp
+ -
+ R/F Reflux
+ control
sp +
xB mv
(a)
Xo mv
Feed F;.reb
+ -
sp---.~----~
+, IJ.T Qnct
Frob
control control
sp + - +
T.''P Thc.lt.sec.
(b)
FIGURE 16-11. (a) Separation performance control. (b) Separation accent control.
Industrial Experience with Double Quality Control 349
input via the intermediate reboiler is taken ence with the top temperature was used. So
into account. Because of the interaction, the Q net/F became the master controller
two separation indicators were used to partly and the Ii temperature became the slave
decouple the influence of reflux and re- controller. See Figure 16-21b.
boiler heat.
Simulations of the column showed fur-
16-13-6 Experience.
ther that much better control could be
achieved if information about changing con- The effect of the new control scheme was
centrations was available sooner. This was evaluated by collecting plant data. The stan-
especially true for Qnet/F control loop. Due dard deviation of the impurities was re-
to the high purity of the column, tempera- duced by a factor of 7 and energy consump-
tures in the bottom and the top hardly tion of the column was reduced by 5%. In
changed with concentration variations. A addition, the intermediate reboiler could be
temperature measurement somewhere in operated for optimizing the refrigerating
the bottom section appeared to be a good system without introducing disturbances in
indication for a moving concentration pro- the column.
file. To make this indication independent of After six years of experience this control
pressure variations, the temperature differ- approach has proven to be robust and is
o 100 200 300 400 500 0.00 0.10 0.20 0.30 0.40 0.50 0.60
Top impurity in ppm ethane Bottom impurity in % ethylene
0.5%
-
X bottom
500 ppm ./"'.... 0% A
r-
15 o
840517
- 8
X ,OP
16
o ppm
o
840518
8
..../ ~
16 o
840519
8 16
,....--/
o
I\.....r
840520
8
(a)
FIGURE 16-22. (a) Real plant data: previous control scheme. (b) Real plant data:
new control scheme.
350 Practical Distillation Control
o 100 200 300 400 500 0.00 0.10 0.20 0.30 0.40 0.50 0.60
Top impurity in ppm ethane Bottom impurity in % ethylene
0.5%
- /'-
X bottom
~
500 ppm 0%
X 10P
.-
o ppm
20 4 12 20 4 12 20 4 12 20 4 12
850706 85707 850708 850709
(b)
well accepted by the operators. Outside of ing, it is important to have a good insight
periods of instrument maintenance, the into the dynamic behavior of the column
control scheme operates on automatic. See itself.
Figure 16-22 for the plant data. Thinking in terms of Qnet/F and the
indicators for separation performance and
separation accent appears to be very useful.
Sometimes they can be applied literally and
16·14 CONCLUSION sometimes they serve as a way of approach-
A cookbook approach will never work for ing and solving the double quality control
solving control problems in general, and problem. Mter our success with these C2
particularly not for more complicated con- splitters, control systems for other columns
trol problems like double quality control of were designed using this approach. In some
distillation columns. No two columns are cases the use of the separation indicators
the same, and even when they are said to be was not needed: the top impurity was con-
the same, the operating conditions, includ- trolled by R/F and the bottom impurity by
ing disturbances affecting the column, are Qnet/F. For operators the approach ap-
not identical. In addition to the knowledge peared to be very well understandable and
of the Gm)possibilities of control engineer- thus well accepted.
17
Control of Distillation
Columns via Distributed
Control Systems
T. L. Tolliver
Monsanto Company
351
352 Practical Distillation Control
trol (Tolliver and Waggoner, 1980). Now, itoring is desirable to provide preventive
feedforward control is routinely applied. maintenance, and on-line fault tree analysis
is also needed to reduce troubleshooting
effort and downtime when problems do oc-
17-1-3 Future Trends
cur.
In spite of this seemingly glowing endorse-
ment of DCS instrumentation, there remain
17·2 CASE I
several areas for improvement, which are
being addressed by major instrumentation This first example of distillation column
manufacturers. One trend is the extension control, implemented via DCS instrumenta-
of the data highway (digital signals on a tion, will be referred to as Column A. An
coaxial cable) out of the control room and overview of the process material balance
into the field. In addition to reduced wiring and column control strategy is provided by
costs, perhaps the main benefit of this ap- Figure 17-1. The process background and
proach will be easier access to additional considerations leading to this particular
control loop information from smart trans- control scheme design, implementation de-
mitters and smart valves. tails, and results are discussed in the sub-
Smart transmitters are transmitters that sections that follow.
perform self-diagnostics, compensate them-
selves for changes in ambient conditions,
17-2-1 Process Background -
and can be automatically recalibrated with
Debottlenecking
varying ranges. In general, they provide
more accurate and reliable data. Likewise, Column A was an eXlstmg column, being
smart valves will also provide self-diagnos- operated via DCS instrumentation installed
tics and true valve positioning and indica- on a prior modernization project. At the
tion. Valve positioning problems have been time the DCS hardware was installed, it was
reported as the primary cause of poor loop believed that a subsequent debottlenecking
performance in basic control loops (Ender, project would soon replace this column;
1990). True valve position, rather than the Therefore, the existing analog control strat-
implied position based on controller output, egy, without improvement, was imple-
will certainly help in control loop trou- mented for the short period of interim
bleshooting. operation. Control of this column was nor-
Other areas of improvement will address mally stable, although somewhat oscillatory,
the ever-growing data base of process infor- and during larger than normal disturbances,
mation and the need for smart alarms, data the column control required intervention
reconciliation, and troubleshooting diagnos- and considerable operator attention.
tics. All of these areas require improved As the debottlenecking project was being
data management methods as the amount developed, an alternative process design was
of process data increases. Critical operating proposed, which made modifications to the
information needs to be brought to the op- existing column rather than replacing it. It
erators attention on an exception basis was determined that the stripping section of
through the use of smart alarms. The exten- the existing column was limiting capacity,
sive process data must be verified by data primarily because a considerable amount of
reconciliation methods before subsequent sensible heat was having to be provided by
integrated business systems propagate er- the reboiler. The alternative design there-
rors. Maintenance of sophisticated control fore provided another source of heat, by
strategies becomes much more complex with partially vaporizing the feed with a new
the increased number of field transmitter feed preheater, and increased capacity by
signals involved. Therefore continuous mon- modifying the trays below the feed for oper-
354 Practical Distillation Control
·
••
·
.$
,.............................................................
.
-..•
•
~•
••
·
~.-....~...- s
11.7
....~.~
................................::OCllUMl
•••••••••••••• ••••••••••••••••••••• ROOD
} I'OIfI' • -E}....
•••••• •••••••••••••••••••• •••••• CM.CUIA1IOII
·:
·
•
.---&...........................,
..
~
.: :
1110
r····.f§/I·········
t.......... CIIOI8
•
~~--~--- ~
FIIIIG
CONnIDI.
,..........0-.
··
··:
.......... -_ ........-....... ................•..
~ .....,
..........! .
IU
DGURE 17·1. Column A.
ation at the resulting new liquid and vapor rates, or that the column could be satisfac-
rates. The modified column would then be torily controlled near its flood point. How-
able to support a 50% capacity increase. ever, the proposed column design was fi-
The proposed design required the col- nally accepted, due to its large reduction in
umn to operate close to its flood point al- capital requirements, and a detailed control
most all the time. There was much skepti- review was recommended as part of the
cism among those with prior experience debottlenecking project. It is doubtful that
operating the existing column that this new this aggressive design would have been ap-
design would really handle the proposed proved without a distributed control system
Control of Distillation Columns via Distributed Control Systems 355
available to handle the difficult control re- as a decanter for separating the more dense
quirements. and immiscible component A from compo-
nent B. This decanter also represents the
only process surge capacity for component
17-2-2 Control Scheme DesIgn
B; therefore, loose level control is generally
Column A represents a truly multicompo- desirable. With these considerations, in ad-
nent distillation with nearly 25 components dition to the loose overhead composition
present in significant quantities in the feed. specification and the relatively low reflux
However, from a control perspective, the ratio, it is logical to establish a controlled
feed can be considered to contain the fol- flow of reflux for optimum column separa-
lowing four lumped components: tion, and to adjust the distillate flow for
decanter level control.
3% component A, lighter than the Lumped components C and D, together
light key consist of more than 15 individual compo-
12% component B, the light key nents, and product D actually becomes
three products with light, medium, and
70% component C, the heavy key
heavy boiler distributions, Hence, this pro-
15% component D, heavier than the cess includes a refining train consisting of
heavy key three additional distillation columns and
other processing equipment fed by the bot-
where components A, B, and C are recy- tom stream of Column A. There are no
cled back to the reaction area of the process intermediate storage tanks in this refining
and component D is the final product. Col- train; therefore, it is highly desirable to
umn A makes the split between the keys, isolate those operations from transients due
taking components A and B overhead with to disturbances in the front end of the pro-
approximately 12 to 15% component C, cess. For this reason, the bottom stream of
whereas components C and D are taken Column A is placed on flow control and its
out the bottom with less than 0.05% compo- setpoint establishes production rate for the
nent B. All compositions and specifications refining process.
are expressed in weight percent. The overall Column A could have been considered
column material balance is given in Figure part of the refining train, and then its feed
17-1 on the basis of 100 Ib/h of feed flow. would have been flow controlled in a more
The bottom specification is considered to typical column control strategy. However,
be the primary control objective because an analysis of the types of disturbances indi-
increases in component B leaving with the cated that the most frequent and severe
bottom stream represent an unacceptable upset to Column A is a variation in the
yield loss. The overhead specification im- amount of component B contained in the
pacts the amount of recycle material, which feed. If the feed to Column A were flow
affects operating costs; however, the operat- controlled, then any such large variation in
ing cost curve is relatively flat near the component B would cause a significant dis-
overhead specification. Although transients turbance to the bottom flow. However, with
in the overhead composition are acceptable, flow control of the bottom stream, the feed
a permanent increase in recycle material is flow to Column A is adjusted as necessary
undesirable because it eventually impacts to satisfy the column material balance, as
column capacity, returning to the column as indicated by the sump level, and the refin-
additional feed. ing rate is thus maintained.
The control scheme arrangement of the The new feed preheater, which partially
column overhead is somewhat dictated by vaporizes the feed, was originally viewed by
the use of the large horizontal reflux drum manufacturing as a potential operating
356 Practical Distillation Control
problem. However, from a control perspec- sible heat requirements under the current
tive, it gave a new opportunity to react to operating conditions.
the most frequent disturbance, feed varia- The sensible heat required is computed
tions in component B, before they affected as
the column. By applying temperature con-
Qs = BCp(Tc - Tb ) (17-1)
trol to the feed leaving the preheater, as the
amount of component B varies, so does the where Qs = sensible heat duty (MBtu/h)
amount flashed immediately upward. This B = bottom flow rate (klb /h)
control strategy maintains a more constant Cp = specific heat constant
flow of the liquid entering the stripping (MBtu/klb- °F)
section. Tc = control tray temperature eF)
The key components in Column A have a Tb = bottom temperature eF)
high relative volatility and therefore sepa-
rate rather easily. As is often the case with At startup, it was found that the measured
easy separations, the boiling point tempera- temperature on the control tray had tran-
ture varies considerably, in this case over sients that significantly affected the sensible
300° F, from top to bottom of Column A. heat calculation. With the control scheme
The majority of the temperature increase shown in Figure 17-1, column temperature
occurs below the feed, in the stripping sec- is controlled by manipulating the feed tem-
tion of the column. The feed preheater is perature setpoint, which in tum adjusts the
able to use steam; however, the reboiler steam flow to the feed preheater. The intent
requires a gas-fired heater due to the higher of this arrangement is to allow the vapor
process temperature. Because the majority rate in the stripping section to remain con-
of the feed leaves with the bottom flow from stant. However, as the temperature on the
the column, a significant portion of the re- control tray would change, the sensible heat
boiler duty is still providing sensible heat. calculation would also respond to the tran-
In order to operate with the stripping sient. The solution to this problem was to
section near its flood point, it is necessary to use the column temperature controller set-
keep the vapor rate constant. Often differ- point, rather than the measured tempera-
ential pressure can be used as a measure of ture.
vapor rate, and for this reason, a pressure After the sensible heat requirement is
sensor was installed just below the feed calculated, a latent heat determined in pro-
tray. However, a reliable measurement could portion to the bottom flow rate is added,
never be demonstrated on Column A. The and then the sum is converted to a gas-firing
constant vapor rate could be maintained rate. The latent heat used has upper and
only if the amount of latent heat provided lower limits that correspond to vapor rates
by the reboiler was held constant. The gas- at the flood and weep points of trays in the
firing rate of the reboiler is thus determined stripping section. The resulting firing rate
by the sum of the sensible heat required by can be manually biased before being used
the external column energy balance and the as a setpoint to the firing control system.
latent heat necessary to provide a constant The manual bias is implemented as a
internal boilup. trim to the computed setpoint in terms of
The sensible heat requirements vary con- percent of the full scale firing rate. At a
siderably as the distribution of components 100% setting, the computed value is used
lumped as components C and D change. with a 0% bias. At 101% or 99%, a ± 1%
For example, the bottom temperature of bias is used. The bias station indicates both
Column A can vary as much as 50°F with the percentage used and the percentage re-
different feed stocks. For this reason, it is quested. This allows the operator to manu-
necessary to measure and compute the sen- ally balance the firing rate before putting
Control of Distillation Columns via Distributed Control Systems 357
the firing system in remote setpoint. Follow- a PID controller with tight setpoint re-
ing startup, the flood calculation was sponse. Later, in order to get the best re-
checked periodically by raising this bias un- sponse for the column temperature
til incipient flooding was observed. controller, the feed temperature loop was
This latter capability proved quite helpful retuned with mostly proportional action,
in gaining the confidence and support of the slow reset, and no derivative. Feedforward
operators. In spite of extensive training with control based on the feed rate provides
regard to the sensible and latent heat as- most of the action required to handle load
pects of the reboiler duty, the operators disturbances, minimizing the need for reset.
often felt that the flood point calculation Also, offset from setpoint is not critical to
was unduly limiting the firing rate. They overall performance because the feed tem-
observed operation under some conditions perature setpoint is being adjusted by the
where the computed reboiler firing rate was column temperature controller.
as much as 30% less than at other times. This triple cascade control1er arrange-
However, every time they attempted to bias ment with two temperatures and a flow was
up the firing rate, they experienced incipi- an operational concern prior to startup.
ent flooding, usually within 3% of the com- However, the feed temperature controller
puted value. proved its value in rejecting disturbances
during startup, when the column is typically
run in a manual mode prior to putting the
17·2·3 Implementation Details
column temperature controller on auto-
During the control study, it was determined matic. During startup conditions, the col-
that the reboiler gas flow had been previ- umn now runs more stably, and manual
ously cycling due to almost 8% hysteresis in adjustments are required much less often
the gas flow control valve. These cycles were than with previous operation.
significant enough to have been causing the The column temperature control loop
oscillatory column temperature experienced maintains a characteristic column tempera-
prior to the debottlenecking project, and ture used to infer the column's separation
which would be unacceptable with post pro- objectives. The prior sensor location was
ject operation near the flood point of the verified using a procedure that computes
column. Installation of a valve positioner parametric cases from a steady state mate-
has since allowed the gas flow to be con- rial and energy balance design program
trolled essentially at setpoint. (Tolliver and McCune, 1980). With steady
The combustion control system on the state temperature profiles at slightly greater
reboiler was also upgraded with a cross limit and less distillate-feed ratio than the design
firing control scheme using an oxygen con- case, the temperature sensor is located at
troller to trim the air flow setpoint. This the stage that exhibits the greatest, most
latter upgrade was necessary to achieve the symmetrical, steady state temperature devi-
maximum heat duty required at the highest ation from the base case. Results of these
rates and temperatures. Prior operating parametric studies are presented in Figures
practice using excessive air flow, in addition 17-2 and 17-3 for the two alternate feed
to being energy inefficient, would have lim- stocks used by this process.
ited the heat duty available to less than that The existing temperature sensor location,
required. on a column tray equivalent to stage 4, was
The feed temperature control loop uses found acceptable for both feed stocks. How-
feedforward control to set the steam flow in ever, the resulting process gain in this tem-
proportion to feed rate with feedback trim perature control loop was significantly dif-
from the feed temperature controller. The ferent between these two feed cases. Hence,
feed temperature loop was at first tuned as it was found necessary to apply alternative
358 Practical Distillation Control
15 15
14 14
13 13
12 12
\
11 \
,,
11
\
10 10
.......
9 ID 9
ID .0 "-
.0
E 8
E
~
c 8 "- ....... \
~
c Q)
"-
\
Q)
7 Feed j 7 Feed
~
en 6
fJ)
6 "\
\
5 5 \
4 Control 4 Control \
\
3 3 \
\
2 2
0 0
200 300 400 500 600 200 300 400 500 600
Temperature (OF) Temperature (OF)
FlGURE 17·2. Column A temperature profile. FlGURE 17·3. Column A alternate feed tempera-
ture profile.
PID tuning, dependent upon the feed stock turbances occur, the operators are less likely
being used. With the prior control scheme, to manually intervene in time.
this loop was tuned to provide control un- After observing three of four of these
der the high process gain shown in Figure events occurring over a six month period, it
17-3; however, the loop then performed became apparent that a major disturbance
poorly under the lower process gain evi- could allow the column temperature to move
dent with the feed in Figure 17-2. The far enough above setpoint that the use of
DeS instrumentation allowed these tuning the temperature setpoint in the flood calcu-
changes to be automated and thus transpar- lations becomes a poor assumption. Flood-
ent to the operator. ing can occur because the flood calculation
Although the performance of the column predicts more sensible heat than is actually
temperature control loop was generally being used during the transient, and hence
much better than before the project, there temporarily allows too much reboiler duty.
were still infrequent disturbances of a mag- For a temperature decrease, the assumption
nitude such that the column would enter presents no problem because it errs in the
flooding conditions if additional corrective conservative direction.
actions were not immediately taken. How- The solution to these infrequent events
ever, because the column runs so much bet- was to implement a high, positive tempera-
ter, the operators now pay less attention to ture deviation interlock. When this inter-
it, so when these infrequent but major dis- lock condition occurs, the reboiler duty is
Control of Distillation Columns via Distributed Control Systems 359
immediately decreased by 5%, and as the column and two single-effect evaporators in
condition clears, the reboiler duty is slowly series. Also, a large intermediate storage
ramped back to the computed firing rate. tank was eliminated between this column
An alarm is used to annunciate that the and the distillation column that follows it.
interlock has occurred, but further operator Column B is now the first column in a three
intervention has not been required. column refining process.
The project to install this column also
17-2-4 Results established inventory management proce-
A direct comparison of column control per- dures aimed at minimizing in-process inven-
formance before and after the project is not tories. As part of those procedures, feed
completely valid because of the process and from the crude area now routinely bypasses
equipment modifications. However, it is ob- the crude storage tank and is fed directly to
vious that in addition to a 50% capacity Column B. The crude tank is still used
increase, tighter control with less operator during start-ups, shutdowns, and to balance
attention is also being achieved. It is also operating rates between the crude area and
reasonable to assume that this degree of refining, but generally it operates with less
complexity could not have been imple- than a 10% level.
mented and maintained with an analog sys- A second area of inventory management
tem. addresses the recovery of material, dilute
Historical trends provided excellent trou- with water, from the periodic washout of
bleshooting tools as performance was fine several batch operations throughout the
tuned and the infrequent mishaps were ad- process. The material from many of these
dressed. The flexibility of the DCS software sources is collected in a recovered material
allowed the implementation to evolve as storage tank. This recovered material is then
operating experience was gained. Early suc- processed along with the feed to Column B.
cess gained operator confidence and sup- A third area of inventory management is
port, which in turn, lead to their taking an the reduction of material processed by a
active role in troubleshooting and suggest- batch recovery column, fed from several
ing several of the subsequent modifications sources of high boiling materials. This re-
that were implemented. duction is accomplished by a recycle feed of
high boiling material from one of the more
17-3 CASE II dilute sources to the reboiler of Column B.
This recycle feed flashes some of the more
This second example of distillation column volatile product into the reboiler vapor and
control, implemented via DCS instrumenta- concentrates the high boiling impurities
tion, will be referred to as Column B. An along with the bottom stream of Column B,
overview of the process material balance which then is fed to the batch recovery
and column control strategy is provided by column.
Figure 17-4. The process background and
considerations leading to this particular
control scheme design, implementation de-
17-3-2 Control Scheme Design
tails, and results are discussed in the sub- Column B operates as a purge column, re-
sections that follow. moving both a low boiling distillate and a
high boiling bottom stream from the prod-
17-3-1 Process Background-
uct, which leaves the column as a liquid
Reduced Capital sidedraw above the feed. (Luyben, 1966)
Column B, as shown in Figure 17-4, is not a The overall column material balance, on the
simple distillation because the column has basis of 100 lb /h feed flow, is given in
multiple feeds and a sidedraw. This column Figure 17-4. The column is designed for the
was designed to replace a simple distillation following feed composition, given in terms
360 Practical Distillation Control
.
1···~··········································7 .
···•
··:
.·
. . .:t. .
~
.... ~..
12
.... --®-..................~
•...--®-.....
r;:;::~~m. ......................
r··········" 81.1
',. ,I
··0. : I
··,• ..,•
, i
i'....
rf~~~~~~------ ............... ,, ,,
,, ·· ....:..................
....................•................
.i
.~
:,,
,,
·f ..: :
....................
IL. __ , : ::
i ::
100
L...~....-......
t· ,, :-------------- ~--::~
:,
:•. 4'11-"_ _....
~:
......~.
15 L1
uct sidedraw must contain less than 1.0% viding a fairly constant feed rate and water
water, primarily due to the problems excess composition to Column B.
water creates in the vacuum system of the A direct material balance approach is
next column. Variations in the amount of taken with the overall column control
water leaving with the sidedraw are not scheme shown in Figure 17-4, as is often
important, as long as total amount leaving is recommended for purge column control
acceptable. (Tolliver, 1986). This approach directly sets
In the bottom section of the column, the the top and bottom purge stream flows to
intermediate and high boiling impurities are maintain compositions, whereas the corre-
separated from the sidedraw product, with a sponding internal flows are manipulated to
specification of less than 15 ppm total impu- maintain the reflux drum and sump levels.
rities in the product. The recycle feed stream The sidedraw flow must then be manipu-
is fed to the reboiler of Column Band lated to maintain the collection tray level
typically contains 99.8% organic product and and thus satisfy the overall column material
0.2% intermediate boiling impurities, which balance. The reflux flow below the sidedraw
is approximately equal to the vapor compo- is set to establish the reflux ratio and sepa-
sition leaving the reboiler. ration in the lower section of the column,
The only requirement for the bottom and because of the energy balance, this
stream is to concentrate the intermediate results in a fairly high reflux ratio and ex-
and high boiling impurities, together with cess separation for the top section of the
the recycle feed, to a minimum flow, with column.
manageable viscosity, for further processing One concern with the proposed direct
by the batch recovery column. These condi- material balance control scheme is the sump
tions are met when the bottom stream con- level controller, which must manipulate the
tains approximately 95% organic product. reboiler heat to maintain level. This control
Inventory management is achieved by scheme, in some instances, can be subject to
blending fresh crude, crude from the crude an "inverse response" resulting in poor level
storage tank, and dilute recovered material, control (Buckley, 1974). Adding to this con-
through an in-line static mixer, to become cern was the process requirement to mini-
the feed to Column B. The fresh feed from mize holdup in the sump. However, three
the crude area is flow controlled at a rate factors indicate that Column B should not
slightly less than the crude being produced. experience an inverse response. The column
Excess crude flows to the crude storage has a forced circulation reboiler, structured
tank. The flow from the crude storage tank packing internals with low liquid holdup,
is flow controlled, with the setpoint adjusted and the liquid in the lower section of the
manually to maintain 10% inventory, using column is nearly all product. Some further
total feed rates within the capacity limits of considerations to address this concern are
Column B. discussed with the implementation details.
The flow of the dilute recovered material
is manipulated to control the water compo-
17-3-3 Implementation Details
sition of the total feed to Column B. The
level in the dilute recovered material tank is At the top of the column, the reflux drum
loosely controlled by manually adjusting the level is tuned with high gain for tight level
setpoint on the total feed water composition control, such that excess boilup is immedi-
controller. If the level is high, then the ately returned to the column. The distillate
water concentration setpoint is increased water purge rate is determined by a feedfor-
and the dilute material is worked off. This ward and feedback control scheme. Feed-
approach compensates for large water dis- forward control, based on the water enter-
turbances in the fresh crude, which occur ing the column, is calculated from the total
periodically during washouts, thereby pro- feed flow and the feed composition. The
362 Practical Distillation Control
total feed flow is computed from the sum of symmetrical, steady state temperature devi-
the three individual flows that combine at ation from the base case. Because of uncer-
the static mixer. The water composition in tainty in the height of a theoretical plate
the combined feed is determined from a (HETP) in the structured packing and in
continuous and reliable density measure- the reflux ratio in the top section of the
ment. Dynamic compensation in the form of column, alternate sensor locations are pro-
a lead-lag plus deadtime is provided for vided one stage above and below stage 21.
each feedforward signal and field tuned to In the top section of Column B, which is
match the process dynamics. separating water from organic material, the
Feedback control uses a characteristic relative volatility is very large and there is a
column temperature to infer overhead and sharp temperature break across the control
sidedraw composition. The temperature stage located with the preceding procedure.
sensor location is determined from para- For this high gain situation, the feedback
metric cases of the steady state material and temperature control scheme actually utilizes
energy balance. Column temperature pro- a weighted average temperature signal as a
files from the base case design and two form of profile position control (Luyben,
parametric cases at slightly greater and less 1972). With this approach, anyone, or all of
distillate-feed ratio are shown in Figure the measured stage temperatures may be
17-5. The temperature sensor is located at utilized to whatever extent is found appro-
stage 21, which exhibits the greatest, most priate.
Also, a notch gain PID control algorithm
is used to compensate for the extreme non-
25 linearity that occurs with this high process
24 gain (Fisher Controls, 1984). The output of
23 the feedback temperature controller is then
22 Control , __ combined with the feedforward signal de-
--
21 ontrol \ --, scribed previously to provide a remote set-
ontrol , _ _ :\
26 point for the distillate flow controller. Be-
19 Side draw cause the feedforward signal is based on
18
both feed flow and composition, the feed-
17
16
back signal is combined as a bias, rather
than as a multiplier (Tolliver, 1991b).
Iii 15
D The internal liquid from the top section
E 14
~
c: 13 of Column B is collected on a chimney tray.
Q)
12 Level control of this tray is achieved by
I 11
10
manipulating the sidedraw flow rate. This
level loop is tuned loosely, with mostly pro-
9 portional action, to minimize flow distur-
8 bances to downstream equipment. A notch
7
gain PID algorithm is again used, in this
6
5
case to allow the gain to be further reduced
4 while the level is close to setpoint. A notch
3 Feed gain factor of 0.5, with the notch defined as
2 10% above and below setpoint, has been
found to be the most effective tuning for
0 averaging out temporary flow disturbances.
100 200 300 400
Temperature (oF)
The recycle feed flow, which is fed to the
column reboiler, is set via an operator ad-
FIGURE 17·5. Column B temperature profile. justable percentage of the sidedraw flow.
Control of Distillation Columns via Distributed Control Systems 363
Typically, the recycle flow is maintained as are applied to the total organic feed flow,
10% of the sidedraw flow. the water flow entering with the feed, the
The reflux below the sidedraw is flow lower section reflux flow, and the recycle
controlled with a setpoint that provides the feed flow. This feedforward signal provides
necessary separation to keep the organic most of the level control action; therefore,
impurities below specification. There is no the feedback loop is tuned with moderate
on-line measurement of these impurities; gain and little reset.
therefore, feedback control is implemented
manually based on periodic off-line labora-
tory analysis. Feedforward control is also
17-3-4 Results
implemented based on a reflux-total or-
ganic feed ratio, which is computed from This control strategy has worked quite well,
the total feed flow, less the water entering in spite of many process problems that were
with the feed. Upper and lower reflux limits encountered during start-up. The ease with
are enforced to avoid flooding and weeping which this column control strategy allows
constraints. total reflux operation and manual composi-
At the bottom of Column B, a similar tion control before commissioning the auto-
feedforward signal establishes a tails-total matic temperature and viscosity feedback
organic feed ratio, and is used to set the loops proved invaluable during the start-up.
bottom purge flow rate. The recycle feed The extensive use of feedforward control
flow is not included as a separate distur- has allowed the column to run smoothly
bance because its impurities are accounted even before the feedback loops were put in
for, at least in the steady state, with the service.
ratio of the total organic feed taken. There Loop checkout and operator training via
is no temperature sensitivity to infer compo- a real-time digital simulation of the process
sitions in the lower section of Column B; also contributed greatly to the success of
rather an on-line viscosity measurement is this project (McMillan, 1985). Inventory
used to provide feedback trim of this ratio. management by adjusting the feed composi-
Viscosity is actually the stated control ob- tion was a difficult concept to explain; how-
jective for the bottom stream. The viscosity ever, it was easily demonstrated by operat-
measurement is continuous, but less reliable ing the simulator. The rejection of water
than the feed density measurement. When disturbances, as a result of this approach,
the viscosity measurement is not available, was probably one of the most significant
the tails-total organic feed ratio is still factors contributing to overall successful op-
maintained and may be manually adjusted eration.
with the viscosity controller output. The main process problem that was en-
As mentioned previously, the sump level countered at start-up was poor separation
is controlled by manipulating the column of the organic impurities from the column
boilup. This is achieved by providing a re- sidedraw. A modification of column inter-
mote setpoint to the reboiler steam flow nals to allow a higher reflux ratio was even-
controller. In order to address concerns tually required to solve this problem. This
about the performance of this loop, feedfor- higher reflux ratio used below the sidedraw
ward control is provided to help correct for also causes a higher reflux ratio to result in
potential disturbances. The remote setpoint the top section of the column; however, no
for the reboiler steam flow is determined control scheme revisions were required. The
from the sum of the level controller output effective location of the characteristic col-
plus a feedforward steam flow calculation umn temperature was easily modified by
based on several measured disturbance sig- adjusting the weighting factors on the three
nals. Individual steam-disturbance ratios individual temperature sensors.
364 Practical Distillation Control
r ~ ----------------------1 ---------------------
··t ~ -t____________1_ _ _ _ _ _ _ _ _ J
I
_
........ _..... _._ .... _._ .... --_ ......................
i --_ .............. _........ _...........---_ .............. --_ .... --- .... -_ ........ -_ ..
nipulated to maintain the sump level in lated for reflux drum level control, and the
Column C. distillate flow, which is manipulated for di-
In order to maintain the energy balance rect material balance composition control.
split between columns, the Column C con- Temperature control is used to infer and
denser-steam boiler duty is set to remove maintain composition in both Column C
approximately 90% of the reboiler duty. Be- and Column D. Steady state temperature
cause this condenser generates steam, the profiles are shown in Figure 17-7 for both
heat removal duty may be set simply by flow columns. Parametric cases with alternate
controlling the low pressure steam leaving values of DIF and B IF are used to deter-
the boiler. The steam pressure inside the mine the sensitivity and best location for
boiler is allowed to float, which establishes these temperature control sensors. The
the temperature difference available for heat parametric cases are shown as dashed lines
transfer. about the solid line for the design case. The
Thus, the steam flow from the condenser deviations toward the top are due entirely
is set in ratio to the steam used at the to changes in D IF, and the deviations at
reboiler. Feedback control of the Column D the bottom are due entirely to changes in
reflux ratio is then used to trim this steam B IF. Thus there is little steady-state inter-
ratio. The Column D reflux ratio is com- action between these two temperature loops,
puted from the reflux flow, which is manipu- with sensors located in Column D equiva-
lent to stage 22, and in Column C in the
vapor space above the sump.
28
27 17-4-3 Implementation Details
26
25 The Column C bottom temperature loop is
24 implemented with both feedforward and
23
22
feedback manipulation of the bottom flow.
21 The feedforward signal is based on a design
20 BIF fraction of the feed flow. The feed-
19 back temperature controller output is used
18 as a bias to this feedforward signal, provid-
~ 17
E 16
ing a higher bottom flow if temperature is
~ 15 above setpoint and less flow if the tempera-
Q) 14 ture is low.
~ 13
i'i5 12 Temperature sensitivity to composition
11 changes is low. However, pressure varia-
10 tions can contribute significantly to the
9 measured temperature; therefore, pressure
8 compensation is applied to the measured
7
6 temperature. Pressure effects on the tem-
5 perature at the bottom of Column Care
4 negated by applying a pressure correction
3 bias to the measurement. If the measured
2
1
pressure is greater than design, then mea-
OL....._...L..._....L..._.....L._.....L_.....J_---J sured temperature is decreased to approxi-
300 320 340 360 380 400 420 mate the equivalent temperature at the
Temperature (OF)
design pressure. Because the inferred com-
FlGURE 17-7. Columns C and D temperature pro- position varies slowly, a significant dynamic
file. lag, with a time constant of several minutes,
Control of Distillation Columns via Distributed Control Systems 367
is used on this pressure correction term in of the valve, without changes from the feed-
order to filter any rapid pressure transients. forward signal. Therefore, in this applica-
The reboiler steam flow receives its re- tion, the feedforward signal is combined
mote setpoint from a feedforward signal with the feedback signal internal to the level
designed to provide optimum separation at controller (Fisher Controls, 1984; Tolliver,
various feed rates. Upper and lower set- 1991b).
point limits are used to avoid flood and The boiler blowdown flow is set in ratio
weep conditions. A manual bias is available to the boiler steam flow, with feedback ad-
to provide more or less heat to the reboiler justment of this ratio provided by a con-
depending on ambient weather conditions. ductivity controller. The blowdown flow
The Column C reflux-feed ratio is a good controller allows a simple flow ratio to be
indicator for manual feedback to the re- maintained when the conductivity measure-
boiler steam flow. ment is not available for feedback.
The Column C base level control loop, The Column D base level control loop,
which manipulates the sidedraw flow, is also which manipulates the reflux flow returning
implemented with a feedforward signal to Column C, is tightly tuned. Likewise, the
based on a design S/ F fraction of the feed Column D reflux drum level controller,
flow. The level controller output is a bias to which manipulates reflux flow returning to
this feedforward signal. The feedback level Column D, is also tightly tuned. For both
controller is tuned as tightly as possible; these level loops, tight tuning is achieved
however, the deadtime, due to the trays with a high gain and little reset.
between the sidedraw and the sump, limits The Column D temperature control loop
the gain that can be used. is implemented with both feedforward and
As mentioned previously, condenser- feedback manipulation of the distillate flow.
boiler steam flow is set to establish the The feedforward signal is based on a design
design energy balance split between Col- D / F fraction of the feed flow. The feed-
umn C and Column D. Because the steam back temperature controller output is used
pressure is allowed to float, both pressure as a bias to this feedforward signal, provid-
and temperature measurements are used to ing a higher distillate flow if the tempera-
calculate the mass steam flow. The remote ture is low, and less flow if the temperature
setpoint for the condenser steam flow is is high.
determined from a ratio times the reboiler
steam flow. This ratio is adjusted by inte-
17-4-4 Results
gral-only control of a computed Column D
reflux ratio, which is a good steady state This control strategy has worked quite well.
indicator of the design energy balance split. The columns are easily started up under
The integral-only control algorithm main- total reflux operation, and by the time feed
tains the desired steady state reflux ratio is introduced, the sidedraw product is typi-
while avoiding dynamic interactions with the cally well within specification and remains
other control loops (Tolliver, 1991a). that way.
The condenser steam flow is used as a Loop checkout and operator training via
feedforward signal added to the level con- a real-time digital simulation of the process
trol feedback signal to position the boiler contributed greatly to the success of this
feedwater valve. This application of feedfor- project. The integral-only control of the re-
ward control does not provide a remote flux ratio in Column D, as a means of
setpoint because there is no boiler feedwa- maintaining the energy balance split, was
ter flow controller. It is important that when the most difficult concept to explain. How-
the operator puts the level controller into ever, the operation and dynamics were eas-
manual mode, he/she has complete control ily demonstrated on the simulator.
368 Practical Distillation Control
During the initial start-up, the only oper- . person weeks. This time included review of
ational problem experienced was related to historical plant data, parametric case stud-
achieving the desired operating pressure. ies, conceptual design and review, detailed
There were many leaks, making the high design, software configuration, and imple-
vacuum unattainable. As a result, the sump mentation.
temperature was higher than design, which Real-time simulations were also devel-
limited the heat input capability. oped during this six week effort and used
At first, the condenser-steam generator for configuration checkout. These models
was not cooling properly, which was allow- were then used to demonstrate the control
ing too much vapor feed to Column D. strategies to plant supervisory personnel,
Once the steam valve was opened manually, who reviewed the implementation, and in
all the inerts that had been trapped vented tum used the models to conduct operator
into the low pressure steam system. Al- training. Time for start-up coverage was not
though this caused problems with all the included, as process and mechanical delays
other low pressure steam users, it resolved obscured those numbers.
the immediate problem with the condenser. It is interesting to note, that although
Then, the opposite problem of too much these three examples included highly instru-
cooling became apparent. The steam pres- mented, complex distillation columns, there
sure increased, as expected, in order to were no composition specific measure-
reduce cooling; however, a miscalibrated ments, such as with gas chromatograph (GC)
pressure relief valve limited the steam pres- analyzers. Compositions were inferred by
sure. With higher than normal column tem- nonspecific property measurements such as
peratures, too much cooling was being temperature, ,density, and viscosity. This
achieved and total condensation was pre- approach is quite typical in the chemical
venting any vapor feed from reaching Col- processing industries where analyzers and
umn D. A shutdown was finally required to sampling systems are quite expensive to im-
fix the vacuum leaks and properly set the plement and maintain. The computational
pressure setting on the relief valve. The capabilities of distributed control systems
following start-up proceeded without prob- are frequently utilized to make these in-
lems. ferred measurements.
Feedforward control was the primary
"advanced" control technique that was used
throughout these examples. Some optimiza-
17-5 SUMMARY
tion, adaptive tuning, constraint, and decou-
As Buckley states in his paper on the status piing control considerations were ad-
of distillation control system design, "In de- dressed. Two point composition control was
signing a process control system we must performed with both examples having side-
consider a number of nontechnical factors draw products, but was not required to sat-
such as: (a) availability of skills in the design isfy the control objectives for Column A.
organization, (b) availability of skills in the The control strategy for each of these exam-
plant organization, (c) time and money bud- ples has withstood the test of time, having
geted for the project, and (d) the ultimate now been in continuous service for more
competitive position of the plant" (Buckley, than two to three years.
1978).
Distributed control system instrumenta- References
tion has had a major impact in allowing Buckley, P. S. (1974). Material balance control in
more sophisticated control strategies to be distillation columns. AIChE Workshop on In-
designed, implemented, operated, and dustrial Process Control, Tampa.
maintained. For each of these examples, the Buckley, P. S. (1978). Status of distillation con-
design effort lasted for approximately six trol system design. ISA Analytical Instrumen-
Control of Distillation Columns via Distributed Control Systems 369
tation Symposium, Houston, Vol. 16, pp. mentation and maximization. ISA Trans. 30,
11-39. No. 2,5-7.
Dallimonti, R. (1980). Experiences with dis- Rademaker, 0., Rijnsdorp, 1. E., and Maar-
tributed systems in process control: 1975- leveld, A. (1975). Dynamics and Control of
1980. Proceedings of the Joint Automatic Continuous Distillation Units. New York: Else-
Control Conference, Vol. 1, paper no. TA-4A. vier.
Ender, D. B. (1990). Dse of personal computers Shinskey, F. G. (1984). Distillation Control for
in control loop analysis. ISA Instrum. Chern. Productivity and Energy Conservation, 2nd ed.
Petroleum Ind. 21, 49-57. New York: McGraw-Hili.
Fisher Controls (1984). Configuring interactive Tolliver, T. L. (1986). Purge column control via a
and computing controllers. PRoVOX Instru- distributed control system. ISA Trans. 25(3),
mentation Regulatory Controllers. Marshall- 49-54.
town, IA: Fisher Controls. Tolliver, T. L. (1991). Continuous control experi-
Hanley, J. P. (1990). How to keep control loops ences. ISA Trans. 30, No.2, 63-68.
in serivce. In Tech 37(10), 30-32. Tolliver, T. L. (1991). Feedforward control im-
Hougen, J. O. and Brockmeier, N. F. (1969). plementation on a distributed control system,
Developing process control strategies 1- ISA Instrum. Chem. Petroleum Ind. 22,
Eleven basic principles. Instrum. Tech. 16(8), 99-117.
45-49. Tolliver, T. L., Dugar, D., Wintjen, D., Ahmad,
Luyben, W. L. (1966). 10 schemes to control P., Rominger, M., and Hill, K. (1989). Panel
distillation columns with sidestream drawoff. discusses control systems upgrading. Hydro-
ISA J. 13(7), 37-42. carbon Process. 68(11), 63-68.
Luyben, W. L. (1972). Profile position control of Tolliver, T. L. and McCune, L. C. (1978). Distil-
distillation columns with sharp temperature lation control design based on steady state
profiles, AIChE J. 18(1), 238-240. simulation. ISA Trans. 17(3),3-10.
Luyben, W. L. (1975). Steady state energy con- Tolliver, T. L. and McCune, L. C. (1980). Find-
servation aspects of distillation column con- ing the optimum temperature control tray for
trol system design. Ind. Eng. Chem. Fund. distillation columns. Instrum. Tech. 27(9),
14(4), 321-325. 75-80.
McMillan, G. K. (1985). Real time simulations in Tolliver, T. L. and Waggoner, R. C. (1980). Dis-
distributed control systems for training and tillation column control; A review and per-
engineering. Proceedings of the American spective from the CPl. ISA Adv. Instrum.
Control Conference, Vol. 1, pp. 1718-1720. 35(1), 83-106.
McMillan, G. K. (1989). Continuous control Thurston, C. W. (1981a). Computer-aided design
techniques for distributed control systems. of distillation column controls, Part 1. Hydro-
ISA Independent Learning Module. Research carbon Process. 60(7), 125-130.
Triangle Park, NC: ISA. Thurston, C. W. (1981b). Computer-aided design
McMillan, G. K. (1991). Editor's viewpoint on of distillation column controls, Part 2. Hydro-
distributed control systems-Selection, imp le- carbon Process. 60(8), 135-140.
18
Process Design and
Control of Extractive
Distillation
Vincent G. Grassi II
Air Products and Chemicals, Inc.
370
Process Design and Control of Extractive Distillation 371
The extractive tower bottoms stream actions. Various methods of design and
feeds a solvent recovery distillation tower. operation have been published, but indus-
The solvent recovery tower overhead prod- trial experience has shown that the process
uct contains the intermediate component. design and control of this process is not
The solvent recycles from the bottom of the obvious.
solvent recovery tower to the extractive dis-
tillation tower.
Figure 18-1 presents a simplified flow
18-1-1 extractive and Azeotroplc
sheet of the extractive distillation process.
Distillation
This process requires very nonideal
vapor-liquid equilibrium and tight process Extractive and azeotropic distillation are in-
integration within the double column sys- dustrially important separation processes.
tem. The process dynamics are highly non- These are used when key components form
linear and multivariable with many inter- an azeotrope or have a sufficiently low rela-
A
S
AlB
B/S
y
FIGURE 18·1. Extractive distillation system.
372 Practical Distillation Control
tive volatility that conventional distillation azeotropic tower as the bottoms product.
becomes impractical. Overhead product containing the minimum
A component not present in the feed is boiling azeotrope is fed to a second distilla-
added to the process. This new component tion tower to recover the second key com-
enhances the relative volatility of the key ponent.
components so the mixture can be sepa- If the overhead product condenses to two
rated efficiently by distillation. Figure 18-2 liquid phases in the reflux accumulator, the
illustrates this effect. entrainer rich phase serves as reflux to the
Extractive distillation is characterized by azeotropic tower. The second phase feeds
the addition of a heavy component, called the recovery tower. This tower recovers the
the solvent, added toward the top of the key component as the bottoms product. If
distillation column. Examples of industrially the overhead of the azeotropic tower con-
significant extractive distillations are: denses to a single phase we can operate the
recovery tower at a different pressure to
1. Methyl acetate from methanol using wa- remove the key component. Examples of
ter. industrially significant azeotropic distilla-
2. Ethanol from water using ethylene gly- tions are:
col.
3. Butene-2 from n-butane using acetone. 1. Acetic acid from water using ethyl ac-
etate.
Addition of a light component, called the 2. Ethanol from water using benzene.
entrainer, differentiates azeotropic distilla- 3. Acetic acid from water using isopropyl
tion from extractive distillation. The en- acetate.
trainer forms a minimum boiling azeotrope
with one of the key components. The re- Figure 18-3 contains a simplified flow dia-
maining component is removed from the gram of a generic heterogeneous azeotropic
system.
Azeotropic distillation is closely related
Extracti't Distillation to extractive distillation. This chapter will
Vapor Liquid EquiUbrium
be confined to the generics of extractive
distillation, but the reader will see many
extensions to azeotropic distillation.
y Extractive distillation has highly nonideal
vapor-liquid equilibrium. It is a dual col-
umn, integrated system and it presents many
difficulties in process simulation, design, and
x control. I will present concepts that improve
.....
our engineering insight and develop meth-
+ Solvenl y odologies for the process design and control
(OR)
of these systems. These concepts will be
applied to the three extractive distillation
case studies. These systems are highly non-
x linear and multivariable. It is important to
y consider the process dynamics during the
process design of these systems. This can be
done easily in a way that is productive and
provides greater process engineering insight
x into the nature of extractive distillation sys-
ftGURE 111-2. Vapor-liquid equilibrium effect. tems.
Process Design and Control of Extractive Distillation 373
AlE
E
Al B
B A
ProcIuctA
SJlbr1JI. Product B
IIlm
SoIvco.
-1"'4_--
Feed
AlB
Solvent
tive distillation system and needed addi- A binary mixture of components A and
tional capacity in another. These systems B feeds the extractive distillation process.
are heterogeneous with multiple azeotropes. These two components either form an
When we started working on these columns, azeotrope or have sufficiently low relative
we could not even simulate the distillation volatility so conventional distillation is im-
columns at steady state. Over the years we practical. The feed is introduced to the mid-
have successfully expanded their capacity dle of the extractive distillation tower. A
almost threefold through process improve- . heavy solvent is chosen that has an affinity
ments without increasing the size of the for component B. We feed the solvent a
distillation towers. These process improve- few trays from the top of the extractive
ments came about only by a fundamental distillation tower. These trays remove sol-
understanding of the distillation process that vent from the overhead product. The distil-
was gained through detailed steady state late contains component A as product. The
and dynamic process simulation verified by bottoms contains the solvent and compo-
plant testing. nent B. This stream feeds the solvent recov-
ery tower. The distillate contains compo-
nent B as product. The bottoms contains
18-1·3 Process Description
the solvent that is recycled to the extraction
Figure 18-4 is a generic process flow dia- tower.
gram of the extractive distillation process The solvent is the highest boiling compo-
we will consider. This figure shows the key nent in the system. We must cool the sol-
manipulative variables and interstream heat vent before returning to the extraction tower
exchangers. Our aim in this chapter is to to prevent it from vaporizing. Much of its
develop a general procedure that optimally heat can be recovered in the process by
designs the extractive distillation process using tower feed preheaters. A final trim
and finds the best control scheme needed to cooler cools the solvent to its required feed
operate it. temperature.
Process Design and Control of Extractive Distillation 375
The solvent tower sump maintains sol- where component A is nonpolar and com-
vent inventory. Solvent tower sump level ponent B is polar. A polar solvent will
controls the fresh solvent to makeup for attract component B greater than compo-
solvent losses through the product streams. nent A. This attraction, or intermolecular
Fresh solvent makeup flow gives us an addi- interaction, will cause the solvent to pull
tional degree of freedom. We can use the component B down the column into the
solvent tower bottoms flow (solvent recycle) bottoms stream.
as a control variable for something other The components of ideal mixtures do not
than sump level control. interact. The solvent is chosen so nonideal
Discounting pressure and inventory con- interaction provides a sufficiently large in-
trols, there are 6 degrees of freedom in this crease in relative volatility between the two
process. These control five compositions and key components.
the temperature of the solvent feed to the
extractive tower.
Extractive distillation affects a key com- 18-2-1 Relative Volatility
ponent by reducing its volatility and remov-
The fugacity of the vapor and liquid phases
ing it in the bottoms stream with the heavy
are equal at equilibrium. Our model for the
extractive solvent. As in conventional distil-
vapor-liquid equilibrium for these systems
lation, both overhead and bottom composi-
is the same as that presented in Chapter 3.
tion controllers interact. Material recycle of
We will consider low pressure, less than 50
the solvent and heat integration of the in-
psia, nonreactive systems. The vapor is es-
tercoolers create dynamic interactions. Rig-
sentially ideal so the vapor phase fugacity
orous process design and control of this
coefficient is equal to 1. The liquid phase
process must consider the entire system as a
activity coefficient is not unity due to inter-
coupled system of distillation towers.
action between components in the liquid
Figure 18-1 shows the manipulative vari-
phase.
ables, that is, control valves for this process.
The activity coefficient is a function of
As in conventional distillation the reflux,
the liquid composition and temperature.
distillate, heat input, and bottom flows can
The relative volatility between any two
be manipulated. However, solvent flow is
components is a function of the activity
also a manipulative variable in extractive
coefficient and the pure components vapor
distillation. Its control is not obvious. Addi-
pressures. Figure 18-5 describes these rela-
tionally, there could be an optimum addi-
tionships.
tion rate when considering process dynam-
The NRTL correlation is used to com-
ics that is different from that by considering
pute the liquid phase activity coefficient for
only steady state effects.
Extractive distillation can produce a het-
erogeneous azeotropic overhead. In this
case, the overheads will condense into two fvapor = !liquid
these case studies. You can choose your sure there are no degrees of freedom left
own model, or equation of state, for your and there must be only one solution that
particular application. Corresponding states fixes temperature. This solution is the
models work well for these nonideal chemi- ternary azeotrope.
cal systems. Good results may also be ob- Similarly, there are only two components
tained using the Wilson and UNIQUAC for each binary pair in the mixture and one
activity coefficient models. constraint, y 1 = X l' Setting pressure fixes
We must consider the relative volatility the system at the azeotrope for the binary
for all of the binary pairs in the extractive pair.
distillation process. The feed contains com- The phase rule is also useful to find
ponents A and B. The solvent component operating conditions when multiple liquid
S has an affinity for component B. The phases can occur. For example, the total
relative volatility between components A number of phases in a distillation process
and B, when solvent is present, determines with two liquid phases is 3. The phase rule
how effective the extraction tower is. The reduces to: F = c + 2 - 3, or F = c - 1. A
relative volatility between components A ternary system has 2 degrees of freedom,
and S determines how well we will be able that is, temperature and pressure. It is
to keep solvent out of the extraction tower therefore possible to have more than one
distillate. The relative volatility between B tray with two liquids on it. However, a bi-
and S determines the effectiveness of the nary system has only 1 degree of freedom.
solvent recovery tower. Because pressure is set, only one composi-
Other phenomena that can occur from tion results in two liquids. It is very unlikely
the addition of solvent to the mixture are that any trays in a binary tower will have
additional azeotropes and two liquid phases. two liquid phases.
A ternary system can have up to four The presence of two liquid phases on
azeotropes. Each binary pair can form an some trays in a distillation tower can create
azeotrope and the mixture can have one operating instabilities. If the two phases
ternary azeotrope. Azeotropes can create have significantly different transport proper-
distillation composition boundaries over ties they can reduce tray efficiency. Small
which a single distillation tower cannot changes in operating variables can cause
cross. trays to switch from single to two liquid
The Gibbs phase rule relates the number phases. Strong operating instabilities, such
of phases and azeotropes in a mixture: as wide temperature swings, can result. Sys-
tems that exhibit two liquid phases have an
F = c +2- Ph (18-1)
increased tendency to foam if large surface
where F = number of degrees of freedom tension gradients exist. We need to be
c = number of components in mix- circumspect of additional azeotropes and
ture heterogeneous phase behavior during the
Ph = number of phases present process design and control of extractive
distillation towers.
Our system contains three components and I recommend that a solvent be chosen so
a homogeneous azeotrope contains two it yields a relative volatility of at least 2
phases. Therefore, F = 3 + 2 - 2, or 3 de- between A and B, 10 between A and S,
grees of freedom. The vapor and liquid and 5 between Band S. The solvent should
compositions are equal at an azeotrope. We not create two liquid phases or any new
therefore have two constraint equations, y 1 azeotropes. The larger the relative volatility
= Xl and yz = x z, at the azeotrope, which is enhanced, the better.
take 2 degrees of freedom; only 1 degree of Sometimes a less desirable solvent is
freedom is left. Therefore at constant pres- more economical than the ideal solvent. A
Process Design and Control of Extractive Distillation 377
less desirable solvent might be one that is The residue curve map for our ternary
already present in the plant, inexpensive, system is drawn as a right triangle. The
nontoxic, or plentiful. We must learn how to liquid composition of component A is the
design a process and control it for these ordinate and the liquid composition of com-
solvents as well. ponent B is the abscissa. The vertices of the
triangle locate the pure components. To
construct the residue curve map we con-
18·2·2 Residue Curves
sider the simple single stage batch distilla-
Residue curve maps are an insightful way to tion process. A mixture is batch charged
visualize the phase equilibrium of a ternary into a pot. Heat is added to boil the mix-
system. They graphically depict azeotropes, ture. Vapor is removed, but no additional
heterogeneous phase location, separation feed is added to the pot. We plot the liquid
feasibility, and distillation paths for the ex- composition trajectory on the residue curve
tractive distillation tower. These curves are map as the boiling process proceeds.
very useful in determining the process de- The residue curve map using S2 has one
sign of extractive distillation systems. Do- binary azeotrope. It is a minimum boiling
herty presents the theory of residue curve azeotrope between components A and B.
maps quite well (Doherty and Perkins, 1979; Pure components and azeotropes are stable
Doherty and Caldarola, 1985). points on the residue curve map. Residue
A residue curve map is the time trajec- curves either start or end at these points.
tory of the liquid phase composition during The residue curve map for this system using
a simple single stage batch distillation with S2 is easy to interpret. The edges of the
no rectification. Figures 18-6a and b illus- residue curve map are the residue curves
trate the residue curve map for the first two for the binary mixtures. The residue curves
of our case studies. These are the residue always proceed from low to higher tempera-
curve maps for methyl acetate (A) and tures. The residue curves end at the pure
methanol (B) using either water (St) or solvent vertex because it is the highest boiler
ethylene glycol (S2) as the extractive sol- in the system. Residue curves diverge from
vent. the minimum boiling azeotrope. No distilla-
Pure Pure
A A
1.0 1.0
0 Binary azeotropes
~ Distillation boundary
0.8 0.8
«
«
j 0.6
~
~
0.6
~
!
~ 0.4
:g 0.4
g"
g"
0.2 0.2
Vapor liquid
equilibria
A/8
1.00 'T'""----------......,.
0.80
0.60
0.40
0.20
Solvent 1 Solvent 2
A/51 A/52
1.00 -r-----------~ 1.00 'T'"",...---------..,.
0.80 0.80
0.60 0.60
0.40 0.40
0.20 0.20
0.00 0.00'
0.00 0.20 '0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
8/51 8/52
1.00 1.00
0.80 0.80
0.60 0.60
0.40 0.40
0.20 0.20
0.00 0.00
0.00 ' 0.20 0.40 0.60 0.90 1.00 0.00 0.20 0.40 0.60 0.80 1.00
(a)
FIGURE 18-7. (a) Systems 1 and 2-VLE. (b) System 3-residue and VLE.
Process Design and Control of Extractive Distillation 379
C/D
1.00 . , . . . . - - - - - - - - - - - - "
0.80
0.60
0.40
0.20
0.00
0.00 0.20 0.40 0.60 0.80 1.00
0.80 0.80
'"
c
.g
0.60 0.60
"<U
<!::
0"
0.40 E 0.40
"'5
<:T
:.:i 0.20
0.20
0.00 0.00
0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
0.80
0.60
0.40
0.20
0.00
0.00 0.20 0.40 0.60 0.80 1.00
(b)
tion boundary or two liquid phase region operate and create higher utility and raw
exist in the residue map for System 2. material costs due to upsets. To get the
Let us now examine a more complex optimal economic design, it is very impor-
residue curve map that results for the same tant to consider the controllability of the
A and B components, but using SI instead extractive distillation process during process
of S2. Solvent 1 (S1) forms a heterogeneous design.
azeotrope with component A. Open circles We will develop a computerized design
label the azeotropes on the residue map. procedure that can be run on a small per-
The presence of the second azeotrope cre- sonal computer. Computer programs that
ates a distillation boundary. This boundary rate an existing distillation tower are more
divides the composition plane into two dis- common than design programs. You can use
tinct regions. The distillation process cannot commercial process simulators to rate an
cross the boundary and therefore cannot existing extraction distillation tower. These
make products in both. The bottoms com- programs are cumbersome to use to design
position of the extractive distillation tower a new extractive distillation system and do
must be a binary mixture of Band S1. It is not leave the process engineer with as much
therefore impossible to make a pure distil- process insight as a design procedure.
late product consisting of component A.
The overhead product cannot lie beyond
18-3-1 Degrees of Freedom
the line passing between the two binary
azeotropes. To begin the design of our process we must
We can see the two liquid phase region first determine which variables we will set
using this residue map. This is a great aid to and what variables will be calculated by the
help us avoid the two liquid phase region design procedure. The feed stream is given
when we design the extractive distillation and the operating pressures of the distilla-
tower. The residue curve map of this com- tion towers are fixed by the plant cooling
plex system helps our understanding of the medium. We will fix a design limitation on
process and aids in the design and control the minimum approach temperatures in our
of extractive distillation systems. interstage heat exchangers.
Figure 18-7a and b contains the residue Given these conditions we need to find
curve maps and binary vapor-liquid equilib- the number of trays, reflux ratio, and con-
rium curves for our three case studies. The denser and reboiler duties for the towers. If
components used in System 3 are ethanol we have a procedure to design the extrac-
(C), water (D), and ethylene glycol (S2). tive distillation tower, we can find the sol-
vent temperature, solvent recycle rate, and
the cross exchanger areas. Conventional
18-3 PROCESS DESIGN
methods can design the binary solvent re-
Our goal is to design an extractive distilla- covery tower. The extractive distillation
tion process at minimum economic cost. An tower is much more complicated because it
optimum process design minimizes total is muiticomponent, has multiple feeds, and
cost: the fixed capital cost of the equipment has nonideal phase equilibrium.
plus the variable cost of operation. Capital Figure 18-8 details the degrees of free-
costs are the total fixed project costs to dom available for the design of the extrac-
design, construct, and start-up the process. tive distillation tower. The degrees of free-
Variable costs are the operating costs of dom are the total number of variables less
utilities, raw materials, and workforce. Sta- the total number of equations describing
bility of process control significantly affects the problem. The extractive distillation
variable cost. If the process is difficult to tower has two feeds and three components.
control it will require additional labor to We have one total material balance equa-
Process Design and Control of Extractive Distillation 381
Equations
Material balance Ne
Energy balance N
Total equations: N(2e + 3)
Variables
6 degrees of freedom
Specify: P
3 compositions: xD(B), XD(S), xB(A)
RID
nF
tion for every stage and one component feed stage locations are variables. This gives
balance for each component for every stage. us a total of N(2c + 3) + 6 variables.
The vapor and liquid mole fractions must The design of the extractive distillation
sum to 1 on every stage. We have one phase tower has 6 degrees of freedom. There are
equilibrium relationship for each compo- six more variables than equations. We must
nent on every stage and we have one total therefore set six variables to design this
energy balance on every stage. This pro- tower. The six variables that I will set in my
duces N(2c + 3) equations, where N is the design procedure are:
total number of stages and c is the number
of components. I include the condenser and 1. The column pressure.
reboiler as stages using this nomenclature. 2. Three product compositions.
The total number of stages is therefore the 3. The column reflux ratio.
total number of trays plus 2. 4. One feed stage location.
We now count the number of variables. Column pressure will be set by plantwide
Each stage contains a liquid and vapor mole utility considerations. The three product
fraction for each component. Each stage compositions that we will set are:
has one liquid and vapor flow rate leaving
the stage. The reboiler and condenser du- 1. The overhead B composition.
ties and column pressure are variables. Fi- 2. The overhead S composition.
nally the total number of trays and the two 3. The bottoms A composition.
382 Practical Distillation Control
Yj+! =
'j +1
Lj HD -Qcl D - HrH
'j = D = Hr+1 - H jL
Solvent Section
Stripping Section
Xk+l =
YkSk + X B L.:+l
Sk + 1 ~+1
Vk
HB - QRIB - Ht+l
Sk=Ji= H[+l- H[
L...----'----+---I~ B
xB
{oo
Extraction tower design total stages in the column. The feed loca-
S/F-1.5
T,-120°F
tion c~n be found by repeating the design
procedure for various feed stage locations
until the best one is found. This completes
the design of the extractive distillation
tower.
j Solvent Tower
.!? The solvent tower can be designed by the
~ 0.40
:!l preceding procedure without the complic~
g" tion of a second feed. The solvent tower IS
designed as a binary distillation tower b~
cause the composition of component A IS
negligible. The Ponchon-Savarit method
can be used.
0.20 0.40 0.60 0.80 1.00
Uquid mole fraction B Optimization
FIGURE 111-10. Extractive distillation design compo-
Now that we have a way to design the
sition profiles. equipment in the process we must optimize
the system considering process economics.
cal methods. We specify the organic feed
You could write all the design equations
stage, choose a sufficiently large maximum
that we have developed so far and plug
number of section stages, and compute the
them into a simulation package that would
minimum reflux ratio by trial and error. I
produce the economic optimum. I person-
usually set the organic feed stage and the ally do not like that general approach. I
maximum section stages at 50. Assume a prefer to break large problems down into
reflux ratio that is reasonable for the sys-
parts and solve each by iteration to reach
tem. A reflux ratio of 1.0 is a good starting
the optimum. I believe that an engineer
estimate for the three case studies.*
learns more about the process by maintain-
If the composition profiles cross, a lower
ing a physical picture of it during the solu-
reflux ratio is selected and the procedure is tion. The engineer learns how design vari-
repeated. If the composition profiles do not ables affect each piece of equipment, rather
cross, a larger reflux ratio is used. The bi- than the mathematics of its solution.
section method can be used to choose the The important stream to optimize in this
next estimate of the reflux ratio between process is the solvent recycle stream. We
iterations. The calculations are very fast and need to find the best solvent stream compo-
can be run in less than 1 minute on a sition temperature, and flow.
personal computer. Once the minimum re- Soivent composition is not a strong opti-
flux ratio is known, you can set the design mization variable. The impurity in the sol-
reflux ratio at a factor larger than the mini- vent stream is component B. Most of com-
mum. The economic optimum factor for ponent B that enters the extractive tower
these case studies is about 1.1 for the ex- will leave in the distillate as an impurity, so
traction and solvent towers. Now that the a good assumption is that all of component
design reflux ratio is known, we need to B that enters the extractive tower leaves in
compute the optimum location of the feed the distillate. Use this criterion to set the
stage. composition of the recycle stream.
The best location for the organic feed is The two important variables to optimize
the stage that results in the least number of are the solvent temperature and flow. Sol-
* Extractive distillation columns also have a maxi- vent flow (solvent recycle ratio) has the
mum reflux ratio. strongest effect on system economics and
386 Practical Distillation Control
§
The design variables with the most signifi-
cant effect on cost are: J 8.00
cost of steam for this basis because this is FIGURE 18-12. Cost impact of solvent recycle com-
the largest component of operating cost. position.
Figures 18-11 and 18-12 illustrate the effect
of these three variables on total cost for one
of our case studies using the procedure de- that in a practical sense the total cost rela-
scribed in the last section. tion is relatively flat. The solvent ratio and
Figure 18-11 is typical of extractive distil- temperature can be specified somewhat
lation processes. Total cost is a function of loosely without a large cost penalty. The
the solvent recycle rate, temperature, and same is true for operating the plant. You
composition. The design procedure gives us can move the solvent ratio and solvent tem-
a way to determine the optimum design perature without a large operating cost
point. Inspection of Figure 18-11 reveals penalty. This can be of great benefit if feed
Process Design and Control of Extractive Distillation 387
composition is altered and you need to bal- must be sized larger than the solvent holdup
ance the tower loadings. in the towers to prevent the sump from
Figure 18-12 shows that we must be sure overfilling or emptying during upsets.
that the solvent recycle composition does Extractive distillation is a highly inte-
not exceed a maximum limit. This limit is grated process. We have seen that changes
easily calculated by assuming that all of the in process design variables have a significant
impurity entering the extraction tower leaves effect on product specifications at steady
in the distillate product. A good rule to state. The controllability for this process is
follow is to design for one-half of this limit.
also sensitive to the choice and pairing of
We can also see there is not a strong cost the manipulative and control variables.
penalty in overpurifying the bottoms prod- Our first step to determine the pairing of
uct in the solvent tower. the control and manipulative variables of
the extractive distillation process is to re-
duce the number of degrees of freedom.
18-4 PROCESS CONTROL
Our objective is to control the product
To the process control engineer a two prod- compositions in this process. Pressure dy-
uct single distillation tower typically has 5 namics are very fast compared to composi-
degrees of freedom. There are usually five tion dynamics in these systems. Therefore
control valves: we will treat pressure control as a static
problem. We can control column pressure
1. The distillate product flow. by manipulating cooling water flow in the
2. Reflux flow. condenser, manipulating the condensate
3. Bottoms product. flow leaving the condenser, or manipulating
4. Heat input. the vapor flow into the condenser. There
5. Pressure control. are many other possibilities, but these are
the most common. Selecting a pressure con-
Our extractive distillation process con- trol scheme in each distillation tower elimi-
tains two distillation towers. The extraction nates 2 degrees of freedom. See Chapter 1
tower bottoms feeds the solvent tower. The for a more detailed discussion of pressure
bottoms stream from the solvent tower is control schemes.
recycled back to feed the extraction tower. Two material inventories that must be
Independently both towers have 10 degrees controlled in conventional distillation tow-
of freedom. When the two are coupled we ers are the material in the reflux accumu-
lose 1 degree of freedom by recycling the lator and bottoms sump. Therefore two
solvent tower bottoms to the extraction control valves must be used for these two
tower. It is the process control engineer's control variables.
job to determine what to control with each The reflux ratios of the three case studies
valve. It is equally important that the pro- are relatively low. Experience (Shinskey,
cess design engineer supplies the process 1983) shows that it is best to use the prod-
control engineer with a system design that is uct streams to control the column invento-
controllable. ries for such systems. The extraction tower
The solvent makeup feed theoretically distillate and bottoms streams control the
creates an additional degree of freedom. extraction tower reflux accumulator level
We can use the solvent makeup to control and bottoms sump, respectively.
the solvent inventory in the system. How- The solvent tower distillate controls the
ever, the solvent makeup flow is very small solvent tower reflux accumulator level.
relative to the solvent recycle flow. Solvent However, the solvent tower bottoms stream
inventory cannot be controlled precisely cannot control the solvent tower sump level
during dynamic upsets; it can only be con- because it is recycled within the process.
trolled long term. The solvent tower sump Solvent makeup flow loosely controls the
388 Practical Distillation Control
solvent tower bottoms sump, as discussed We can use multivariable or model based
previously. controls if none of these control schemes is
Inventory control eliminates 3 degrees of satisfactory to control the product specifi-
freedom. This leaves 2 degrees of freedom cations. They are, however, rarely needed in
for each tower with an additional degree of industrial practice.
freedom for control of the solvent recycle The extraction column contains 3 de-
stream. Extractive distillation is a process grees of freedom. We will therefore con-
with 5 degrees of freedom, therefore five sider controlling, at most, three composi-
variables to control. These could be five tions.
compositions or any combination of five The distillate is a product stream from
flows, compositions, or tray temperatures. the system. Composition control of this
The solvent column is a binary distilla- stream is important to meet customer prod-
tion tower. The distillate stream is a prod- uct specifications. The bottoms stream is an
uct from our system. The bottoms stream is intraprocess stream. All the light material
an intraprocess stream. Composition con- (component A) in this stream will leave in
trol of the distillate is necessary to meet the distillate of the solvent tower. The light
customer product specifications, and com- composition of the bottoms stream directly
position control of the bottoms stream is affects our customer product specifications,
critical to ensure operability of the system so we must control it. The remaining two
and to meet the extraction tower overhead components in the bottoms stream do not
specification. directly affect customer products, but in-
At most, we will control two composi- fluences the solvent tower and, hence, plant
tions in the solvent tower. The simplest operability .
combination of control pairings is to use the These considerations lead us to conclude
reflux or distillate to control the overhead that there are two combinations of three
composition. If reflux is chosen to control composition control specifications for the
overhead composition, then the distillate extraction tower. We can either control two
will control the reflux accumulator level. compositions in the extraction tower over-
Similarly, if the distillate is chosen to con- head and one composition in the bottoms,
trol the overhead composition, then the re- or one composition in the overhead and two
flux will control the accumulator level. compositions in the bottoms. We can con-
Heat input controls the bottoms compo- trol the compositions of components Band
sition of the solvent tower. The solvent tower S in the overhead and the composition of
bottoms flow should not control the bottoms component A in the bottoms stream if the
composition because this stream is recycled extraction tower overhead product specifi-
within the process. cations are important. If we need to mini-
Using reflux to control the overhead mize feed composition disturbances to the
composition and heat input to control the solvent tower (because it is operating very
bottoms composition on the solvent tower is close to flood), we can control one impurity
common. The solvent is usually much heav- in the extraction overhead, the composi-
ier than component B, which leads to a tions of components A and B in the bot-
moderate to low reflux ratio in the solvent toms.
tower. Reflux is usually the best control
variable when the reflux ratio is low to
18-4-1 Control System Economics
moderate. The solvent makeup rate is usu-
ally too small to use for composition con- The engineer must consider overall plant
trol. Therefore, heat input is usually the economics to find the best control scheme
best variable to control the solvent tower for the extractive distillation system. We
bottoms composition. found that we can control five variables in
Process Design and Control of Extractive Distillation 389
the extractive distillation system. We now any light component in the bottoms stream
need to determine how many of these vari- will end up concentrated in the distillate
ables and which ones are needed to control product of the solvent tower. The extractive
the system. We do not need to control vari- distillation tower has four possible schemes:
ables that are insensitive to plant operation.
My approach is to find the least number of
1. Control the bottoms A composition and
control variables. This has the advantage of
the overhead B composition.
reducing project cost, simplifying the opera-
2. Control the bottoms A composition and
tion, and providing greater process integrity
the bottoms B composition.
by reducing the chance of failure.
3. Control the bottoms A composition, the
The first step in determining control sys-
overhead B composition, and the bot-
tem economics is to define our control ob-
toms B composition.
jectives for the plant. For our case studies
they are to control the product composi-
tions at 99 ± 1 mol%. We must meet these Control schemes that control the amount
specifications for feed rate and feed compo- of solvent in the overhead product of the
sition disturbances of ± 10%. This is a very extractive tower will not be considered. This
ambitious goal, so it serves well to test this is because the solvent is a very heavy com-
design methodology. ponent in the extractive tower and is in
We will examine the steady state effect of trace composition in the overhead stream.
various control configurations by holding The composition of component B is the
manipulated or control variables constant important consideration for the overhead
when we change the feed composition. stream.
Holding a controlled variable constant as The three systems presented in this chap-
feed composition changes quantifies the ter all have the same economic control
value of placing that variable in a control scheme; in each, the best plantwide control
loop. Holding a manipulative variable con- scheme is E1-S1. The bottoms composition
stant quantifies the value of leaving a loop of both the extraction and solvent towers
off control. Figure 18-13a contains the pos- are controlled. Figure 18-14 contains four
sible control schemes for extractive distilla- possible control schemes for the System 1
tion systems. Figures I8-13b through e give extractive distillation process. The total re-
simplified instrument diagrams for the four boiler energy requirements are calculated
most common control schemes. for each scheme over the range of feed
We assume that at least one composition compositions for which we want to design.
will be controlled in each distillation tower. The heat duties are calculated so the prod-
The solvent tower has three possibilities: uct specifications are met. We accomplished
this by fixing the composition if it is a con-
1. Control the bottoms composition. trol variable or fixing the manipulative vari-
2. Control the overhead composition. able if the composition is not controlled.
3. Control both bottoms and overhead com- The manipulative variable is at a constant
positions. value over the feed composition range so
that the desired product composition is al-
It is very important to control solvent ways met. This means that the product is
tower bottoms composition because we need overpurified over most of the feed composi-
to prevent the extraction tower overhead tion range. Overpurification is not as effi-
product from being affected during an up- cient as maintaining the product specifica-
set. We must control the extraction tower tion over the entire feed composition range
bottoms composition of the light compo- by direct control. However, this analysis
nent. If this composition is not controlled, quantifies the cost of inefficiency.
390 Practical Distillation Control
El xS<A) QR
E2d xS<A) and xD(B) QR and R
E2b xS<A) and xS<B) QR and S
E3 xB(A), xD(B), and xB(B) QR' R, and S
Sl xB(B)
S2 xS<B) and xD(B)
(b)
(c)
(d)
We can control most of the product com- bottoms stream. Stripping section tempera-
positions in our extractive distillation pro- tures are sensitive to component A compo·
cess by measuring temperature. The solvent sition because it is very light relative to the
is much less volatile than either key compo- solvent. Higher concentrations of the light
nent. The solvent tower product composi- component in the bottoms stream lower the
tions are usually sensitive to reflux and heat stripping section temperature profile con-
input, so we will control it by enriching and siderably. Therefore stripping section tern·
stripping tray temperatures. peratures are good measurements for con-
The stripping section of the extraction trol of the bottoms light component.
tower is rich in solvent and intermediate The overhead composition of the extrac-
component, and by specification, the com- tion tower is not well suited to temperature
position of component A is very low in the control. The solvent section contains a sig·
394 Practical Distillation Control
design extractive distillation processes to tive to the manipulative variables. The sol-
vary the stage equilibrium by solvent flow. vent is very heavy in the extractive distilla-
In this respect, extractive distillation is more tion tower and moves down the tower
nonlinear than ordinary distillation. Nonlin- quickly. Extraction tower overhead solvent
earities in extractive distillation can lead to composition control is not important.
counterintuitive behavior of the column. The Reflux or solvent flow can control the
process control engineer must deal with overhead B component. I prefer to use
these nonlinearities carefully in the design reflux rather than solvent flow to minimize
and tuning controls for extractive distilla- the interactive effects of solvent flow.
tion. Further inspection of Figure 18-16 re-
Figure 18-16 contains the steady state veals a good example of a counterintuitive
process gains between the manipulative and process gain in extractive distillation. The
control variables for our three case studies. process gain between the extraction tower
These gains are linear and are calculated overhead B composition and solvent flow is
for very small changes in the manipulative negative for Systems 1 and 3. This is what
variable. The results are scaled to instru- we expect. Increasing solvent flow in the
ment ranges so they are dimensionless and extraction tower will extract more of com-
are similar to what you would find in the ponent B at the bottoms of the column,
field. which will lower the overhead composition
These gains illustrate many of the gen- of component B and produce a negative
eral cause and effect variables in the extrac- gain. However, System 2 has a positive gain!
tive distillation column. A first observation This is not what we would expect; it is
is that heat input has a large effect on the counterintuitive.
bottoms composition of the lightest compo- Further inspection of the nature of Sys-
nent. This agrees with our previous conclu- tem 2 reveals why this process gain is posi-
sion that a stripping section temperature tive. The positive gain results from an im-
can control the bottoms composition by ma- pure solvent recycle stream composition and
nipulating heat. a high purity extraction tower distillate. Sys-
A second important observation is that tems 1 and 3 use a solvent recycle stream
the overhead solvent composition is insensi- composition of 0.001 mole fraction B. The
System 1
System 2
System 3
extraction tower overhead compositions are This process gain for System 2 is very
0.01 mole fraction B. System 2 operates nonlinear. It is positive for small changes in
with a solvent recycle composition and ex- solvent flow but becomes negative for large
traction tower overhead B composition of changes in solvent flow. This occurs as a
0.005 mole fraction B. Our design guideline result of two competing effects: An increase
says that the recycle composition should be in solvent flow rate lowers the volatility of
about 0.001 mole fraction B. If we operate component B and, at the same time, in-
the solvent tower bottoms composition at creases the amount of impurities fed into
0.001 mole fraction B, the gain is negative. the top of the extraction tower. When we
Increasing the bottoms purity from 0.005 to design the system too close to the limiting
0.001 mole fraction B only increases the solvent impurity, the later effect becomes
reboiler duty 0.5%. very pronounced. Therefore, when we
0.1
0.0
'f
4~
-0.1
-0.2 -
-0.3
(I)
tC ....
-0.4
-'$ EI System 2
........
m~
-0.5
Q
)(
• System 2 (Alt)
tC
-0.&
-0.7
-O.B
-
y
)
-0.9 -
-1.0
-50 -40 -30 -20 -10 o 10 20
os (o/e)
slightly increase the solvent, the effect of process gains for all changes in solvent flow.
lowering the volatility of component B is I designed the alternate based on an extrac-
less than the effect of feeding more impuri- tion tower overhead composition and a sol-
ties into the extraction tower, which results vent recycle composition of 0.01 and 0.001
in an increase in the distillate composition mole fraction B, respectively.
of component B. The effect on volatility is If the recycle composition in System 2 is
greater than the effect of increased impuri- lowered to 0.001 mole fraction B, or if the
ties when we make a large change in solvent extraction tower overhead composition is
flow. raised to 0.01 mole fraction B, a curve
Figure 18-17 contains a plot of the pro- similar to System 2-Alt results. These ef-
cess gain between the extraction tower's fects are qualitatively valid for any extrac-
overhead B composition and solvent flow tive distillation system with impure solvent
for two systems. The results for System 2 recycle and tight extraction tower overhead
show that the process gain is positive for composition specifications.
small changes in solvent flow and negative I have indicated many ways to avoid this
for large decreases in solvent flow. Addi- nonlinearity. The solvent tower can be oper-
tionally, Figure 18-17 includes a second case, ated with a lower impurity in the bottoms,
labelled "System 2 (AlO," which is a re- or the extraction tower can be operated
design of the system that produces negative with a higher overhead specification. In this
0.0140 I System2~
/'
--
/
0.0130
.......
-.U
!XI l!
oY-
M~
2: 0.0120
.....
0.0110
System 2 Alt
0.0100 + - - - - - / - - - - - - + - - - - - + - - - - - - 1
case I decided to redesign the process, not TABLE 18-1 Case Study Designs
only for the steady state gain effect but also System System System System
for dynamic considerations. 1 2 2-Alt 3
Extraction Tower
18-4-4 Extraction Tower Open
Loop Dynamics Enriching trays 8 5 5 7
Solvent trays 9 33 11 48
Distillation towers usually have well be- Stripping trays 13 14 14 14
haved open loop characteristics. The open Diameter (ft) 6 6 5 5.5
Reboiler duty
loop eigenvalues are negative and real. This (MMBtu/h) 21 19.6 19 26.8
results in open loop dynamics that decay Reflux ratio 0.9 1.5 1.5 1.4
exponentially without oscillation. Extractive Solvent/feed 1.0 0.8 1.0 1.2
distillation columns have similar open loop Weir height (in) 2.5 2.5 2.5 2.5
dynamics. One important exception is a Sump holdup
(lb mol) 150 150 170
caveat that the process engineer needs to Accum. holdup
consider during the design of an extractive (Jb mol) 30 33 25
distillation system: Too many trays in the Avg. tray holdup
solvent section of the extraction tower can (mol) 10 7 6
lead to slow overhead composition response
Solvent Tower
for changes in solvent flow. Figure 18-18
illustrates this effect. Enriching trays 9 5 5 3
The curves in Figure 18-18 are the open Stripping trays 11 3 3 8
loop responses of the extraction tower over- Diameter (ft) 6 4 3.5 3.5
Reboiler duty
head B composition for a 1% decrease in (MMBtu/h) 23 10.6 15 15.4
solvent flow. The two curves represent dif- Reflux ratio 2.0 0.28 0.1 0.16
ferent designs for System 2. The design Weir height (in) 2.5 2.5 2.5 2.5
overhead B composition is 0.01 mol%, the Sump holdup
(Ib mol) 300 250 250
solvent impurity is 0.005, and the solvent to
Accum. holdup
feed ratio is 1 in both cases. The extraction (lb mol) 40 40 60
tower for System 2 has 52 trays with 33 Avg. tray holdup
solvent section trays and the extraction (mol) 6 4 4
tower for System 2-Alt has 30 trays with
only 11 solvent section trays. The alterna-
tive design has considerably fewer trays! It
is important to note that neither system has Figure 18-18 shows that the alternative
the positive steady state gain effect de- design has better solvent process dynamics
scribed in the previous section. Table 18-1 and is the design of choice. System 2 takes
contains a summary of the process design almost 3 h to reach steady state from a
specifications for the three case studies. change in solvent flow. System 2 Alt reaches
Each system is designed with a feed compo- steady state in 30 min. Even if overhead
sition of 50 mol% A and B, and tray Mur- composition control is not important for
phree vapor phase efficiencies in all towers this tower, distillate disturbances from this
are 80%. Both systems are design for the tower will upset a downstream tower.
same product specifications. Although the
alternative system consumes about 3% addi-
18-4-5 Control Schemes
tional energy because it has less trays, the
total cost of both systems is essentially equal. El-Sl
We save in lower initial capital cost what we We found that the most economical control
pay in additional energy. scheme for our case studies was El-Sl. We
Process Design and Control of Extractive Distillation 399
control the stripping temperature in both peratures do not affect each other signifi-
the extraction and solvent towers by manip- cantly. This control scheme is treated as a
ulating the heat input of the tower's respec- diagonal control structure without any cross
tive reboiler. The solvent recycle and both terms. This was discussed in detailed in
reflux flows are constant. Chapter 11.
We held the solvent and reflux flows con- We select the desired control trays for
stant based on changes in feed composition. these control loops by the procedure de-
We want to also design our control system scribed in Section 18-4-2. We can then tune
for feed rate changes. Product compositions the two temperature control loops individu-
are controlled during feed rate disturbances ally because interaction is not significant. I
by fixing a constant solvent-to-feed and re- have found that the Ziegler-Nichols tuning
flux-to-feed ratio. Figures 18-13b through e rules for PI controllers work well for these
show schematics of these control schemes. types of control loops. The ultimate gain
The extraction and solvent towers inter- and period for the Ziegler-Nichols settings
act dynamically because the solvent recycle are found by relay testing (Luyben, 1990).
stream couples the distillation towers. The The loops are easy to tune, so you can use
EI-Sl control scheme is a single-input- any other single-input-single-output tuning
single-output control scheme in which the method with which you are comfortable.
two controllers do not interact. The solvent I used a simulation based on model 2A
recycle changes slowly relative to the strip- from Chapter 3 to simulate the towers and
ping temperatures and the stripping tem- control system. The simulation includes both
Extraction Tower
StrIpping Temperature Response
~
1\ A
170.6
!!
.~ ~ ~170.5
~ ; ~ 170.4 ~------'. ~----' ' - - - - - - -
.t; g" c
.,; ~ ~ 170.3
I-
170.2 WL--_ _ _ _ _ _ _ _ _ _ _ _ _ _ __
Solvent Tower
Stripping Temperature Response
204
!! 203.8
~a,.....
'8; !!! ~ 203.6
E E8..!l 203.4
III ~
~ 203.2
. 203~-------------------~
O.OOE+OO 2.00E+00 4.00E+00 6.00E+00 8.00E+00 1.00E+01
Time [Hrs.)
(a)
the extraction and solvent towers as a cou- the extraction tower is decreased from 0.5
pled system using the E1-S1 control scheme to 0.4 mole fraction of component A. The
shown in Figure 18-13b. simulation is run for 3 h. Then the feed
Figures 18-19a and b contain sample re- composition is changed back to 0.5. The run
sults for System 1 using this control scheme. continues for another 3 h before the feed
They illustrate the closed-loop response of composition is increased to 0.6.
the E1-S1 control scheme for extractive dis- These are significant disturbances to the
tillation systems. Figure 18-19a contains the column. Disturbances in feed composition
temperature response for the extraction and are much more difficult to control than dis-
solvent tower temperature control loops. turbances in feed rate. Feed rate distur-
Figure 18-19b contains the composition re- bance are controlled by the ratio controls
sponse of the important streams. These discussed before. Feed composition is usu-
results are generically representative of ally not measured, and cannot be controlled
extractive distillation systems. All three by feedforward schemes. Feed composition,
systems presented in this chapter respond unlike feed rate, also affects the steady state
similarly. temperature profile.
Three feed composition disturbances Inspection of the response curves in Fig-
were used to test the control system perfor- ures 18-19a and 18-19b reveals that E1-S1
mance. Initially the feed composition to provides excellent control for this system.
1.50E·03 1.50E·03
.....
t: 1.00E·03
"""I!
~~ 1.00E·03
)(15 5.00E·04
!.
5.00E·04 +------+---+---+---+-----< O.OOE+OO l - - - _ - _ - _ - t - - - - - o
O.OOE 2.00E 4.00E 6.00E B.OOE 1.00E O.OOE 2.00E 4.00E 6.00E B.OOE 1.00E
+00 +00 +00 +00 +00 +01 +00 +00 +00 +00 +00 +01
1.20E·02 0.993
0.992
.....
t: B.00E·03 t? 0.991
aI~ aI~
o 0.99
)( !
1/
01/
)('0
!. 4.00E·03 0.9B9
0.9BB
O.OOE+OO l - - - - + - - - - t------o-- 0.9B7 +--------+---+-----+----<
O.OOE 2.00E 4.00E 6.00E B.OOE 1.00E O.OOE 2.00E 4.00E 6.00E B.OOE 1.00E
+00 +00 +00 +00 +00 +01 +00 +00 +00 +00 +00 +01
There are steady state offsets in the product compositions to vary within 1% of design. If
compositions because product compositions operating costs increase significantly or
are not controlled directly, but they meet there is a need for tighter product composi-
our control system requirements. tion specifications, higher order control
schemes will be necessary. The higher order
Higher Order Control Schemes control schemes are E1-S2, E2d-S1, E2b-S1,
We have seen that E1-S1 is the most eco- E2d-S2, E2b-S2, E3-S2. These are diagonal
nomical control scheme. It is very easy to multivariable control schemes.
implement and provides good control. Our Figure 18-13 reveals that there are seven
economics are based on minimizing the to- higher order control schemes for extractive
tal cost of operation and control scheme distillation. All of these schemes are con-
implementation. I have shown that E1-S1 trollable if you follow the guidelines given
provides good economic control. Current at the introduction to process control of
operating costs do not justify higher order extractive distillation. It is beyond the scope
control schemes for these case studies. The of this chapter to detail each of these alter-
control objectives used allow the product natives. A more complete description can
0
Extraction Tower Stripping Temperature
A
II 170.6
~! ~
Ji E 170.5
170.4
;\
,--------'- ~---'- - - - - -
~ i ..... 170.3
I- 170.2 ' - " - - - - - - - - - - - - - - - -
O.OOE+OO 2.00E+00 4.00E+00 6.00E+00 6.00E+00 1.00E+01
Time IHr..)
1.50E.02,
~~ 1.00E·02 '-----
.. ~ 5.00E·03
!.
O.OOE+OO 1 - - - - - - - - - - - - - - - - -
O.OOE+OO 2.00E+00 4.00E+00 B.OOE+OO 8.00E+00 1.00E+01
Time [Hr..)
·1 i8. ~E ~~:
'C-
~ i ...
I-
203
202
t:----. \ /
V
2011------------------
\ ~
V
O.OOE+OO 2.00E+00 4.00E+00 B.OOE+OO 8.00E+00 1.00E+01
Time IHr••)
(a)
FIGURE 18·20. System 1 E2d·S2 (a) temperature response and (b)
composition response.
402 Practical Distillation Control
1.50E·03 1.50E·03
..
......
---l! ---i 1.00E·03
~~ 1.00E·03 !e1O-
"!
III "
"15 5.00E·04
!.
5.00E·04 O.OOE+OO
0 2 4 8 8 10 0 2 4 8 8 10
Time [HrLI Time IHrLI
1.20E·02 0.993
0.992
'ff 8.00E·03 'ff
iii I! 0.991
---I!
!e1O- ...... 10-
0 2 4 8 8 10 0 2 4 8 8 10
be found in Grassi (1991). Figures IS-20a I tuned the loops using the ATV and
and b illustrate the performance of the BLT methods (Luyben, 1990). There is sig-
E2d-S2 control scheme for System 1. This nificant interaction between the two control
system is a 4 x 4 multivariable control prob- loops within each tower. These must be
lem. We control two variables on the extrac- tuned as a multivariable diagonal controller.
tion tower and two on the solvent tower. A more detailed description of this can be
The two variables controlled on the ex- found in Chapter 11.
traction tower are a stripping section tem- The temperature control loops respond
perature and the overhead composition of very well. The extraction tower overhead
component B. Heat input controls the strip- composition loop is slow, but meets our
ping section temperature. Reflux controls control objective very well. Control loop in-
the overhead composition. An analyzer with teraction between the distillation towers
a 5 min deadtime is used to control the does not significantly affect the tuning of
overhead composition of component B. the temperature loops. However, intercol-
One stripping and one enriching section umn interaction must be taken into account
temperature are measured in the solvent to tune the extraction tower overhead com-
tower. Reflux flow and heat input controls position loop. I took this interaction into
the enriching and stripping section temper- account by identifying transfer functions for
atures, respectively. the extraction tower control loops with both
Process Design and Control of Extractive Distillation 403
towers coupled. I then used the BLT method of accurately representing these systems
to tune the 2 x 2 extraction tower control without unnecessary model complication.
system. Intercolumn interaction is not sig- Model 2A is also suitable for a personal
nificant for the solvent tower. Transfer computer. The columns presented in these
functions for the solvent tower can be iden- case studies run about 30 times faster than
tified with the tower decoupled from the real time on a 80386 based PC.
extraction tower. The results show that extractive distilla-
Intercolumn interaction can be recog- tion processes are controllable. The pre-
nized by examining the steady state process ferred control scheme is El-Sl. This control
gains for the entire extractive distillation scheme is very robust, easy to implement,
process. First compute the steady state pro- operate, and maintain. Higher order control
cess gains for the control system with the schemes are not needed unless you have
two towers coupled. Next compute the unique process requirements.
steady state gains for each column by itself, It is very important that recycle composi-
that is, uncoupled. If the gains are similar, tion and flow are given enough conservatism
then the control system can be tuned with- during the process design to ensure that the
out considering intercolumn interaction. system is controllable. These considerations
However, if there are significant differences are very important when you want to imple-
you will need to consider intercolumn inter- ment control schemes of higher order than
action. For the case studies presented in El-Sl. Effective design of extractive distilla-
this chapter, System 1 has significant inter- tion processes relies on a strong interaction
column interaction; Systems 2 and 3 do not. between process design and process control.
The E2d-S2 control scheme is the highest
order control scheme needed to control the
process. Because all product compositions References
Andersen, H. W., Laroche, L., and Morari, M.
are under control with E2d-S2 we do not
(1991). Dynamics of homogeneous azeotropic
need E3-S2. E3-S2 is more difficult to con-
distillation columns. Ind. Eng. Chem. Res.
trol than E2d-S2 because it adds a control 30(8), 1846-1855.
variable (the solvent flow) to the multivari- Atkins, G. and Boyer, C. M. (1949). Application
able control problem. I have found for each of the McCabe-Thiele method to extractive
of these systems that the E3-S2 control distillation columns. Chem. Eng. Prog. 49(9),
scheme is not needed, and El-Sl is eco- 553-562.
nomically the best. Benedict, M. and Rubin, L. C. (1945). Extractive
and azeotropic distillation-Parts I and II.
Trans. AIChE 41, 353-392.
18-5 CONCLUSIONS Black, C. and Ditsler, D. E. (1972). Azeotropic
A methodology to design and control ex- and Extractive Distillation. Advances in Chem-
tractive distillation systems has been pre- istry Series. Washington, DC: American
sented. Three general extractive distillation Chemical Society, Vol. 115, pp. 1-15.
systems illustrate the methods and highlight Chambers, J. (1951). Extractive distillation de-
sign and applications. Chem. Eng. Prog.
the important insights. I hope that these
47(11), 555-565.
insights will prove useful to you in your
Doherty, M. F. and Caldarola, G. A (1985).
future work with extractive distillation sys- Design and synthesis of homogeneous
tems. azeotropic distillations III. The sequencing of
The procedures are suitable for applica- towers for azeotropic and extractive distilla-
tion on small personal computer systems. tions. Topical Report, University of Mas-
The process design procedure runs in less sachusetts, Amherst, MA.
than 1 min on a Pc. A dynamic simulation Doherty, M. F. and Perkins, J. D. On the dynam-
using model 2A from Chapter 3 is capable ics of distillation processes-III. The topolog-
404 Practical Distillation Control
ical structure of ternary residue CUlVe maps. Lee, F. and Coombs, D. M. (1987). Two-liquid-
Chem. Eng. Sci. 34, 1401-1414. phase extractive distillation for aromatics re-
Gerster, J. A. (1969). Azeotropic and extractive covery. Ind. Eng. Chem. Res. 26(3), 564-573.
distillation. Chem. Eng. Prog. 65(9),43-46. Luyben, W. L. (1990). Process Modeling, Simula-
Grassi, V. G. (1991). Process design and control tion and Control for Chemical Engineers, 2nd
of extractive distillation. Ph.D. Dissertation, ed. New York: McGraw-Hill.
Lehigh University. Prokopakis, G. J. and Seider, W. E. (1983). Dy-
Kumar, S. and Taylor, P. A. (1986). Experimen- namic simulation of azeotropic distillation
tal evaluation of compartmental and bilinear towers. A/ChE J. 29(6), 1017-1029.
models of an extractive distillation column. Shinskey, F. G. (1983). Distillation Control. New
Report, Department of Chemical Engineer- York: McGraw-Hill.
ing, McMasters University.
19
Control by Tray
Temperature of Extractive
Distillation
John E. Anderson
Hoechst Celanese
405
406 Practical Distillation Control
S 60
~ ~i.tillate
Solvent I~ I
I ,I lCD.%~
I
c?-S/F
I
I
II
I
I'F
I
Feed I
,I
Tsp
HAc
-,
I
Product
I
.!.Q
T1
lCB,%!oFo
Solvent
Bottom
nGURE 19-1. Old controls.
XD is the weight percent of HAc in the 19-2 and 19-3 show the resulting contour
main column distillate. plots of XD and XB on a grid of the
XB is the weight percent of HFo in the manipulated variables: reflux flow versus
main column bottom product on a sol- distillate flow and solvent flow versus distil-
vent-free basis. The HFo analyzer on the late flow. From Figure 19-2 we see that
distillate of the recovery column gives the reflux flow has almost no effect on either
same value but is delayed. XD or XB over the range of reflux rates
examined. Reflux flow would be of little to
To get a better understanding of how XD no use as a manipulated variable. Thus the
and XB change with reflux, solvent, and practice of operating with a constant reflux
distillate rates, a number of simulation cases rate is a reasonable one.
were run using a tray to tray steady state The choice of the pairing of control and
simulation of the main column. A constant manipulative variables is not so obvious from
feed rate was assumed for all cases. Figures Figure 19-3. However, a graphical calcula-
Control by Tray Temperature of Extractive Distillation 407
,, ,,
40
\ \
I
I I
I
,,,
tJ!.' tJ!.1 tJ!.1
tJ!.1
CII' 11)1 ,.. .... 1
..;' ci'
,,
III
36 ~l,
1 ,
,,,
I
f0-
,,r,
,,
tJ!. tJ!. I
, tJ!.
II) ~
M
tJ!. N I II
28 ~
,.. CII I
I
I ~
I
I
I
I
24 I I I , I I, I
I
6.6 6.8 7 7.2 7.4
Distillate flow
110
90
70~--------~~------~L---~----~~------~
6.6 6.8 7 7.2 7.4
Distillate flow
FIGURE 19-3. XD and XB contours on solvent versus distillate.
408 Practical Distillation Control
tion will help. Step from point A to point B where XB = percent HFo in the overhead
along a horizontal line by changing distillate of the solvent recovery col-
flow while keeping the solvent flow con- umn
stant. Step from point A to point C using T = tray temperature
the same size change in distillate flow but R/S = reflux flow to solvent flow ra-
move along an XD contour line. Project tio
point C up to point D on the horizontal a, b, c = constants
line through A by moving along an XB
contour line. The line segments AB and The ratio of reflux flow to solvent flow was
AD are the changes in XB for a given used as an indicator of the ratio of the light
change in distillate flow at constant solvent components to solvent on a tray. Other
rate and constant XD, respectively. Then models could be developed from theory.
the ratio AB / AD is the element of the However, the empirical model of Equation
relative gain array corresponding to the 19-1 was quite adequate over the normal
pairing of XB and distillate flow. Because range of operations. Regressions of plant
the ratio is greater than 0.5, changing distil- operating data to Equation 19-1 gave both
late flow is the better way to move between good fits of the data and the values of the
XB contour lines. Stated another way: XB model constants.
is best controlled by manipulating distillate Before doing the regressions on the oper-
flow. However, under the constraint of con- ating data, it was necessary to shift the XB
stant reflux flow any change in steam flow is data forward in time so as to reconcile it in
translated into a corresponding change in time with the T and R / S data from the
distillate flow. Thus manipulating steam flow main column. Of the existing tray tempera-
is equivalent to manipulating distillate flow tures on the main column it was tray 10
when the reflux flow is held constant. temperature that gave the best correlation.
In order to develop a tray temperature as Rearranging Equation 19-1, a better state-
an inferential measurement of XB, it helps ment of the relation is obtained for our
to think of the liquid on a tray as being a purposes:
mixture of solvent and light components.
XB =a + beT + (c/b) *(R/S» (19-2)
The light components are the mixture of
HFo, water, and HAc in the tray liquid. The TC == T + (c/b)(R/S) (19-3)
r<§) @
!"'R I~i.tillate
I
S 60
Solvent ~, ,I lCD.%HAc
1I
~-S/F
1I
I
!"'F
I
II
Feed
II
10
4: 1
I
I"'T
HAc
Product
. . %tFo
Solvent
Bottom
FIGURE 19-4. New controls.
the old controls except that the measure- would be to automatically update a bias to
ment to the temperature controller has been the compensation equation based on the
replaced by a compensated temperature as HFo analyzer output.
calculated by Equation 19-3. The operator An examination of some operating data
is able to select a setpoint for the compen- shows both the need for and the benefits of
sated temperature corresponding to the de- using a compensated temperature in the
sired XB. Then the column will control at control scheme. Figure 19-5 contains plots
that compensated temperature and its cor- of data collected before the new controls
responding XB even though the feed and were applied. The solid curve in Figure
the solvent rates change. If the compensa- 19-5b is the measured temperature on tray
tion Equation 19-3 is not exact, the operator 10 and the dotted curve is the calculated
may find it necessary to make small adjust- compensated temperature. Solvent flow is
ments to the temperature setpoint based on the solid curve in Figure 19-5c. Note that
the average output of the HFo analyzer. the measured temperature changed during
Another method to correct for model misfit the four days of operation even though the
410 Practical Distillation Control
5~------~------~------~------~------~------,
4 ........................................................................... .
-
~ 2 ···················1···················
·· ............................... ······t··················· .................. .
...
ai' · ....
X ···· .
0
(0)
128
. . .
. .. .
. ../.<:~.~~.l.. . . . . . . ~. . +.-.~:\~~ . . .l.........nnoc
_____ , •• , •• : •••••••••••••...••• : ••••••••.•••••••••• : ••••...••••••••••••
126 o~ ~.-::,.#Io
..,. . .,
[/:·.\<J::.,
••••••• o •• •• o.o . : •• ' _ ' •• 0 .0 •• 0 •••••
U
(!) 124 .;..;..;;0_ __
I.&J ~ ~ ~ ~
0
I \:' \ _ . ___ : _____ _
..
0 ••••••••••• '0 •• ; • • • • 0 • • • • • • _0 • • • • • •
U
0
. · .. ·· . .. ...
...
.-
I-- 120 ..................................................................................................................
·· ..
·· .
.. . ....
cItS . .
......... ; ................... ; ........... ····j···················i···················
0 118 ·· . . ..
.-
I-- ,..-:---"", :.. :.. .:
............ ·, ...................•...................
116 ···················t···············
····
.. ··!···············
....
...
··· ..
·· .
··· .....
114
(b)
150~------~------~~------~------~------~------~
UI
125 ..........................................................................................
·· .. ...
~
... Solv
....
o
c;: ·· .... ..
··
· .
C 100 ............................................................................
··· ... ... ----_ ............ -
Rflx
~ ··· .. .
.
> 75 .................. :................... .:. ................... .. .................. ................... ................. ..
(5 ~ ~ ~
· . .· ..
r
Vl
··· ..
.
cItS 50 ................... :.......,............ :................... : ................... ;................... :.................. .
x
::s ·~ .~ .~ ....... --:: ~
'; 25 .~ =:.~: .-.~ ::.: ~.: ~r :.~ =:.: :.~ =:.: ~.: ~.:: ~.: ~:.:~.: ~.: r~.:~.: :~.:: ....... ·1 .... ······· ....... '1' ................ ,.
c::
O~-------r-------T------~------~r-----~r-----~
o 24 48 72 96 120 144
Time. Hours
(c)
s~------~------~--------~------~------~------~
4 .................................................................... .
m
X
. .. .... ...·................... . .. .. .. ..
~ ~. ...
0
(0)
128
126 ·· ....
.................................................
U ··
<.!) 124
w
0
122 ................... ................... :................... :.................... Tl0C
~
U
0 ··· ... ...
.. .
.-
t-
· . .
....
120 .......................................................................................................................
···· .... .... ....
·· .. .. .. ...
118 ................... ,· ................... .; .................. .; ................... .·; ...................•...................
~
0...... ..
t- ··· .
...
·
116 .................. ·! ................... .: .................. , .................. ' ..................................... .
·· ...
·· .
114
(b)
150,---------~------~--------~------~~------~------~
75 ................... ~ ................... !................... ~ ............ .. ·· .. ·: .......... ·.. ·.. ···i··· .. ········ .. ·.. ·
"
413
414 Practical Distillation Control
addressed during the design of plantwide the process but have no obvious source or
control strategies. The interaction of unit sink. The rate of formation and of con-
operation control strategy design with the sumption of these components is set equal
overall plantwide control strategy is illus- by mechanisms such as reaction equilibrium
trated using examples. A case study is then or concentration dependent irreversible re-
presented to show in detail the relationship actions. For components that occur in sig-
between column control strategies and an nificant quantities, the control strategy for
overall plantwide control strategy. the unit operations must take into account
how the inventory of each component will
20-2 PLANTWIDE COMPONENT be maintained.
The material balance equation for each
INVENTORY CONTROL
component can be written as
20-2-1 Component Inventory Control
input + generation - output
for a Tank
- consumption = rate of accumulation
The control of the overall material balance
for unit operations is usually taken for Based on the terms of the material balance
granted because of the obviousness of an equation one can identify the stream flows
imbalance of material flows. For the overall and the process operating conditions that
control of inventories that are liquids, we affect the accumulation or depletion (Le.,
typically control liquid levels; for gas inven- inventory) of each component.
tories we control pressures; for solid inven-
tories we control weights. When the unit Example 20·1
operation equipment overfills or goes empty, Consider the tank shown in Figure 20-1.
the underlying cause must be found and Water is the feed to the tank and the water
corrected for the unit operation to have any leaving the tank is determined by the height
chance of operating as intended. However, of liquid in the tank, which follows the
a much more subtle problem is the control
of component inventories. Each component,
whether important or insignificant, must
have its inventory controlled within each
unit operation and within the whole pro- water Feed
cess.
The control of component inventories
may occur by process self-regulation or by
manipulating process flows or conditions.
For example, many processes have innumer-
able trace components that "recycle to ex- Water Outlet
tinction." That is, they are ever present in FIGURE 20-1. Water tank with gravity outlet flow.
Enters Leaves
Through Produced Through Consumed
System Within the System Within the Inventory
Component Boundaries System Boundaries System Controlled by
equation
Fout = kv(height) water Feed
Each term of the material balance equation
can be identified as shown in Table 20-1.
For this example the inventory of water
in the tank is "self-regulating." For an in-
crease in water feed, the inventory of water water Outlet
(41)
increases, the flow out of the tank increases,
and, as a result, the balance between the
~
SP__________ ~~FI
Enters Leaves
Through Produced Through Consumed
System Within the System Within the Inventory
Component Boundaries System Boundaries System Controlled by
Enters Leaves
Through Produced Through Consumed
System Within the System Within the Inventory
Component Boundaries System Boundaries System Controlled by
........... ............:
r···· ..
.•
o
::~..~ An ...!l
~ ..... ~
.
cl>...........
o
(>
rE\····0
y • • cy....... (.)
•••• -_.. I• . . . . . . . .
N
y
418
<r. . . . . . . 0
I:
o
•
I:
.w
....................... o ......
~ L.
• 01
:...... : ~ ....
~ ..... \.V
:::.. ~ ~..... Ji
1:.01: UO"':SI':
.
cl>...........
o
o
roo ................ . ....................&
'l-'
.
...
.
t•....JU\ ..
~.-
~
o <y····0
419
420 Practical Distillation Control
Enters Leaves
Through Produced Through Consumed
System Within the System Within the Inventory
Component Boundaries System Boundaries System Controlled by
control of the individual unit operations may If the flow of stream 1, Fl , and stream 2,
be satisfactory. F2 , are fixed, then there is no value of D
Neglecting the small amount of water that will satisfy both equations unless lOFl
that leaves in the vent, the outlet flows for = 1.11F2xMeOH2' that is, the feed rates of
methanol and water are given by methanol and ~ater are in a 90/10 ratio.
Because of the fractionating ability of the
(flow methanol out) = (distillate flow)(90%) distillation unit operation, the methanol will
= 0.9D concentrate in the top of the distillation
(flow water out) = (distillate flow)(lO%) column. The column control strategy will
ensure that distillate stream flow will satisfy
= O.lD the removal of the methanol from the pro-
The inlet flows of each component are given cess. The distillate rate will follow the equa-
by tion D = 1.11 F 2 X MeOH 2 (= 11.1 mol/h).
For a 10 mol/h feed iate of water and a
(flow methanol in) water removal rate of 1.11 mol/h (= 0.1
D), the water will accumulate in the process
= (flow of stream 2) at the rate of 8.89 mol/h.
x (fraction methanol in stream 2) Consider the control strategy illustrated
in Figure 20-5b. By virtue of the level con-
= F2 x Me O H ,2 troller on the base of the scrubber being
(flow water in) = (flow of stream 1) = Fl linked to the feed rate of water to the
process, the water inventory is self-regulat-
The overall component balances that result ing. As more methanol enters the process
for each component are and thus drags with it more water from the
process in the column distillate stream, less
Water: water is returned to the scrubber from the
base of the column. As a result of less water
Fl = O.lD => D = lOFl being fed to the scrubber, the scrubber base
Methanol: level falls. The level controller for the
scrubber base increases the amount of wa-
F 2 xMeOH,2 = 0.9D => D = 1.11F2 xMeOH,2
ter being fed to the process. The balance of
Distillation Control in a Plantwide Control Environment 421
Enters Leaves
Through Produced Through Consumed
System Within System Within Inventory
Component Boundaries the System Boundaries the System Controlled by
production rate in the reactor, then B will A is no longer self-regulating and the feed
tend to accumulate in the tops of the distil- rate of A must now be manipulated to
lation columns due to the fractionation of control the inventory of A in the process.
A, B, and carrier. The removal rate of B The carrier component inventory is not
from the column is typically adjusted to self-regulating. Because it is not entering or
achieve a temperature or composition tar- leaving the process, its inventory in the pro-
get in the column. As more B is produced, cess should be constant. However, there is
the column control scheme must result in no mechanism to prevent small amounts
more B being removed as product to satisfy being lost in vents, the product stream,
plantwide component inventory require- sludging operations, and so forth. In prac-
ments. tice, small amounts of captive components
The alternative to using the product rate are lost over time and the inventory of the
to control the plant inventory of component carrier may be made up batchwise on a
B is to manipulate the production rate of B daily or weekly basis.
by reaction. The reactor conditions of level,
temperature, or concentration of A can be
Example 20-6
changed to alter the formation rate of B to
match the product rate. When this choice is Consider the process described in Example
made, the process inventory of component 20-5, Figure 20-6. To illustrate the interplay
~ •
r-.....
1
i
,
_-t---.l
~-------------------!
!L __
i
i
i
L : ! --.@------------l
I :
[3-:-8 : !
Feed A
---< Steam
I
[3":-8
I
I
[3J-8
Recycle
------I Cond >
(a)
01
rFCL--,
i i
i !
_-.--....'
~-------------------,
i i
L__
Y
i i',
I -
L
I
I --.@-----·------l
[3+eJ
I
!
Feed A I
[3~
I
I
---< Steam
[3J-8
Recycle
------I Cond >
(b)
,..
••
II'
ow 0 . . . . . . . . . . . ..
ow
o
...
..
III
. ...
424
Distillation Control in a Plantwide Control Environment 425
Stream Number 2 3 4 5
Stream Label Air feed Ox vapor Vent Water feed Scrubber
product to scrub underflow
Flow (Ib-mol/h) 120.000 124.947 102.288 2.58855 25.2475
Temp (deg C) 25.0000 127.065 20.0000 25.0000 20.0000
Stream Number 6 7 8 9 10
Stream Label HAc feed LBcol Ox LBcol Water col
distillate underflow underflow underflow
Flow (Ib-mol/h) 40.0000 66.3519 107.373 41.0207 38.5594
Temp (deg C) 25.0000 40.0000 127.065 120.239 125.000
Stream Number 11
Stream Label Water col
distillate
Flow (Ib-mol/h) 2.46127
Temp (deg C) 100.000
and 193.4 psia with a liquid holdup of 20 loop and cooler at the bottom. Most of the
lb-mol 02.1 ft 3 of liquid). recovery is due to the cooling of the vapor
The vapor stream leaving the reactor product to condense the acid and acetalde-
flows to a water scrubber to remove acetic hyde.
acid and some acetaldehyde from the vapor The liquid product leaving the reactor is
stream before the stream leaves the process the feed to the low boiler (LB) removal
boundaries. The scrubber uses water as a column. The primary purpose of the LB
scrubbing medium and has a recirculation column is to recover the unreacted ac-
~
66.4 aol/br
0.0415 Acat Aid
0.0176 Acat Acid
0.9331 W.tar
22L--1Vl
17L.-lV\
---~cXJ .1 11
41.02 aol/br
0.94 Acat Acid
0.06 Watar
FIGURE 20·9. Low boiler (LB) distillation column.
Distillation Control in a Plantwide Control Environment 427
etaldehyde and recycle it to the oxidizer. In ficult to state an operating objective. Com-
addition, water is recycled to the oxidizer to mon perception is to assume that the
maintain a water concentration of 0.6 mol operating objective is to separate the feed
fraction in the oxidizer. The acid product components into two streams having com-
and some water leave the bottom of the LB positions equal to the compositions given in
column and feed the water column where the base case or design basis.
the water and acid are separated into rela- The control system design needs to
tively pure product streams. provide disturbance rejection properties ap-
The focus of this case study is to develop propriate to the type and nature of the
an understanding of how the control of the disturbances. Typically, feed rate and feed
LB distillation column affects the overall composition are common disturbances that
plantwide control strategy. the control system must deal with. In addi-
tion, variability in the utility supply, envi-
ronmental upsets, and slow degradation of
column internals are common as well.
20-3-2 Low Boller Column Analysis
The low boiler (LB) distillation column Steady State Analysis
sketched in Figure 20-9 is used to separate The steady state gains relating the column
acetaldehyde (nbp 20 e), water (nbp 100 e),
0 0
temperatures to the manipulated variables
and acetic acid (nbp 115°C). The column are determined using a rigorous steady state
has 24 crossflow trays, a thermosyphon re- simulation of the column. An example cal-
boiler, and an overhead total condenser. culation is the determination of the gains
The column is operated at atmospheric relating temperature on tray j to the reflux
pressure by venting the primary condenser rate while holding bottom product rate, feed
to a dedicated column vent scrubber and rate, and feed composition constant. Being
then to the atmosphere. Measurements on a steady state calculation implies that the
the column consist of temperatures on trays distillate rate and steam rate to the reboiler
4, 8, 17, and 22. No composition measure- are used to control reflux drum level and
ments are available except by lab sample base level, respectively. The gains are found
once per 8 h shift. Differential pressure type using the equation
level transmitters provide level measure-
ments for the base of the column and the (aL
a~) "" ~(L + I1L) - ~(L - I1L)
reflux drum. B,F'XF 211L
The column material balance is listed in
Table 20-6. The vapor-liquid equilibrium where Tj(L + ilL) is the temperature on
relationships for this column reveal that the tray j corresponding to a reflux rate of
acetaldehyde has a high relative volatility to L + ilL and similarly for Tj(L - ilL). The
water and acetic acid and as a result it is value of ilL used for this numerical calcula-
removed exclusively in the distillate prod- tion of the gains is 1 x 1O- 6L. Steady state
uct. The water has a low relative volatility to process gains between the column tempera-
acetic acid of about 1.15. tures and the manipulated variables reflux
Base case temperature and composition L and bottom product rate B are plotted in
profiles illustrated in Figure 20-10 indicate Figure 20-11. Steady state gains for other
that the key split is between the water and combinations of manipulated variables are
acetic acid, and that the temperature found in a similar manner.
changes from tray to tray are the greatest in For each pair of manipulated variables a
the stripping section of the column. matrix of steady state gains relating the
Without knowing the purpose of this col- temperatures to the manipulated variables
umn from a plantwide viewpoint, it is dif- is calculated. Each matrix has 25 rows (tem-
428 Practical Distillation Control
125~--------------------------------------------~
120
115
110
105
100
1 5 10 15 20 25
fEMPERATURE (DEG C) VS. STAGE NUMBER
(a)
0.8
0.6
0.4
1 5 10 15 20 25
0.4
0.2
E-07
-0.2
-0.4
1 5 10 15 20 25
1 =L 2. B 3 .Fl
FIGURE 20·11. Low boiler column steady state gains.
peratures) and 2 columns (manipulated peak in the curve for !l at trays 4 and 5
variables). Singular value decomposition can indicates that the temperature on those trays
be used to estimate good locations for tem- will be the most sensitive to changes in the
perature measurements. A detailed descrip- manipulated variables. The peak in the
tion of singular value decomposition and curve for !2 occurring on trays 15 to 18
how it is used in the design of distillation indicates that this zone is the second most
column control strategies is presented in sensitive location in the column whose tem-
Chapter 8. perature movement is independent of trays
The singular value decomposition of the 4 and 5. The two ! vectors define output
gain matrices can be written as directions that the column temperature pro-
file will follow when changes are made to
the manipulated variables. The! vectors for
other sets of manipulated variables [(L, V),
(D, V), (LID, VIB), and (LID, B)] are the
same as those for (L, B).
The SVD results suggest using tray 4 for
single point temperature control and trays 4
and 17 for two point temperature control.
The results for the manipulated variable set The choice of trays 4 and 17 for two point
(L, B) are shown in Figure 20-12. The val- control is based on the attempt to use col-
ues of the elements of!h and !2 are plotted umn temperature measurements that con-
versus tray number as curves 1 and 2. The tain the most independent information
430 Practical Distillation Control
0.4
0.2
0.0
1 5 10 15 20 25
FIGURE 20-12. Low boiler column SVD !! vectors.
about the column. Selection of temperature 20-7 lists four steady state measures evalu-
measurements based on this criteria does ated for five sets of manipulated variables.
not include the issue of how well these The RGA is listed for the pairing (T17-m t )
temperatures correlate with end composi- and (T4-m 2 ). It is desirable to have the
tions. relative gain All close to 1, the condition
After choosing the location of column number (CN) small, the second singular
temperature measurements, the many value U 2 large, and the intersivity index
choices of manipulated variables to be used u2/CN large. Using these guidelines the
for temperature control can be screened choice of (L, B), (LID, VIB), and
using various steady state measures. Table (LID, B) appears to be better than (L, V)
O'zlCN
ml m2 Au CN 0'2 (X 106 )
101. 4ELB_C_O_L_T-_1_7_ _ _ __
101.1-1_f--
_ _, -_ _---._ _---, 125.0-+-_ _-,-_ _-..,._ _---,
I
116'ILB
COL T-' STEAM FLOW
115 .O~--------
I I
7.000 LB COL UF X-H20
xlO- 2
TIME (HRS)
FIGURE 20-13. Low boiler column response using single-end temperature control.
432 Practical Distillation Control
E
-2
x10
1.5000
T-17 135.
4
lO1.
101.1 I 125.0
7.0::fl UF X-H2O
115.0 I
I
x10
5.0000 I 40.00
I I
0.00 1. 33 2.67 4.00 0.00 1.33 2.67 4.00
TIME (HRS)
FIGURE 20-14. Low boiler column response using dual-end temperature control.
than the single end control strategy would 20-3-3 Component Inventory
be preferable. However, if the top composi- Control Analysis
tion were not critical, then the simplicity of
A candidate overall control strategy for the
the single end control strategy may be de-
acetaldehyde oxidation process is illustrated
sired. The key point is that the choice of
in Figure 20-15 using the single point tem-
strategy in this case is more dependent on
perature control on the LB column that
operating objectives than on control pedor-
resulted from the LB column analysis. The
mance. The determination of the overall
tray 4 temperature controller for the LB
plant operating objectives and how they in-
column does a good job holding the bottom
fluence the objectives of the distillation
composition of the LB column constant.
columns within a process is key to the col-
Consider the material balance envelope
umn strategy design.
shown in Figure 20-15. Based on the candi-
From the local analysis of the LB column
date plantwide control strategy, the mate-
we have determined that both single- and
rial balance for each component can be
double-ended temperature control are
listed as in Table 20-8 for the envelope
workable. Because single-end control works
shown.
about as well as dual-end control and is
Provided that the reactor is sized large
simpler to implement, we will consider the
enough, the inventory of acetaldehyde w~ll
use of single point temperature control us-
self-regulate similarly to component A In
ing tray 4 and manipulating the column heat
Example 20-5. The oxygen and the nitrogen
input. If the designer of the column control
will self-regulate similarly to the water and
system is not aware of the overall plant
objectives, this strategy is a likely design. methanol in Example 20-3. The pressure
r---
I
I r»@ I ,OU, ••
I :, :'
I _____ / Material Bal ance Envelope
····43 L__
---- -,
.~
o
"i
~ .. j •
i PC
~L..j
"•
, '
:I
r •..
r··<;9 [.......-8... ! 1-4·.1 C
0
C 1
o ..
1 I
~ I : 1-------
r·-€> I
1(00
&..
4 •• 3 &' PC •
I
I
'
i
1
I '' r--: I i
I i
- - - - - - - - - - - - - - - r. ............. _
: L ________________________ ,
- - - ______ 1
FIGURE 20-15. Acetaldehyde oxidation process control strategy.
~
434 Practical Distillation Control
TABLE 20-8 Component Material Balance for the Acetaldehyde Oxidation Process
Enters Leaves
Through Produced Through Consumed
System Within System Within Inventory
Component Boundaries the System Boundaries the System Controlled by
controller is analogous to the level con- cess will change. In particular this change
troller in that example. The acetic acid will will manifest itself in the composition of the
self-regulate via the temperature controller oxidizer.
on the LB column similar to the methanol For example, assume an increase in the
in Example 20-4. The water is not self-regu- scrubber water feed. The inventory of water
lating. in the process will begin to increase in the
The temperature controller at the bot- oxidizer. As a result the feed composition to
tom of the LB column provides good com- the LB column will become richer in water.
position control of the bottom product However, the tray 4 temperature controller
stream. If we assume that the bottoms com- in maintaining the 4/96 water to acid ratio
position is held constant by the temperature in the bottom stream forces the extra water
controller at the base case value of 4% in the feed out the distillate stream and
water and 96% acid, consider the conse- back to the oxidizer. The oxidizer then con-
quences of changing either the scrubber wa- tinues to become a higher and higher frac-
ter feed or the acetaldehyde feed rate. If tion water. In practice the water concentra-
either rate is changed, then the ratio of tion becomes so high as to stop the oxida-
water to acid that must exit the process is tion reaction and results in a plant shut-
no longer 4/96. Because the LB column down. A similar scenario occurs following a
temperature controller maintains this ratio change in acetaldehyde feed rate.
in its bottoms stream (and hence the pro- From a plantwide viewpoint it becomes
cess exit streams leaving the water column), obvious that the composition of the bottoms
the inventory of water and acid in the pro- stream leaving the LB column must be al-
Offg ••
.... ~
..~
o
RAcl ~
•i
d .. ~
&. 1 i 'r€f
• .~ ..:
=
t I
•r
e'J
f"~ PC .~
L
AirJ ; Y .1 ~ C
• : 'l'C 0
C 1
~ . . .-e. . o ......... J
1 : 1.......
..
•
D 1
&1 : !-_"
<y"
~--
. ......... ---_ ... --- ...... -.-.--- ....
FIGURE 20-16. Acetaldehyde oxidation process control strategies without control of LB column tray 4 temperature.
~
Vl
.. ~~
...~ r···········
....... u ......
... ~ ··
-r ··
". uo . . . . . ·
·
¢¢:.........•
@...
&>...........j
f································
0 . . . . . . . . . ..
. . <p <p.
: .
::(1 err·
I. •••••
u
c ...c.
•
436
0.615 0.700
0.610
0.675
0.605
0.650
0.6001 9 ~~ 9 9 0
0.625
0.595
0.600
0.590
0.585 1 I I 0.575+1---,--.,----.-----r--r---.-----r--.---.--...;
0.850 1.000 1.150 0.850 1.000 1.150
OX X-H20 vs. Fractional Changes in: Ox X-H20 vs. Fractional Changes in:
A - AIR FEEO 120.00 B - HAC FEED 40.000 A - AIR FEEO 120.00 B - HAC FEED 40.000
C - WATER FEED TO SCRUB • 2.5886 D - REACTION FOR CON, AF. 49437. C - WATER FEED ToJ SCRUB 2.5886 0 - REACTION FOR CON, AF. 49437.
(a) (c)
3.5 3.5
3.0 3.0
2.5 2.5
2.0 2.0
1.5 1.5
1.0 1.0
0.541--~---r--._--~-.--~--~--~--~-. 0.5'1---.---.r--.r---r---r---r---r---r---r-~
0.B50 1.000 1.150 0.850 1.000 1.150
HAC LOSS vs. Fractional Changes in: HAC LOSS vs. Fractional Changes in:
A - AIR FEED 120.00 B - HAC FEED 40.000 A - AIR FEEO 120.00 B - HAC FEED 40.000
C - WATER FEED TO SCRUB • 2.5886 D - REACTION FOR CON, AF. 49437. C - WATER FEEO TO SCRUB • 2.5886 o - REACTION FOR CON, AF • 49437.
(b) (d)
~ FIGURE 20-17. Acetaldehyde oxidation process disturbance analysis.
438 Practical Distillation Control
lowed to float to match the ratio of water distillate stream that recycles back to the
being fed and acid being produced in the oxidizer. The control strategy shown in Fig-
process. The control and operation of the ure 20-16b has no feedback control on the
LB column is the key to controlling the LB column. The reflux and the reboiler
inventory of the water and acid. duty are simply ratioed to the feed rate.
This strategy accounts for process rate
changes but allows the top and bottom com-
20-3-4 Control of the Acetaldehyde
positions to float.
Oxidation Proc888
A disturbance analysis as described in
To control the inventory of acid and water Chapter 6 is carried out to determine the
in the process, the component inventory effect of four disturbances on the acetalde-
analysis suggests that the composition of the hyde losses and the oxidizer water composi-
bottom product of the LB column must be tion. In practice other disturbance analyses
allowed to vary as the feed ratio of water are done for other disturbances and other
and acetaldehyde changes. To illustrate how process variables as well. Figure 20-17 illus-
the control of the LB column influences trates the results for the two control strate-
overall plantwide control, two control gies. The strategy that controls tray 17 tem-
strategies are compared from a steady state perature has a greater sensitivity of oxidizer
disturbance rejection viewpoint. water composition and acetaldehyde loss to
Figure 20-16 illustrates two plant control the disturbances than the LB column ratio
strategies that differ only in how the LB strategy. In particular, for a change in air
column is controlled. In Figure 2O-16a, the feed rate to the oxidizer, the ratio strategy
top composition of the LB column is con- results in much less sensitivity of the oxi-
trolled by means of the tray 17 temperature dizer water composition. Other variables
controller, which manipulates the reflux rate such as the sensitivity of acetaldehyde loss
to the column. This strategy attempts to to acetaldehyde feed rate are not affected
hold a relatively constant composition in the much by the choice of the LB column COD-
70.0
127.~P""UR'
0.6101 ox X-_H-20-_ _ _ __
1.850
3.0_40Eox
X-HAC 8.0:0pLBCOL UF X-H20
x10 2 xl0 2
100.0-j-_ _-r-_ _, - - _ - - ,
0.00 2.67 S.33 8.00
101. 1 +
0.00
I
2.67
I
5.33
I
8.00
TIME (HRS)
FIGURE »18. Acetaldehyde oxidation process response for a change in the water
feed rate to the scrubber. Control strategy shown in Figure 20-16a.
Distillation Control in a Plantwide Control Environment 439
126.5~ 65.00~
J :;:::S'-HOAC
I I I
0.610 OX X-H20
o. 6000_L~-------
I
1.6500~ I
3.0_40EOX
X-HAC 8.0:0ECOL
LUFX-H20
B
IOu57____
xlO 2 xlO 2
TIME (HRS)
FlGURE 20-19. Acetaldehyde oxidation process response for a change in the water
feed rate to the scrubber. Control strategy shown in Figure 20-16b.
trol strategy. This limited disturbance analy- strategies will admit different LB column
sis suggests that the ratio strategy for the control choices. Each strategy will have
LB column may be the better choice if strengths and weaknesses when faced with
oxidizer water composition is important. different process disturbances. The job of
A dynamic simulation is used to compare the control system designer is to ferret out
the dynamic performance of the alterna- the strengths and weaknesses of each strat-
tives. Figures 20-18 and 20-19 illustrate the egy while ensuring that the component in-
response of the process to a water feed rate ventory control is maintained.
change to the scrubber for each candidate
strategy. For this disturbance, the strategy
that has no temperature or composition
20-4 CONCLUSIONS
control on the LB column performs as well
as the strategy where tray 17 temperature is The development of control strategies for
controlled. Other dynamic simulations can distillation columns is highly dependent on
be performed to compare results for other the overall process inventory control objec-
disturbances. As a result of the limited dis- tives. Distillation column control strategies
turbance analysis and dynamic simulations, must be designed using an understanding of
the ratio strategy with no temperature feed- how column control affects the overall com-
back looks to be a good choice for the LB ponent inventory control picture. By looking
column. at the overall picture and understanding
For this case study the control of the LB how each component's inventory is con-
column turned out to be a simple ratio trolled, one can develop criteria for the
control strategy even though dual-ended objectives of the distillation column control
control is feasible and performs satisfacto- strategy. Mter this perspective is developed,
rily. The dependence of the LB column the tools and techniques for the design of
control strategy on the overall plantwide particular column control strategies can be
strategy is the key point. Different overall used effectively.
21
Model-Based Control
F. F. Rhiel
BayerAG
440
Model-Based Control 441
I
chapter deals essentially with this problem
~ and its solution.
As already mentioned in the introduc-
lRC?o~.
~ ~~.50ppmLM
tion, an adaptive multivariable controller,
operating on the basis of a self-tuning prin-
ciple, is used. This controller takes into
account the control circuit coupling be-
tween the rectifying and stripping sections
~_15 of the column, and automatically tunes it-
H~ F self to the column dynamics.
LM It is relatively simple to control the con-
HS
centrations in the bottom products. Control
of these concentrations is achieved via tem-
peratures and pressures of plates in the
F: FHdRate
R: Reflux Rate stripping section (sixth plate in the case of
D: DiIIlIIa1e-Ramovai Rate the solvent column as shown in Figure 21-4).
F1GURE 21-1. Solvent column. LM denotes solvent; The so-called pressure-compensated tem-
HS denotes high boilers. perature, which is correlated with the con-
centration of the bottom product (water
concentration in the case of the solvent
In the conventional control concept the column), is formed from the two variables.
water concentration of the bottom product In contrast to the control of the bottom
has been controlled via a temperature in product concentrations, it is much more
the stripping section of the column. By this difficult to control the concentrations of the
concept the water concentration could be top products because their correlations with
kept within the required limits without the temperatures and pressures in the recti-
problems. On the other hand, it was consid- fying section are not sufficient for control.
erably more difficult to control the solvent Therefore, in the conventional control con-
concentration in the top product. Samples cept the concentrations of the top products
of the top product were taken off at 4 h have to be analyzed by gas chromatography.
intelVals and analyzed by gas chromatogra- The concentrations are only measured at
phy. The operating personnel obtained the fairly long intelVals and the results are not
analysis results 1 h after sampling, and could immediately available to the operating per-
then carry out possible corrections by sonnel. This disadvantage is to be elimi-
changing the reflux and steam rates. nated with an obselVer model that estimates
The second column (isomeric column) has the concentration of the top product from
40 sieve trays and separates isomeric com- the process variables capable of direct mea-
ponents with the following product specifi- surements and immediately conveys this in-
cations. In the bottom product the concen- formation both to the operating personnel
tration of the low boiling key component is and to the controller.
not allowed to exceed 1.5 wt%. In the top
product the sum of the two high boiling key Static Column Behavior
components should be 300 ppm. The con- The static behavior of the solvent column is
ventional control concept of this column is illustrated in Figure 21-2. The CUlVes in this
similar to the concept of the solvent col- figure are based on measurements recorded
umn. at the plant column and on simulation cal-
Because avoidance of fluctuations in the culations. The temperature CUlVes in the
top product concentrations is by far the diagram were obtained by variation of the
greater problem in the present case, this steam and reflux rates. This diagram shows
442 Practical Distillation Control
that the top temperature is unsuitable for 21·2·2 o.slgn of the Observer Model
the control of the solvent concentration in Prindpk 0/ the Observer Model
the top product because this temperature The solvent concentration in the top prod-
remains practically constant over a change uct of the solvent column cannot be con-
in the solvent concentration from 50 to 200 trolled directly by the temperature of the
ppm. Only the temperatures of the 21st and 24th plate, as an example. This would also
24th plates are sensitive to a change in be true if the effect of pressure on tempera-
solvent concentration. ture were taken into account. Only the con-
centration on the 24th plate can be kept
Dynamic Column Beluwior constant by this method. The concentration
The dynamic column behavior is illustrated in the top product is additionally deter-
in Figure 21-3. The diagram shows the tran- mined by the separation efficiency of the
sient response of the temperatures when column in the region between the 24th plate
~re
80 of
2111 Plate
~~=======24th
30th Plate
80
-t--..~----------- TopliNnp.
40
1 2 3 4 5 6 7hr
1
ncr.. In the _ Ra by 1 metric tonIhr
TIme
and the top. This efficiency varies with the varies with the hydrodynamic load. The re-
other process variables. An observer model lationship between efficiency and hydrody-
has to include these variables for the esti- namic load must still be determined by em-
mation of solvent concentration in the top pirical tests. The formulas describing the
product. relationship between the solvent concentra-
In the present case, a steady state ob- tion of the 24th plate and the top of the
server model is developed for the upper column are published in Rhiel and Krahl
region of the rectifying section of the col- (1988a). These formulas are simple and can
umn. In the case of the solvent column only easily be implemented in process computers
the section between the 24th plate and the or process control systems.
top, which contains solely water and sol- In the case of the isomeric column there
vent, is considered. Thus solvent concentra- are three components in the upper region
tion at this plate can be calculated from of the rectifying section. Therefore, it is
measured temperature and pressure on this necessary to measure temperature and pres-
plate. The solvent concentration in the top sure at two different plates in order to en-
product can be regarded as a function of able the determination of the concentration
the concentration at the 24th plate and the in the top product. Also, the algorithm for
separation efficiency between the plate and the calculation of the top product concen-
the column top. Using a steady state model, trations is more complex compared to the
this separation can be calculated from the formulas for the solvent column.
liquid to vapor ratio, the number of plates,
and the tray efficiency in this section of the Properties of the Observer Model
column. The liquid to vapor ratio can be The observer model is based on a static
determined by the reflux rate, the distilla- column model. It follows from this that the
tion removal rate, the reflux temperature, values estimated by the observer model
and the top temperature of the column. agree with the actual, that is, analytically
These values can be measured directly. The determined values, only when the column is
tray efficiency does not remain constant, but at a steady state. If the column is disturbed
F: Feed Rate
R: Reflux Rate
D: DilliU&te-Removaf Rate
..
100
Feed Concentration
80 I
40
20
O~~""~~~""r-~~""~~~""~
o 10 20 30 40 50 60 70 80 90 100 110hr
TIme
FIGURE ll-S. Rate disturbance: 15% change in feed rate by
manual adjustment.
Model-Based Control 445
%
100
40
Pressure-Comp. Temp., 8th Plate
20 Top Concentration
(calculated with Observer)
Set Point value of Top Concentration ....-
O~~~~~~~~~~~~~
o 10 20 30 40 50 60 70 80 90 100 110hr
Time
FIGURE 21-6. Reaction of the observer: coordinated manipula-
tion of steam and reflux rates by adaptive multivariable controller.
promptly, as is apparent from the increase 21-3-2 Introducing the New Control
in the reflux rate. Concept In the Production Plant
Figure 21-7 shows a comparison between
the observer and analytical results. It is ap- The implementation of the new control con-
parent that the two sets of results agree cept in the production plant has been done
when the column is at steady state. Only in three steps. This is advisable in order to
when the disturbance occurs does the ob- avoid any risk as far as possible.
server model produce a higher value, which In the first step, only the observer model
causes the controller to make prompt cor- is installed and the column is operated with
rections. As a result, a very flat curve of the conventional control concept. The ob-
solvent concentration in the top product is server values are recorded by the computer,
obtained in spite of this disturbance. but they are not transmitted to the operat-
ppm
600 _ .. Top Concentration
(Determined Analytically With Ge)
400
Top Concentration
(calculated with ObIerver)
200
/
+2 ~P4 •• ~"r~~11
o ir_-r----Ti----,-~,----"T"----,r--r----Ti----,_~i----__r_i
o 20 40 60 80 100 hr
Time
FIGURE 21-7. Results: The top concentration (analytical value)
remains constant in spite of large rate disturbance.
446 Practical Distillation Control
Concentration
ppm
- Analytical value
1000 Observer Value
600
400
200
Time
FIGURE 21-8. Conventional operation: Observer values are rwt
displayed to operating staff.
ing staff. Figure 21-8 shows the observer clear from the diagram that the observer
and analytical values of the solvent concen- values represent satisfactory predictions.
tration in the top product. In this way, the
correlation between the observer and the
analytically measured values can be tested. 21-4 COMPARISON BETWEEN
Additionally this diagram illustrates that the CONVENTIONAL AND NEW
fluctuations in column operation are shown
CONTROL CONCEPT
even more clearly by the observer values.
In the second step, the column also is Figure 21-10 reproduces the analytical val-
operated with the conventional control con- ues of the solvent concentration in the top
cept, but, in contrast to the first step, the product for the column operation with the
observer values are continuously displayed conventional and the new control concept.
on a screen for the operating staff. This The comparison between the two curves
leads, as shown in Figure 21-9, to a consid- highlights the much smoother curve ob-
erably smoother concentration curve. It is tained with the observer model, .combined
Concentration
ppm
1000
800
600
400
200
o ~i~--~'--r~--~'-~~~i~~~~~i---
o 100 140hr
TIme
i i I I I
20 .to 60 ao 100 hr
TIme
DGURE 21-10. Comparison between conventional and new
control concept (solvent column).
with the adaptive multivariable controller. . peratures, particularly in the stripping sec-
In this way, the fluctuations could be re- tion, could be substantially lOwered. Lower-
duced to a standard deviation of 5 ppm, at ing of the temperatures would permit a
an average value of 50 ppm. reduction in the reflux ratio and thus in the
The temperatures plotted in Figure 21-11 energy consumption.
are averages recorded over a lengthy period The improvement of the top product
of time. It is apparent that the column tem- quality of the isomeric column by the new
35
30
25
20
Feed-15 -
10
1~~~~~~~~~-+~
40 50 60 7D 80 90 100 110 120 ere
Temperature
100 i
21-5 CONCLUSIONS YM
H20- 90
'10
Concentration
On the basis of operating tests, the advan- In the Vapor 80
40
• More uniform product quality.
• Higher yields of products (lower solvent 30
losses as an example). 20
m + (1 - m);;a - 1
21-6 APPENDIX - -1
m
Formulas of the Model for the Solvent
Column
Principle of the observer model:
In[(1 _m)
a X
mx_D + ma ]
_YM_-
D - mXD
n = ~~---~~~---~
In(: )
Clo~
(21-2)
m = tanJ3 (21-3)
t~ t
v+
np
Controlled ~
Xo-Io~t·~
T \ / L
"M. PM Variable
~M, PM 0
R a=-=tana (21-4)
(calculated by np' TJ
V
QF the Observer) , ,
data, measured
at the column without delay
L RjD
Determination of n: (21-5)
n -fo (Xo, YM , t) V 1 +RjD
Velue, .T 11 n = n p 71 (21-6)
analyticattyJ ,~M' PM 0 J
determined data, measured
at the column
71 = 71(F) (21-7)
448 Practical Distillation Control
400
200
100
o I I i i i I i I
24 48 72 96 120 144 168 192 216hr
TIme
400 • Conventional
• New
.COntrol Concept
300
200
100
o i i i I i i I i
48 72 96 120 144 168 192 216 hr
Tillie
22·1 OCCURRENCE AND 135"F). This means that waste heat often
IMPORTANCE can be used. Other industrial examples in-
clude many types of isomer separations (e.g.,
Distillation columns that perform difficult isopentane-normal pentane, ortho-xylene-
separations are called superfractionators. By meta-xylene, etc.).
difficult separations we mean that the rela- It is important to note that the separa-
tive volatility between the chemical compo- tions that use superfractionators are almost
nents is quite low: less than 1.2 and ap- always binary, that is, only two components
proaching 1.05. Such low relative volatilities are present. This is because it is more effi-
mean that there is very little difference be- cient to make these difficult separations with
tween the composition of the liquid phase binary systems than when other heavier-
and the composition of the vapor phase in than-heavy key or lighter-than-light key
equilibrium with it. Hence many equilib- components are present. Thus superfrac-
rium stages are required to accomplish the tionators usually are positioned at a point in
separation. the process where other lighter and heavier
There are a fair number of these types of components have been removed in previous
columns. The separation of heavy water processing steps. This means that distur-
from regular water is probably the most bances from upstream units are usually fre-
striking example. The relative volatility is quent because the superfractionator is at
only 1.05 and multiple columns are required the end of the line. The control system must
with a total of 750 trays. be able to effectively reject these distur-
The most common industrial example is bances.
the separation of propylene from propane.
This separation is performed in almost ev-
ery petroleum refinery in the world and in
22·2 FEATURES
many chemical plants. Relative volatility is
about 1.1 and columns contain typically 150 Because of the low relative volatilities, su-
to 200 trays. The propylene-propane sepa- perfractionators have the following fea-
ration is a major consumer of energy. How- tures: many trays, high reflux ratios (large
ever, it should be emphasized that this en- energy consumptions), flat temperature pro-
ergy is at a very low temperature level (about files, and very slow dynamic responses.
451
452 Practical Distillation Control
very slow (time constants that are many the schemes are sketched in Figure 22-1.
hours). This makes control more difficult. Because of the high reflux ratio, control of
the reftux drum is almost always achieved by
manipulating reflux and not distillate.
22-3 ALTERNATIVE CONTROL
STRUCTURES
There are a number of alternative control 1. D-V: Figure 22-1a shows the D-V
structures for superfractionators. Some of scheme where distillate flow controls dis-
(a)
SP
(c) (d)
FIGURE 21-1. Alternative control structures: (a) D-V; (b) D-BR; (c) RR-BR;
(d) D-B. ·
454 Practical Distillation Control
tillate composition and heat input con- bottoms composition. This scheme re-
trols bottoms composition. quires the measurement of four flows
2. D-BR: Figure 22-1b shows a scheme and the calculation of two ratios. There-
where distillate flow rate controls distil- fore, it is more complex and more sus-
late composition and boilup ratio (heat ceptible to problems with noisy flow
input over bottoms flow rate) controls signals.
bottoms compositioIi. 4. D-B: Figure 22-1d shows an unusual
3. RR-BR: Figure 22-1c shows a scheme scheme where distillate composition is
where reflux ratio (reflux flow rate over controlled by distillate flow rate and bot-
distillate flow rate) controls distillate toms composition is controlled by bot-
composition and boilup ratio controls toms flow rate.
80 ....... u.l~gaVmin. CW
230bph 44
70%C:i 162 bph
99.6%C:i
66.2MMl
40paig
66bph
3.2%Ci
Chapter 7).
It is worth noting that obtaining transfer
function models for the D-B scheme is not
a straightforward procedure. This is be-
cause the steady-state gains are indetermi- 6
nant: One cannot find the steady-state gain
between distillate flow rate and any compo- 2520~
r -
sition with bottoms flow rate fixed since D ~ I
plus B must add up to F at steady state. 8 I
a: I I
Dynamically D + B does not have to equal
F. We will present a method for determin- 2440 0
~I
20 40 60 80 100 120
ing the dynamic D-B transfer functions later Time (min)
in this chapter.
"GURE 22-3. Plant tests compared to dynamic
model predictions (Sun Oil column),
22-4 INDUSTRIAL EXAMPLE
At this point a specific example might be enough to proceed with simulation compar-
useful. isons of alternative control structures.
I"M[~-==>~~ 1....
~
99.45
[ ~
99.45
i ". f
3.50[
!-~ <====--
~ ~
r"t=
0.0 0.0
v
ft:=:--------......;:"-"- R
i"~rt----
~
99.45
[
i '~[ r===:=--
~
1.50
r"~----:=
o.ot-
Cl
rr=
v
R
FIGURE 22-4. Simulation results for alternative control structures (Sun Oil column):
(a) D·V; (b) RR-BR; (c) D·B.
Superfractionator Control 457
composition. The RR-BR and the D-B Combining Equations 22-1 through 22-3
structures give significantly better responses gives
than the traditional D-V structure. Because
the RR-BR requires more instrumentation V= GLR-B
complexity, the D-B structure looks promis- V=GL(V-D)-B
ing.
(1- GL)V= GLD-B
22-5-1 Transformations
The transfer functions for the D-B struc-
[ ::] = ~M ~] = [~:: ~:: ][ ~ ]
[
[:: 1~ gM[:1
overflow, perfect level control in reflux drum
and base, and liquid flow dynamics de-
scribed by
(22-1)
1 0 D
=~M -GL -1
The assumption of perfect level control in
the base gives B
1 - GL 1 - GL
(22-2)
-37.7e- 68 0.647e- 68
XD D
(72OOs + 1)(2s + 1) (159s + 1)(2.86s + 1)2
(22-8)
-43.8e- 68 -1.71e- 68
xB V
(2338s + 1)(12.7s + 1)2 (26s + 1)(3.82s + 1)2
(22-9) j
~
~.O
J -20.0
I
180.0
total mass balance for the tray:
0.0 ~ _ _J:'"D-V
t
1- 180.0
------D.i----:::-----
t
(22-10) - 380.~ o......_=-.--:":10~-';---:-:10J....,-'---:-:-: 0.=--
10L:-.---:1..,J.0--:-,_..L....J.1..1.
W
where wL = weir length
D = column diameter F1GURE 22·5. Bode plot of the (1, 1) element of the
D-V (solid line) and (dashed line) D-B transfer func-
mN = molar holdup on the nth tray. tion matrices.
Superfractionator Control 459
8,x,
uat) for a feed flow rate disturbance. Note following through the sequence of events
the unusual open-loop underdamped behav- that occur following a rise in base level. The
ior. The presence of the boilup ratio con- level controller increases bottoms flow rate
troller in a tall column that has significant and brings the base level back down.
liquid hydraulic lags produces an oscillatory But the increase in bottoms flow also will
open-loop system. This can be explained by cause an increase in heat input. This vapor
0.4.---.....,....~--r-~ ........--.....--.......,
1 0.0
~
-0.4 ~_~-=-_~~--:±-:-_-=:-::-_~
-4.0,~_~~_~~--:±-:-_~~~"!"!!. 0.0 20.0 40.0 60.0 80.0 100.0
0.0· 20.0 40.0 60.0 80.0 100.0 Time (min)
Time (min)
~ 20.0r---........- -.......---r-----.--.....,
'-~:mill]~:,- I
~
0.0
!
~
L 20.0 ~-~~-~---:±-:--~....,...-.,.~
0. 0.0 20.0 40.0 60.0 80.0 100.0
cB 0.0 20.0 40.0 60.0 eO.o 100.0 Time (min)
Time (min) ATV test on the xD·BA loop
EI[t;-----.j'I
lF--l f~:~E::~:: I
0.0 50.0 100.0 150.0 200.0 250.0
50.0 100.0 150.0 200.0 250.0 Time (min)
Time (min)
0
!1o.
~ 5.0
g
:>. 0.0 0.0 50.0 100.0 150.0 200.0 250.0
Time (min)
0.0 50.0 100.0 150.0 200.0 250.0
Time (min) Open loop response of Ihe column for a feed flow disturbance
quickly goes overhead and begins to in- not so obvious: location of the sidestream
crease reflux drum holdup. The reflux drum drawoff tray (NS).
level controller then increases reflux. Re-
member, distillate flow rate is fixed because
the composition controllers are on manual. 22·7-1 Process
Mter some significant time due to the liquid The column studied was based on an actual
hydraulic lags, the increase in reflux will column at the Exxon Chemical Baytown
cause an increase in the liquid entering the chemical plant. The column separates a 70
base of the column. This causes another mol% propylene, 30 mol% propane feed
temporary increase in B and V, and the into a 99.66 mol% propylene polymer-grade
cycle repeats itself. distillate product, a 94 mol% propylene
The open-loop damping coefficient be- chemical-grade liquid side stream product,
comes smaller as the number of trays in the and 0.02 mol% propylene bottoms product
column increases and as the tray hydraulic that goes to fuel (LPG). The column has
time constant increases. 190 trays and operates with a reflux ratio of
This unusual dynamic behavior means 13.3. Because a vapor recompression system
that it is difficult to obtain the open-loop is used, the operating pressure in the col-
transfer functions and tune the controllers. umn is 110 psia. This pressure is lower than
would be required in a conventional column
22·7 SUPERFRACTIONATOR WITH with a water-cooled condenser. Therefore,
the relative volatility is increased, making
SIDESTREAM EXAMPLE
the separation easier. Table 22-1 summa·
A second and somewhat more complex ex· rizes the steady-state design and operating
ample is presented in the following text. parameters for the side stream column, Fig-
This column produces a side stream in addi- ure 22-10 shows the system, and Table 22-2
tion to distillate and bottoms products. This gives the properties of the various streams
means that there is one more controlled numbered in Figure 22-10.
variable: side stream composition. There is Rigorous, nonlinear steady-state and dy-
one more manipulated variable that is obvi- namic models were used. The tray efficiency
ous: sidestream flow rate (S). There is in the steady-state model was adjusted until
an additional manipulated variable that is the steady-state model matched the plant
462 Practical Distillation Control
TABLE 22-1 Parameter Values for Sidestream the input to the model was not the se-
Column quence of pulses in distillate flow found
F, lb-moljh 1000
from the plant data. Thus the autotune test
S,lb-moljh 196 was a particularly sensitive evaluation of
D, lb-moljh 535 model fidelity. These results demonstrated
z, mole fraction propylene 0.70 that the dynamic mathematical model was
x D, mole fraction propylene 0.995 accurate enough to use for comparative
xs, mole fraction propylene 0.93
x B , mole fraction propylene 0.0002 studies of alternative control structures.
~ 200
lVS 150
lVF 60 22-7-3 Steady-State Analysis
P, psia 110 The steady-state model was used to calcu-
MB,lb-mol 1060
MD' lb-mol 780 late steady-state gains for a large number
Mn, lb-mol 29 of alternative control structures for the
R, lb-moljh 7450 sidestream column. An efficient procedure
RR 13.3 was developed to determine these gains for
d,ft 10
any choice of manipulated variables with
~,~ 1
wL' ft 6.41
only three runs of the rating program
(Papastathopoulou and Luyben 199Oa).
The controlled variables were x D' x s,
and x B. A wide variety of manipulated vari-
operating data. An efficiency of 95% was ables were explored in the 3 x 3 multivari-
found. able system. The basic manipulated vari-
ables were D, S, and V or D, NS, and V.
Ratio schemes were also included such
22-7-2 Plant Dynamic Tests
as boilup ratio (BR = VjB), reflux ratio
The dynamic model was validated by plant (RR = RjD), and sidestream-to-vapor ratio
tests. Several types of tests were conducted. (SV= SjV).
Figure 22-11a compares the plant experi- Figure 22-12 shows the D-NS-V control
mental data with the model predictions for structure. Three sidestream drawoff trays
a dual pulse on distillate flow. Figure 22-11b are used with five trays separating them
shows the same comparison for an autotune (NS 3 - NS 2 = NS 2 - NS 1 = l1NS = 5).
test on the x D-D loop (see Chapter 7). It Liquid streams from two of the three trays
should be noted that the switching times in are mixed together. The signal from the
both the model and the plant were deter- composition controller (CO) sets this split at
mined by x D crossing the setpoint, that is, any point in time.
D.xo
s.
B.xs
Several steady-state indices were calcu- divided by the product of its diagonal
lated for the alternative control structures elements. This index must be positive.
when all three compositions were controlled 4. CN: The condition number is the ratio of
using various pairings and various choices of the maximum singular value of the
manipulated variables. The indices used are steady-state gain matrix to its minimum
summarized as follows: singular value. The smaller this number,
the better the control.
1. MRI: The Morari resiliency index is the 5. JEC: The Jacobi eigenvalue criterion is
minimum singular value of the steady- the largest eigenvalue of the matrix A,
state gain matrix. The larger its value, where A satisfies the equation
the better the control.
2. RGA: The relative gain array is the
Hadamard product of the steady-state K= D(I -A) (22-11)
500.0 1000.0
...-,
I 375.0
r
.
,
! ...
.~
......
500.0
.:in
B 250.0 .! i
~ ...
,, 8 0.0 -
B : ~-
J
-
~ ~
-500.0
0.0
-125.01 -1000.01.:-:-~~~~_~~~~~
0.0 25.0 SO.O 75.0 100.0 125.0 0.0 25.0 SO.O 75.0! 100.0 125.00
Time (min) Time (min)
(tI)
Plant Model
4OO.0~-"---"---'----' 400.0
8'-200.0
"""
II' ~
\I II'
\I
I
- 400.0 I:-::--:-:~-=:±-:--=:::-:~~ -4OO.01:-::-~~~~~~~~
0.0 100.0 200.0 300.0 400.0 0.0 100.0 200.0 300.0 400.0
Time (min) Time (min)
1000.0,-------.----, 1000.0~-..---...---....--.,
i...... 500.0
I
r: :-' :1 r1 :: r-: r: r:
8 .! I! !! !! !! I! !! ! j 1
0.0
I o.oh! i! !! ! j j i ! j j! ! ! !
I !I !! !I!! !I!i III
-500.0 i5 -500.0 l.i U U U U U U U L.
II'IGURE 2Z-1l. Plant tests compared to dynamic model predictions (Exxon column):
OD xD-D loop.
(tI) Dual pulse in distillate flow rate; (b) autotune test
Superfractionator Control 465
NT D,xD
...............-oIlCl-----"""'---I.I----DI!Q---""'-"
•••••••••••• +-£:Ja.;"
s,xs
•••••••••••• +--lI)lCiI-+-~iI--fJ--__1IX1t_--='---
•••••••••••• +-£;:11(1-.;;.1
F,z NF
- - - - - t ........... .
,,
The values of JEe should be between abIes have no direct effect on these con-
-1 and 1, and the best pairing is that trolled variables. Of the remaining schemes,
with the smallest JEe. the D-S·V and the D-NS-V were select-
ed for further evaluation using dynamic
All schemes with negative Nls or RGAs analysis.
were eliminated. All schemes with very large One scheme that cannot be analyzed us-
CNs were eliminated. This still left a large ing the steady-state indices is the D-S-B
number of possible schemes. Heuristic dy- structure because the steady·state gains are
namic considerations were used to reduce indeterminant. However, the transfer func-
the number of cases. All schemes that use B tions for this structure can be obtained at
to control sidestream composition or S (or higher frequencies by transformations that
NS) to control overhead composition were are similar to those discuss earlier for the
eliminated because these manipulated vari- D·B structure.
466 Practical Distillation Control
Xs O.076(325s + l)e- 6S
(22-12)
CO 81s +1 40.0
iii
~
Because the sidestream composition loop
was much faster than the distillate and bot- ~ 3O.0!
20.0
1
1 O.O!
D-S-V 20.0
c
0
1-20.0
i
~ -20.0
D-S-V
8 D-S-B
l-4O·0
~
2400.0 - 4O.0!:-:-......,.,~"""""-::::::!-::,.-.,.~:--'":-=!-:-=O,-:-~
0.0 400.0 800.0 1200.0 1800.0 2000.0 0.0 400.0 800.0 1200.0 1800.0 2000.0 2400.0
Time (min) Time (min)
Z
c
0.04 Z 0.04 r-~-.----.----r-~--r--...-...--,
i E
8
0.00'
-- j
E
~ 0.00 I\J(\
i\:-=--
i --
i -0.04: 8 ,:"
1-0.08
IE _ 0.04 \i
\_~
D-S-V
D-S-B
Z 0.005 Z 0.005r---r-~-'--~-"""T""---r-----,
c
.2
f·-~·
'"8.
III
~
j -0.005 III
!5
-0.005 D-S-V
-0.010'
0.0 400.0 800.0 1200.0 1800.0 2000.0 2400.0 ~ - 0.010 0.0
!:-:-.......,,=-=-=---=~,......,.~:--'":-=!-::--~~-=-!
400.0 800.0 1200.0 1600.0 2000.0 2400.0
Time (mini Time (min)
(.J (bJ
23·2 DESIGN
FIGURE 23-2. VRC with compression of overhead The choice of column operating pressure on
vapor. the design of the vapor recompression col-
470 Practical Distillation Control
The behavior of pressure and boilup also systems considered for vapor recompression
appears to be different. In a conventional represent difficult separations (low relative
column, pressure dynamics are normally fast volatility), the column will always have a
and can be controlled by manipulating con- large number of trays, which results in slow
denser duty. Vapor boilup can be directly composition dynamics. Composition control
and independently manipulated by heat in- loop dynamics will also be slowed by com-
put to the reboiler. position sensors (typically chromatographs)
In a vapor recompression column, the because it is often not possible to use tem-
effect of compressor speed on pressure peratures to infer compositions due to the
seems to be complex. For example, one flat temperature profile.
would expect an increase in compressor Pressure dynamics are fast in almost all
speed to increase the flow through the com- distillation columns because the vapor resi-
pressor. This should pull more vapor from dence time (time constant) of the column is
the column and decrease the pressure in the small. Changes in the rate of vapor boilup
column. However, the same increase in or condensation rapidly affect pressure.
speed will increase the heat added to the The second condition means that if the
column, and thus increase vapor rates. The dynamics of the reboiler are fast in the
latter, might produce an increase in column conventional columns, as is usually the case,
pressure. the dynamics of the reboiler in the vapor
Despite the complexity added by the va- recompression column must also be fairly
por recompression design, the process does fast. This is indeed the typical situation in
not require a complex control structure most vapor recompression systems. Vapor
(Muhrer, 1989). We can use the conven- holdup in the compressor is relatively small
tional column control strategy in the corre- compared to the vapor holdup in the col-
sponding vapor recompression column by umn (ratio 1/20); thus, changes in compres-
simply replacing heat input control with sor speed (or other compressor control vari-
compressor control. The pressure loop is ables) rapidly affect the vapor rates in the
faster than the composition loops, so we can column, emulating the effect of steam in
handle it independently. For this to be valid, conventional columns.
two conditions are necessary: These conditions normally exist in most
vapor recompression distillation systems for
1. The composition time constants of the reasons discussed in the preceding text.
column must be at least 5 times larger Thus, I believe this is a generic solution to
than the pressure time constant. the control of vapor recompression columns.
2. The time constants of the reboilers of
the conventional and vapor recompres-
23-4 ALTERNATIVE COMPRESSOR
sion columns should be approximately
CONTROL SYSTEMS
the same.
The control strategies for vapor recompres-
In terms of a distillation design parame- sion towers presented so far have been de-
ter, the first condition corresponds to hav- signed with a variable-speed compressor. If
ing a temperature difference between the the compressor is driven by a steam turbine,
top and the bottom of the column that is no changing speed is straightforward. However,
greater than approximately 60°F, which rep- variable-speed electric motor drives are
resents the upper limit for vapor recompres- expensive. Compressor control can be ac-
sion under today's economic conditions. complished in several other ways: suction
The inherent composition dynamics are throttling, bypassing flow back into the com-
primarily determined by the holdup of ma- pressor suction, or varying the heat-transfer
terial in the column. Because almost all area in the reboiler-condenser using a
472 Practical Distillation Control
Overflead
Vapor
Reflux Reflux
Flow Flow
B B
R. R.
D o
(6) (~)
Overflead
Vapor
Vepor
Reflux Reflux
Flow Flow
8 8
R R
D D
(c)
DGURE 23-6. Compressor control alternatives. (a) Variable speed; (b) suction
throttling; (c) variable heat-transfer area; (d) bypassing.
flooded reboiler design (condensate throt- sition control by using four different com-
tling). pressor control structures.
In this section, these alternatives will be
compared in terms of both steady-state de-
23-4-1 Compressor Performance
sign (energy consumption) and dynamic
Curves
controllability. The compressor control reg-
ulates heat input to the column, thus con- The performance of a centrifugal compres-
trolling bottoms composition. Figure 23-6 sor cannot be exactly predicted from theo-
shows control schemes for bottoms compo- retical analysis. Instead it must be experi-
Control of Vapor Recompression Distillation Columns 473
E
.... B.O
.!!
UJ
~ 6.0
UJ
a:
UJ
IL
IL
M
0 4.0
~
~
(I)
(I)
UJ
a: 2.0
Go
High
pressure
~--------------~----~~~
steem Discharge
Steem
turbine
Verleble
speed
EKheust compressor
L-______ ~ __ ~ ____ ~ ____ ~~ F.
steem
Suction
(bl
(23-1)
60Ap TB - a
ap=PD -P = --V+ -Ps
s UAfJ fJ
where QR =heat-transfer rate in reboiler- (23-5)
condenser (Btu/h)
A = heat-transfer area (ft2)
U = overall heat-transfer coeffi- where Ps is the compressor suction pressure
cient (Btu/h ft2 OF) or the column pressure.
TR = saturation temperature eF) Equation 23-5 is the equation of a straight
of distillate product (mostly line:
propylene) at discharge pres-
sure ap = mV+ b (23-6)
Control of Vapor Recompression Distillation Columns 475
where m and b are both constants: This type of almost linear plant charac-
teristic curve should be found in most vapor
60Ap
(23-7) recompression distillation systems. For the
m = UA8 propylene-propane system, the values for
T - a m and bare 0.0014 min atm/ft 3 and 3.72
b= B -Ps (23-8)
8 atm. These values compare very well with
the values obtained by fitting data from a
This explains the linear shape of the plant rigorous rating program: 0.0012 and 3.75.
characteristic curve. At the zero flow condi-
tion, the temperature differential driving
force in the reboiler-condenser must be 23-4-3 Description of AHernative
zero. That means that the temperature of Compressor Controls
the condensing distillate product must be
equal to the temperature of the boiling bot-
Variable Speed
toms product. Because the bottoms is less
Compressor speed is adjusted to compen-
volatile than the distillate, the pressure in
sate for process disturbances. Figure 23-8a
the column (determined by bottoms compo-
shows how the compressor speed must
sition and temperature) will be less than the
change to maintain compositions when ad-
compressor discharge pressure (determined
by distillate composition and temperature). ditional heat is required by the column. The
This can be seen mathematically by look- discharge pressure must increase to provide
the higher temperature differential driving
ing at the intercept of the linear relation-
ship given in Equation 23-6: force in the reboiler and the flow must
increase. This is achieved by moving the
TB - a operating point to a higher speed curve.
lim ilP = b = - P s = Pd(TB ) - Ps
V--->O 8
(23-9) Suction Throttling
The suction throttling strategy adjusts com-
where Pd(TB ) = pressure that the distillate pressor suction pressure to compensate for
product (mostly propylene) process disturbances. Figure 23-8b shows
exerts at the base tempera- the case when additional heat is required by
ture T8 the column. The operating point moves
Ps = column pressure along a constant speed curve to a higher
flow by reducing the pressure difference
If we neglect the pressure drop through across the compressor. The higher tempera-
the column (which will be small at zero ture differential driving force in the re-
vapor boilup), the column pressure will be boiler-condenser increases the discharge
the pressure Ps(TB ) that the bottoms product pressure. The increased opening of the suc-
(mostly propane) exerts at temperature TB • tion valve decreases the pressure drop over
Distillate product is more volatile than bot- the valve and increases suction pressure
toms product, so it will exert a higher pres- enough to provide the needed increase in
sure Pd(T ) at the same temperature. There- both discharge pressure and flow. The pres-
fore, Pd B> Ps and a pressure differential sure differential across the compressor goes
exists at zero flow. down.
Assuming pure propylene and propane
products,
Variable Heat-Transfer Area
lim [flp] = [p~roPYlenelTB - [p~ropanelTB The heat-transfer area of the reboiler-con-
V--->O
denser is adjusted to compensate for pro-
(23-10)
cess disturbances. As shown in Figure 23-8c,
476 Practical Distillation Control
12.0 12.0
'I' 'I'
...
c ...
c old operating po int
x x
8.0 8.0
Z Z
~ ~
0
4.0 0
4.0
« «
IJJ IJJ
J: J:
o. o.
o. 8.0 16.0 24.0 32.0 40.0 O. 8.0 16.0 24.0 32.0 40.0
INLET FLOW (Icfm) ( X 10"') INLET FLOW (Icfm) (X 10"')
(a) (b)
12.0 12.0
'I'c 'I'
... old operating point
\ ...
c ~etl operating po int
x x
8.0 8.0
Z Z
~ ~
4.0 A > A DESIGN
0
4.0
0
« «
IJJ IJJ
J: J:
o. o.
o. 8.0 16.0 24.0 32.0 40.0 o. 8.0 16.0 24.0 32.0 40.0
INLET FLOW (acfal) ( X 10"') INLET FLOW (lcflD) (X 10"')
(e) (d)
FIGURE 23·8. Operation mode. (a) Variable speed; (b) suction throttling; (c)
variable heat-transfer area; (d) bypassing.
the heat input to the column is regulated by able-speed compressor case, but a larger
changing the setpoint of the level controller reboiler-condenser must be used to make it
in the reboiler-condenser, thus varying the possible to handle disturbances.
heat-transfer area available.
This flooded-condenser design changes
the plant characteristic curve. Figure 23-8c Bypassing (SpiU Back or Recycle)
shows how the operating point of the com- The bypass control scheme operates with
pressor moves to a higher flow rate by in- more flow through the compressor than is
creasing the heat-transfer area when the necessary to meet the column heat input
column requires more heat. The new oper- requirements. The excess flow is recycled to
ating point is farther down the constant the compressor suction. Figure 23-8d shows
speed curve, allowing more flow to go that to increase the heat input to the col-
through the compressor. The compressor umn the bypass flow must decrease, causing
head is lower because increasing the avail- a corresponding increase in the compressor
able heat-transfer area decreases the tem- head. The heat input increases due to the
perature differential driving force in the higher temperature differential driving force
reboiler-condenser. This means that the in the reboiler-condenser. The net effect is
compressor discharge pressure decreases. increased flow through the reboiler, less flow
At the design operating point, the pro- through the compressor, and much less flow
cess conditions are the same as in the vari- through the bypass.
Control of Vapor Recompression Distillation Columns 477
4.4
10000
20 . . . . . .1
110F
nD------------------~
ISIF
IIH
(8)
10000
2O ••~1
110F
FEm
ISIF
4.4 .Im
2SOF
173F
(b)
FIGURE 23·11. Ethanol-water column design. (a) Conventional
column. (b) Vapor recompression column (numbers inside heat ex-
changers are 106 Btu/h; numbers without units along lines are flow
rates in pound-moles per hour; compositions are in mole percent).
480 Practical Distillation Control
32!
RR·ll.6
600.
6011CI
110 F
FEED !!..
19.
1 Itm st.
(a)
7IF
10.htm
5125
RR-ll.1 1197hp
12IF
600
6011 C· 17.6Itm
I
110F
FEED _ _ _ _ _ _ _.... 41
12
(b)
FIGURE 23·12. Propylene-propane column design. (a) Conventional column. (b)
Vapor recompression column (numbers inside heat exchangers are 10 6 Btujh; num·
bers without units along lines are flow rates in pound-moles per hour; composition are
in mole percent).
Control of Vapor Recompression Distillation Columns 481
The vapor recompression design for the If colder cooling water were available,
same separation is presented in Figure allowing condensation at 90°F, the pressure
23-11b. The column itself is identical to the could be reduced to 15.6 atm, resulting in a
conventional one because the operating 4.4% reduction in energy consumption.
pressure and reflux ratio are the same. The The same separation using vapor recom-
energy consumption is 8000 hp or 20.4 X 10 6 pression is presented in Figure 23-12b. In
Btu/h. Because the top temperature is high this case, it is possible to decrease the oper-
enough to use water to condense the excess ating pressure and hence increase the rela-
vapor (the vapor not required to match re- tive volatility. A high-pressure reflux drum
boiler duty), it is possible to use a low-pres- scheme is used because of the low top tem-
sure reflux drum scheme. perature. The temperature in the reflux
The capital investments for the conven- drum is fixed to provide a 15°F temperature
tional and vapor recompression designs are difference in the reboiler-condenser.
estimated to be $1.3 and $5.5 X 10 6 • Using This design has a column with 54 fewer
utility costs of $5 per 1000 pounds of steam trays and an operating pressure 7 atm lower
and $0.07 per kilowatt hour of electricity, than the conventional column. The energy
the annual energy costs are $5.6 and $3.8 X consumption drops from 27.3 X 10 6 Btu/h
10 6 . These numbers give a 2.3 yr payout (steam) to 3.0 x 10 6 Btu/h (electricity).
time for the incremental investment in go- The base case design of the conventional
ing from conventional to vapor recompres- column has a capital cost of $1.5 x 10 6 and
sion systems. The very large capital cost of an annual energy cost of $0.56 x 10 6 . The
the vapor recompression system is mostly payout time for the incremental investment
due to the cost of the compressor. Com- in going from conventional to vapor recom-
pressor cost is a function of the compres- pression in this system is only 0.25 yr. This
sion work required and the volume of vapor is because of the small column ~ T and high
that must be handled. For this chemical pressure.
system, the relatively large temperature dif-
ference (62°F) between the top and bottom
23-5-2 Dynamic Models
of the column indicates the need for a large
compression ratio. For systems with smaller The distillation columns are simulated by
temperature differences, less energy is re- using standard column modelling equations.
quired and capital cost and payout time For ethanol-water, the model includes two
decrease (as in the propylene-propane dynamic component balances and one en-
system). ergy balance per tray. For propylene-pro-
pane, equimolal overflow was assumed,
Propylene-Propane System eliminating the energy balances. The liquid
The base case for the conventional pro- flow rate from each tray is calculated from
pylene-propane distillation is shown in Fig- the Francis weir formula. Details of
ure 23-12a. It is necessary to operate at vapor-liquid equilibrium relationships,
17.6 atm (244 psig) to use water in the physical properties, and column dimensions
condenser. This pressure is based on a de- are presented by Muhrer (1989).
sign overhead vapor temperature of llO°F. To model the variable pressure in the
This high pressure reduces relative volatil- column and in the compressor piping, the
ity, resulting in a very large column (208 system is divided into three parts: suction
trays, 100% efficiency) and high reflux ratio piping, discharge piping, and the compres-
(13.6). It is not possible to preheat the feed sor proper. Mass balance equations are
using a column product because the tem- written for the suction and discharge vol-
peratures of these streams are close to- umes (Davis and Corripio, 1974). These
gether. differential equations give the suction and
482 Practical Distillation Control
and Luyben (1986). Steady-state "rating" The curves for ethanol-water show that
programs are developed for all columns. both vapor boilup and reflux flow must be
They are used to determine incentives for adjusted to compensate for feed composi-
dual composition control by plotting the tion disturbances, but the reflux ratio re-
steady-state changes required in all manipu- mains almost constant over the range con-
lated variables to hold the controlled vari- sidered. This means that single-end compo-
ables constant for disturbances in feed com- sition control could be used with very little
position (Luyben, 1975). Disturbances in penalty in energy consumption. Singular
feed flow rate do not have to be considered value decomposition (Moore, 1986) recom-
because they can be handled with ratio mends the use of tray 6 temperature. The
schemes. Figure 23-13 gives these results for recommended control system has heat input
both systems. controlling tray 6 temperature and reflux
2.0
....... RR
UJ
1.6
J
CD
e
H
a: 1.2
e
> V!F
0
UJ
l-
e 0.8
J
=>
Q.
H
Z
e
% 0.4
O.
0.12 0.16 0.20 0.24 0.28 0.32
FEED COMPOSITION (MOL FRACTION)
(al
15.0
UJ
~
~ 13.0
e
>
~ 11.0
I-
e
~ V/F
j
H
9.0
7. 0 ~"""--L_~"""---1_.L-"""---1_.L-"""-..J
ratio fixed. Although dual composition is and Polleck (1989) studied the control of
not required for the ethanol-water system, conventional propylene-propane columns
it is examined in the interest of increasing and recommended an unorthodox D-B con-
our understanding of the generic behavior trol structure: distillate flow controls distil-
of vapor recompression columns. An ana- late composition, bottoms flow controls bot-
lyzer is used to measure the top composi- toms composition, reflux flow controls reflux
tion instead of inferring it from the temper- drum level, and heat input controls bottoms
ature because of the poor sensitivity of level. Unfortunately, this structure has poor
the temperature on the top trays. Figure integrity and cannot be analyzed by steady-
23-14 shows the control system for the state methods. The RR-BR structure (reflux
ethanol-water separation with vapor re- ratio controls distillate composition and
compression. boilup ratio controls bottoms composition)
The curves in Figure 23-13b for pro- is equivalent to the D-B structure assuming
pylene-propane show that all manipulated tight level control. It requires more complex
variables change significantly as feed com- instrumentation and is more difficult to op-
position changes, so dual composition con- erate (Waller et aI., 1988).
trol is required. Several choices of manipu- The D-B, RR-BR, and the more conven-
lated variables are explored. Finco, Luyben, tional D-Q R structures are all studied. Con-
F
10
trollers are designed for each case. Figure speed (D-Sp )' Figure 23-16 shows the con-
23-15 shows that all control schemes per- trol system with vapor recompression.
form similarly. Therefore, the widely used The steady-state rating programs are also
D-QR structure is chosen for the conven- used to calculate steady-state gains. These
tional column. The corresponding control are used in obtaining transfer functions so
structure is used for the vapor recompres- that manipulated variables can be selected
sion column, distillate flow and compressor and unworkable pairings can be eliminated.
so.O 3:
0 360.0
.J
"-
-so.o D-QR W
.
~
I
<C
.J 340.0
..
.J
.I!l.
~
II)
! 0 320.0
~
<C
SO.O 2B.O
>
..
Q. 10.0
0 15.0
t-
o
~
<C
-10.0 D-S a: 14.0
)(
:I
.J
It;
... ~
a: 13.0
.
,... ~
0
~
g D-a. <C
a:
4.4 Q.
z 3...
......
ClI 0 8.0
...
III
360.0
In
ClI 5.2 ~
~ Ii
u RR-BR ...
\II 340.0
...~ ~.O0v-
...t-~
4.4 ~
...t-
..J
..J
51 ...a
In 320.0
"'~
4.4
D-S
51 I
o. 2.0 4.0 11.0 8.0 10.0 12.0 O. 2.0 4.0 6.0 8.0 10.0 12.0
TIME CMINI ( X 10-1) TIME (MIN) ( X 10-1)
(a) (b)
154M----.......
F
41
Transfer functions of the processes are shows that the only pairings that do not give
identified by using the auto-tune (ATV) negative RGA elements on the diagonal are
identification procedure (Luyben, 1987). those that pair cooling water with pressure:
Table 23-1 gives the transfer functions for
the conventional and vapor recompression 0.240 0.893 -0.137]
columns for both chemical systems. RGA = [ 0.754 0.255 -0.010
Note that the transfer functions obtained 0.001 -0.148 1.15
between manipulated variables and pres-
sure have time constants that are much It is interesting to note that the Nieder-
smaller than those for the compositions (see linski index does not show that pressure
the last entry row of Table 23-1 for pro- must be paired with cooling water. It gives a
pylene-propane). Note also that compres- positive index of 36.3 for the pairings XD-Sp'
sor speed has little effect on pressure; it xB-Fw , and PoD.
mostly affects compositions. The Niederlin- The multiple SISO controllers are tuned
ski index is + 3.6 for the pairings xD-D, by using the BLT technique (Luyben, 1986).
xB-Sp , and P-Fw • The RGA for this system Table 23-2 gives tuning parameters.
TABLE 23-1 Transfer Functions
Ethanol-Water
Conventional column
[ 1[
XD = (6.9s
3.7e-1.2s
+ 1)(162s + 1) -5.8e-1.3s
(1.8s + 1)(1190s + 1) 1[ 1RR
-170e- O•2s 323e- O•2s
~
(0.6s + 1)(127s + 1)2
[ 1~ [ 1[ 1
1.6e-s
s
: (0.02<+ 1~~93' + I)' (0.8s +-26.6e-
1)(1510s + 1) RR
2130
6 (0.6s + 1)(I40s + 1)2 --:(1-2-11-s-+---:1)--:('-0.-08-s-;;"2-+-0-.6-s-+-1) Sp
Propylene-Propane
Conventional column
+5~;;:;:s +
-6.84e-5s
[
XD
[ = [ (1.55s + 1)(431Os + 1) (0.8s 1)2 [[ D
-10.8e- lOs -1.91e- 5s
------------~2 QR
(3.2s + 1)(7.2s + 1)
·Controller gains are dimensionless. Flow spans are twice the steady-state value and transmit-
ter spans are: for ethanol-water, XD = 0.20, tray 6 tOO"F; for propylene-propane, XD = 0.05,
xB = 0.20.
487
488 Practical Distillation Control
23·5-4 Results for Specific Systems loop behavior of the vapor recompression
system is slightly faster than the conven-
EtIumol-Waler
tional column (smaller deadtimes and time
The performance is evaluated with the non-
constants). This behavior could be due to
linear model by introducing step distur-
the effect of pressure transients.
bances in the feed composition. The results
These results indicate that the control
are shown in Figure 23-17a for the conven-
structure used in the conventional column
tional column and in Figure 23-17b for the
can also operate successfully in the vapor
vapor recompression column. Constant
recompression column if compressor con-
pressure is assumed in the conventional col-
trol is substituted for steam flow.
umn because it would typically be vented to
the atmosphere through a vent condenser.
In the vapor recompression column, the Propylene-Propane
pressure loop is quite fast. Therefore, it is The performance of the conventional pro-
tuned independently from the composition pylene-propane column has already been
loops. shown in Figure 23-15. The same basic
The control strategy used on the conven- structure is used on the vapor recompres-
tional column works quite well on the vapor sion column. Compressor speed is used in-
recompression column. Controller settings stead of steam flow. Controllers are retuned
are somewhat different because the open- because the dynamics of the vapor recom-
0.08
£c: 0.04
-4
.£
,~::::::::- 0--..
1 --._._........_-_.
~~ "'1
0.0 ... +2 -2 ----::::-.:.
......
§ +4
CL
~ -0.04
& 0.05
E
~.-.-.-.-.-.-.~
,;,-------__ ---.-.•
-0.08 8 ...·-·-·_·-·_- • *------
8'-0.05
'""
3.0 '\ 3.0
_, ..,r., -4%~z (20% -+ 16%)
',\
'... -.....
"
., t
i i'i
1.0
- 2 ..............:'!:=:::;:::-.
~~~--
,.... _.,..... -.-
.~
t
.",,;
E
~
I
-3.0 iI;1\'
+4 ,;
.'". .I
t
E
-3.0
~
, - 5.0 L..-..L..-..L..-..........-'---'---'--....L...........................................I
-5.0
0.0 20.0 40.0 60 ..0 80.0 100.0 120.0 0.0 20.0 40.0 60.0 80.0 100.0 120.0
Time (min) Time (min)
(a). (b)
SO.O
I..9- / ...-....... _._._.-
c:: 30.0 460.0
.2
1ii I
~ ~ 420.0 i
.~
10.0 "0
E 360.0 i'·-·_· __·-·-·_·_·_·-
~
III g
8..-10.0 ~ 340.0
~
c.
q::
~ 300.0
{!!. -30.0
~ 260.0
220.0
6.8
E
86.0
.I
.,.,
6.0 g-
N"'
84.0 I '-._._.-
I
Jr· ....... . . _._._.-._._.
I
0 82.0
5.2 x
"t:J 80.0 - - z: from 0.6 to 0.4
3l
c.
----- z: from 0.6 to 0.5
4.4 III 78.0 _.- z: from 0.6 to 0.7
_._._. z: from 0.6 to 0.8
- - z: from 0.6 to 0.4 0
III
III 76.0
----- z: from 0.6 to 0.5
3.6
-.-.- z: from 0.6 to 0.7 ~
E 74.0
_ . - z: from 0.6 to 0.8 0
0
2.8 L..-.L--'--...L..---'----L_ _'--.L..-""'"---'---'----'----' 72.0
0.0 2.0 4.0 8.0
6.0 10.0 12.0 0.0 2.0 4.0 6.0 8.0 10.0 12.0
Time (min) (X10- 2 ) Time (min) (x 10- 2 )
(a) (b)
FIGURE 23·18. Feed composition disturbances, propylene-propane vapor recom-
pression column. (a) Controlled variables; (b) manipulated variables.
pression column are slightly faster than the tional column. The manipulated variable
conventional. This can be explained by the movements indicate that the range of con-
lower operating pressure, higher relative trollability could be wider (bigger distur-
volatility, and fewer trays. bances could be handled) without running
Cooling water flow rate to the trim con- into valve saturation.
denser is used to control the pressure. As
shown in Table 23-2, the time constants of
23-6 CONCLUSION
the pressure-to-cooling water loop are much
faster than the composition dynamics. Both of the separations studied yielded the
Therefore, the pressure loop is tuned tightly same conclusion: The control system used
and the composition loops are treated as a on the conventional column can be applied
2 X 2 system in the BLT tuning method. to the vapor recompression column with the
The performance of the vapor recom- simple substitution of compressor speed for
pression system for step changes in feed steam flow. The effect of a change in com-
composition is given in Figure 23-18. The pressor speed is primarily to change the
response is slightly faster than the conven- vapor rate through the column, not to make
490 Practical Distillation Control
Moore, C. (1986). Application of singular value column with a heat pump. Presented at IFAC
decomposition to the design analysis and con- Symposium on Adaptive Control of Chemical
trol of industrial processes. Presented at the Processes, ADCHEM-88, Lyngby, Denmark.
American Control Conference, Seattle, WA Null, H. R. (1976). Heat pumps in distillation.
Mosler, H. A (1974). Control of sidestream and Chern. Eng. Prog. 73, 58-64.
energy conservation distillation towers. Pre- Quadri, G. P. (1981). Use of heat pump in P-P
sented at AIChE Continuing Education splitter. Part 1: Process Design. Part 2: Pro-
Meeting on Industrial Process Control, Tam- cess Optimization. Hydrocarbon Process. 60,
pa, FL. 119-126,147-151.
Mostafa, H. A (1981). Thermodynamic availabil-
Robinson, C. S. and Gilliland, E. R. (1950).
ity analysis of fractional distillation with vapor
Elements of Fractional Distillation. New York:
recompression. Can. 1. Chern. Eng. 59,
McGraw-Hili.
487-491.
Muhrer, C. A (1989). Study of the dynamics and Shinskey, F. G. (1979). Process Control Systems.
control of vapor recompression columns. New York: McGraw-HilI.
Ph.D. Thesis, Lehigh University, Bethlehem, Waller, K. V., Finnerman, D. H., Sandelin,
PA P. M., Haggblom, K. E., and Gustafsson, S. E.
Muhrer, C. A, Collura, M. A, and Luyben, (1988). An experimental comparison of four
W. L. (1990). Control of vapor recompression control structures for two-point control of dis-
distillation columns. Ind. Eng. Chern. Res. 29, tillation. Ind. Eng. Chern. Res. 27, 624.
59-71. Yu, C. C. and Luyben, W. L. (1986). Design of
Nielsen, C. S., Andersen, H. W., Branbrand, H., multiloop SISO controllers in multivariable
and Jorgensen, S. B. (1988). Adaptive dual processes. Ind. Eng. Chern. Process Des. Dev.
composition control of a binary distillation 25, 498-503.
24
Heat-Integrated Columns
William L. Luyben
Lehigh University
24-1 INTRODUCTION
feature multiple columns and intercon-
The rapid increase in energy prices in the nected energy systems. This markedly in-
1970s spurred a flurry of activity in energy creases the complexity of the process, pro-
conservation programs in the chemical in- duces higher-order systems, and introduces
dustry. Because distillation columns are ma- significant interactions. Thus, these heat-
jor energy consumers in many chemical integrated columns pose some challenging
plants, more energy-efficient designs for dis- control problems.
tillation systems were actively studied. Heat integration is used in distillation
One of the most successful energy con- systems in two basic forms:
servation measures for distillation columns
is the use of heat integration: the overhead 1. Energy integration only (columns that are
vapor from one column is used to provide separating different mixtures and making
heat to a second column. The idea is similar four or more different products).
to the use of multiple effects in evaporation. 2. Energy and material integration (multi-
Major reductions in energy consumption ple heat-integrated columns are used in
(about 30 to 40%) can be achieved in some the separation of a single feedstream into
separation systems. Therefore the economic two or more products).
incentives can be very large (payback peri-
ods of a few months). In this chapter we will
24-2·1 Energy Integration Only
discuss some of the steady-state design as-
pects of heat-integrated distillation columns Two Binary Feeds
and will explore methods for developing Figure 24-1 illustrates the case where there
effective control systems. are two feedstreams, Fl and F 2 , that con-
tain different chemical components. The
first column separates components A and
24-2 TYPES OF SYSTEMS
B; the second column separates components
There are a large variety of heat integration C and D. Figure 24-1a shows the conven-
schemes. Some of the more common ones tional non-heat-integrated case where the
are discussed in a fair amount of detail in two columns· operate completely indepen-
the following text. All of these schemes dently, with each having its own source of
492
Heat-Integrated Columns 493
energy and cooling. Because each uses cool- pressure (condenser pressure plus tray pres-
ing water in its condenser, the reflux-drum sure drop). The figure shows some hypo-
temperatures are about HO"F. Operating thetical but typical temperatures and pres-
pressures are the bubble-point pressure of sures. Some specific real examples will be
the overhead product at HO"F. Base tem- given later in this chapter.
peratures are the bubble-point tempera- Figure 24-1b shows the case where these
tures of the bottoms product at the base two columns have now been heat-
(R) (D)
(8)
F1
(~)
(8)
(b)
FIGURE 24-1. Energy integration, two binary feedstreams. (a) Conventional; (b)
heat-integrated.
494 Practical Distillation Control
integrated. The operating pressure in the In Figure 24-2b the columns have been
first column is raised from 15 to 120 psia so heat-integrated by raising the pressure in
that the temperature in its reflux drum is the first column high enough to make its
raised from 110 to 180"F. Because this tem- reflux-drum temperature higher than the
perature is now higher than the base tem- bottoms temperature of the second column.
perature of the second column, the over- Figure 24-2c shows another alternative
head vapor from the first column can be where the columns have been heat-
used to reboil the second column. The en- integrated by raising the pressure in the
ergy that is put into the reboiler on the first second column high enough to make its
column is being used twice, so very signifi- reflux-drum temperature higher than the
cant reductions in energy consumption are bottoms temperature of the first column.
realized. We call the first type forward integration
Note that the temperatures and pressure because energy flows in the same direction
in the second column are unchanged, but as the process flows (feed and energy are
the temperatures and the pressure in the both introduced into the first column). We
first column are higher. This has two re- call the second type reverse integration be-
sults: cause the flow of energy is in the opposite
1. If the relative volatility between compo- direction to the process flows (energy is
nents A and B decreases significantly as introduced into the second column while
temperature is increased, the column in the fresh feed is introduced in the first
the heat-integrated system will require column). These configurations are some-
more trays and/or a higher reflux ratio. what analogous to multiple-effect evapora-
2. The base temperature of the reboiler in tors with forward and reverse feeds.
the first column is increased, so higher- Which of these alternatives is better de-
pressure steam or other higher tempera- pends on the chemical system (temperature
ture heat sources must be used. Thus we dependence of vapor pressures, critical tem-
end up using less energy (fewer British peratures and pressures, effect of pressure
thermal units per hour or pounds per on relative volatilities, thermal sensitivity of
hour of steam) but requiring a higher components, corrosiveness, etc.) and on the
temperature source of heat (higher-pres- specific plant and process environment
sure steam). (steam pressures available, existence of
waste energy at some fixed temperature
Single TertUII'J Feed level, safety consequences of leaks in the
Another variation of this type of system is condenser-reboiler, etc.).
when there is a single feedstream with three Many other alternatives are also possible.
components (A, B, and C), and two columns For example, we could take component C
are used to make the separation into three out the bottom of the first column and take
products. As shown in Figure 24-2 a, the the A / B mixture out the top_ This distillate
conventional non-heat-integrated separa- would be fed into the second column where
tion scheme may consist of two columns, the A / B separation would be achieved.
each with its own sources of both cooling Heat integration could be in either the re-
and heating. The lightest component (A) verse or the forward directions. Figure 24-2d
would be taken overhead in the first col- shows an interesting possibility in which the
umn, and the mixture of components B and second column is a complex configuration
C would be taken out the bottom of the (two feeds and a sidestream). The first col-
first column. This bottoms stream is fed into umn is a "prefractionator," which produces
the second column, which separates B an overhead that contains almost all of the
and C. lightest component A, about half of the
Heat-Integrated Columns 495
C·)
' - - - - (B)
(A)
(b)
FlGURE 24·2. Energy integration, one ternary feedstreams. (a) Conventional; (b)
heat-integrated (forward); (c) heat-integrated (reverse); (d) prefractionator heat-
integrated reverse.
496 Practical Distillation Control
(i)
(0)
(d)
In almost all situations there will be an be used, even under steady-state conditions.
imbalance between the reboiler heat-input These auxiliary heat exchangers can then be
requirements of the second column and the used dynamically to control the energy in-
condenser heat-removal requirements of the put to the second column. Figure 24-3 shows
first column. Therefore an auxiliary reboiler the two cases, one involving an auxiliary
or an auxiliary condenser will almost always reboiler (when the first column does not
(.)
(b)
produce enough vapor to provide heat for umn, which operates at higher pressure.
the reboiler of the second column), and the About half of the lighter component is
second involving an auxiliary condenser removed as the distillate product (D t ).
(when the second column does not require The bottoms B t is a mixture of both
all the vapor from the first column over- components and is fed into the second
head). In the event that the columns are column, which operates at lower pres-
almost exactly balanced in terms of heat sure. The other half of the lighter com-
requirements (condenser heat removal re- ponent is removed in D2 and almost all
quirements in the high-pressure column of the heavier component leaves in the
equal the reboiler heat-input requirement bottoms product B 2 • There are only three
in the low-pressure column), both an auxil- product streams, so there are only three
iary condenser and an auxiliary reboiler compositions to be controlled. Steam is
must be used to handle dynamic transients. used in only the first column.
We will discuss this in more detail later.
Both an auxiliary condenser and an auxil- Many other alternatives exist such as us-
iary reboiler may be needed for start-up or ing reverse heat integration in the light-split
so that each column can be operated inde- case (operate the second column as the
pendently under certain abnormal condi- high-pressure column), taking about half of
tions. the heavier component out the bottom of
the first column and feeding the distillate
AlB mixture to the second column (heavy
24-2-2 Energy and Process Integration
split), and so forth.
In the configurations discussed in the pre-
ceding text, each column had different feeds
24-3 ECONOMIC INCENTIVES
and produced different products. Another
very common type of application of heat Table 24-1 gives a fairly detailed compari-
integration is the situation where we have a son of the steady-state designs of several
single binary feedstream. Conventionally we alternative configurations for the separation
would use one column to separate the two of a specific binary system: methanol-water.
components (see Figure 24-4a) with a The single conventional column consumes
steam-heated reboiler and a water-cooled 369,000 kcaljmin of energy. The feed-split
condenser. (FS) heat-integrated configuration con-
Instead of using just one column, it is sumes only 227,000 kcaljmin. This is a 38%
sometimes economical to use two (or more) reduction in energy consumption! The
columns that are heat-integrated. There are light-split forward (LSF) and the light-split
many alternative configurations. Two of the reverse (LSR) configurations consume about
more common are shown in Figures 24-4b the same amount of energy as the feed-
and c. split configuration: 241,000 and 226,000
kcaljmin, respectively, so all of these heat
1. Split feed (FS). The total feed is split integration schemes offer very significant
between a high-pressure column and a savings in energy.
low-pressure column. Each column Building two columns will cost more than
makes specification distillate product (D t building one column, even though the diam-
and D 2 ) and bottoms product (B t and eters of the two columns are smaller than
B 2 ). Thus there are four product streams the diameter of the single column. Typically
and there are potentially four composi- any of the heat-integrated schemes will re-
tions to be controlled. quire higher capital investment. But the
2. Light split with forward integration. All payback periods for the incremental invest-
the fresh feed is fed into the first col- ment are usually quite short.
Heat-Integrated Columns 499
'----_(A)
(b)
(e)
Single Column
Feed Split
Column 1 Column 2
Light-Split Forward
Column 1 Column 2
Light-Split Reverse
Column 1 Column 2
Note that pressures in the high-pressure used in the condenser. It is also seldom
columns of the FS and the LSF configura- economical to lower pressures too far into
tions are the same (3900 mm Hg), but the the vacuum range (except in the case of
base temperature in the high-pressure col- thermally sensitive material) because the low
umn of the LSF configuration is lower vapor densities increase column, condenser,
(125.6°C versus 146.7°C). This occurs be- and vapor-piping sizes quite significantly.
cause the material in the base of the column There is a limit to how high pressures
in the FS configuration is mostly water, can be raised to achieve heat integration.
whereas in the LSF it is a mixture of These limits can be imposed by several
methanol and water. The lower base tem- factors:
perature means that lower-pressure steam
can be used in the LSF configuration. The 1. Temperature sensitivity. The temperature
single column has the lowest base tempera- in the reboiler of the high-pressure col-
ture (94SC) but consumes much more en- umn may be limited to some maximum
ergy. value because of several possible factors:
Note also that the pressure in the high- thermal degradation, polymerization, de-
pressure column of the LSR is the lowest of composition, fouling, corrosion, and so
all the heat-integrated configurations (2280 forth.
mm Hg). This configuration should be the 2. Effect of pressure on vapor-liquid equilib-
most economical in those chemical systems rium. Raising pressure reduces relative
where relative volatilities decrease more volatilities in some systems. This requires
quickly with increasing pressure. more trays and higher reflux ratios.
Therefore pressures cannot be raised so
high that the separation becomes uneco-
24-4 LIMITATIONS
nomically difficult. Note that some chem-
Heat integration is quite attractive in many ical systems show increases in relative
systems, but it has some limitations. It is volatility as pressures are increased, so
seldom economical to lower pressures be- the energy efficiency of these systems
low the point where cooling water can be would be improved by raising pressures.
502 Practical Distillation Control
Knapp and Doherty (1990) discuss this severe because auxiliary reboilers and/or
problem for azeotropic separations. auxiliary condensers can be used to isolate
3. Critical temperatures or pressures. The the energy inputs for each column. As shown
pressure must be kept well below the in Figure 24-3a the auxiliary reboiler on the
critical temperatures and pressures of the second column would provide the additional
principal components so that the densi- energy that the second column requires. As
ties of the vapor and liquid phases are shown in Figure 24-3b the auxiliary con-
appreciably different. The hydraulics of denser removes the additional energy that
the column (liquid flowing downward the second column does not need.
against a positive pressure gradient)
would not function properly if the densi-
ties of the phases became too close. Energy and Process Integration
4. Temperature level of heat. The tempera- The control problem for those systems that
ture of the heating medium may be lim- have both energy and process integration is
ited. For example, steam in the plant more difficult. These systems feature
may only be available at pressures up to higher-order multivariable dynamics and
150 psig (saturation temperature 366"F). more severe interaction.
Assuming a 40°F temperature difference The feed-split configuration shown in
in the high-pressure column reboiler, Figure 24-7 is a 4 X 4 multivariable process.
base temperatures above 326°F could not The four controlled variables are the puri-
be used. ties of the two distillate products and the
two bottoms products. The four manipu-
Heat integration is seldom used in multi- lated variables are the heat input to the
component systems. If lighter-than-light key reboiler in the high-pressure column, the
components exist in the feed, they will re- two reflux flows, and the feed split (what
quire a higher pressure for the same fraction of the total feed is fed into each
reflux-drum temperature. This will raise the column). In this multivariable environment,
base temperature in the low-pressure col- the methods for diagonal controllers dis-
umn and require a much higher pressure in cussed in Chapter 11 or multivariable con-
the high-pressure column. If heavier-than- trollers should be used.
heavy key components exist in the feed, The light-split forward configuration
they will raise the base temperatures. This (Figure 24-4c) and the light-split reverse
will also increase the required pressure in configuration have only three final products
the high-pressure column. Thus most heat to control: compositions of two distillate
integration applications are in binary or products and one bottoms product. The
ternary systems with components that do three manipulated variables are heat input
not have extremely different boiling points. to the reboiler of the high-pressure column
and the two reflux flow rates. This is a 3 X 3
multivariable process.
24-5 CONTROL PROBLEM
Heat-integrated systems involve two col-
umns that are interconnected by energy
24-6 TOTAL HEAT-INPUT
flows and sometimes by process flows. They
CONTROL
are usually multivariable.
The use of a "total heat-input" controller is
Energy Integration Only recommended in those heat-integrated sys-
The control problem for those systems that tems that contain auxiliary reboiler and/or
have only energy integration is usually not condensers. Figure 24-5 illustrates the idea.
Heat-Integrated Columns 503
The first high-pressure column is controlled total heat input at its setpoint. The setpoint
in a conventional way: a composition con- of the total Q controller is the output signal
troller (CC) adjusts reflux to control distil- from the bottoms composition controller.
late purity, and a composition controller Note that the pressure in the high-pres-
adjusts steam to the reboiler to control bot- sure column is not controlled in the scheme
toms purity. In the low-pressure column the shown. It simply floats up and down as
flow rates of the steam to the auxiliary re- dictated by the required differential temper-
boiler and the vapor from the high-pressure ature for heat transfer in the reboiler. Thus
column to the condenser-reboiler are both at low throughput rates the pressure in the
measured. These flow rates are multiplied high-pressure column will be lower than at
by their appropriate heats of condensation high throughputs because a smaller IlT is
and the results are added together to obtain needed to transfer the smaller amount of
the total heat removed from both streams as energy. The pressure in the high-pressure
they condense in the reboiler of the low- column will also change if the compositions
pressure column. This signal is fed into a change in either the base of the low-pres-
"total Q" controller that adjusts the steam sure column or the reflux drum of the high-
valve on the auxiliary reboiler to keep the pressure column. The variable pressure
504 Practical Distillation Control
makes it vital to use pressure compensation ure 24-6, the auxiliary condenser heat re-
of temperature signals if they are being moval is adjusted by the control valve on
used to infer composition. the liquid line from the bottom of the con-
If an auxiliary condenser is being used, denser. The area for condensation is varied
the heat removed in the condenser (cooling by covering or exposing more area for heat
water flow rate times the temperature in- transfer.
crease of the cooling water) can be sub- This total heat-input controller decou-
tracted from the total heat available in the pies the heat inputs for the two columns
vapor from the high-pressure column to de- and eliminates some of the interactions
termine the "net" heat input to the re- present in the system.
boiler. A simple and effective alternative is
to place a control valve in the vapor line to 24-7 INCENTIVES FOR
the condenser-reboiler and use it to control COMPOSITION CONTROL
the heat input to the second column. The
OF ALL PRODUCTS
heat removal rate in the auxiliary condenser
could be used to control the pressure in the We mentioned earlier in this book that for a
high-pressure column. As sketched in Fig- single distillation column the minimum en-
Heat-Integrated Columns 505
ergy is consumed by controlling both prod- curve of the feed split (FH/FL ) are fairly
ucts at their specified purity levels. This flat, there may be very little energy saved by
"dual composition" control strategy is, how- using the 2-2 scheme (4 X 4 multivariable).
ever, frequently not economically justified The control problem can be reduced to
because the energy savings compared to sin- 3 X 3 by not controlling bottoms composi-
gle-end control are often fairly small in many tion in the low-pressure column and simply
binary separations. The control engineer fixing the feed split (see Figure 24-9a; called
needs to quantify the economic incentives the 2-1 scheme). The first column is just a
for dual composition control. 2 X 2 system. The second column is dis-
The same question needs to be answered turbed by changes in the vapor entering its
in heat-integrated systems. For binary sys- reboiler from the first column, but the con-
tems is it often possible to simplify the trol problem in the second column is just
control problem by not controlling the com- single-input-single-output (SISO). The in-
positions of all product streams. Figure 24-7 teraction between the columns has been
shows the feed-split configuration with the broken.
compositions of both products in both Figure 24-9b shows an even further sim-
columns controlled (called the 2-2 scheme). plification (the 1-1 scheme) in which only
This is a full-blown 4 x 4 multivariable con- the distillate compositions are controlled.
trol problem. The feed-split ratio and the reflux to the
Figure 24-8 illustrates some typical high-pressure column are fixed. Heat input
steady-state results for many binary systems. in the high-pressure column and distillate
The specific chemical system is isobutane- flow rate in the low-pressure column are the
n-butane. Feed composition is varied over manipulated variables. Each column is a
the expected range and the compositions of simple SISO control problem.
all four products are held constant. The These simpler approaches will not be
required steady-state changes in several of economically attractive in all cases, particu-
the manipulated variables are shown. Be- larly multicomponent systems and systems
cause the curves of reflux-to-feed in the with low relative volatilities because too
high- and low-pressure columns and the much energy will be wasted.
506 Practical Distillation Control
27
1
ae
U~ 25
24
ae
25
20
J J 10
115 I 9
5.5
4.5
12 3.5
5.5 2
4.5 1
3.5 '---...,..--..,----,,.---..,
.2 .3 .4 .2 .3 .4
FIGURE 24-8. FS changes in system variables with feed composition for 2-2
scheme.
Heat-Integrated Columns 507
I
.I I
I I
!I
l ___________________ JI
® (b)
Ii1GURE 24-9. (a) 2-1 scheme (FS). (b) 1-1 scheme (FS).
William L. Luyben
Lehigh University
In binary separations, there are two ing. Figure 25-2 shows some typical changes
products and one slop cut (the material in distillate composition (x D) and bottoms
removed as distillate during the period when composition (x B) for a binary system. Dur-
the distillate contains too much heavy com- ing the start-up period when the column is
ponent to include in the light product and on total reflux, the top composition gets
the contents of the still pot and column richer and richer in the more volatile com-
contain too much light component). In ponent. When the purity of the material in
ternary separations, there are three prod- the column overhead reaches its specified
ucts and two slop cuts; in general we obtain level, distillate flow is begun at some appro-
at most NC - 1 slop cuts and NC products priate flow rate (which reduces the reflux
(where NC is the number of components). ratio from infinity to some finite value).
Figure 25-1 shows the configuration and There are a number of ways to adjust the
nomenclature for a ternary batch distilla-
tion.
1.0
Mter the fresh feed and any recycled
E
slop cuts have been charged to the still pot, ~
there is an initial start-up period during ~08
which the column operates at total reflux 8
E
until a distillate of the desired purity in the .51'
-;: 0.6
most volatile component can be withdrawn. 0
~
Batch time is established by the time it .I:
~ 0.4
takes to produce NC - 1 distillate products E
.,
and to purify the heavy product left in the c
1
.2
still pot. 0 .2
~
0
0 2 3 4 5
25-2-2 Composition Profiles Time (hours)
In batch distillation the compositions at all FIGURE 25-2. Typical composition profiles for bi-
points in the column are continually chang- nary batch distillation.
510 Practical Distillation Control
./ \\
that it is assumed that the liquid on the
trays (or packing) in the column drains into
\
.\
, I \ the still pot. This slop cut will contain both
\
\\ \
\ components. It is usually recycled back to
\\. \
\
\
the next charge to the still pot.
__ ________ ____- u
OL-~\~ ~~ ~~
For ternary systems, there can be two
Time
slop cuts. The first will contain mostly the
FIGURE 25-3. Typical distillate composition profiles light component and the intermediate com-
for ternary batch distillation. ponent. The second slop cut will contain
mostly intermediate component and the
heavy component. The size of these cuts
reflux ratio during the course of the batch. depends on the composition of the initial
The easiest and most commonly used charge, the number of trays in the column,
method is to simply hold the reflux ratio the reflux ratio used, and the degree of
constant (at some optimum constant value). difficulty of the different separations (rela-
This policy produces distillate whose com- tive volatilities and product purities).
position varies during the batch.
The distillate is stored in a product tank.
When the average composition of the mate- 25-3 ASSESSMENT OF
rial in the tank drops to the specified purity PERFORMANCE: CAPACITY
level xi5c , the distillate stream is diverted FACTOR
to another tank and the slop cut S is pro- In a batch column we can have many objec-
duced. When the material in the still pot tives, such as minimize batch time, maxi-
and in the column (not including the reflux mize production in a fix period of time, and
drum because this is added to the slop cut) so forth. One quantitative measure of per-
meets its specification for the heavy prod-
r
uct, x c , the binary batch cycle is stopped.
formance is the capacity factor defined by
Luyben (1971). The capacity factor (CAP)
In a ternary system (see Figure 25-3) the of the batch still is defined as the total
lightest component is removed in the first specification products produced divided by
distillate product (pl). Then the first slop the total time of the batch. Extending the
cut (Sl) is produced until the distillate pu- original equation to a multicomponent sys-
rity (in terms of the intermediate compo- tem and assuming 0.5 h to empty and
nent) reaches the specification value. The recharge the still pot
second distillate product (P2) is produced
until its average purity drops to its specified
value. At that point the purity of the last Ef:llj
CAP = (25-1)
product (P3) may be better than specified. tF + 0.5
If not, a second slop cut (S2) is produced as
required to purify the heavy product. In this definition it is assumed that all prod-
ucts have the same economic value. A weight
factor can be easily included to account for
25-2-3 Slop Cuts
differences in prices. The total batch time t F
As already mentioned, in a binary distilla- (hours) includes the time at total reflux and
tion the slop cut is the distillate that is the time producing the products and slop
Batch Distillation 511
cuts. Note that the total products produced steady-state distillation, investigated theo-
also will be equal to the net fresh feed retically the stability characteristics of sev-
charged to the still pot, which is equal to eral numerical methods, and predicted the
the total still pot charge (HBO ) minus the maximum increment size to be used in each
sum of the slop cuts: one. Holland and Liapis (1983) solved the
equations for multicomponent batch distil-
NC NC-l lation using the two-point implicit method
E lj = H BO - E Sj (25-2) and the theta method convergence, as well
j=l j=l as a combination of the two implicit method
and the 2N Newton-Raphson method.
The capacity factor can be used to deter- Azeotropic multicomponent batch distilla-
mine the optimum column design. As one tion was treated by Van Dongen and
would intuitively expect, CAP increases as Doherty (1985), who developed a model for
more trays are added to the column for the ternary azeotropic columns operating at high
same vapor rate in the column. This corre- reflux ratios and with a large number of
sponds to the same energy consumption and plates. The model agreed well with expe-
the same column diameter. Thus, more feed rimental results. Bernot, Doherty, and
can be processed for the same energy cost. Malone (1991) predicted the possible set of
Alternatively, a smaller diameter column fractions that can be obtained for any initial
and less energy could be used to handle the composition of an azeotropic mixture.
same amount of material per unit time. Of
course, the capital cost of the column in-
creases as more trays are added. Therefore, 25-4·1 Differential Distillation
the capacity factor can be used to perform
The simplest type of batch distillation is
the necessary economic trade-off calcula-
called differential distillation, shown in Fig-
tions between capital and energy cost to
ure 25-4. The vapor is generated by boiling
determine the optimum number of trays
the liquid in the still pot. No rectifying
and optimum reflux ratio.
column is used. Rayleigh developed an
equation for differential distillation. Con-
25·4 MODELS sider a binary batch of Ho moles of liquid.
At any instant during the distillation there
Since the early 1960s when large digital
are H moles of liquid left in the still pot. At
computers become available, rigorous com-
that instant, the mole fraction of the more
puter-based calculation procedures have
been generated for binary and multi compo-
nent batch distillation. Meadows (1963) pre-
sented a multicomponent batch distillation y (t )
model, Distefano (1968a, b) extended the
model that was used to simulate several
commercial batch distillation columns, and
Boston et al. (1981) further extended the
model to efficiently handle the nonlinear
nature of the problem. The stiff nature of
H (t )
multicomponent batch distillation equations
sometimes requires the use of special nu- x U)
Q
merical methods. Distefano (1968b) pre-
sented the cause of instabilities that arise
during the numerical solutions of ordinary
equations, showed it in the solution of un- FIGURE 25-4. Differential distillation.
512 Practical Distillation Control
By applying the proper limits, we obtain the operated at constant reflux ratio. The oper-
well-known Rayleigh equation (Coates and ating lines at different times are a series of
Pressburg, 1961; Block, 1961): parallel lines, the slopes being the same
because L/V is constant. As the light com-
(HI)
In - ltl - -1 d x
= (25-5)
ponent is depleted from the system, the
composition of the heavy component in-
Ho t=oY-X
creases on all trays and in the still pot. The
This simple example illustrates the "dy- dashed lines give the construction for a later
namic" nature of the batch distillation pro- time when the overhead composition is x D 2'
cess. As a result, a steady-state analysis This type of pseudo-steady-state model
of the type employed for continuous dis- can only be used in those situations where
tillation cannot be made of the column the liquid holdups on the trays in the col-
behavior. umn and in the reflux drum are negligible
compared to the volume of the still pot.
Chiotti and Iribarren (1991) used this type
25-4-2 Pseudo-Steacly-State Models of steady-state model for optimization
Several attempts to study the process as a studies.
variation of a continuous column have been
made. Galindez and Fredenslund (1988)
modelled the unsteady-state process by sim-
25-4-3 Rigorous Dynamic Models
ulating a succession of short periods of time
in which a continuous, steady-state distilla- To describe batch distillation columns accu-
tion process is assumed. To illustrate the rately, we must use a dynamic model that
procedure let us assume a binary system includes tray and reflux-drum holdup dy-
with equimolal overflow where we can use namics. Based on the assumptions of con-
the McCabe-Thiele diagram to relate com- stant tray and reflux drum holdups, constant
positions. Figure 25-5 shows the McCabe- relative volatilities, and equimolal overflow,
Thiele analysis for a two tray column that is the equations describing a multicomponent
Batch Distillation 513
ajX Bj
YBj = ~~---=- (25-8) These equations can be solved using the
E~~lakXBk numerical integration methods discussed in
Tray n: Chapter 3.
dx nj
MnTt = R[xn+l,j - x nj ] 25-4-4 Fitting Models to Experimental
Batch Distillation Data
+ V[Yn-l,j - ynjJ (25-9)
Muhrer, Grassi, and Silowka (1990) pre-
(25-10) sented an industrial case study that illus-
trated how a dynamic model can be used in
Tray NT (top tray): conjunction with plant data to increase pro-
cess productivity. It was assumed that the
dx N .
MN ----I:..!... = R[x D · - x N
dt J
.J vapor-liquid equilibrium was reasonably
T T,}
well known. Simulations using the model
+ V[ YNT-I,j - YNT,j J (25-11) were run and tray efficiency and holdup
were adjusted until the model matched the
(25-12) plant data. Figure 25-6 gives a comparison
of plant data and model predictions.
c:
1.0
I
, .... ~
J
,(-. iy--.. . . '\
\ .
I
I
I 1 ,
I , 'I
CD
(5 I
I
II
\.;, . \\
:E I • Comp 1
I ,J \ • Comp 2
",
4Comp3
: .
/~
tComp 4
.Comp 5
./
.I
'\
\
°0~~~~-'~-~-~~~~-~1~5---~~~L----~25
\ ,
Time (h)
FIGURE 25-6. Model verification (distillate compositions).
514 Practical Distillation Control
The model was then used to test various 25-6 REFLUX RATIO
control schemes, investigate the effect of TRAJECTORIES
operating procedures on batch cycle, and
determine the necessary measurements One of the most interesting aspects of the
needed to infer product and slop cuts. In operation of a batch distillation column is
one typical example process, the amount of the question of how to best set the reflux
slop cuts and the cycle time for the batch ratio during the course of the batch cycle.
were reduced significantly. Cycle time was There have been a variety of approaches.
reduced from 21 to 16 h by using model The optimum theoretical solution to the
analysis. problem involves the use of the calculus of
The model was also used to design new variations to solve for the optimum time
columns by applying first principles and trajectory of reflux ratio that extremizes
augmenting the analysis with engineering some objective function (Coward, 1967). The
experience obtained from working systems mathematical complexity of this elegant ap-
at similar plants. proach makes it very difficult to apply in
realistically complex industrial batch distil-
lation columns.
The two most frequently used ap-
25-5 COMPARISON WITH proaches are to operate at a constant reflux
ratio (with distillate composition varying) or
CONTINUOUS DISTILLATION
to operate with distillate composition (or
Continuous distillation is normally more en- overhead temperature) held constant (with
ergy efficient than batch distillation, but reflux ratio varying). The former strategy is
batch distillation requires only one distilla- the more commonly used because of its
tion column, even for the separation of a simplicity. Several investigators have stud-
multicomponent mixture, whereas in contin- ied the differences between these two
uous distillation a sequence of columns is approaches. The results reported are con-
usually required. To illustrate the differ- flicting: Some researchers claim significant
ences let us compare the two alternative differences and others have found only slight
processes for a specific ternary separation: differences. This remains an open question.
relative volatilities of 9/3/1, product puri- Results are probably strongly dependent on
ties of 95 mol%, and feed composition 30 the size of the column, the difficulty of
mol% light, 30 mol% intermediate, and 40 separation (relative volatility), and the prod-
mol% heavy. By use of standard short-cut uct purities.
design procedures for continuous distilla- As one specific example, Luyben (1988)
tion with reflux ratios 1.1 times the mini- studied a ternary separation with relative
mum, a total of 100 mol/h of vapor boilup volatilities of 9/3/1 in two different
in two continuous columns would corre- columns (20 trays and 40 trays). Product
spond to a feed rate (or capacity) of about purities were 95 mol%. Capacity factors for
80 mOl/h. Each column would require 12 the optimum fixed reflux operation were
trays. 38.6 and 45.4 for the 20 and 40 tray columns.
Making the same separation in a 20-tray Capacity factors for a variable reflux opera-
batch distillation with the same energy con- tion were 33.8 and 44.4. These results indi-
sumption (100 mol/h of vapor boilup) yields cate no major differences.
a capacity of about 37 mol/h. If 40 trays As a conflicting example, Farhat et aI.
were used in the batch column, the capacity (1990) studied fixed, linear, and exponential
would be about 45 mOl/h. If 80 trays were reflux trajectories. They claim a 10% im-
used, capacity would increase to about 55 provement, which they felt was significant.
mol/h. However, they used a very small column
Batch Distillation 515
(nine trays) and a very simplified model nents come overhead. This may be desirable
(neglected tray and reflux-drum holdups). from the standpoint of temperature sensi-
More work is needed to clear up some of tive material. It also may make the separa-
these questions. As of now, a reasonable tion easier if relative volatilities increase as
mode of operation would appear to be the pressure decreases.
use of optimal fixed reflux ratios, which However, there is a question of capacity.
could be different for each product cut if Lower pressure results in lower vapor den-
relative volatilities are different between the sities and higher vapor velocities. If the
various components. column is vapor-flooding limited, the
throughput of the column may be reduced if
pressure is reduced. In this situation, there
25-7 PRESSURE TRAJECTORIES is probably an optimum operating strategy
There has been very little investigation of that varies both heat input and column
how column pressure should be varied over )ressure during the batch cycle to maximize
the course of the batch. Some of the possi- capacity.
ble strategies are now briefly discussed.
adding additional trays. This is not always view. We are mixing together streams that
the case, particularly in vacuum distillation have already been partially separated. Each
systems because of the increased pressure slop cut is approximately a binary mixture.
drop. Quintero-Marmol and Luyben (1990)
However, using more trays also increases studied several slop-handling strategies.
the height of the column and therefore its Some typical results are summarized in the
capital cost. So the optimum design of a following text for two of the most interest-
batch distillation column must balance the ing alternatives and are compared with the
economic and operational aspects of all total slop recycle strategy.
these considerations. AI-Tuwaim and
Luyben (1991) gives some simple guidelines
to aid in the shortcut design of batch distil- 25-9-1 Alternatives
lation columns for different cost of energy,
materials of construction, relative volatili- Three operating strategies are considered:
ties, and product purities.
For example, Table 25-1 defines several 1. Total slop recycle. All slop cuts are mixed
cases with typical materials of construction, with fresh feed to make the initial charge
product purities, and energy costs. For each to the still pot for the next batch. The
of these cases, Figure 25-6 gives the number capacity of the column is total product
of trays that gives the minimum cost (capital streams produced, which is equal to the
plus energy) as a function of relative volatil- total still-pot charge minus the slop cuts
ity. produced. This is the most simple proce-
For a specific example, take a batch sep- dure and the most commonly used. How-
aration of a ternary mixture with 95% prod- ever, its thermodynamic efficiency is low.
uct purities, Monel materials, and $2.50/10 6 2. Multicomponent-binary batches. A num-
Btu energy cost (case 9 in Table 25-1). Fig- ber of batch distillations are run produc-
ure 27-7 indicates that 70 trays should be ing the NC product streams and the NC
used if relative volatilities are 4/2/1, but 30 - 1 slop cuts. All slop cuts are stored in
trays should be used if relative volatilities their own segregated tankage until
are 9/3/1. Figure 25-8 gives the reflux ratio enough material has been accumulated
that gives the maximum capacity factor. The of each slop cut to run a batch distilla-
optimum reflux ratios for the two cases con- tion of each slop cuts. Because these slop
sidered above are 2.2 and 1.5, respectively. cuts are essentially binary mixtures, these
batch distillations are binary. The slop
cut from each binary distillation is recy-
cled to its own slop cut tank. Products of
25-9 SLOP CUT PROCESSING
the binary distillations are mixed with
As discussed earlier,- there can be up to the products of the multicomponent dis-
NC - 1 slop cuts produced when a mixture tillations.
of NC components is separated in a batch This procedure is the most efficient,
distillation column. The best way to repro- but it requires more tankage and more
cess these slop cuts is an interesting prob- complex operating procedures.
lem. There are a host of alternative proce- 3. Fed batch. The charge to the still pot is
dures. all fresh feed. During the course of the
The simplest is to recycle all the slop cuts batch distillation each slop cut is fed into
to the next batch. The initial charge to the the middle of the column when the com-
still pot is then all the slop cuts plus fresh position on the feed tray is approxi-
feed. However this simple strategy makes mately equal to the composition of the
little sense from a thermodynamic point of slop cut. This procedure requires less
Batch Distillation 517
1.3.7
II)
~
~
15 80 2.4.6.8.12
...
Q)
.0
E
:J
10
c::
E
:J 5.11
E 60
a
0
~1L-----------~2~----------~3----------~4
Relative volatility
10
N =10
8
N =30
0
~ 6
)(
:J
N =50
!
E
:J N =70
E N =80
8- 4
o~----------~------------------------~
1 2 3 4
Relative volatility
tankage and is more efficient than slop (two product purities and two relative
recycle. However there are a large num· volatilities). The multicomponent-binary
ber of parameters to optimize: slop feed slop-handling strategy gives the highest ca-
time and flow rate for each slop cut, feed pacity factor. Figure 25-9 compares the two
tray location, reflux ratio, and so forth. slop-handling strategies for different prod-
uct purities and different numbers of trays.
The increases in capacity factors in going
25·9·2 Results
from the easy slop recycle operation to the
Table 25·2 gives typical results (in terms of more complex binary·multicomponent oper-
capacity factors) for four different systems ation are on the order of 30 to 50%. This
50
a=4/2/1
40
20
could be very significant in some batch dis- tual measured temperatures can be com-
tillation applications. However, in some pared with the temperatures predicted by
plants the increased complexity may not be the mathematical model, and any differ-
justified by the economic gain. ences can be used to correct the model
predictions.
This procedure in the control field is
25-10 INFERENTIAL CONTROL called constructing a state estimator or ob-
OF BATCH DISTILLATION servers. A Kalman filter is one example of
such a state estimator. It can be used when
25-10-1 Problem signals are noisy (contain a lot of uncer-
In theory, the control of batch distillation tainty). Some very elegant statistical mathe-
columns should be fairly simple. If a compo- matics and optimal control theory are used
sition analyzer is available that gives the in Kalman filter design.
instantaneous distillate compositions during A more simple observer is a Luenberger
the course of the batch cycle, one could observer that uses deterministic models.
simply switch from product to slop cuts at Other types of observers (models) have also
the appropriate times while following the been used. Quintero-Marmol and Luyben
optimal reflux ratio and pressure trajecto- (1992) and Quintero-Marmol, Luyben, and
ries. However, composition analyzers are Georgakis (1991) explored the application
often not available and seldom instanta- of several observers to the problem of the
neous. control of batch distillations. Some of the
As an alternative, a mathematical model highlights of this work are discussed in the
could be used to predict how compositions following text.
will change with time and could predict
switching times. However, two requirements
would have to be met if this "open loop" 25-10-2 Basic Insight
model is to be successfully used:
One of the most interesting and important
findings of the work on inferential control
1. The model must be perfect. It must very
of batch distillation is the way the composi-
accurately describe the system, including
tion fronts move through the column during
such parameters as the vapor-liquid
the course of the batch cycle. Figure 25-10
equilibrium, tray efficiencies, heat losses,
illustrates this phenomenon for a ternary
and so on.
system. At the beginning of the batch
2. The initial conditions in the still pot and
(time = 0) the compositions on all trays and
in the column and reflux drum must be
in the still pot are assumed to be the same
accurately known. If the model starts
(flat composition profiles). While the col-
integrating from the incorrect initial con-
umn is operated under total reflux condi-
ditions, the entire trajectory will be in-
tions, the concentration of the lightest com-
correct.
ponent (Xl) builds up; on the upper trays in
the column and the concentrations of the
Almost never will these two requirements intermediate component (x 2 ) and the heavi-
be satisfied. Therefore, if a mathematical est component (x 3 ) decrease in the top of
model is to be used to predict compositions, the column but increase in the still pot. At
some type of feedback correction must be the time when withdrawal of the first prod-
used. Temperature information is usually uct (p1) is begun (time = 1.5) there is a
available from a number of locations in the composition front located in the lower part
column during the entire batch. These ac- of the column that separates the lightest
520 Practical Distillation Control
·..X3
........
....... ....,
10 10
0 0
0 0.5 0 0.5
Composition Composition
30 30
--- ---~ ...,
ci ci
,
c: c: I
x,
!
20
-- ...
"',, ~
~
20 X2 1
I
I
I
10 I 10 I
I
--- --'"
I
--'"
/
X3 I
".
0
0 0.5 0.5
Composition Composition
",'
~
,,
10 10 '" '.
".
.... ..':.:'.;,;»-.!'":-;:----
X3 _, (
0 ". 0
0 0.5 0 0.5
Composition Composition
!
20 x3 : >- 20
!!!
~
10 10
I
I'
I
0 0
0 0.5 0 0,5
Composition Composition
and intermediate components; this ~front The significance of this behavior is that
moves up the column (time = 2.25) as light simple temperature and pressure measure-
product is removed. ments can be used to estimate compositions
When this front nears the top of the in these binary systems. The components
column, the first slop cut (Sl) is begun. It is change during the batch cycle, but for any
essentially a binary mixture of components specific binary mixture, compositions can be
1 and 2. When the purity of component 2 inferred from temperature and pressure
reaches its specification, the withdrawal of measurements.
the intermediate product (P2) is begun.
Now a second composition front forms
near the base of the column, which repre-
25-10-3 Quasldynamlc Model
sents the interface between components 2
and 3. This front sweeps up the column as This binary wave-like behavior of batch
intermediate component 2 is removed in distillation columns is the basis for the
the distillate. The shapes of these fronts de- "quasidynamic" ad hoc estimator studied by
pend on the reflux ratio and the relative Quintero-Marmol and Luyben (1992). A
volatilities. temperature is measured in the middle sec-
Thus, what is in the upper part of the tion of the column. The composition of the
batch distillation column at anyone time vapor phase of the appropriate binary mix-
during most of the batch cycle is some bi- ture at that point in time and at that loca-
nary mixture. Initially it is the binary mix- tion is calculated from the measured tem-
ture of components 1 and 2. Then it be- perature and pressure and the vapor-liquid
comes components 2 and 3. This continues equilibrium relationships.
for as many components as there are in the As illustrated in Figure 25-11, this calcu-
initial feed to the unit. lated binary vapor is fed into a dynamic
Reflux drum
Column
Product receivers
0.9
0.8
0.7
I - - - s.s. estimator
0.6
XD 0.5
0.4
0.3
0.2
0.1
2 3 4 5 6
Time (h)
0.9
0.8
0.7 --Actual
_. - Quasidynamic
0.6
XD 0.5
0.4
0.3
0.2
0.1
2 3 4 5 6
Time (h)
an approximate idea of the composition of bottom half of the figure actual distillate
the initial charge to the still pot. The distil- compositions are compared to the predic-
late compositions predicted by the model tions of the quasidynamic estimator. The
converge fairly quickly to the actual distil- predictions are quite good.
late compositions during the period of total
reflux operation.
Figure 25-12 gives some typical results.
25·10-4 Extended Luenberger
In the top part of the figure actual distillate
Observer
compositions are compared with the predic-
tions of a simple steady-state estimator Most of the work on observers in the con-
(holdups on trays and in the reflux drum are trol literature has been limited to low-order
neglected). The steady-state estimator lags systems. Our batch distillation processes is
the real column by 5 to 10 min. In the quite high order (126 differential equations
Process:
Measured
temperatures
Observer:
Predicted Predicted
compositions temperatures
f 40
f26
7;3
fs
FIGURE 25-13. Extended Luenberger observer.
524 Practical Distillation Control
for a 40 tray ternary column). There have observer will converge the predicted tem-
been very few studies of the application of peratures toward the measured tempera-
observers to complex, nonlinear chemical tures. Then the distillate composition pre-
engineering systems. dicted by the model can be used to control
A Luenberger observer uses a linear the batch cycle.
model of the process to predict its states. The rate of convergence of the observer
An extended Luenberger observer uses a depends on the gains ~ used. Making these
nonlinear model of the process. Because gains large produces small closed-loop time
this is a batch process, compositions vary constants (large negative eigenvalues) and
drastically during the course of the batch. rapid convergence. However, if the gains
Therefore, a linear model gives very poor are too large, the observer can become too
predictions of the actual behavior and a underdamped or even closed-loop unstable.
rigorous nonlinear model must be used. Therefore, the design of the ~ matrix is
The nonlinear dynamic model has the crucial to the successful operation of the
form observer. "Pole-placement" methods are
usually employed using a linear model of
(25-15) the system for the local operating point.
The observer will only be useful in the
(25-16) batch distillation process if it is able to
converge to the correct distillate composi-
tions (starting from some assumed initial
where ! = states of the process (composi- conditions because we seldom know the ini-
tions on all trays, reflux drum, tial compositions in the charge to the still
and still pot; for a 4O-tray pot) before the end of the total reflux oper-
ternary column there are 126 ating period (before the withdrawal of the
states) first product P1 has begun). Therefore, the
! = measurements (temperatures observer must converge fairly rapidly, but it
on several trays in the column, must be stable.
for example, four temperature
measurements)
1.0
The structure of the extended Luenberger
\
observer is sketched in Figure 25-13. At the
top of the figure the process is shown with 0.8 \\ ~
~
the measured inputs R and V (reflux flow \ ~
\ ~
rate and steam flow rate) and the measured
0.6
\
\ ~
'.
temperatures (in the still pot and on trays \ ~
. \ ~
40,26, and 13). These process data are fed I \
I ~
into the observer. The differences between \ \
..
0.4
the measured temperatures and the temper- \
\
~
There are a number of important practi- the region of convergence is small, in-
cal design questions to be answered about crease the number of measurements.
the design of the extended Luenberger ob-
server: Quintero-Marmol et al. (1991) found that
it was not necessary to update the ~ matrix
1. How many temperature measurements of gains during the batch. The gams were
are needed? fairly insensitive to the composition profile
2. Where should they be located in the in the column, and once convergence was
column? obtained during the initial total reflux oper-
3. Where should the closed-loop poles be ation, it did not matter much what gains
located? were used.
4. Does the ~ matrix of gains have to A typical number of temperature mea-
be changed as the nonlinear process surements was four to five. Figure 25-14
changes? illustrates how the region of convergence
5. How close must the guessed initial con- increases as the number of temperature
ditions be to the actual initial conditions measurements (M) is increased.
in order to get successful convergence of Figure 25-15 illustrates how distillate and
the observer? bottoms compositions (actual and predicted
by the observer) vary with time. The specific
Quintero-Marmol et al. (1991) studied numerical case is a ternary 20-tray column
most of these questions and proposed the with five temperature measurements (base
following design procedure: and trays 5, 10, 15, and 20). The actual
initial conditions were 0.25/0.65/0.15,
whereas the initial conditions guessed were
1. Guess the number of temperature mea-
0.333/0.333/0.334. Note that the observer
surements (M).
converges in about 15 min, which is well
2. Locate these temperature measurements
before the PI withdrawal begins. Figure
at the top of the column, in the still pot,
25-16 shows what happens when only three
and equally spaced up the column.
temperatures were measured (base and trays
3. Use a linear model (linearized around
10 and 20). The response is slower and
the guessed initial conditions) and a ro-
much more oscillatory, particularly in the
bust pole-placement algorithm (for ex-
reboiler where they last for almost 2 h.
ample, the MATLAB package, which
The extended Luenberger observer was
uses the work of Kautsky, Nichols, and
found to be somewhat better than the
van Dooren, 1985). Specify the closed-
quasidynamic, but it is more difficult to de-
loop poles to be equal to the open-loop
sign. Most batch distillation applications can
poles with the smallest of the open-loop
probably use the quasidynamic observer, but
poles moved to the left in the s plane.
others may require a more complex ob-
These should be moved to the left until
server.
the closed-loop response of the observer
is as fast as possible but still has a rea-
sonable damping coefficient (greater than
0.3).
25-11 CONCLUSION
4. If the observer does not converge quickly This chapter has attempted to discuss briefly
enough, increase the number of mea- some of the many important and interesting
surements and repeat steps 2 and 3. aspects of the design and operation of batch
5. Test the sensitivity of convergence to the distillation columns. Much more study in
error in the guessed initial conditions. If this area is needed.
526 Practical DistiUation Control
~------S1------~~--
0.9
0.8
0.7 -Actual
0.6; - - - -088-5
XD 0.5
0.4
0.3
0.2
0.1
00 2 3 5 6
Time (h)
0.8r----"'T-"--~----r---___r---..._--...,
0.7
-Actual
Xs 0.4 ----Oa5-5
0.3 \
2 3 4 5 6
Time (h).
DGURE 25·15. Observer with five temperature measurements.
Batch Distillation 527
P2
0.9
0.8 1
0.7 -Actual
0.6 -'-'--085-3
0.5
Xo
0.4
0.2
0.1
2 3 4 5 6
Time (h)
0.9 ~
0.8 ;";:". j\
0.7;; ;;.
!
0.6 ;
Xs 0.5 -Actual
0.4 ----085-3
0.3
ii
2 3 4 5 6
Time (h)
Pairing, 244
Material-balance structure, 192, 196, 314 Parameter estimation, 99
Mathematical model, 340 Parametrization, 81
~Tld\B, 101, 123, 127, 149 Parseval's theorem, 104
Maximizing recovery, 284 Parsimonious model, 97, 113
Maximum peak criterion, 291, 292 Partial least squares, 97, 114
Maximum reflux ratio, 385 Performance index, 279
Measurement noise, 98 Persistent excitation, 99
Message, 74 Perturbation signal design, 99
Minimum approach temperature, 380 Phase margin, 291, 292
Minimum reflux ratio, 384 Piece-wise deterministic disturbance, 98
Minimum variance control, 185, 444 Plant characteristic curve, 474
Model-based control, 101, 248, 440 Plant dynamic tests, 455, 462
Model-plant mismatch, 291 Plantwide control, 413
Model uncertainty, 291 Plantwide process control simulator, 86
Model validation, 99, 119 Pole-placement, 524
Morari resiliency index, 233, 235, 463 Polymorphism, 73, 75, 79
Move suppression, 111,254 Ponchon-Savaritt method, 385
MTBE (methyl tertiary butyl ether), 333 Positive zero, 9
Multivariable closed-loop characteristic equation, 235 Power spectral density, 104
Multivariable identification, 257 Power spectrum, 101, 104
532 Index