Professional Documents
Culture Documents
Robust Control of An Industrial Distillation Column
Robust Control of An Industrial Distillation Column
10666
Robust Control of an
Industrial High-Purity
Distillation Column
presented by
HANS-EUGEN MUSCH
Dipl. Chem.-Ing. ETH
born June 19, 1965
citizen of Germany
1994
3
To my grandparents
4
5
Acknowledgments
This Ph. D. thesis was written during my years as a research and educa-
tional assistant of the Measurement and Control Laboratory at the
Swiss Federal Institute of Technology (ETH) at Zurich. I would like to
take this opportunity to thank the numerous persons who have
supported this project.
First of all I express my gratitude to Prof. M. Steiner. He arranged this
project and helped to overcome many difficulties with the industrial
environment. Many thanks are also due to him and to Prof. D. W. T.
Rippin for the critical examination of this thesis, which essentially
improved its clarity.
The numerous discussions with my colleagues and their uncountable
suggestions gave rise to important contributions to this work. In this
context, E. Baumann, U. Christen, and S. Menzi must be specially
mentioned.
Last but not least I should emphasize the support of B. Rohrbach. She
never lost her patience with my never ending questions concerning the
English language. Without her willingness to correct the manuscript,
the choice of the English language for this thesis would have been
impossible.
6
7
Content
Abstract ......................................................................................... 15
Kurzfassung .................................................................................. 17
1 Introduction ......................................................................... 19
10 Conclusions and
Recommendations ............................................................... 219
Symbols
r Reference signals
Se(s) Sensitivity function at e, S e ( s ) = [ I + G ( s )K ( s ) ] –1
Su(s) Sensitivity function at u, S u ( s ) = [ I + K ( s )G ( s ) ] – 1
T Temperature (°C)
T r → y Transfer function from reference signals to output signals
u Control signals
V51 Boilup (mol/s)
W(s) Diagonal matrix of weighting transfer functions
w(s) Weighting transfer function
xD Top product composition (mol/mol)
xB Bottom product composition (mol/mol)
xF Feed composition (mol/mol)
y Output signals
15
Abstract
This thesis discusses a new approach for the dual composition control
design, which takes the entire operating range of a distillation column
into account. With the example of an industrial binary distillation
column, a structured uncertainty model is developed which describes
quite well the nonlinear column dynamics with several simultaneous
model uncertainties. This uncertainty model forms the basis for feed-
back controller designs by μ-synthesis or μ-optimization. The resulting
controllers are distinguished by a high controller performance and high
robustness guaranteed for the entire operating range. This method
enables the synthesis of state-space controllers as well as the μ-optimal
tuning of advanced PID control structures.
Kurzfassung
Diese Arbeit beschreibt einen neuen Ansatz für den Entwurf von
Konzentrationsregelungen, der den gesamten Arbeitsbereich einer
Rektifikationskolonne berücksichtigt. Am Beispiel einer industriellen
binären Rektifikationskolonne wird ein strukturiertes Unsicherheits-
modell entwickelt, welches das nichtlineare dynamische Verhalten der
Rektifikationskolonne durch mehrere Modell-Unsicherheiten gut
beschreibt. Dieses Unsicherheitsmodell bildet die Basis für den
Entwurf von Reglern mittels μ-Synthese oder μ-Optimierung. Die resul-
tierenden Regler zeichnen sich durch eine – über den gesamten
Betriebsbereich garantierte – hohe Regelqualität bei sehr grosser
Robustheit aus. Dieses Vorgehen erlaubt sowohl den Entwurf von
Zustandsregelungen als auch die Berechnung μ-optimaler Einstel-
lungen für erweiterte PID-Regelstrukturen.
Chapter 1
Introduction
Ten years later the situation has not changed. The modern control theo-
ries provide the process control engineer with increasingly sophisticated
tools for a robust, model-based controller design. The advantages of
these controllers over the PID control structures which are usually
tuned on-line, have been shown in numerous publications. Neverthe-
less, more than 90% of all control loops in the process industry use PID
control, while only a few applications of the modern control theories can
be reported [1.10]. Therefore the mismatch between theory and practice
is still evident. Some of the reasons for this situation are discussed
below.
20 1 Introduction
For a control engineer in the process industry, process control in the first
place is a hardware problem. His perspective is the installation and
configuration of a Distributed Control System (DCS) [1.1]. Even the
modern DCS are often limited to PID and advanced PID control. For the
DCS, an implementation of modern state space controllers requires
either the coupling with an external computer or the programming of
software modules. Both ways are troublesome and expensive. The
university research pays little attention to this situation. The design of
robust controllers with fixed structures (e.g., PID control structures) is
a largely unexplored field.
Dynamic Models
Economic benefits
Distillation is one of the most widely used unit operations in the process
industry. In the simplest case, a distillation column separates a feed of
two components into a top product stream (with a high fraction of the
low-boiling component) and a bottom product stream (with a high frac-
tion of the high-boiling component). In an industrial setting, the feed
flow rate and the feed composition may vary within a wide range of oper-
ating conditions.
Without any doubt the distillation process is the most studied unit oper-
ation in terms of control. Skogestad estimates that new papers in this
field appear at a rate of at least 50 each year [1.7]. It is practically
impossible to give a review of all these publications. The interested
reader is advised to consult the reviews of Tolliver and Waggoner [1.8],
Waller [1.9], MacAvoy and Wang [1.3], and the recent review of
Skogestad [1.7].
mizes the controller performance for the entire operating range has not
been addressed as yet.
This thesis presents a new approach for the composition control design
of a binary distillation column (Figure 1.1). The design concept is based
on a structured uncertainty model which describes the column dynamics
for the entire operating range quite well. The resulting controller
designs using μ-synthesis (for state-space controller) or μ-optimization
(for controllers with fixed structure), respectively, lead to results which
guarantee robust performance and robust stability for the entire oper-
ating range of the distillation column. Special emphasis is placed on the
optimal tuning of easy-to-realize PID-control structures. It will be
shown that extraordinary controller performance can be achieved even
with these relatively simple controller structures.
1.3 Structure of the dissertation 23
Uncertainty model
Linear model for a
describing column dynamics
single operating point
for entire operating range
Nonlinear Model
Linear Models
Uncertainty structure
Parameter
Controller synthesis
adaptations
Nonlinear simulations
Tests on plant
Implementation in DCS
The following chapter consists of three parts: The first part describes
the design and operating data of the distillation column, followed by an
overview of the steady-state and dynamic column behavior. The second
part discusses the control objectives and control configuration for this
column, while the third part describes the use of pressure-compensated
temperatures as controlled outputs.
Rigorous nonlinear dynamic models are the basis for simulation studies
and for linearization. They are discussed in Chapter 3.
1.3 Structure of the dissertation 25
1.4 References
[1.1] Brisk, M.L.: “Process Control: Theories and Profits,” Preprints of
the 12th World Congress of the International Federation of Auto-
matic Control, Sydney, July 18-23, 7, 241-250 (1993)
[1.4] Pearson, R. K.: “Modern Control: Why Don’t We Use It?,” InTech,
34, 47-49 (1984)
Chapter 2
2.1 Introduction
A distillation column is not just any mass-produced article such as a
toaster or a washing-machine. Each distillation column is a unique
process unit, specially designed for the separation of a particular
substance mixture. Nevertheless, the thermodynamic principles and
basic dynamics are always the same. Therefore it is possible to demon-
strate ideas for the controller design by the example of one column
without extensive loss of generality.
First in this chapter, the design and operating data of the industrial
distillation column are outlined, followed by a brief description of the
composition dynamics. The further two sections outline the control
objectives, the control structures, and the use of tray temperatures as
controlled outputs. The literature references terminate the chapter.
lation column is given in Table 2.1. The distillation column (Fig. 2.1) is
equipped with 50 sieve trays, a total condenser, and a steam-heated
reboiler. The subcooled feed F enters the column on tray 20 (counted
from the top) and for the greater part consists of a mixture of two
substances. Because of the small fraction of impurities, these are
neglected and the distillation column is considered to be a binary distil-
lation column. The desired product compositions are 0.99 mol/mol (low
boiling component) for the top product D and 0.015 mol/mol for the
bottom product B. As these product purities are relatively high, this
distillation column can be classified as a “high purity distillation
column.”
Column data
No. of trays 50
Column diameter (m) 0.8
Feed tray 20
Murphree tray efficiency ≈0.4
Relative volatility α 1.61
Operating data
Top composition xD (mol/mol) 0.99
Bottom composition xB (mol/mol) 0.015
Feed composition xF (mol/mol) 0.7-0.9
Feed flow rate F (mol/min) 20-46
Top pressure (mbar) 60
Vacuum
VT Condenser
Feed
20
F, xF
47
48
49
50
Boilup
V51
Reboiler
B, xB
Bottom product
Feed disturbances
The boiling points of the entering substances are high at standard atmo-
spheric pressure. Because of a thermal decomposition of the light
component at higher temperatures, the column is operated under
vacuum. Correspondingly, the cooling water flow rate for the condenser
is kept constant and the top pressure is controlled by a vacuum pump.
10
20
xF = 0.7 mol/mol
No.
TrayNo.
xF = 0.8 mol/mol
Tray
xF = 0.9 mol/mol
30 F = 20 mol/min
F = 33 mol/min
F = 46 mol/min
40
50
0.0 0.2 0.4 0.6 0.8 1.0
Composition
Composition (mol/mol)
Figure 2.2: Simulated composition profiles for the industrial distillation column
34 2 The Distillation Process — An Industrial Example
Liquid flow
10
Vapor flow
20
Tray No.
40
50
60 80 100 120
Flow rate (mol/min)
2.4 Composition dynamics 35
discontinuities of the vapor flow profile at trays 1/2 and 20/21 result
therefrom. The reason for the slopes of the two profiles within the strip-
ping and rectifying section of the column is the different heat of evapo-
ration of the two substances.
0 0
10 10
t
20 20
t = 20 h
Tray No.
Tray No.
t=0h
Δt= 30 min
30 30
Δt= 30 min
t=0h
40 40
t
t = 20 h
50 50
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Composition
Composition (mol/mol) Composition
Composition (mol/mol)
a) b)
Figure 2.4: Simulated composition profiles (light component) for a step change of
the reflux. Reboiler heat duty, feed flow rate and composition are kept at their
nominal values (see Table 2.1)
a) L0=1.05*L0,nom b) L0=0.95*L0,nom
2.5 Control objectives and configurations 37
• Ill-conditioned behavior.
σ max { G ( jω ) }
κ = ----------------------------------
- (2.1)
σ min { G ( jω ) }
• Satisfaction of constraints
The first objective includes the control of the vapor holdup (top pres-
sure), the reflux accumulator level, and the reboiler level. Generally,
these control objectives are easily achieved by simple PI controllers.
38 2 The Distillation Process — An Industrial Example
0 0
10 10
Δt = 4h
t
20 Δt = 10h 20
t
Tray No.
Tray No.
t = 40 h
30 30 t=0h
t
t = 40 h
t=0h
40 40
Composition
Composition
50 50
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Composition
Composition (mol/mol) Composition
Composition (mol/mol)
a) b)
Figure 2.5: Simulated composition profiles (light component) for a step change of
the reflux and the reboiler heat duty. The feed is kept at nominal condition (see
Table 2.1).
a) L0=1.265*L0,nom b) L0=1.260*L0,nom
V51=1.19*V51,nom V51=1.19*V51,nom
2.5 Control objectives and configurations 39
• Top composition
• Bottom composition
• Reflux accumulator level
• Reboiler level
• Top pressure
• Reflux
• Boilup (indirectly controlled by reboiler duty)
• Top product flow rate
• Bottom product flow rate
• Cooling water flow rate (or vapor flow rate to vacuum)
In principle, the design of a MIMO controller for the 5×5 or in this case
the 3×3 control problem does not cause any particular difficulties.
However, the failure of just one actuator or sensor disables all control
loops. Due to the high sensitivity of MIMO controllers to sensor or actu-
ator failure, the inventory control and the composition control usually
are independently designed, thus improving the robustness of the
control system and simplifying the controller design. The corresponding
design approach consists of three steps [2.5]:
2.5 Control objectives and configurations 41
In a first step the two manipulated variables for the composition control
are to be chosen. This choice names the control configuration. For
example, if the top composition xD is controlled by reflux L and the
bottom composition xB is controlled by boilup V, the control configura-
tion is called L,V control configuration. After the choice of the manipu-
lated variables for composition control, the remaining three
manipulated variables are available for level and pressure control.
T = f ( p, x ) (2.1)
∂T
T Compensated = T + ------ ( p – pN ) (2.2)
∂p N
2
∂T 1 ∂
T Compensated = T + ------ ( p – p N ) + --- 2 T ( p – pN ) 2 (2.3)
∂p N
2∂p
N
2.6 Tray temperatures as controlled outputs 43
x = θ 1 + θ 2 ( T + T Corr ) + θ 3 p + θ 4 p 2 (2.4)
x = θ1 + θ2 ( T – TN ) + θ3 ( p – pN ) + θ4 ( p – pN )2 (2.5)
The sensitivity of the tray temperatures near the ends of the column to
changes of the product compositions is very small. To make the temper-
2.7 References 45
2.7 References
[2.1] Ariburnu, D., C. Ozge, and T. Gurkan: “Selection of the Best
Control Configuration for an Industrial Distillation Column,”
Preprints of 3rd IFAC Symposium on Dynamics and Control of
Chemical Reactors, Distillation Columns and Batch Processes,
April 26-29, 1992, College Park, MD, 387-392 (1992)
Chapter 3
3.1 Introduction
The rigorous dynamic process simulation has become an accepted and
widespread tool in process and even more so in controller design [3.11].
Increasing competition and environmental protection provisions
require an optimization of process and control structures, which can be
obtained only by a substantial knowledge of process dynamics. At the
same time, dynamic experiments on a running plant are less and less
desired. Rigorous dynamic modelling and simulation can replace such
expensive and time-consuming measurements. This has special signifi-
cance for high-purity distillation columns. Due to their long time
constants and varying feed flow rates and feed compositions, reproduc-
ible operating conditions are difficult to guarantee. Therefore, new
controllers are usually tested thoroughly by dynamic simulation for the
full operating range of the distillation column. The rigorous models of
distillation columns used for that purpose match the reality to a large
extent [3.17].
3.2 Conventions
Figure 3.1 shows a schematic representation of a distillation column
equipped with nt trays. The column top (condenser and reflux accumu-
lator) is denoted by the index 0, the trays with the indices 1, 2, … nt, and
the column bottom (including the reboiler) with the index nt+1. To
simplify the formal mathematical description the reflux stream R is
designated as liquid flow (L0).
Condenser
0
Q0
R D
1 (=L0)
2
3
Sl,4
4
Vj
Fj j
Lj
Fnt-2
nt-2 Sv,nt-1
nt-1
nt Vnt+1
………… Qnt+1
…… Reboiler
nt+1
the trays, for the dead time caused by the transport time of vapor flow
from one tray to the next one above, or for the heat exchange with the
environment, the resulting model would be of very high order. As
mentioned earlier, the aim of modelling the distillation column
dynamics is a sufficient description of the real macroscopic behavior.
This means that we are interested primarily in the dynamics of tray
compositions, temperatures, and pressures etc. rather than in the fluid
streams on the trays. Experience shows that no substantial improve-
ment can be achieved with models including effects with more micro-
scopic characteristics. Hence the following assumptions are usually
introduced in order to achieve a compromise between model accuracy
and order ([3.3], [3.13], [3.17]):
• The holdup of the vapor phase is negligible compared to the
holdup of the liquid phase.
• Liquid phase and vapor phase are each well mixed on all trays,
i.e., the composition of the liquid and of the vapor phase are inde-
pendent of the position on the tray.
• The residence time of the liquid in the downcomer is neglected.
• The variation of the liquid enthalpy on a tray can be neglected on
all trays. (This assumption is not applicable to the evaporator.)
In the literature so far, uniform liquid flows and constant holdups for all
trays have often been assumed (equimolar overflow). This assumption
is problematic because it implies a neglect of flow dynamics. Essential
dynamic effects may remain unmodelled, e.g., a non-minimum phase
behavior (inverse response) of the reboiler level and the tray composi-
tions in the lower section of the column to an increase in reboiler heat
supply.
3.5 Balance equations 51
Lj-2
xk,j-2
. .
Vj-1
. . . ..
.
. .. .
yk,j-1
. .. . Sl,j-1
. . .
.
. . . j-1 Lj-1 xk,j-1
. ..
xk,j-1
Vj
.
yk,j . .. . ..
Fj . . .. . Sl,j
. . .
xF,k,j j
.. . .
.. . .
xk,j
Lj .
xk,j Sv,j+1
yk,j+1
. Vj+1
. .. . . .
.
. .. . yk,j+1
. . .
. . .
.
. . .
. .
j+1
Lj+1
xk,j+1
Vj+2
yk,j+2
dn k, j d ( n j x k, j )
- = F j x F, k, j + L j – 1 x k, j – 1 – ( L j + S l, j )x k, j
-------------- = -----------------------
dt dt (3.1)
+ ( V j + 1 – S v, j + 1 )y k, j + 1 – V j y k, j
52 3 A Rigorous Dynamic Model of Distillation Columns
In the same way, the balance equations for the column top and the
column bottom are formulated:
dn k, 0 d ( n 0 x k, 0 )
- = ( V 1 – S v, 1 )y k, 1 – ( L 0 + D )x k, 0
- = -------------------------
-------------- (3.2)
dt dt
Usually the liquid phases in the column bottom and the reboiler are
mixed either by natural convection or by a pump. Assuming perfect
mixing we obtain
dn k, nt + 1 d ( n nt + 1 x k, nt + 1 )
------------------------- = ---------------------------------------------
-
dt dt (3.3)
= L nt x k, nt – Bx k, nt + 1 – V nt + 1 y k, nt + 1
The total holdup on tray j equals the sum of the holdups of the indi-
vidual substances:
nc
nj = ∑ n k, j (3.4)
k=1
d
------ ( n j h' j ) = F j h' F, j + L j – 1 h' j – 1 + ( V j + 1 – S v, j + 1 )h'' j + 1
dt (3.5)
– ( S l, j + L j )h' j – V j h'' j
d dn dh'
------ ( n j h' j ) = h' j ---------j + n j ----------j (3.6)
dt dt dt
dn
---------j = F j + L j – 1 + V j + 1 – S v, j + 1 – S l, j – L j – V j (3.7)
dt
dh' j
n j ---------- = [ F j ( h' F, j – h' j ) + L j – 1 ( h' j – 1 – h' j )
dt (3.8)
+ ( V j + 1 – S v, j + 1 ) ( h'' j + 1 – h' j ) – V j ( h'' j – h' j )]
1
V j = -------------------- [ F j ( h' F, j – h' j ) + L j – 1 ( h' j – 1 – h' j )
h'' j – h' j (3.9)
+ ( V j + 1 – S v, j + 1 ) ( h'' j + 1 – h' j )]
Hajdu et al. [3.9] present a model for this vapor flow lag. We can imagine
that an energy stream Q supplied to the evaporator is subdivided into
two fractions: One part causes an evaporation of liquid, the other
increases the bottom temperature. Written as a differential equation we
obtain the energy balance equation
54 3 A Rigorous Dynamic Model of Distillation Columns
ΔQ = ΔH v, nt + 1 ΔV nt + 1
dΔT nt + 1 (3.10)
+ n nt + 1 ν' nt + 1 ρ' nt + 1 c' p, nt + 1 ------------------------
dt
∂Δp j
Δ ( Δp j ) = ⎛ ----------------⎞ ΔV j + 1 (3.11)
⎝ ∂V j + 1⎠
nj
∂Δp j
Δ ( p nt + 1 ) = nt ⎛ ----------------⎞ ΔV nt + 1 (3.12)
⎝ ∂V j + 1⎠
nj
∂T nt + 1
ΔT nt + 1 = ⎛ -------------------⎞ Δ ( p nt + 1 ) (3.13)
⎝ ∂p nt + 1 ⎠
x k, nt + 1
ΔQ – ΔH v, nt + 1 ΔV nt + 1 =
∂T nt + 1 ∂Δp j dΔV nt + 1 (3.14)
n nt + 1 ν' nt + 1 ρ' nt + 1 c' p, nt + 1 ⎛ -------------------⎞ nt ⎛ ----------------⎞ ------------------------
⎝ ∂p nt + 1 ⎠ ⎝ ∂V j + 1⎠ dt
Therefore, the vapor flow lag at an increase in reboiler heat supply can
be described by the first-order lag
3.6 Fluid dynamics 55
dQ lag 1
--------------- = ----------- ( Q – Q lag ) (3.15)
dt T lag
∂T nt + 1 ∂Δp j
n nt + 1 ν' nt + 1 ρ' nt + 1 c' p, nt + 1 ⎛ -------------------⎞ nt ⎛ ----------------⎞
⎝ ∂p nt + 1 ⎠ ⎝ ∂V j + 1⎠
T lag = ------------------------------------------------------------------------------------------------------------------------------
- (3.16)
ΔH v, nt + 1
dn nt + 1
h' nt + 1 -------------------- = L nt h' nt + Q lag – Bh' nt + 1 – V nt + 1 h'' nt + 1 (3.17)
dt
2 3/2
L V, j = μ 2g --- b W h LW ,j
(3.19)
3
pj – pj – 1
Hydrostatic liquid level in downcomer = ----------------------- (3.20)
ρ' j g
The liquid head h L, j of the pure liquid on a tray (without a vapor phase
fraction) is equal to the total liquid volume on the tray n j ν' j minus the
pj-1
pj
h LW, j
---------------
εj pj – pj – 1
-----------------------
h L, j ρ' j g
----------
hW εj hL,j
L V, j
AA AB
pj – pj – 1
n j ν' j – ----------------------- A B
ρ' j g
h L, j = ------------------------------------------------
- (3.21)
AA + AB
For the application of the Francis weir formula, we have to evaluate the
liquid level of the pure liquid (liquid without vapor phase fraction). For
that purpose, first the height of the two-phase layer is to be evaluated
and second the liquid phase fraction εj must be taken into account. The
effective liquid level becomes
pj – pj – 1
n j ν' j – ----------------------- A B
h L, j ρ' j g
h L, W, j = ⎛ ---------- – h W⎞ ε j = ------------------------------------------------
- – εj hW (3.22)
⎝ ε ⎠ A + A
j A B
Substituting (3.22) into the Francis weir formula (3.19), we obtain the
volumetric liquid flow rate of the two-phase mixture. The flow rate from
tray j in molal units is calculated by:
pj – pj – 1 3/2
⎛ n ν' – ---------------------- - AB ⎞
2 ⎜ j j ρ' j g ⎟
μ 2g --- b W ⎜ ------------------------------------------------- – ε j h W⎟
3 ⎜ AA + AB ⎟
⎝ ⎠
L j = ------------------------------------------------------------------------------------------------------------ (3.23)
ν' j
• Dry pressure drop occurring at the flow through the tray without
liquid ( Δp tr, j )
ρ'' j ⎛ V V, j + 1⎞ 2
Δp tr, j = ξ ( Re ) ------- ------------------- (3.24)
2 ⎝ A0 ⎠
A 2 s –1.27 ρ'' j ⎛ V V, j + 1⎞ 2
Δp tr, j = ⎛ 1 – -------0 ⎞ + 0.21 ⎛ ------⎞ ------- ------------------- (3.25)
⎝ A A⎠ ⎝ d 0⎠ 2 ⎝ A0 ⎠
The hydrostatic pressure drop results from the liquid head and the
liquid density. We have to take the liquid volume in the downcomer into
account (see 3.6.1).
3.7 Phase equilibrium 59
pj – pj – 1
n j ν' j – ----------------------- A B
ρ' j g
Δp L, j = ------------------------------------------------- ρ' j g (3.26)
AA + AB
The total pressure drop consists of the sum of the two parts dry pressure
drop and hydrostatic pressure drop:
Δp j = p j + 1 – p j = Δp tr, j + Δp L, j (3.27)
γ k, j p k0, j
y kEquilibrium
,j = -------------------- x k, j = K k, j x k, j (3.28)
pj
0 B
log [ p k ] = A – ------------- (3.29)
-
T+C
60 3 A Rigorous Dynamic Model of Distillation Columns
In a distillation column only little contact time exists on each tray for
the mass transfer between liquid and vapor phase. Therefore no perfect
phase equilibrium can be achieved, and the tray efficiency will deviate
from the unit value. This effect can be modelled by the Murphree tray
efficiency for the vapor phase.
y k, j – y k, j + 1
η = ------------------------------------------------- (3.30)
y kEquilibrium
,j – y k, j + 1
nc nc
∑ ykEquilibrium
,j = ∑ Kk, j ( Tj, pj, xk, j )xk, j = 1 (3.31)
k=1 k=1
The Murphree tray efficiency is not considered for the boiling point
calculation, because it relates to the mass transfer between vapor and
liquid phase rather than to the equilibrium composition.
3.8.2 Density
The densities of liquid and vapor phase can be computed from the molar
volume, the molar mass, and the mole fractions.
nc
∑ x k, j M k
Liquid phase density: ρ' j = -----------------------------
k=1 - (3.33)
ν' j
nc
∑ y k, j M k
Vapor phase density: ρ'' j = -----------------------------
k=1 - (3.34)
ν'' j
62 3 A Rigorous Dynamic Model of Distillation Columns
3.9 Enthalpies
The quantity not discussed so far is the enthalpy of a substance mixture
in liquid or vapor phase. The enthalpy of a real fluid is estimated by the
real
sum of an ideal part and the value of a departure function Δh T, P
describing the deviation of the enthalpy from the enthalpy of the ideal
gas state:
∫
0 id real
h = h + c p dT + Δh T, P (3.35)
T0
The ideal part can be calculated by summing the ideal parts for each
component:
T
⎛ T
nc
⎞
∫ c pid dT = ∑ x k ∫ c pid, k dT⎟
⎜ (3.36)
⎜ ⎟
T0 k=1 ⎝ T ⎠
0
id
The ideal heat capacities c p are often approximated by a third-order
polynomial for each component:
c pid, k = A k + B k T + C k T 2 + D k T 3 (3.37)
If measurement data for the heat capacities and for the heat of vapor-
ization are available, a simple solution is possible in a manner similar
to that mentioned in section 3.8.1:
3.10 Numerical solution 63
T
⎛
nc j
⎞
Liquid phase enthalpy: h' j = ∑ x k, j ∫ c p, j, k dT⎟ + h 0
⎜ (3.38)
⎜ ⎟
k=1 ⎝ T ⎠
0
T
nc
⎛ j ⎞
Vapor phase enthalpy: h'' j = ∑ y k, j ∫ c p, j, k dT + ΔH v, j, k⎟ + h 0 (3.39)
⎜
⎜ ⎟
k=1 ⎝ T ⎠
0
×
Δpj-2 h’j-2-h’j-1
pj-1
×
y*k,j-1 yk,j-1
h’’j-1-h’j-1
Murphree / Vj-1
BP Tj-1 States
dn
k, j – 1
----------------
dt - Eq. of state
xk,j-1
nk,j-1 nk,j-1
∫ ∑nk,j-1 nj-1 Lj-1
Flow dynamics
h’’j-h’j-1
×
×
Δpj-1 h’j-1-h’j
pj
×
×
y*k,j yk,j
Murphree h’’j-h’j / Vj
BP Tj States
dn
k, j Eq. of state
----------
dt xk,j
nk,j nk,j
∫ nj Lj
∑nk,j
nc Flow dynamics
h’’j+1-h’j
×
h’j-h’j+1
×
Δpj
pj+1
×
×
y*k,j+1 yk,j+1
h’’j+1-h’j+1
Murphree / Vj+1
BP Tj+1 States
dn
k, j + 1
-----------------
dt - Eq. of state
xk,j+1
nk,j+1 nk,j+1
∫ ∑nk,j+1 nj+1 Lj+1
Flow dynamics
pj+2
×
Vj+2
3 A Rigorous Dynamic Model of Distillation Columns
3.10 Numerical solution 65
The Jacobian matrix of the equation system (as described below) corre-
sponding to these dependent variables has a numerically advantageous
block diagonal dominant structure.
Differential equations
nc material balance equations (3.1)
Algebraic equations
1 equation for vapor flow rate (3.9)
1 equation for liquid flow rate (3.23)
1 equation for boiling point (3.31)
1 equation for pressure drop (3.27)
and in addition the equations for the evaporator, the condenser, and the
control system. Considering industrial distillation columns which are
often equipped with more than 50 trays, the resulting algebraic differ-
ential equations amount to several hundred equations. The model for
the industrial binary distillation equipped with 50 trays gives an
impression of these numbers: It consists of a system of 107 differential
and 210 algebraic equations.
66 3 A Rigorous Dynamic Model of Distillation Columns
dn
------- = f ( t, n ( t ), z ( t ) ) n ( t0 ) = n0 (3.40a)
dt
0 = g ( t, n ( t ), z ( t ) ) z ( t0 ) = z0 (3.40b)
The vector n consists of all tray holdups (for all components), while the
vector z contains all other dependent variables. A different but equiva-
lent formal representation is the implicit form:
F ( t, y ( t ), y' ( t ) ) = 0 y ( t0 ) = y0 (3.41)
d ( F ( t, y ( t ), y' ( t ) ) ) d m ( F ( t, y ( t ), y' ( t ) ) )
------------------------------------------------ = 0, …, ---------------------------------------------------
m
- = 0
dt dt
During the integration, the right-hand sides of the differential and alge-
braic equations repeatedly have to be evaluated for a given vector y of
the dependent variables and for a given time t. The algebraic equations,
and often the differential equations as well are solved in an implicit
manner. The equation errors, which have to be supplied to the integra-
tion, are the difference between the right-hand sides of the equations
68 3 A Rigorous Dynamic Model of Distillation Columns
(that means the calculated vapor flow rates, liquid flow rates, etc.) and
the corresponding values within the vector y.
Using the flow rates and compositions calculated in the previous steps
rather than the data supplied by the integration routine, the differential
terms (left-hand sides) of the equations describing the vapor flow lag
(Step k) and the holdup of the substances in liquid phase (Step l) are
calculated. In a last step (Step m), all differential terms, the errors
3.10 Numerical solution 69
between supplied and calculated flow rates and pressures, and the
errors of the boiling point equations are combined in one vector and
supplied back to the integration routine.
3.11 Notation
A0 [m2] Hole area in tray
AA [m2] Tray area without downcomer area
AB [m2] Downcomer area
bW [m] Length of weir
c pid [J/mol·K] Ideal gas heat capacity
cp,l [J/kg·K] Liquid heat capacity
d0 [m] Diameter of holes of sieve tray
Fj [mol/s] Feed flow rate to tray j
h [J/mol] Molar enthalpy
h' j [J/mol] Molar enthalpy of liquid phase
h'' j [J/mol] Molar enthalpy of vapor phase
hL [m] Liquid level above upper edge of weir
hW [m] Weir height
ΔHv,k,j [J/mol] Heat of evaporation of component k on tray j
ΔHv,j [J/mol] Heat of evaporation of liquid on tray j
Kk,j [mol/mol] Distribution coefficient for comp. k on tray j
Lj [mol/s] Liquid flow leaving tray j
LV,j [m3/s] Volumetric flow from tray j
Mk [g/mol] Molar mass of component k
nt [–] Number of trays in column
72 3 A Rigorous Dynamic Model of Distillation Columns
3.12 References
[3.1] Brenan, K. E., S. L. Campbell, and L. R. Petzold, Numerical so-
lution of initial-value problems in differential-algebraic equa-
tions, North-Holland, New York (1989)
[3.9] Hajdu, H., A. Borus, and P. Földes: “Vapor Flow Lag in Distilla-
tion Columns,” Chem. Eng. Sc., 33, 1-8 (1978)
Chapter 4
Linear Models
4.1 Introduction
Robust controllers are designed on the basis of linear process models.
Therefore the elaboration of linear dynamic models for the distillation
column is a central part of control system synthesis. These models
should describe the dynamic behavior of the process within a wide
frequency range. They can be obtained in two ways:
• System identification
• Linearization of a nonlinear model
Two linear models are evaluated within this chapter, which are based
on the linearization of different nonlinear column models. The first
linear model is obtained by an analytical linearization of a simplified
nonlinear model neglecting flow dynamics. The other model is obtained
by a numerical linearization of the rigorous model presented in Chapter
3. In further sections the accuracy of these linear models and the role of
the flow dynamics are discussed. Different mathematically order reduc-
tion methods are compared at the end of this chapter. The notation is
listed in section 4.9 on page 101, the literature references are collected
in section 4.10.
dx 0
dx 1
dx 0
…
dx 1 x =
x = or dx 51 (4.1)
… n
dx 51 dn 1
…
dn 50
dx F
d = dF
(4.2)
u dL 0
dV 51
dT P 10
y = dT P 44 (4.3)
dT P 24
Idealizing assumptions
Of course, all these assumptions do not agree with the real conditions.
The first assumption means a neglect of the correlation between tray
pressures, holdups, and boilup rates. The second assumption implies
uniform vapor flows within the stripping section and within the recti-
fying section of the column. The assumption of a constant tray holdup
contains a neglect of flow dynamics. The error in the high frequency
range introduced by that is discussed in section 4.5.
dn
---------j = L j – 1 + V j + 1 + F j – L j – V j = 0 (4.4)
dt
Similar balances are obtained for the reboiler and the condenser. Substi-
tuting these balance equations in the material balances (3.1)-(3.3), we
obtain the following differential equations describing the composition
dynamics:
82 4 Linear Models
Condenser
dx 0 1
--------- = ------ V 1 ( y 1 – x 0 ) (4.5)
dt n0
dx 1
--------j = ----- [ L j – 1 ( x j – 1 – x j ) + V j + 1 ( y j + 1 – x j )
dt nj (4.6)
– V j ( y j – x j ) + F j ( x F, j – x j ) ]
Evaporator
dx 51 1
----------- = -------- [ L 50 ( x 50 – x 51 ) – V 51 ( y 51 – x 51 ) ] (4.7)
dt n 51
Feed and reflux of the distillation column are subcooled. A portion of the
vapor flow is condensed at the trays where these two streams enter the
column. The effect of an additional condensation of the vapor stream
caused by increasing the flow rates of these streams must be considered
to avoid large model errors. The two energy balance equations for the
reflux tray (tray 1) and for the feed tray (tray 20) become part of the
nonlinear model:
The model outputs are the deviations of the pressure compensated tray
temperatures. These temperatures are correlated with the tray compo-
sition by the boiling point equation
4.3 Linearization of a simplified nonlinear model 83
∑ ( yk* ) – 1 = 0 (4.10)
k=1
Consequently the boiling point equation for tray 10, 44, and 24 are part
of the simplified column model.
y j = ( 1 – η )y j + 1 + ηy * j (4.11)
αj xj
y * j = --------------------------------- (4.12)
1 + ( α j – 1 )x j
y 50 = ( 1 – η )y * 51 + ηy * 50 (4.13)
84 4 Linear Models
y 49 = ( 1 – η )y 50 + ηy * 49
(4.14)
= ( 1 – η ) 2 y * 51 + η ( 1 – η )y * 50 + ηy * 49
y 48 = ( 1 – η )y 49 + ηy * 48
(4.15)
= ( 1 – η ) 3 y * 51 + η ( 1 – η ) 2 y * 50 + η ( 1 – η )y * 49 + ηy * 48
y j = ( 1 – η ) nt + 1 – j y * nt + 1
nt + 1 – j
(4.16)
+ ∑ η ( 1 – η ) nt + 1 – j – n y * nt + 1 – n
n=1
dx̃ ( t )
--------------- = f [ x̃ ( t ), ũ ( t ), d˜ ( t ) ] (4.17a)
dt
ỹ ( t ) = g [ x̃ ( t ) ] (4.17b)
dx = Ax + B d (4.18a)
------
dt u
y = Cx (4.18b)
4.3 Linearization of a simplified nonlinear model 85
∂f ∂f ∂f 0
-------0- -------0- … ----------
-
∂x 0 ∂x 1 ∂x 51
∂f ∂f
∂f -------1- -------1- … …
A = ------ = ∂x 0 ∂x 1 (4.19)
∂x̃ OP
… … …
∂f 51 ∂f 51
--------- … … ----------
-
∂x 0 ∂x 51 OP
∂f ∂f ∂f ∂f 0
--------0- ------0- --------0- -----------
-
∂x F ∂F ∂L 0 ∂V 51
∂f ∂f ∂f ∂f 1
∂f --------1- ------1- --------1- -----------
-
B = ----------
- = ∂x F ∂F ∂L 0 ∂V 51 (4.20)
˜
∂ d … … … …
ũ
OP ∂f 51 ∂f 51 ∂f 51 ∂f 51
--------- --------- --------- -----------
-
∂x F ∂F ∂L 0 ∂V 51 OP
∂g 1 ∂g 1 ∂g 1
-------- -------- … ----------
-
∂x 0 ∂x 1 ∂x 51
∂g ∂g ∂g 2
C = ∂-----
g
- = -------2- -------2- … ----------
- (4.21)
∂x̃ OP ∂x 0 ∂x 1 ∂x 51
∂g 3 ∂g 3 ∂g 3
-------- -------- … ----------
-
∂x 0 ∂x 1 ∂x 51 OP
86 4 Linear Models
dL j = dL 0
(4.22)
dV j = dV 51
The desired outputs of the linear model are the deviations of pressure
compensated tray temperatures, which are functions of the tray compo-
sitions, but not of a component’s holdups. Therefore the material bal-
ance equations (3.1)-(3.3) are replaced by the following, equivalent
differential equations:
Condenser
dx 0 1
--------- = ------ V 1 ( y 1 – x 0 ) (4.23)
dt n0
4.4 Linearization of the rigorous model 87
Tray holdups:
dn
---------j = L j – 1 – L j + V j + 1 – V j + F j (4.24)
dt
dx j 1
-------- = ----- [ L j – 1 ( x j – 1 – x j ) + V j + 1 ( y j + 1 – x j )
dt nj (4.25)
– V j ( y j – x j ) + F j ( x F, j – x j ) ]
Evaporator
dx 51 1
----------- = -------- [ L 50 ( x 50 – x 51 ) – V 51 ( y 51 – x 51 ) ] (4.26)
dt n 51
The second modification of the rigorous model follows from the assump-
tion of a perfect level control in the reflux accumulator and in the evap-
orator. This assumption is justified by the fact that the level control
loops for the condenser and the evaporator can be tuned much faster
than the composition control loops. With perfect level control and there-
fore constant holdup for the reflux accumulator and the evaporator, the
top and bottom product streams are calculated according to
D = V1 – L0 (4.27)
and
B = L 50 – V 51 (4.28)
dñ ( t )
--------------- = l [ x̃ ( t ), ñ ( t ), ũ ( t ), z̃ ( t ), ṽ ( x̃, ñ, ũ, z̃ ) ] (4.29b)
dt
y ( t ) = g [ x̃ ( t ) ] (4.29c)
The vector ṽ ( x̃, ñ, ũ, z̃ ) represents the solution of the algebraic equa-
tion system k and consists of the tray pressures, the vapor flow rates,
the liquid flow rates, and the boiling points.
The matrices of the linear state space model including flow dynamics
d
------ x = A x + B d (4.30a)
dt n n u
y = Cx (4.30b)
⎛ dx dx ⎞
⎜ ------- – ------- ⎟
⎝ dt x + Δx dt OP⎠
j j
-----------------------------------------------------
Δx j
a i=1…102, j + 1 = for j = 0…51 (4.31a)
⎛ dn ⎞
⎜ ------- – dn ------- ⎟
⎝ dt x + Δx dt OP⎠
j j
------------------------------------------------------
Δx j
4.5 Comparison of the linear models 89
⎛ dx ⎞
⎜ ------- – dx ------- ⎟
⎝ dt n + Δn dt OP⎠
j j
------------------------------------------------------
Δn j
a i=1…102, j + 52 = for j = 1…50 (4.31b)
⎛ dn dn ⎞
⎜ ------- – ------- ⎟
⎝ dt n + Δn dt OP⎠
j j
-------------------------------------------------------
Δn j
A simple method to compare the two linear models with the complete
rigorous model is the simulation of step responses to the model inputs
(reflux L0, boilup V51, feed composition xF, and feed flow rate F). These
are shown by the Figures 4.1- 4.4. During the nonlinear simulations, the
bottom level was controlled by the bottom product flow rate B. Except
for the denoted input, all other column inputs are kept constant at their
steady-state values. The changes of flow rates and feed composition are
very small to maintain the column close to the steady-state and to avoid
large nonlinearities.
The coincidence of the step responses with the rigorous nonlinear model
is acceptable for both linear models. However, the linear model obtained
by a linearization of the rigorous model is distinguished by a somewhat
better representation of the low-frequency gains.
90 4 Linear Models
Tray 10 Tray 44
0 0
-0.05 -0.2
Temp. dev. (K)
-0.15 -0.6
-0.2 -0.8
0 2 4 6 8 10 0 2 4 6 8 10
Time (h) Time (h)
Tray 24
0 0
-0.1 -0.2
Nonlinear model
Temp. dev. (K)
-0.2 -0.4
Anal. linearized model
-0.3 -0.6
Num. linearized model
-0.4 -0.8
-0.5 -1
0 2 4 6 8 10 0 2 4 6 8 10
Time (h)
Figure 4.1: Step response to a 0.3 mol/min (0.46%) increase in reflux
Tray 10 Tray 44
0.2 0.5
0.4
0.15
Temp. dev. (K)
0.3
0.1
0.2
0.05
0.1
0 0
0 2 4 6 8 10 0 2 4 6 8 10
Time (h) Time (h)
Tray 24
0.5 0
0.4 -0.2
Nonlinear model
Temp. dev. (K)
0.3 -0.4
Anal. linearized model
0.2 -0.6
Num. linearized model
0.1 -0.8
0 -1
0 2 4 6 8 10 0 2 4 6 8 10
Time (h)
Figure 4.2: Step response to a 0.3 mol/min (0.29%) increase in boilup
4.5 Comparison of the linear models 91
Tray 10 Tray 44
0 0.1
-0.02
0
Temp. dev. (K)
-0.1 -0.3
0 2 4 6 8 10 0 2 4 6 8 10
Time (h) Time (h)
Tray 24
0 0
-0.05 -0.2
Nonlinear model
Temp. dev. (K)
-0.1 -0.4
Anal. linearized model
-0.15 -0.6
Num. linearized model
-0.2 -0.8
-0.25 -1
0 2 4 6 8 10 0 2 4 6 8 10
Time (h)
Figure 4.3: Step response to a 0.005 mol/mol increase in feed composition
Tray 10 Tray 44
0
-0.02 -0.1
Temp. dev. (K)
-0.04 -0.2
-0.06 -0.3
-0.08 -0.4
-0.1 -0.5
-0.12 -0.6
0 2 4 6 8 10 0 2 4 6 8 10
Time (h) Time (h)
Tray 24
0 0
-0.2
-0.1 Nonlinear model
Temp. dev. (K)
-0.4
-0.2 Anal. linearized model
-0.6
-0.3 Num. linearized model
-0.8
-0.4 -1
0 2 4 6 8 10 0 2 4 6 8 10
Time (h)
Figure 4.4: Step response to a 0.3 mol/min (0.91%) increase in feed flow rate
92 4 Linear Models
3 Disturbance inputs
10
Magnitude
0
10
-3
10
-6
10 -4 -3 -2 -1 0 1
10 10 10 10 10 10
Frequency (rad/min)
3 Control inputs
10
0
Magnitude
10
-3
10
-6
10 -4 -3 -2 -1 0 1
10 10 10 10 10 10
Frequency (rad/min)
Figure 4.5: Singular values of the linear models
Upper plots: G d → y ( jω ) , lower plots: G u → y ( jω )
Solid lines: Analytically linearized model
Dashed lines: Numerically linearized model
4.5 Comparison of the linear models 93
Both models show the typical course of the singular values for a high
purity distillation column. In the low frequency range, the maximum
and minimum singular values of the transfer functions G u → y ( jω ) are
very different and the condition numbers
σ max { G u → y ( jω ) }
κ ( jω ) = ----------------------------------------------
- (4.32)
σ min { G u → y ( jω ) }
The numerically linearized model takes the flow dynamics into account.
Thus, considering the reflux as column input, additional lags for the
composition dynamics are introduced, and for the minimum singular
value a negative slope of several magnitudes per decade for the
frequency range above 1 rad/min results therefrom. The effect of the
flow dynamics considering the boilup as the column input is different:
An increase of the boilup increases the vapor fraction on the tray, which
94 4 Linear Models
3 Singular values
10
Magnitude
0
10
-3
10
-6
10 -4 -3 -2 -1 0 1
10 10 10 10 10 10
Frequency (rad/min)
3 Condition number
10
Magnitude
2
10
1
10
0
10 -4 -3 -2 -1 0 1
10 10 10 10 10 10
Frequency (rad/min)
Figure 4.6: Upper plot: Singular values of the transfer function from the control
inputs u to the tray temperature T10 and T44
Lower plot: Condition number κ of the same transfer function
Solid lines: Analytically linearized model
Dashes lines: Numerically linearized model
causes higher liquid flow rates leaving the trays. Because the composi-
tion of the light component is higher in the upper part of the column,
more of the light component is transported down, and the expected
decrease of the light component’s composition is retarded in the
frequency range between 0.2 and 1 rad/min.
3
10
0
10
Magnitude
Method III
Method III
Method IV
Method IV
-6
10 -4 -3 -2 -1 0 1
10 10 10 10 10 10
Frequency (rad/min)
plots of the models. Figure 4.7 shows that all reduced order models
approximate the low and medium frequency behaviors up to 1 rad/min
very well. However, in the high frequency range the singular values are
best approximated by the models derived with a Balanced Truncated
Approximation (Method I or II).
4.7 Summary
This chapter presented two methods to obtain linear models for the
industrial distillation column. The first model is derived by an analyt-
ical linearization of a simplified nonlinear model neglecting flow
dynamics and most of the energy balance equations. The second linear
4.8 Appendix: Model coefficients 97
∂ *
kj = y (4.33)
∂ xj j
A-Matrix
Condenser (k=1…50)
∂f –V
a 1, 1 = -------0- = ---------1- (4.34)
∂x 0 n0
∂f 0 η ( 1 – η ) k – 1 V 1 k k
a 1, k + 1 = --------- = --------------------------------------------- (4.35)
∂x k n0
∂f 0 ( 1 – η ) 50 V 1 k 51
a 1, 52 = ----------
- = -------------------------------------
- (4.36)
∂x 51 n0
98 4 Linear Models
∂f j Lj – 1
a j + 1, j = -------------- = -----------
- (4.37)
∂x j – 1 nj
∂f – ( L j + V j + 1 – V j + V j ηk j + F j )
a j + 1, j + 1 = -------j = ------------------------------------------------------------------------------
- (4.38)
∂x j nj
∂f ( V j + 1 – ( 1 – η )V j )
a j + 1, k + 1 = --------j- = η ( 1 – η ) k – j – 1 k k ----------------------------------------------- (4.39)
∂x k nj
∂f j ( V j + 1 – ( 1 – η )V j )
- = ( 1 – η ) 50 – j k k -----------------------------------------------
a j + 1, 52 = ---------- (4.40)
∂x 51 nj
Evaporator
∂f 51 L 50
a 52, 51 = ----------
- = -------- (4.41)
∂x 50 n 51
∂f 51 – ( B + V 51 k 51 )
a 52, 52 = ----------
- = ------------------------------------ (4.42)
∂x 51 n 51
B-Matrix
h' 20 – h' F
dV = – ⎛ --------------------------⎞ dF (4.43)
⎝ h'' 20 – h' 20⎠
The liquid flow rate in the stripping section of the column is increased
by the same amount.
4.8 Appendix: Model coefficients 99
Condenser
∂f
b 1, 1 = --------0- = 0 (4.44)
∂x F
∂f h' 20 – h' F y 1 – x 0
b 1, 2 = ------0- = – ⎛ --------------------------⎞ ----------------- (4.45)
∂F ⎝ h'' 20 – h' 20⎠ n 0
∂f
b j + 1, 1 = --------j- = 0 (4.46)
∂x F
∂f h' 20 – h' F y j + 1 – y j
b j + 1, 2 = ------j = – ⎛ --------------------------⎞ ----------------------- (4.47)
∂F ⎝ h'' 20 – h' 20⎠ nj
∂f 20 F 20
b 21, 1 = --------- = -------- (4.48)
∂x F n 20
h' 20 – h' F
x F – x 20 + -------------------------- ( y 20 – x 20 )
∂f 20 h'' 20 – h' 20
b 21, 2 = --------- = --------------------------------------------------------------------------------- (4.49)
∂F n 20
∂f
b j + 1, 1 = --------j- = 0 (4.50)
∂x F
∂f h' 20 – h' F ⎞ x j – 1 – x j
b j + 1, 2 = ------j = ⎛ 1 + -------------------------- ---------------------- (4.51)
∂F ⎝ h'' 20 – h' 20⎠ nj
h' 1 – h' 0
dV 1 = – ⎛ ----------------------⎞ dL 0 (4.52)
⎝ h'' 1 – h' 1⎠
h' 1 – h' 0 ⎞
dL 1 = ⎛ 1 + ---------------------
- dL (4.53)
⎝ h'' – h' ⎠ 0 1 1
With (4.52) and (4.53) there follows for the elements of the B matrix:
Condenser
h' 1 – h' 0
- ( x – y1 )
---------------------
∂f 0 h'' 1 – h' 1 0
b 1, 3 = ------- = --------------------------------------------- (4.54)
∂L n0
∂f y1 – x0
b 1, 4 = ------0- = ----------------
- (4.55)
∂V n0
Tray 1
h' 1 – h' 0
x 0 – x 1 + ( y 1 – x 1 ) ----------------------
∂f h'' 1 – h' 1
b 2, 3 = ------1- = --------------------------------------------------------------------
- (4.56)
∂L n1
∂f y2 – y1
b 2, 4 = ------1- = ----------------
- (4.57)
∂V n1
Trays (j = 2 … 50)
∂f xj – 1 – xj h' 1 – h' 0 ⎞
b j + 1, 3 = ------j = ---------------------- ⎛ 1 + ---------------------
- (4.58)
∂L nj ⎝ h'' 1 – h' 1⎠
∂f yj + 1 – yj
b j + 1, 4 = -------j = ----------------------
- (4.59)
∂V nj
4.9 Notation 101
Evaporator
∂f 51 x 50 – x 51 ⎛ h' 1 – h' 0 ⎞
b 52, 3 = --------- = ---------------------- 1 + ---------------------
- (4.60)
∂L n 51 ⎝ h'' 1 – h' 1⎠
∂f 51 – ( y 51 – x 51 )
b 52, 4 = --------- = ------------------------------ (4.61)
∂V n 51
C-Matrix
4.9 Notation
4.10 References
[4.1] Balas, G. J., J. C. Doyle, K. Glover, A. Packard, and R. Smith: μ-
Analysis and Synthesis Toolbox (μ-Tools), The MathWorks, Inc.,
Natick, MA (1991)
Chapter 5
5.1 Introduction
Each linear or nonlinear dynamic model can only approximately
describe the behavior of a real distillation column. While a nonlinear
model may be valid for a wide range of operating conditions, the error of
a linear model rapidly increases with the distance from its steady-state
design point due to process nonlinearity. Since stochastic effects influ-
ence the process behavior as well, the error of a linear model compared
to the real process can never be exactly determined. Lacking an exact
error description, the error between the process model and the process
itself is modelled as a single frequency-dependent uncertainty bound
(unstructured uncertainty) or as several frequency-dependent uncer-
tainty bounds (structured uncertainties).
However, the experience shows that for most cases uncertainty descrip-
tions with frequency dependent and hence time-invariant uncertainty
bounds are sufficient.
This holds especially for our distillation column: The main disturbances
are step changes of the feed flow rate. Each step change alters the
steady state operating point and defines a new linear model describing
the dynamic behavior up to the next step change. Each of these linear
5.3 Input uncertainty 107
models is one of the models within the set of all models. This set is
defined by the specified uncertainty bounds.
ũ ( jω ) = { I + Δ u ( jω )W u ( jω ) }u ( jω ) (5.1)
Wu Δu
u ũ
with
Δ u ( jω ) ∞
≤1 (5.2)
w u ( jω ) 0
W u ( jω ) = L
(5.3)
0 w u ( jω )
V
If we assume that the reflux and the boilup errors are independent, the
matrix Δ u ( jω ) becomes a diagonal matrix with two single perturbations
δ yielding the following uncertainty model:
⎧ ⎫
⎪ δ u ( jω ) 0 ⎪
ũ ( jω ) = ⎨ I + L
W u ( jω ) ⎬u ( jω ) (5.4)
⎪ 0 δ u ( jω ) ⎪
⎩ V
⎭
with
δ u ( jω ) ≤1 (5.5)
i ∞
Both models have been used for μ-synthesis with very similar results.
For two reasons, the model (5.1) is preferred in this study:
• Any change in reflux may cause a change of the vapor flow rate
within the column and vice versa. The interactions due to flow
dynamics and to the energy balance are to be considered here.
5.3 Input uncertainty 109
It has been shown by Skogestad et al. [5.4] that the controlled system’s
performance for a high purity distillation column is very sensitive to
errors in the manipulated variables. For controller design or analysis
the error bounds W u should be estimated as exactly as possible. This
holds especially for the low-frequency range, where the condition
number of the column models is high. Otherwise potential controller
performance is given away in case of an overestimation.
1 + s ⁄ ωN
G ( s ) = K ------------------------ with ω N < ω D (5.6)
1 + s ⁄ ωD
The gain K represents the steady-state error. The cut-off frequencies are
typically chosen according to ω D > 10ω N .
110 5 A Structured Uncertainty Model
dx j
n j ⎛ --------⎞ = L j – 1 ( x j – 1 – x j ) + V j + 1 ( y j + 1 – x j ) – V j ( y j – x j ) (5.7)
⎝ dt ⎠
Any control system for a distillation column must exhibit large gains in
the low-frequency range to achieve small control errors at steady-state.
Therefore, at steady-state both product compositions (or the tempera-
tures on trays 10 and 44) can be kept at their setpoints. Thus transients
have no significant influence in the low-frequency range and the
internal vapor and liquid flow rates as well as the composition profile
within a column become a function of feed flow rate and composition
only.
smallest feed composition and largest feed flow rate and, vice versa,
smallest internal flow rates for the largest feed composition and
smallest feed flow rates. The composition profiles for these two steady
states bound the domain of all steady state composition profiles (see
Figure 2.2).
The feed data of the different models are listed in Table 5.1.
δy 0 0
1
Wy 0 δy 0
2
0 0 δy
3
y ỹ
⎧ δ y ( jω ) 0 0 ⎫
⎪ 1 ⎪
⎪ ⎪
ỹ ( jω ) = ⎨ I + 0 δ y ( jω ) 0 W y ( jω ) ⎬y ( jω ) (5.8)
⎪ 2
⎪
⎪ 0 0 δ y ( jω ) ⎪
⎩ 3 ⎭
with
δy ≤1
i ∞
and
y ( jω ) = G N ( jω ) d ( jω )
u ( jω )
ΔG I ( jω ) = [ G I ( jω ) – G N ( jω ) ]G N
– 1 ( jω ) (5.9)
5.4 Model uncertainty 113
ΔG R ( jω ) = [ G R ( jω ) – G N ( jω ) ]G N
– 1 ( jω ) (5.10)
The upper bound for the uncertainty weights w y is the maximum of all
j
standardized errors for the output y j .
G I ( jω ) + G R ( jω ) G I ( jω ) – G R ( jω )
- + δ G ( jω ) --------------------------------------------
G ( jω ) = ------------------------------------------- (5.11)
2 2
with
δG ∞
≤1 δ G ∈ C or δ G ∈ R
114 5 A Structured Uncertainty Model
Standardized error
Standardized errorofofT-10,
T10,OP-I
GI(s) Standardized error
Standardized errorofofT-44,
T44,OP-I
GI(s)
1 1
0.5 0.5
Magnitude
Magnitude
0 0
-0.5 -0.5
-1 -5 -4 -3 -2 -1 0
-1 -5 -4 -3 -2 -1 0
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/min) Frequency (rad/min)
Standardized errorofofT-24,
Standardized error T24,OP-I
GI(s) Legend
1 1
xF
0.5 0.5
Magnitude
F
0 0
L
-0.5 -0.5
V
-1 -5 -4 -3 -2 -1 0
-1 -5 0
10 10 10 10 10 10 10 10
Frequency (rad/min)
Standardized error
Standardized error ofofT-10,
T10,OP-R
GR(s) Standardized errorofofT-44,
Standardized error T44,OP-R
GR(s)
1 1
0.5 0.5
Magnitude
Magnitude
0 0
-0.5 -0.5
-1 -5 -4 -3 -2 -1 0
-1 -5 -4 -3 -2 -1 0
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/min) Frequency (rad/min)
Standardized
Standardized error T24,OP-R
errorofofT-24, GR(s) Legend
Legend
1 1
x
xFF
Magnitude
0.5 0.5
FF
0 0
LL
-0.5 -0.5
VV
-1 -5 -4 -3 -2 -1 0
-1 -5 0
10 10 10 10 10 10 10 10
Frequency (rad/min)
Figure 5.3: Standardized model errors at operating points OP-I and OP-R
5.4 Model uncertainty 115
T ( jω )
G L →L-T10
10
GV → T ( jω )
V-T10
10
0 0
-100 -100
0 200 0 200
T ( jω )
G L →L-T44
44
GV → T ( jω )
V-T44
44
0 0
-400 -300
0 700 0 600
T ( jω )
G L →L-T24
24
GV → T ( jω )
V-T24
24
0 0
-250 -250
0 500 0 500
d
GI 1/2
+ +
+
δG I3
+ y
–
GR 1/2
u
In this way we generate a plant which covers the properties of the distil-
lation column at low and at high feed compositions, and at different feed
flow rates without introducing additional conservatism. It is impossible
to model such a behavior with an unstructured uncertainty description!
Effect of transients
earity in the middle and higher frequency range, which can be described
with a multiplicative uncertainty description as in equation (5.8). The
uncertainty weights w y are chosen to have large singular values in the
i
higher-frequency range and low singular values in the low-frequency
range. It is not possible to calculate these uncertainty bounds exactly.
Each disturbance input will cause a variation of the operating point, but
the magnitude of the deviation from a steady-state operating point
cannot be predicted. The selection of appropriate transfer functions is
discussed in Chapter 6.
It has been shown in Chapter 2 that flow dynamics affect the high-
frequency behavior of distillation columns. If linear models which
neglect the flow dynamics are used for control design, an appropriate
uncertainty model is necessary.
Most authors treat the effect of flow dynamics in the same way as the
effect of an input time delay τ with 0 < τ < 1 minute ([5.1], [5.3], [5.4],
[5.5]). The corresponding input uncertainty is often modelled with a
multiplicative uncertainty, using a first order Padé approximation for
the uncertainty bound ([5.4], [5.5]):
1 – τs -----
τs 2
e ≅ ---------------
– (5.12)
τs
1 + -----
2
KT
G T ( s ) = ------------------- (5.13)
1 + sT T
We(s)
d e
Wd(s) GΔ(s)
r
K(s)
+ - T24
T10,T44
sup W e ( jω )e ( jω ) 2
≤1 ∀ω ∈ R + (5.14)
d ≤1
r 2
T ( jω ) ≤1 (5.15)
d →p
r ∞
5.7 Summary
The complete uncertainty model is shown in Figure 5.7. It consists of the
input uncertainty (5.1), the model uncertainties (5.8) and (5.11), and the
performance specifications. Simple dynamic models of the temperature
sensors are included in the column models. This relatively complex
uncertainty model has the advantage that the entire operating range of
the distillation column is covered. The large conservatism of an unstruc-
tured uncertainty description is avoided. Therefore, with design proce-
dures based on the μ-synthesis or μ-optimization, we can expect high
controller performance for the entire operating range.
5.7 Summary
p
We ηo δy 0 0
1 ηi
Wy 0 δy 0
2
0 0 δy
1/2 3
ρi
GI +
ρo + ξ
o ξi + +
d Wu Δu δGΙ
+ - + +
Wd e +
1/2
r + K GR
– ~u
T24
T10, T44
The input uncertainty bounds w u i are easily shaped. Only a few reflec-
tions are necessary about the steady-state error and the frequency
where a 100% error is to be expected. However, the output uncertainties
w y j are more difficult to shape. During the controller design procedure it
is often necessary to adjust them iteratively until nonlinear simulations
show a satisfactory closed-loop dynamics. This problem is discussed
further in Chapter 6.
5.8 References
[5.1] Lundström, P., S. Skogestad, and Z.-Q. Wang: “Uncertainty
Weight Selection for H-Infinity and Mu-Control Methods,” Proc.
30th Conference on Decision and Control, Brighton, U. K. (1991)
[5.3] Postlethwaite, I., J.-L. Lin and D.-W. Gu: “Robust Control of a
High Purity Distillation Column Using μ-K Iteration,” Proc. 30th
Conference on Decision and Control, Brighton, U. K (1991)
[5.4] Skogestad, S., M. Morari, and J. C. Doyle: “Robust Control of Ill-
Conditioned Plants: High-Purity Distillation,” IEEE Trans. Auto-
matic Control, 33, 12, 1092-1105 (1988)
Chapter 6
6.1 Introduction
While the synthesis and analysis of controllers using the structured
singular value μ (SSV) has attracted considerable attention among aero-
space and electrical engineers (e.g., [6.8], [6.9]), it has been less
commonly considered by process control engineers. One reason for that
might be the lack of adequate structured uncertainty models for chem-
ical processes. The uncertainty model discussed in the previous chapter
forms a suitable basis for a μ-optimal controller design. Since this uncer-
tainty model covers the dynamic behavior of the industrial distillation
column for the entire operating range, the resulting controllers guar-
antee stability and performance for all operating points.
This chapter presents the results of a μ-optimal controller design for the
LV control structure of the distillation column. After a summary of the
most useful aspects of the SSV, the design of state-space controllers by
μ-synthesis is demonstrated. Because the implementation of state-space
controllers in a distributed control system is a troublesome project, the
design of controllers with fixed and easy-to-implement structures (PID
control structures) is considered in a special section. A comparison of the
controller’s performances in the time-domain terminates this chapter.
124 6 μ-Optimal Controller Design
The first set of inputs and outputs is highly important. Within the
uncertainty model, this set represent the output and input signals of the
uncertainty blocks. In our case the inputs to the uncertainty blocks are
the signals ρ o, ξ o and η o . The corresponding outputs are the signals
ρ i, ξ i , and η i .
ρi Δ ρo
ξi = z Uncertainties v = ξo
ηi ηo
d P p
r Plant
u y
K
Controller
Δ = diag ( Δ u, δ G I 3, δ y , δ y , δ y Δ u ∈ C 2 × 2, δ G ∈ C, δ y ∈ C )
1 2 3 i
or, alternatively,
Δ = diag ( Δ u, δ G I 3, δ y , δ y , δ y Δ u ∈ C 2 × 2, δ G ∈ R, δ y ∈ C ) (6.1)
1 2 3 i
with Δj ∞
≤ 1, δ j ∞
≤1
⎧ m × mj ⎫
X = ⎨ diag ( δ 1 I r , …, δ s I r , Δ 1, …, Δ f ) δ i ∈ C, Δ j ∈ C j ⎬ (6.2)
⎩ 1 s
⎭
⎧ 1
⎪ ---------------------------------------------------------------------------------------------
⎪ min { σ max ( Δ ) ( det ( I + MΔ )=0 ) }
⎪
μΔ ( M ) = ⎨ Δ ∈ X (6.3)
⎪
⎪
⎪ 0 if no ( Δ ∈ X ) solves det ( I + MΔ )=0
⎩
M
Figure 6.2: M-Δ feedback connection
6.2 The structured singular value 127
D = { diag D 1 , …, D S, d S + 1 I m , …, d S + F I m
1 F
(6.4)
ri × ri
Di ∈ C , D i =D i* > 0, d S + j ∈ R , d S + j > 0 }
⎧ k ×k ⎫
U = ⎨ diag ( U 1, U 2, …, U n ) U i ∈ C i i, U i* U i =I k ⎬ (6.5)
⎩ i
⎭
ρ ( M ) ≤ μ ( M ) ≤ σ max ( M ) (6.6)
μ ( DMD –1 ) = μ ( M ) (6.7)
The optimization problem for the lower bound is not convex and its
calculation may converge to local maxima. Nevertheless numerical
experience indicates that usually the difference between upper and
lower bounds is within 5%, and almost always within 15% ([6.12],
[6.14]).
M ( P, K ) = F l ( P, K ) = P 11 + P 12 K ( I – P 22 K ) – 1 P 12 (6.9)
z
v M 11 M 12
= d (6.10)
p M 21 M 22
r
Δ
z v
d
r
M(P,K) p
Figure 6.3: Representation of uncertain control system with controller K and plant
P combined into the system M
6.2 The structured singular value 129
The input sets are (1) the outputs from the uncertainty block Δ , and (2)
the disturbance and reference inputs. The outputs, in turn, are the
inputs to the uncertainty block Δ and the set of performance measures
p.
BX = { diag ( δ 1 I r , …, δ s I r , Δ 1, …, Δ f )
1 s
mj × mj ⎫ (6.11)
δ i ∈ C, Δ j ∈ C , δ i ∞ ≤ 1, Δ j ∞ ≤ 1 ⎬
⎭
The system shown in Figure 6.3 remains stable for all Δ ∈ BX if and
only if
Theorem 6.1 allows the stability analysis of control systems with struc-
tured uncertainties. If we plot the upper and lower bounds of μ ( M 11 )
for enough frequency points in the frequency range of interest and find
that the maximum value of μ is smaller than one, the control system is
stable for the uncertainties specified with the assumption δ i ∞ ≤ 1 ,
Δ j ∞ ≤ 1 . If μ ( M ) exceeds one for any frequency, the control system is
not guaranteed to be stable. However, for smaller uncertainties with
δ i ∞ ≤ 1 ⁄ ( sup μ Δ ) and Δ j ∞ ≤ 1 ⁄ ( sup μ Δ ) stability is guaranteed.
ω ω
For the application of theorem 6.2 we have to add one uncertainty block
Δ P to the uncertainty structure Δ (Fig. 6.4). Imagine that the perform-
ance specification of the control system is met for all allowed distur-
bance matrices Δ in the set BX. In this case the output p is bounded by
p ∞ ≤ 1 for all inputs [ d T, r T ] T with [ d T, r T ] T ∞ ≤ 1 . If we close the
loop from p to [ d T, r T ] T by introduction of the block Δ P with Δ P ∞
≤ 1,
the system will be stable. But if any block Δ P with Δ P ∞
≤ 1 destabilizes
the loop p → [ d T, r T ] T , the specified performance cannot be achieved
for all possible plants within the specified set. Therefore, in the frame-
work of the SSV, the robust performance problem is handled like a sta-
bility problem. A test for robust performance will be similar to a test for
robust stability. Because the test for robust performance includes robust
stability, it usually will be sufficient.
Δ 0
z v
0 ΔP
d
r
M(P,K) p
• input uncertainties,
• output uncertainties,
• reference inputs, disturbance inputs, and
• controller performance
are to be selected.
All weighting functions are chosen as diagonal matrices:
W d ( s ) = diag [ w x ( s ), w F ( s ), w r ( s ), w r ( s ) ] (6.14)
F 10 44
W u ( s ) = diag [ w u ( s ), w u ( s ) ] (6.15)
L V
W y ( s ) = diag [ w y ( s ), w y ( s ), w y ( s ) ] (6.16)
10 44 24
W e ( s ) = diag [ w e ( s ), w e ( s ) ] (6.17)
10 44
1
w x ( s ) = K x ---------------------- with K x = 0.1 mol/mol, T x = 180 min (6.18)
F1 + T
F
F x s F F
1
w F ( s ) = K F ------------------- with K F = 6 mol/min, T F = 120 min (6.19)
1 + TF s
w r ( s ) = w r ( s ) = 0.2 (6.20)
10 44
1 + 20s
w u ( s ) = w u ( s ) = 0.1 ----------------------- (6.21)
L V 1 + 0.02s
1 + 167s
w y ( s ) = w y ( s ) = w y ( s ) = 0.1 ----------------------- (6.22)
10 44 24 1 + 1.67s
1
w e ( s ) = w e ( s ) = 100 --------------------------- (6.23)
10 44 1 + 27800s
6.4 Controller design with μ-synthesis 133
wF
Magnitude
Magnitude
1 0
10 10 wr
wy
0 wu -2 wx
10 10 F
-1 -4
10 -4 -2 0 2
10 -6 -4 -2 0
10 10 10 10 10 10 10 10
Frequency (rad/min) Frequency (rad/min)
As it is not possible to calculate the SSV exactly, the design task (6.24)
is usually approximated by the upper bound for μ (6.8)
The μ-synthesis is not a trivial task. Yet no algorithms have been devel-
oped which allow a one-step solution of the μ-synthesis problem (6.24).
The known algorithms require the repeated calculation of an H ∞
problem, alternating with a scaling of the plant. These algorithms
cannot guarantee convergence of the iteration.
DK-Iteration
The synthesis problem (6.25) is a simultaneous optimization problem of
the frequency-dependent scaling matrices D and the controller K.
Because no direct solutions exist, Doyle [6.7] proposes an iterative
approach: If we keep the diagonal scaling matrices D constant, the mini-
mization of inf σ max ( DMD –1 ) ∞ forms the convex H ∞ problem
D∈D
K 0 = arg inf F l ( P, K ) ∞
K
˜ (s)
Fit D(s) with stable, min. phase transfer functions D
˜ ( s )F ( P, K )D̃ –1 ( s )
K 1 = arg inf D l ∞
K
.............
μK-Iteration
A new algorithm for μ-synthesis has been proposed by Lin et al. [6.11].
Instead of fitting the scaling matrices D, this algorithm is based on a fit
of the frequency-dependent SSV with a stable rational transfer func-
tion. At each iteration step, the plant is premultiplied with a diagonal
matrix of the μ-approximating transfer function. Thus the peaks of μ
within the frequency range of interest are punished, and the algorithm
tries to flatten the μ-curve. The convergence of this algorithm is not
proved. The authors present “a reasoned argument for believing that
the sequence will converge” [6.11]. A scheme of the μK-algorithm is
136 6 μ-Optimal Controller Design
K 0 = arg inf F l ( P, K ) ∞
K
μ ( jω ) = μ ˜ [ F l ( P, K 0 ) ]
Δ
K 1 = arg inf μ̃ 0 ( s )F l ( P, K ) ∞
K
μ ( jω ) = μ ˜ [ F l ( P, K 1 ) ]
Δ
μ ( jω )
Fit μ̃ 1 ( jω ) ≈ -----------------------
μ ( jω ) ∞
..........
Experiences
1.5
Structured singular value (RP)
3
4
5
1
6
0.5
0 -6 -5 -4 -3 -2 -1 0
10 10 10 10 10 10 10
Frequency (rad/min)
The input vector of the controller may consist either of the pressure-
compensated tray temperatures T10 and T44, or of all three tempera-
tures. The temperature T24 is close to the feed tray and its response to
feed flow disturbances is faster than that of the other two temperatures.
Therefore, an improved controller design should result from this addi-
tional temperature measurement.
1
w e = 100 --------------------------- (6.28)
i 1 + 20580s
1
Structured singular value
RP
Figure 6.9: Robust performance
and stability for the μ-optimal
0.5 RS state-space controller (controller
inputs: T10, T44, T24)
0 -5 -3 -1 1
10 10 10 10
Frequency (rad/min)
1
Tr → y 1
Td → y
10 10
Magnitude
Magnitude
0 0
10 10
-1 -1
10 10
-2 -2
10 -5 -3 -1 1
10 -5 -3 -1 1
10 10 10 10 10 10 10 10
Frequency (rad/min) Frequency (rad/min)
a) b)
Figure 6.10: Singular values for the nominal closed-loop system with the μ-optimal
state-space controller (controller inputs: T10, T44, T24)
a) Transfer function from reference signals to controlled output signals
b) Transfer functions from disturbance signals to controlled output signals
Dash-dotted line: T F → y , solid line: T x → y
F
1
10
0
10
Magnitude
-1
10
-2
10
-3
10 -5 -4 -3 -2 -1 0 1
10 10 10 10 10 10 10
Frequency (rad/min)
Figure 6.11: Singular values of the sensitivity function at u for the nominal closed-
loop system with the μ-optimal state-space controller (controller inputs: T10,
T44, T24)
142 6 μ-Optimal Controller Design
S u ( s ) = [ I + K ( s )G ( s ) ] –1 (6.29)
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.020 0.020
(mol/mol)
(mol/mol)
0.015 0.015
Composition
Composition
Composition
Composition
0.010 0.010
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.4 0.4
0.2 0.2
(K)
Temperature (K)
Temperature
Temperature
Temperature
0.0 0.0
-0.2 -0.2
-0.4 -0.4
Control error T-10 Control error T-10
Control error T-44 Control error T-44
-0.6 -0.6
0 10 20 30 40 0 10 20 30 40
Time (h)
[h] Time [h]
(h)
Figure 6.12: Simulation results with μ-optimal state space controller (controller
inputs: T10, T44, T24) for an increase in feed composition (0.8 → 0.9 mol/mol)
at t=0 h and an increase of feed flow rate (+ 3.6 mol/min) at t=20 h
Upper plots: Product composition
Lower plots: Control error
L, V equal to controller output
ΔL with +10% error, ΔV with –10% error
144 6 μ-Optimal Controller Design
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
70
140
60
(mol/min)
(mol/min)
Flow rate [mol/min]
Flow
30 80
20 60
0 10 20 30 40 0 10 20 30 40
Time
Time (h)
[h] Time
Time (h)
[h]
Figure 6.13: Simulation results with μ-optimal state-space controller (controller in-
puts: T10, T44, T24) for an increase in feed composition (0.8 → 0.9 mol/mol) at
t=0 h and an increase of feed flow rate (+ 3.6 mol/min) at t=20 h
L, V equal to controller output
ΔL with +10% error, ΔV with –10% error
vapor flow thus increases the liquid flow rates below the corresponding
trays. Second, the major part of the feed leaves the column as top
product.
If we compare the plots for the minimum and maximum feed flow rate,
we recognize an essentially slower rejection of the feed composition
disturbance at the maximum feed flow rate. A distinct improvement of
the performance at maximum feed flow rate is not possible using a
linear time-invariant feedback controller. Higher controller gains would
improve the disturbance compensation at this operating point, but
simultaneously destabilize the control loop at the minimum feed flow
rate. A closer look at these figures demonstrates that especially at high
feed flow rates the controller response is more sluggish for changes in
6.4 Controller design with μ-synthesis 145
feed composition but not for disturbances in the feed flow rate. This fact
is explained by the course of the manipulated variables in Figure 6.13.
An increase in feed composition at minimum feed flow rate forces the
controller to reduce the reflux and the boilup by ≈11 mol/min, while at
maximum feed flow both flow rates have to be reduced by ≈30 mol/min!
Since in practice a step change of feed composition is improbable, the
rejection of feed flow variations has much higher significance.
RP
Figure 6.14: Robust performance
and stability for the μ-optimal
0.5 RS state-space controller (controller
inputs: T10, T44)
0 -5 -3 -1 1
10 10 10 10
Frequency (rad/min)
Sensitivity at e
1
10
0
10
Magnitude
-1
10
-2
10
-3
10 -5 -4 -3 -2 -1 0 1
10 10 10 10 10 10 10
Frequency (rad/min)
1
Sensitivity at u
10
0
10
Magnitude
-1
10
-2
10
-3
10 -5 -4 -3 -2 -1 0 1
10 10 10 10 10 10 10
Frequency (rad/min)
Figure 6.15: Singular values of the sensitivity functions at e (upper plot) and at u
(lower plot) for the nominal closed-loop system with the μ-optimal state-space
controller (controller inputs: T10, T44)
6.4 Controller design with μ-synthesis 147
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.020 0.020
(mol/mol)
(mol/mol)
0.015 0.015
Composition
Composition
Composition
Composition
0.010 0.010
θ = arg inf
θ
∑ μΔ3˜ { Fl [ P, K ( θ ) ] } (6.31)
i=1
Summing the cube, large values of the SSV have much more weight and
the design objective becomes closer to (6.30).
μ Δ { F l [ P, K ( θ ) ] } ∞
<1 (6.33)
–
r10 L Distillation T10
PID1
+ Column
with
r44 V inventory T44
PID2
+ control
–
on tray 24. The weighting functions are the same transfer functions as
discussed in section 6.3.
1 + TI 1 s
KR 1 --------------------- 0
L(s) = TI 1 s e 10 ( s )
(6.34)
V(s) 1 + TI 2 s e 44 ( s )
0 KR 2 ---------------------
TI 2 s
Table 6.1 summarizes the results of the parameter optimization for the
analytically and numerically linearized column models, as well as for a
complex and mixed real/complex μ-analysis.
design model, the design can be improved. This leads to results compa-
rable to those obtained with the numerically linearized models. These
experiences corresponds to those of the μ-synthesis.
0.5 RS
0 -5 -3 -1 1
10 10 10 10
Frequency (rad/min)
The transfer functions from the reference and disturbance inputs to the
temperature measurements for the nominal closed-loop systems (Figure
6.19) shows a high condition number for the tracking behavior within
the most important frequency range. This means a high sensitivity of
the tracking behavior to the direction of the reference inputs.
152 6 μ-Optimal Controller Design
Tr → y Td → y
1 1
10 10
0 0
Magnitude
Magnitude
10 10
-1 -1
10 10
-2 -2
10 -5 -3 -1
10 -5 -3 -1 1
10 10 10 10 10 10 10
Frequency (rad/min) Frequency (rad/min)
a) b)
Figure 6.19: Singular values for the nominal closed-loop with
diagonal PI-controller
a) Transfer function from reference to output signals
b) Transfer functions from disturbance to output signals
Dash-dotted line: T F → y , solid line: T x → y
F
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.020 0.020
(mol/mol)
(mol/mol)
0.015 0.015
Composition
Composition
Composition
Composition
0.010 0.010
F(t=0)=20 mol/min
Ft=0=20 mol/min F(t=0)=46 mol/min
Ft=0=46 mol/min
0.4 0.4
0.2 0.2
(K)
Temperature (K)
Temperature
Temperature
Temperature
0.0 0.0
-0.2 -0.2
-0.4 -0.4
Control error T-10 Control error T-10
Control error T-44 Control error T-44
-0.6 -0.6
0 10 20 30 40 0 10 20 30 40
Time [h]
(h) Time [h]
(h)
Figure 6.20: Simulation results with diagonal PI control for an increase in feed
composition (0.8 → 0.9 mol/mol) at t=0 h and an increase of feed flow rate
(+ 3.6 mol/min) at t=20 h
Upper plots: Product composition
Lower plots: Control error
L, V equal to controller output
ΔL with +10% error, ΔV with –10% error
154 6 μ-Optimal Controller Design
L ( s ) = G K1 0 e 10 ( s )
(6.35)
V(s) 0 G K2 e 44 ( s )
with
1 + TI i s + TI i TD i s 2
G Ki ( s ) = KR i ---------------------------------------------------- (6.36)
TI i s ( 1 + sTL i )
Table 6.2: Results for the diagonal real PID control structure
Controller KR TI TD TL
mol/min/°C (min) (min) (min)
PID 1 –15.97 101 7.41 2.00
PID 2 4.40 39.0 15.2 7.16
The μ-plots (Figure 6.21) for the diagonal PID control structure show an
improvement of the robust performance. However, the design objective
of robust performance
6.5 Design of controllers with fixed structure 155
Structured Singular Value
1.5
RP
Figure 6.21: Robust performance
1 and stability for diagonal real
PID control
0.5 RS
0 -5 -3 -1 1
10 10 10 10
Frequency (rad/min)
μ ˜ { F l [ P, K ( θ ) ] } <1 (6.37)
Δ ∞
is by far not reached. The simulation results in Figure 6.22 illustrate the
same sluggish behavior as was obtained with the PI controllers. The im-
provement is a slight reduction of the settling time.
An analysis of the controller’s singular values shows large high-
frequency gains despite a limitation of the minimum filter time constant
TL (Figure 6.23). This makes a first-order filter for the reference inputs
2
10
Magnitude
10
0 Figure 6.23: Singular values of the
diagonal PID controller
-2
10 -5 -3 -1 1
10 10 10 10
Frequency (rad/min)
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.020 0.020
(mol/mol)
(mol/mol)
0.015 0.015
Composition
Composition
Composition
Composition
0.010 0.010
Figure 6.22: Simulation results with diagonal PID control for an increase in feed
composition (0.8 → 0.9 mol/mol) at t=0 h and an increase of feed flow rate
(+ 3.6 mol/min) at t=20 h
L, V equal to controller output
ΔL with +10% error, ΔV with –10% error
r10 –
L T10
PID1
+ Distillation
C2 Column
with
C1 inventory
r44 control T44
V
PID2
+
–
Table 6.3 summarizes the μ-optimal parameters for PI and real PID
Table 6.3: μ-optimal parameters for PI(D) control with
static decoupling
Controller or KR TI TD TL C
decoupler No. (mol/min/°C) (min) (min) (min) (–)
1 –5.21 22.8 – – –0.0240
2 3.71 46.8 – – 1.11
1 –13.1 51.6 7.83 8.43 –0.217
2 4.56 62.1 5.11 3.07 1.03
The results for the decouplers are somewhat surprising. They indicate
that the optimal decoupling is very close to a one-way decoupling! Let us
examine this control structure in detail:
This special behavior of the control system has its significant advan-
tages for the closed-loop behavior. The μ-plots in Figure 6.25 demon-
Structured singular value
0.5 RS 0.5 RS
0 -5 -3 -1 1
0 -5 -3 -1 1
10 10 10 10 10 10 10 10
Frequency (rad/min) Frequency (rad/min)
Figure 6.25: Robust performance and stability for PI control (left) and
real PID control (right) with static decoupling
The singular value plots of the loop transfers from the reference and
disturbance signals to the output signals (Figure 6.26) illustrate the
better controller performance as well. The condition numbers of T r → y
are much smaller than those of the diagonal PI(D) control structure, and
the tracking behavior is significantly improved.
Tr → y Td → y
1 1
10 10
Magnitude
Magnitude
0 0
10 10
-1 -1
10 10
-2 -2
10 -5 -3 -1
10 -5 -3 -1 1
10 10 10 10 10 10 10
Frequency (rad/min) Frequency (rad/min)
a) b)
Figure 6.26: Singular values for the nominal closed-loop system for PID control
with static decoupling
a) Transfer function from reference to output signals
b) Transfer functions from disturbance to output signals
Dash-dotted line: T F → y , solid line: T x → y
F
1 + TDs
C i ( s ) = KC --------------------- (6.38)
1 + TLs
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.020 0.020
(mol/mol)
(mol/mol)
0.015 0.015
Composition
Composition
Composition
Composition
0.010 0.010
Figure 6.27: Simulation results for PID control with static decoupling for an in-
crease in feed composition (0.8 → 0.9 mol/mol) at t=0 h and an increase of feed
flow rate (+ 3.6 mol/min) at t=20 h
L, V equal controller output
ΔL with +10% error, ΔV with –10% error
Controller KR KC TI TD TL
(mol/min/°C) (–) (min) (min) (min)
PID 1 –22.2 – 80.2 19.6 44.8
PID 2 5.68 – 59.4 12.6 24.7
C1 – –0.138 – 117 7.42
C2 – 1.07 – 53.0 71.43
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.020 0.020
(mol/mol)
(mol/mol)
0.015 0.015
Composition
Composition
Composition
Composition
0.010 0.010
Figure 6.28: Simulation results for PID control with dynamic decoupling for an in-
crease in feed composition (0.8 → 0.9 mol/mol) at t=0 h and an increase of feed
flow rate (+ 3.6 mol/min) at t=20 h
L, V equal controller output
ΔL with +10% error, ΔV with –10% error
The results for two-way decoupling have shown optimal results for
decoupling structures which are close to one-way decoupling. In this
section the optimal tuning results for one-way decoupling are discussed.
This control structure is particularly easy to implement in a distributed
control system and simple to initialize. In order to keep the decoupler as
simple as possible, the discussion is limited to static one-way decou-
pling.
Controller or KR TI TD TL C
decoupler No. (mol/min/°C) (min) (min) (min) (–)
1 –10.5 45.7 2.18 5.01 0
2 5.35 67.4 13.4 13.9 1.05
1
Structured singular value
RP
0 -5 -3 -1 1
10 10 10 10
Frequency (rad/min)
for the two-way decoupling lead us to expect a good performance for the
first case ( C 1 = 0 ), no inference is possible for the second case ( C 2 = 0 ).
In fact, the optimization results show insufficient performance for the
second case ( C 2 = 0 ). Therefore a reversal of the decoupling control
structure with shaping of the composition profile by the bottom compo-
sition controller and moving the composition profiles position by the top
composition controller does not lead to results comparable to those
obtained with the other decoupling structure.
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.020 0.020
(mol/mol)
(mol/mol)
0.015 0.015
Composition
Composition
Composition
Composition
0.010 0.010
6.6 Summary
The comparison of the state-space controllers obtained by μ-synthesis
with PID control structures obtained by μ-optimization leads to
surprising results. The frequently heard opinion that state-space
controllers are much superior to PID control structures apparently is
not true for this distillation column. The PID control structures with
decoupling exhibit nearly the same performance as that achieved with
state-space controller of a higher order, provided that the PID control
structures are optimally tuned. The visual results of the μ-curves and
simulation plots shall be supported by numerical measures. For
purposes of comparison, the integral square of the control errors
t=40h
ISE = ∫ [ e 10
2 ( t ) + e 2 ( t ) ]dt ,
44 (6.39)
0
t=40h
ITAE = ∫ [ e 10 ( t ) + e 44 ( t ) ] t dt (6.40)
0
have been calculated and summed up for both operating points and all
controllers. While ISE punishes especially large control errors, the ITAE
performance measure has a higher importance for the process industry
because it punishes any undesirably sluggish disturbance rejection.
Both criteria, relative to the result for the state-space controller using 3
temperature measurements, can be found in Table 6.6. The last two
columns in this table state the maximum absolute value of the SSV (RP)
and the value of the optimization criterion
f(θ) = ∑ μΔ˜ { Fl [ P, K ( θ ) ] }
3
(6.41)
i=1
166 6 μ-Optimal Controller Design
6.7 References
[6.11] Lin, J.-L., I. Postlethwaite, and D.-W. Gu: “μ-K Iteration: A New
Algorithm for μ-synthesis,” Automatica, 29, 219-224 (1993)
Chapter 7
7.1 Introduction
A controller design for the entire operating range of the distillation
column (see Chapter 6) requires a structured uncertainty model incor-
porating two linear models, and a huge computational effort. Naturally,
the question arises what controller performance and robustness proper-
ties can be achieved if we use simpler design methods, based on just one
plant model for the nominal operating point (Model GN) and classical
design methods or simple unstructured uncertainty bounds.
A few of these simpler methods are discussed in this chapter. They are
applied in a straightforward manner, and the design results are not
guaranteed to represent the optimum achievable controller perfor-
mance. However, the results give an impression of the limits and
inherent problems of the application of design methods based on simpler
uncertainty concepts, and they allow a comparison with the μ-optimal
results presented in the previous chapter. The weighting functions of
172 7 Controller Design for Unstructured Uncertainty — A Comparison
the structured uncertainty model used for the μ-analysis are the same
as those used in the previous chapter.
While the following two simple and model based tuning methods make
use of the classical design methods, they try to pay attention to the loop
interactions. Both methods lead to a nominally stable controller design.
However, sufficient stability margins for the closed-loop system at all
possible operating points cannot be guaranteed.
y ( s ) = [ I + G ( s )K ( s ) ] – 1 G ( s )K ( s )r ( s ) (7.1)
det ( I + G ( s )K ( s ) ) = 0 (7.2)
7.2 Diagonal PI-control 173
W ( s ) = – 1 + det ( I + G ( s )K ( s ) ) (7.3)
The closer this function approaches the (-1,0) point in the Nyquist plot,
the closer the MIMO system is to closed-loop instability. The design
objective is defined as
W ( jω )
L lm = 20 log ------------------------- ≤ 2p ∀ω ∈ R + (7.4)
-
1 + W ( jω )
K ZN
K i = -------------i TI i = F TI ZN (7.5)
F i
Controller KR TI
(mol/min/°C) (min)
PI 1 –47.1 95.2
PI 2 6.74 171.8
174 7 Controller Design for Unstructured Uncertainty — A Comparison
3
Structured singular value
RP
2
Figure 7.1: μ-plots for a diagonal
PI-control law tuned with BLT-
1 method
RS
0 -5 -3 -1 1
10 10 10 10
Frequency (rad/min)
The idea of the sequential loop closing was introduced by Mayne ([7.8],
[7.9]). First, a SISO controller is designed for one pair of input and
output variables. When this design has been completed, the corre-
sponding control loop is closed and the next pair of input and output
variables is chosen. Thus the interaction between the control loops is
taken into account. This design procedure is illustrated in Figure 7.2.
–
u1 y1
K 1 (s)
–
u2 y2
K 2 (s)
u3 G(s) y3
.....
.....
G ii ( jω ) > G ij ( jω ) ∀ω ∈ R + (7.6)
Design Results
The sequential loop closing idea has been applied to the composition
control problem represented by the nominal model G N ( s ) . For each
SISO loop, a phase margin of at least 60 degrees and for both controllers
a maximum high frequency gain of 18 mol/min/°C has been required.
The results of both possible design sequences and with a minimal inte-
gral absolute error (IAE) for the rejection of feed composition and feed
flow disturbances (with respect to the linear model) are summarized in
Table 7.2.
Design sequence KR 1 TI 1 KR 2 TI 2
(mol/min/ (min) (mol/min/ (min)
°C) °C)
Top → Bottom –18.0 101.9 10.09 55.8
Bottom → Top –18.0 52.6 8.78 214.5
Top → Bottom design sequence, and of Se=2.4 and Su=1.9 for the
sequence Bottom → Top. These stability margins are insufficient. The
results of the analysis using the structured uncertainty model are illus-
trated by the μ-plots in Figure 7.3. Both controller designs can neither
guarantee robust performance nor robust stability.
Structured singular value
RP
2 2
RP
1 1
RS RS
0 -5 -3 -1 1
0 -5 -3 -1 1
10 10 10 10 10 10 10 10
Frequency (rad/min) Frequency (rad/min)
a) b)
Figure 7.3: μ-plots for the sequential loop closing designs
a) Top → Bottom design sequence
b) Bottom → Top design sequence
T End
The optimal parameters which minimize the IAE criterion are found
either by trial and error or by a constrained parameter optimization.
Results
The results for this design approach are given in Table 7.3. The corre-
sponding μ-plots (Fig. 7.4) illustrate the improved robust stability prop-
erties compared to the previous two methods. While design guarantees
robust stability, the robust performance is substantially worse than the
Controller KR TI
(mol/min/°C) (min)
PI 1 –5.10 600.0
PI 2 4.92 86.2
178 7 Controller Design for Unstructured Uncertainty — A Comparison
6
RP Figure 7.4: μ-plots for diagonal
4 PI-controller designed by op-
timizing method
2
RS
0 -5 -3 -1 1
10 10 10 10
Frequency (rad/min)
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.020 0.020
(mol/mol)
(mol/mol)
0.015 0.015
Composition
Composition
Composition
Composition
Top composition Top composition
Bottom composition Bottom composition
0.010 0.010
0.005 0.005
0 10 20 30 40 0 10 20 30 40
[h]
Time (h) Time (h)
[h]
Figure 7.5: Simulation results with diagonal PI controller (designed by optimiz-
ing method) for an increase in feed composition (0.8 → 0.9 mol/mol) at t=0 h
and an increase of feed flow rate (+ 3.6 mol/min) at t=20 h
Upper plots: Product composition
Lower plots: Control error
L, V equal to controller output
ΔL with +10% error, ΔV with –10% error
G * u → y ( s ) = G u → y ( s )G –1 u → y ( 0 )
or (7.10)
G * u → y ( s ) = G – 1 u → y ( 0 )G u → y ( s )
G – 1 u → y ( 0 ) = ( CA –1 B ) –1 (7.11)
180 7 Controller Design for Unstructured Uncertainty — A Comparison
G – 1 u → y ( j∞ ) = ( CB ) – 1 (7.12)
G u → y ( 0 ) = UΣV T , (7.13)
ˆ = VΣ –1 U T
D (7.14)
F = αI + ( 1 – α )Σ (7.15)
˜ as
and define a new compensation matrix D
˜ = VFΣ – 1 U T
D (7.16)
7.3 PI-control with decoupling 181
Design results
RP
1
0 -5 -3 -1 1
10 10 10 10
Frequency (rad/min)
min/°C) are given in Table 7.5. The μ-plots for this controller design
Controller KR TI
(mol/min/°C) (min)
PI 1 –18.0 47.4
PI 2 18.0 116.0
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.020 0.020
(mol/mol)
(mol/mol)
0.015 0.015
Composition
Composition
Composition
Composition
0.010 0.010
Figure 7.7: Simulation results with SVD based compensator and diagonal PI con-
trol for an increase in feed composition (0.8 → 0.9 mol/mol) at t=0 h and an
increase of feed flow rate (+ 3.6 mol/min) at t=20 h
Upper plots: Product composition
Lower plots: Control error
L, V equal to controller output
ΔL with +10% error, ΔV with –10% error
184 7 Controller Design for Unstructured Uncertainty — A Comparison
H∞ Design specification
The closed-loop system with the plant G(s) and the controller K(s),
augmented with the weighting functions W d ( s ) , W e ( s ) , W u ( s ) , and
W y ( s ) is outlined in Figure 7.8. This scheme is often called S/KS/T-
weighting scheme. The matrix W d ( s ) is a diagonal matrix of transfer
functions and represents the frequency content of the feed composition,
feed flow rate, and reference input signals. The selection of these input
weights is discussed in section 6.3. The same weighting functions are
applied here.
We(s) ze(s)
Wu(s) zu(s)
Wy(s) zy(s)
d(s)
Wd(s) G(s)
r(s) K(s) y(s)
+ e(s) u(s)
–
W e ( s ) = diag [ w e ( s ), w e ( s ) ] (7.17)
W u ( s ) = diag [ w u ( s ), w u ( s ) ] (7.18)
W y ( s ) = diag [ w y ( s ), w y ( s ) ] (7.19)
The poles and zeros of the weighting functions were adjusted until the
sensitivity functions at e and at u of the closed-loop system had attained
approximately the same peak values as the μ-optimal controller design
(with 2 temperature measurements), a high performance, and
T ( jω ) ≈1 (7.20)
d →z
r ∞
1
w e ( s ) = 100 ------------------------ (7.21)
1 + 9520s
186 7 Controller Design for Unstructured Uncertainty — A Comparison
1 + 520s
w u ( s ) = 0.5 --------------------- (7.22)
1 + 13s
1 + 1500s
w y ( s ) = 0.1 ------------------------ (7.23)
1 + 1.5s
Design results
Despite the fact that the singular values of sensitivity functions for the
H∞- design (Figure 7.10) and for the μ-synthesis (Figure 6.15) are nearly
identical, the μ-analysis shows significant differences. The μ-plots of the
H∞ design (Figure 7.9) show much higher peak values in the low and
mid-frequency ranges. The simulation results (Figure 7.11) allow a
conclusion with respect to the larger structured singular values: The
sensitivity of the closed-loop performance to errors in the manipulated
variables is large. A reduction of this sensitivity to plant input errors
was not possible using the common S/KS/T weighting scheme.
1.5
Structured singular value
1 RP
Figure 7.9: μ-plots for H∞ op-
RS timal controller
0.5
0 -5 -3 -1 1
10 10 10 10
Frequency (rad/min)
H∞ optimal design 187
1
Sensitivity at e
10
0
10
Magnitude
-1
10
-2
10
-3
10 -5 -4 -3 -2 -1 0 1
10 10 10 10 10 10 10
Frequency (rad/min)
1
Sensitivity at u
10
0
10
Magnitude
-1
10
-2
10
-3
10 -5 -4 -3 -2 -1 0 1
10 10 10 10 10 10 10
Frequency (rad/min)
Figure 7.10: Singular values of the sensitivity functions at e (upper plot) and at u
(lower plot) for the nominal closed-loop system with the H∞ controller
188 7 Controller Design for Unstructured Uncertainty — A Comparison
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.020 0.020
(mol/mol)
(mol/mol)
0.015 0.015
Composition
Composition
Composition
Composition
0.010 0.010
7.5 Summary
The application of design methods for unstructured uncertainty to the
composition (or temperature) control problem shows that it is extraordi-
narily difficult to obtain performances which are comparable to those of
the μ-optimal controllers. Despite the high effort for a robust tuning of
the PI-control structures, it was not possible to achieve any satisfactory
result.
7.6 References
[7.1] Brambilla, A., and L. D’Elia: “Multivariable Controller for Distil-
lation Column in the Presence of Strong Directionality and Mod-
el Errors,” Ind. Eng. Chem. Res., 31, 536-543 (1992)
[7.2] Dailey, R. L.: “Lecture Notes for the Workshop on H∞ and μ Meth-
ods for Robust Control,” 1991 IEEE Conference on Decision and
Control, Brighton, December 9-10 (1991)
Chapter 8
8.1 Introduction
It is a drawback of feedback control that a corrective action necessitates
a deviation of the controlled variables from their setpoints. This disad-
vantage can be overcome by the use of feedforward control. A major and
probably the most frequent disturbance of a distillation column is a
change in the feed flow rate. Because the feed flow rate is always
measured, it can be used as a controller input. An appropriately
designed feedforward controller takes most of the necessary corrective
action before the product compositions and the controlled tray tempera-
tures change. However, because of model errors and other unmeasured
disturbances a feedforward controller alone will never be able to yield
perfect control so that feedback control will still be needed.
Within this chapter, the design of linear time-invariant feedforward
controllers for our distillation column is discussed. The proposed design
methods take into account the wide operating range of the distillation
column and the unmeasured feed composition.
192 8 Feedforward Controller Design
a) One-step design Δ
xF
F
P p
r
wF(s) KF(s)
K(s)
b) Two-step design
Δ
Step 1: Feedback design
d p
r P
K(s)
Δ
Step 2: Feedforward design
F
r P* p
wF(s) KF(s)
+
+
K(s)
weighted feed flow signal. However, this approach has certain draw-
backs:
8.3 H∞-minimization
The H∞-minimization [8.4] is well suited for a feedforward controller
design. Before we use the numerical tools available (e.g., [8.1], [8.2]), we
have to build up a closed-loop plant with a previously designed feedback
controller K(s). As an example, the μ-optimal state-space controller
using all 3 temperature measurements is selected (see section 6.4.3). If
we wish to improve the compensation of feed flow disturbances for the
plant models G R ( s ) as well as for G I ( s ) , we have to close the feedback
loops for both models separately, define the desired performance, and
limit the high-frequency output of the feedforward controller K F ( s ) .
The design plant is outlined in Figure 8.2.
zu
Wu ( s )
F
GR ( s )
u
–
+
K (s)
F uF + ze
KF ( s )
We ( s )
F
GI ( s )
u
+ –
K (s)
+
W e ( s ) = diag [ w e ( s ), w e ( s ), w e ( s ), w e ( s ) ] (8.1)
It demands the same performance for both column models and both
controlled temperatures. The transfer function w e ( s ) is chosen as a
first-order lag with a high static gain. The pole of w e ( s ) is adjusted
until T F → z ∞ ≈ 1 is achieved. The final transfer function becomes
1
w e ( s ) = 100 ------------------------ (8.2)
1 + 2380s
W u ( s ) = diag [ w u ( s ), w u ( s ) ] (8.3)
F F
1 + 104s
w u ( s ) = 0.5 --------------------- (8.4)
F 1 + 2.5s
1
KF 1
Td → y
10 10
Magnitude
0 0
Magnitude
10 10
-1 -1
10 10
-2 -2
10 -5 -3 -1 1
10 -5 -3 -1 1
10 10 10 10 10 10 10 10
Frequency (rad/min) Frequency (rad/min)
a) b)
Figure 8.3: a) Singular values of the feedforward controller
b) Singular values of the transfer function from the disturbance inputs d to the
controlled output signals y for the nominal model GN with feedforward and
feedback control. Solid line: T x → y , dash-dotted line: T F → y
F
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.020 0.020
(mol/mol)
(mol/mol)
0.015 0.015
Composition
Composition
Composition
Composition
0.010 0.010
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.4 0.4
0.2 0.2
Temperature(K)
Temperature(K)
Temperature
Temperature
0.0 0.0
-0.2 -0.2
-0.4 -0.4
Control error T-10 Control error T-10
Control error T-44 Control error T-44
-0.6 -0.6
0 10 20 30 40 0 10 20 30 40
Time [h]
(h) Time [h]
(h)
Figure 8.4: Simulation results with μ-optimal state space controller (controller in-
puts: T10, T44, T24) and feedforward controller for an increase in feed composi-
tion (0.8 → 0.9 mol/mol) at t=0 h and an increase of feed flow rate (+ 3.6 mol/
min) at t=20 h
Upper plots: Product composition
Lower plots: Control error
L, V equal to controller output
ΔL with +10% error, ΔV with –10% error
198 8 Feedforward Controller Design
KR L 1
KF ( s ) = ---------------- (8.5)
KR V 1 + Ts
Most of the feed flow disturbances entering this distillation column are
step changes. Consequently, we are able to define an appropriate design
objective in the time domain. It is the minimum absolute control error
for a step change in the feed flow rate. The design objective becomes
with
Te
E = ∫ { e10R ( t ) + e 44 ( t ) + e 10 ( t ) + e 44 ( t ) }dt .
R I I
(8.7)
0
F
GR ( s )
u
– e 10
R
+
K (s) e 44
R
F uF +
KF ( s )
F
GI ( s )
u
+ – e 10
I
K (s) e 44
+ I
1
K F ( s ) = 1.5 -------------------- (8.8)
2.6 1 + 5.0s
1
KF 1
Td → y
10 10
Magnitude
Magnitude
0 0
10 10
-1 -1
10 10
-2 -2
10 -5 -3 -1 1
10 -5 -3 -1 1
10 10 10 10 10 10 10 10
Frequency (rad/min) Frequency (rad/min)
a) b)
Figure 8.6: a) Singular values of the feedforward controller with fixed structure
b) Singular values of the transfer functions for the nominal closed loop system
from the disturbances inputs d to the controlled output signals y (Feedback
and feedforward control). Solid line: T x → y , dash-dotted line: T F → y
F
Ft=0=20 mol/min
F(t=0)=20 mol/min Ft=0=46 mol/min
F(t=0)=46 mol/min
0.020 0.020
(mol/mol)
(mol/mol)
0.015 0.015
Composition
Composition
Composition
Figure 8.7: Simulation results with μ-optimal PID controller with one-way decou-
pling and a simple feedforward controller for an increase in feed composition
(0.8 → 0.9 mol/mol) at t=0 h and an increase of feed flow rate (+ 3.6 mol/min)
at t=20 h
L, V equal to controller output
ΔL with +10% error, ΔV with –10% error
8.5 Summary
The compensation of feed flow disturbances can be improved by using
feedforward controllers. H∞-norm minimization and the minimization of
the control errors in the time domain (for feedforward controllers with
fixed structure) are efficient design methods. Frequency domain as well
as time-domain results demonstrate the pleasing improvements which
are obtained even by a feedforward controller of order one. A comparison
of the ISE and ITAE criteria (see section 6.6) in Table 8.1 demonstrates
202 8 Feedforward Controller Design
8.6 References
Chapter 9
Practical Experiences
9.1 Introduction
In simulations the performance of controllers is tested in a sterile envi-
ronment. Lacking measurement noise, operator actions, and varying
environmental conditions, the results of these simulations represent a
well established basis for a comparison of different controller designs.
However, only the implementation of a controller in the real plant
proves its performance. While in the literature a great number of design
methods has been proposed and the resulting controllers have been
tested by simulations, only very few results of an implementation at a
real industrial distillation column have been reported.
This chapter complements the simulation results presented in previous
chapters with the results of a controller implementation in the distrib-
uted control system (DCS) which is coupled with this distillation
column. The first section describes the implementation including the
handling of constraints. Further sections discuss the use of pressure
compensated temperatures, the controller performance observed, and
economic aspects. A short summary concludes the chapter.
206 9 Practical Experiences
Robustness: The control design must guarantee stability for the entire
operating range of the column, including time variations due to corro-
sion of trays, transmitter drifts, etc.
This control scheme has been implemented in the DCS installed at the
plant considered here (i.e., an Eckardt PLS 80E). The controller inputs
are estimated tray compositions x̂ i , which for the operators have proved
to be easier to understand than pressure compensated temperatures.
The proportional gains of the controllers are easily converted for these
controller inputs.
T 10
θ 1 + θ 2 ( T + T Corr ) x̂ 10 R FB1
PID 1 rR
% Valve
+ θ3 p +
PID R
R FB1, max
R max – R FF R
p0 9 p̂ 10
R FB1, min
p 0 + ------ ( p 51 – p 0 )
50 Composition R min – R FF KQ/R
estimation
Composition Q FB1
p 51
43 p̂ 44
estimation
p 0 + ------ ( p 51 – p 0 ) Feedback
50 controllers Decoupling
with
constraints
Θ 1 + Θ 2 ( T + T Corr ) x̂ 44 Q FB2 rQ
T 44 PID 2 + % Valve
+ Θ3 p + Θ4 p2 PID Q
Q FB2, max
Q max – Q FF – Q FB1 Q
Q FB2, min
Q min – Q FF – Q FB1
F R FF
Figure 9.1: Controller
LAG
implementation
Q FF
KQF/RF
Feedforward
controller
9 Practical Experiences
9.3 Composition estimators 209
If the reflux as well as the reboiler heat duty reach their minimum
constraints, both products become purer than desired.
Hence operating data were recorded for two weeks. Since the feed
composition was almost constant, it was possible to compare these
measurement data with tray compositions calculated by steady-state
simulations. Minimizing the errors between the estimated and the
calculated tray compositions, a correction of the estimated tray pres-
sures by 20% was necessary to correct the estimates. Since pressure
sensors on tray 10 and 44 are not installed, the pressures on these trays
are calculated by a linear interpolation between top pressure and
bottom pressure (see Fig. 9.1). The error in the pressure compensation
210 9 Practical Experiences
might have been caused by this interpolation. Other error sources could
have been incorrect {Tpx} data or pressure measurements. Once the
parameters of the estimators had been adjusted, these simple estima-
tors worked fairly satisfactorily. Nevertheless, the compensation of the
pressure variations’ influence on the tray temperatures is the limiting
factor for the overall performance of the control scheme. This will be
shown in more detail in the following section.
During the recording of these operating data, the setpoints were kept
constant. In Figure 9.2 the feed flow rate was increased by 40 l/h in four
steps. The feed flow rate at t=0 h was 260 l/h (49 mol/min), while the feed
composition was approximately 0.85 mol/mol. Although the feed flow
9.4 Controller performance 211
Feed
20
flow rate (l/h)
Deviation of
10
-10
0 10 20 30 40 50 60 70
Time (h)
Tray 10
Deviation of compensated
Deviation of estimated
5
composition (mol-%)
-0.5
temperature (°C)
0
0.5
-5
0 10 20 30 40 50 60 70
Time (h)
6
composition (mol-%)
temperature (°C)
-0.5
0.5
-6
0 10 20 30 40 50 60 70
Time (h)
Figure 9.3: Recorded operating data with installed PID control scheme including
one-way decoupling and feedforward control.
Top: Deviation of feed flow rate from 170 l/h (32 mol/min)
Middle: Deviations of estimated tray composition and of pressure
compensated temperature from setpoint on tray 10
Bottom: Deviations of estimated tray composition and of pressure
compensated temperature on tray 44
212 9 Practical Experiences
Feed
40
flow rate (l/h)
Deviation of
20
0 20 40 60 80 100
Time (h)
Tray 10
Deviation of compensated
Deviation of estimated
5
composition (mol-%)
-0.5
temperature (°C)
0
0.5
-5
0 20 40 60 80 100
Time (h)
Tray 44
Deviation of compensated
Deviation of estimated
6
composition (mol-%)
temperature (°C)
-0.5
0.5
-6
0 20 40 60 80 100
Time (h)
Figure 9.2: Recorded operating data with installed PID control scheme including
one-way decoupling and feedforward control.
Top: Deviation of feed flow rate from 260 l/h (49 mol/min)
Middle: Deviations of estimated tray composition and of pressure
compensated temperature from setpoint on tray 10
Bottom: Deviations of estimated tray composition and of pressure
compensated temperature on tray 44
9.4 Controller performance 213
rate was out of the design range, the reflux and boilup remained within
the range covered by the controller design. In Figure 9.3, the feed flow
rate at t=0 h was 170 l/h (32 mol/min) and it was increased only once by
10 l/h.
The control errors in presence of these feed flow disturbances remain
extraordinary small. In fact, it is almost impossible to separate the
control error from the measurement noise and the effect of all other
unknown disturbances. This proves the high performance of this simple
advanced PID control scheme.
The advantages of the controller implementation are demonstrated best
by a comparison of the product compositions analyzed once a day before
and after the installation. At the beginning of this project, top and
bottom composition were controlled manually. The results are shown on
the left-hand sides of Figure 9.4 and 9.5. Obviously, the average product
compositions are found far from their setpoints, and the variations of
the product compositions are very large.
The right-hand sides of Figure 9.4 and 9.5 show the analysis results
beginning after the adjustment of the composition estimators. Clearly,
the variations of the product compositions are much smaller and the
average product compositions are close to the desired results. However,
despite the high performance of the control scheme as illustrated by
Figure 9.2 and 9.3, significant variations of the product compositions
can still be detected. Please remember that pressure measurements of
tray 10 and 44 are lacking. Therefore the influence of the large pressure
variations (bottom pressure: 120-190 mbar) to the tray temperatures
cannot be perfectly compensated and an adjustment of the controller
setpoints depending on the feed flow rate is necessary. Since the results
presented are achieved with almost constant setpoints, the results will
improve even further as the operators gain more extensive experience
with the setpoints.
214 9 Practical Experiences
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 35 70 0 22 44
Days Days
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 35 70 0 22 44
Days Days
0.5 0.5
Relative frequency
Relative frequency
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.1 0.2 0.3 0 0.1 0.2 0.3
Top composition (mol/mol) Top composition (mol/mol)
0.5 0.5
Relative frequency
Relative frequency
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.1 0.2 0.3 0 0.1 0.2 0.3
Bottom composition (mol/mol) Bottom composition (mol/mol)
9.6 Summary
The results of the implementation of the PID control structure with one-
way decoupling and feedforward control on the real plant confirm the
high performance of this simple control scheme indicated by simula-
tions. The main problem of the implementation was, except for over-
coming high psychological resistances, the correct parametrization of
the composition estimators. A solution of this problem never would have
been possible without an extensive comparison of simulation and oper-
ating data. Nevertheless, the use of pressure compensated tempera-
9.6 Summary 217
Chapter 10
Conclusions and
Recommendations
10.1 Introduction
This thesis treats all the necessary steps for a composition control
design for an industrial binary distillation column. Each of these steps
produced new insights into various aspects of the control design. Since
a chronological discussion of these steps would lead to a thematic confu-
sion, they are summarized in the four sections
• Controller synthesis
• Column models
This thesis does not presume to present a final solution to all distillation
control problems. The ideas presented come up against many gaps in
research, limits of distributed control systems, and problems of cooper-
ation between industry and university. In the last section the most
important aspects of these topics are discussed.
220 10 Conclusions and Recommendations
This thesis discusses the design of robust controllers for the dual compo-
sition control problem of an industrial binary distillation column. Distil-
lation columns are usually operated over a wide range of feed
compositions and feed flow rates. Consequently, a controller must guar-
antee stability and a high performance not only at a single operating
point, but for the entire operating range of the distillation column.
The optimal tuning of PID control structures with decoupling for this
distillation column caused an additional insight. The optimal controller
performance is achieved with an implicit decoupling scheme where in
essence
222 10 Conclusions and Recommendations
All results of this thesis are based directly or indirectly on models of the
distillation column. Especially the model-based adjustment of the
composition estimators clearly proved that such process models are
absolutely necessary. However, the control design may be based on
linear models that include or exclude flow dynamics.
10.6 Recommendations
Curriculum vitae