Dynamics and Control of

Distillation Columns
1. Distillation Dynamics (45%)……….p.2
2. Distillation Control (30%)……….....p.23
3. Some basic control theory (25%)…..p.37

SIGURD SKOGESTAD
NDF Workshop I
Lyngby, Denmark, May 24, 1989

Revised for: NDF Workshop III

Trondheim, Norway, Aug. 22, 1991
Additional Literature: Dynamics: p. 75 in NDF Notes 1990
Control: p.57 in NDF Notes 1990
(AIChE, May 1990) 1
1. Distillation Dynamics
1.1 Introduction 3
1.2 Degrees of freedom/steady state 4
1.3 Dynamic Equations 5
1.4 composition Dynamics 7

1.4.1 Dominant time constant 8τ1 

1.1.4 Internal flows 17

1.5 Flow dynamics 19

1.6 Overall Dynamics 21
1.7 Nonlinearity 21A
1.8 Linearization 22

2
 1.1 Introduction
Typical column: 1 feed, 2 products, no intermediate cooling

Overhead Condenser
Vapor VT

P Condenser Holdup
MD
(mα, m3)

Reflux Distillate
N
N-1 L
D, yD
Feed
 mol 
F, zF   1
 mol 
  1
 s 
 s 
Mole fraction 3
light component 2
Boilup
Reboiler 1
holdup MB (mol) V=QD/λ
Reboiler

QB (kN)
Bottom Product

B, xB ? 0
(mol/s) 3
Manipulated Variables (Valves): (μ)
V (indirectly), L, D, B, VT (indirectly) (5)

Disturbances: (d)
F, zF, qF (= fraction liquid in feed, feed enthalpy), rain
shower, + all u’s above

Controlled Variables (y)

MO, MB, P, YP, XB (5)

Inventory Separation
(composition, temperature)

4
 Inventories
M D ,M B :holdup liquid in condenser and reboiler
RT
p=M V  -pressure; M V :mol vapor insidecolumn
V
VV :volume vapor inside column.
That is, pressureis a measure for holdup vapor (M V )

(1.2) STEADY-STATE OPERATION

Must keep holdups (MD,MB,MV) constant

This uses 3 degrees of freedom (u’s); only two left.
The two degrees of freedom may be used to specify
two product specifications

e.g., YD and xB
Recall specifications for
Ttop and Tatm steady-state simulations
D (product rate) and YD etc. (PROCESS)
5
 Steady-state behavior (design)

y D /(1-y D )
Overall separation, binary S=
x B /(1-x B )

Infinite reflux, exact (Fenske): S=α N

 L /V 
NT

Finite reflux, good approximation: S=α N T T N

 LB VB  B

holds for columns with feed

at optimal location

(slight modification of Jafarey, Douglas, McAvoy, 1979)

Example: 3-Stage Column

Constant p
Constant holdup liquid and level
Negligible vapor h.
Constant molar flows
Constant α=10
2 Vicor stages + total condenser
Mi = 1 kmol (i = 1,2,3) 6
 D = 0.5

M2 YD=0.9

L=3

C = 1kmol/min
M2
ZF=0.5
V=3.5

M2 B = 0.5
XB = 0.1

Stage i Li Vi Xi yi
Condenser 1 3.05 0.9000
Feedstage 2 4.05 3.55 0.4737 0.9000
Reboiler 3 33.55 0.1000 0.5263

7
Get:
Total reflux: S=α N102 =100

Actual column: S=
1-YD  /YD = 0.9/0.1 =81
x B /(1x B ) 0.1/0.9

 LT /VT   3.05/3.55
NT 1

Approximation: S=α
N
=100 =75
 LB VB   4.05/3.55
NB 1

8
 Derivation of S=α N  LT VT  i+1
 LB VB 
NB
L V

Bottom Section XL+1 Yi

Lxi+1 =VYi +Bx B (Mass Balance) XB(x0)

YNi
Yi =α  x i  (VLE) (Constant Relative Volatiliy)
X Ni
Combine: <1 <1 x1
X i+1 X Ni Vα YNi B X B X Ni
= +
X i X Ni+1 L X Ni+1 L X i X Ni+1

=0
xi Si+1 Vα
Si = 
x Ni Si L
stripping factor bottom 9
 1.3 Dynamics of Distillation Columns
Balance equations

in
Accumulated = in – out
=D/DT (inventory)
out

Assumptions (always used)

A1. Perfect mixing on all stages
A2. Equilibrium between vapor on liquid on each stage
(adjust total no. of stages to match actual column)
A3. Neglect heat loss from column, neglect heat capacity of
wall and trays

Additional Assumptions (not always)

A4. Neglect vapor holdup (Mvi ≈ 0)
A5. Constant pressure (vapor holdup constant)
A6. Flow dynamics immediate (Mvi constant)
A8. Constant molar flow
A9. Linear tray hydraulics
10
 stage
Mi+1
i+1 Li+1 Balance equations:
Vi
d IN OUT
i
Mi  inventory  = -
dt sec sec
"state"
Vi-1 Li
i-1

Balance equation for stage without feed/side draw

INDEP.
DIFF. Component balance (index for component not
Shown EQ.

d
N Î -1  Mi x i +M vi Yi  =Li +1x i+1 +Vi-1Yi-1 -Li x i -Vi Yi
dt

(Allgebraic) VLE:
Yi = K i1  x i, p i 
FLASH
Ti =K i2  x i ,p i 
(Strongly Nonlinear)
h ,H =K i3  x i p i 
L V 11
L i
Total balance (sum of component balances)
d
1  M i +M vi  = Li+1 +Vi-1 -Li -Vi
dt
Energy balance
d
0or1  Mi Vi +M vi U ri  =Li+1h i+1
L v
+Vi-1Hi-1 -Li h iL -Vi H iV
dt

Assume Mvi*0 (negl. vapor holdup)
Tray Hydraulics
(Algebraic) A9.Simplified (linearized):

+λ  Vi9 -Vi-1,0 
M i -M io
Li =Lio +
τL
τL: time constant for change in liquid holdup (≈2-10sek.)
λ: effect of increase in vapor rate on L
Li0,Mi0,Vi0: steady-state values (t=0).
(ALGEBRAIC) Pressure drop: Δpi = f(Mi,Vi,…) 12
 Numerical Solution (Integration)
More details: p.82 in NDF Notes 1990

1.Rigorous Approach
State variables:
n i =μ i x i +μvi yi  N C on each tray 
Moles of
component (note:Sum of n i is give μ i +μ vi )
on tray i
U toti =μ i U i +μ vi U vi (1 on each tray)
total internal energy on tray i

Solution (at given time)

Given value of state variable
Perform constant nU-flash (given internal energy and phase
split (MLi, Mvi), compositions (xi, yi), temperature, pressure
(ρi), spec energy (hi, Vi)
Compute Li and Vi from tray hydraulics and pressure drop
relations

13
2. Simplified Approach, neglect vapor holdup (Mvi = 0)
State Variables:
(NL on each tray)
Mole fraction → Xi (note: ni = Mixi)
total holdup → Mi

Solution:

1. Given value of state variables, guess pressure, Pi

2. Perform bubble point flash (given xi,pi) → yi,Ti,
enthalpies
3. Compute Vi from energy balance (gives “index
problem”: LHS (derivative) is known) Common
simplification: Use dVi 1dt = dhil /dt from
previous step. (Do not set d/dt (MiVi) ≈ 0
constant molar flows
4. Compute Li and Pi from tray hydraulics and
pressure drop.

14
 (1.4) Composition Dynamics

1(when M v  0)

SUM: (NC independent differential equations) xN

1 (when MLconst.no flow dynamic)

No. of components No. of trays

Might expect very high-order complicated behavior.

Surprise: The dominant composition dynamics is

approximately 1 order!

Δy D
(or temp., etc.) Δyo  t  = 1-l-t/t1  Δy o   

100% Δyo   

yD  t 
1-l-1 =63%

time
τ1 15
Fig. Response in YD to step change
τ1: approximately independent of what we step (reflux,
feed rate, boilup,…) and what and where we measure
(YD, Ttop, Tbtm, etc,…)

EXAMPLE: 3-stage column (see p.4B)

 Neglect flow dynamics (Mu = 1 = constant)
 1 state on each tray
 Constant molar flows (V2 = V3)

∆X2(t)

Dist.m: (matlab subroutine)

Feed tray

63% VLE
reboiler
∆X3(t)

condenser
∆X1(t) Step in ZF from 0.50
50 0.51

τ1=4.5 min. Time (min)

16
Fig. Response for 3-stage column to feed composition
change.

Note: Composition change inside column much larger

than at column ends. This is the main reason for the
“slow” composition response

17
 Composition Response of an Individual Tray

xi+1
i+1
Component material balance, tray i
i
Mixi d
 Mi x i  =Li+1x i+1 +Vi-1yi-1 -Li x i -Vi yi
i-1 dt

Assume the column is at steady-state, and consider the

effect of an increase in xifi to xi+1+∆xi+1. Assume flows
constant, and neglect interactions between the trays (yi-
1: constant). In terms of deviation variables

dΔx i
Mi =Li+1Δx i+1 +0-Li Δx i - Vi K i Δx i
dt
Δyi
Where the linearized VLE-constant is
α
Ki   average Ki  1
1   α 1 x 
2
i

18
Collecting ∆Ki terms we get a 1st order response with
time constant
Mi
τx =
Li +K i Vi

Overall response time from top to bottom of column

(neglecting “vapor” interactions) total HOLDUP
Inside Column

MI
θ x  Nτ x   0.5 M I /L
LV

Example: 3-stage Column

M1 M2 M3
θx = + + =0.282+0.171+0.097=0.55 min
L1 +D L 2 +V2 V2 B+k 3V3
condenser reboiler

Conclusion: Do not yet correct overall response time

(4.5 min) by simply adding together individual trays.

19
 1.4.1 τ1 Dominant Time Constant
Objectives:
 Understand why overall response 1st order
 Develop formula for τ1
 When does τ1 not apply?
Assumptions: Use A1-A6
A4.Mvi is negligible (OK when pressure is low; at 10 bar
Mvi will be about 10% of liquid holdup)
A6. Flow dynamics much faster than composition dynamics.
L
xB

LB

τ1 LB
time
θL xB
Response to step change in reflux

)This does not imply that flow dynamics are not

important for composition control!)
20
In particular, assume liquid holdup (Mi) constant
 Column at Something New Steady-State
Steady-State at Happens at t = 0 t= 
t≤0 (not steady-state) (subscript f=final)

D,YD D(t) Df
YD(t) YDf

F1zF Ft ,zFt
Ft ,zFt

Bf
B,xB B(t)
xB(t) xBf

t=0 t >0 t =

Assumption 6:
D(t) = Df
B(t) = Bf

21
 Component balance whole column;

d N 
t  0 :   M i x i (t)   Ff ZFf -Bf x B  t  -Df y D  t 
dt  i=1  imbalance in supply
change in components of component
holdup in column

d 
Subtract:  
dt  i
M i Δx i  t  

=-Bf xB  t  -D f Δy D  t 

t>0:
where Δy D  t  =y D  t  -y D f

22
Assumption
A7. “All trays have some dynamic response”, that is,
Δx i  t  =Δx i  0   k  t  ; k  0  =1 (4)
k(t) is the same for all trays, for example:
Δy D  t  =Δy D  0   k  t 
Δx B  t  =Δx B  0   k  t 

Justification: Large interaction between trays

because of liquid and vapor streams. (Reasonable if

Substitute (4) into (3):

  dk  t 
  M i Δx i  0     dt    Bf Δx B  0   D f Δy D  0   k  t 
 i 
1st order differential equation for k(t).
 dk  t  1 -t τ1 1 
Solution: k  t   l-t τ1
  =- l =- k  t  
 dt τ1 τ1 
where
 M Δx  0 i i
(5)
τ1  i
Bf Δx B  0  +Df Δy D  0  23
Physically
"change in holdup of component (kmol)"
t1 =
"Imbalance in supply of component (kmol/min)"

(5) was first "derived" by Davidson (1956). Rederived Moczek

(1963), Wahl and Harriot (1970), Waller (1969)
SMALL D'sonly.

24
 Comments on τ1-Formula (5)

1. No linearization (change may be large)!

2. More convenient formula for denominator.
Use: t  0 : Fo z F0 =BO x BO +DO yOO
t= : Ff z Ff =Bf x Bf +Df y Df
Subtract: Δ  Fz F  = Bf Δx B  0  +ΔDyOO -ΔBx OO
denominator in

that is Denominator in (5)    Fz F   ΔDyOO  Bx OO (6)

//example : change in feed composition  Δz F ¹0, but ΔF=0, ΔD=0,ΔB=0 
Denominator  Fz F //

3. Only steady-state data needed! (+holdups) Need

steady-state before (t=0) and after (t=∞) upset.

25
4. (5) applies to any given component

5. τ1 may be extremely large if both products pure

(Reason: Numerator>>Denominator because
compositions inside column change a lot, while
product compositions change very little).

6. Limitation: τ1 Does not apply to changes in

INTERNAL FLOWS ONLY, that is, L and V
increase with ∆D=0 and ∆B=0. Reason:
Denominator (τ)=0, (will find τ2<τ1!)q

26
Example (Continued): Three-stage column

Compositions
Flows with ZF=0.50 with ZF=0.51
Stage i Li Vi Xi Yi Xi Yi
Condenser 1 3.05 0.9000 0.9091
Feed Stage 2 4.05 3.55 0.4737 0.900 0.5001 0.9091
Reboiler 3 3.55 0.1000 0.5263 0.1109 0.5549

Estimated Dominant Time Constant

N+1

 M Δxi i 0.091  0.0264  0.0109

τ1 = i=1
Δ  FZF  - y D ΔD-x BΔB
10.01 =0(flows constant)

0.0464
  4.46 min
0.01
Excellent agreement with observed 4.5 minute

Check of Assumption (7):

“All trays have same dynamic response”(because of
tray interactions)
27
 Reasonable if

Nτx θx τ1 Overall response time

(incl. Tray interactions)
Response time neglecting tray
interactions

In our case θx  0.55min 4.5min OK!

In generalθx τ1 for  i large reflux (θ x small)


ii  both products pure 1 large 

28
 Example 2. Propane-propylene splitter

110 theoretical stages

 = 1.12 (relative volatility)
Assume constant molar flows
L/D = 19, D/F = 0.614
M tot 
Mi
= i
=111min  have neglected holdup in reboiler and condenser 
F F
Find τ 1 when ZF decreases from 0.65 to 0.60

All flows kept constant

zF yD xB x tot =
M x i i

M tot

Simulation 1 0.65 0.995 0.100 0.714

(t=0)

Simulation 2 0.60 0.958 0.030 0.495

(t=∞ )

τ1 =
 M Δx i
=iM tot 111min  0.714  0.495
=
Df Δy D +Bf Δx B FΔz F 0.05
 480 min  8hours
29
 Response:

0.05

ΔyD

.63

0 time
0 1h 8h 10h

MI 111
Check of Assumptions 7:θ x    4.6 min 480 min
L+v 11.8  12.3 OK!

30
 τ1: Shortcut Formula
Make some simplifying assumptions which hold best
for columns with
large reflux
pure products
Small changes
Binary separation (or use pseudobinary)

M I1F
Derive τSCI =  contribution from holdup inside column
Is lnS
τSC  τSCI  τSCD  τSCB τSCD =
M D /F
1yD  yD  contribution from holdup in condenser
IS
M B /F
τSCB = 1-x B  x B  contribution from holdup in reboiler
IS
Varies with B D
oper. Cond! IS = x B 1-x B  + y D 1y D  "impurity sum"
F F
y /1-y D
S= D "separation factor "
x B /1-K B

(ln is typically from 4 to 15)

Columns with pure products 

IS is small
 τCI is large (may neglect τCD , τCB ) 31
Example :Propane - Propylene B
=0.386
M
Asssume : I = 111min  insideholdup column 
F
F y D =0.995
MD M
= 60min, B = 20min x B =0.100
F F
For small changes assume (*) applies.
Have
I S =0.386 ×0.1×0.9 +0.614×0.005 ×0.995 =0.035 +0.003 =0.038
0.995/0.005
S= =1791ÞlnS=7.49
0.1/0.9
Then
111
t SCI = min =390 min
0.038 ×7.49
60
t SCD = ×0.005×0.995 min =8 min
0.638
20
t SCB = ×0.1×0.9 min =47 min
0.038
Sum :t =445 min

Reasonable agreement:
NOTE: Contribution small from condenser because
purity is high so absolute changes in compositions are 32
small
 Variation in τ1 with operating point

τ1 Peak is large if both

products are pure

1 xB
1-y D

Typical plot. May be derived from shortcut formula.

Conclusion: Time constant depends on operating

conditions – mainly on purity of least pure product (IS).
-Get “asymmetric” behavior:
a.) Least pure product gets purer: Time constant gets
longer (slow response)
b.) Least pure product gets less pure: Time constant
gets smaller (fast response)
33
Example
ΔZF

“Fast”
getting pore
Pure “slow”
time

34
 (1.4.2) τ2 External and Internal Flows

Steady -state

Steady -state Steady -state

tray

0 COMP1 0 COMP

AB=ΔD AB=ΔD=0

Large effect on Small effect on

composition composition
(large “gain”) (small “gain”)
Effect on composition Effect on composition
obtained by assuming obtained by considering
separation factor change in S:
constant 35
 y D /1-y D y D /1-y D
S= S=
x B /1-x B x B /1-x B

(“separation unchanged (“separation changed

split changed”) split unchanged”)

MAIN EFFECT ON COMPOSITION BY ADJUSTING D/F;

“FINE TUNE” WITH INTERNAL FLOWS

36
 Dynamics
External Flows Internal Flows
ΔxB

63% ΔyD
ΔxB
63% 0
ΔyD

0
τ1 time τ2 time

Step ΔB = -ΔD Step ΔL = -ΔV

Conclusion:
Large S.S. effect Small S.S. Effect
Slow (τ1) Faster (τ2)

37
 Initial Response: Not as Different

Recall:
τ2  MI /F
M I /F
τ1  (see more accurate formulas in
ISlnS
Skogestas & Morari, 1988 I E E
MI /F C Res, 27, 1848-1862)

For columns with pure But: Derived when flow

products dynamics neglected (doubtful
since τ2 is relatively small)

38
 FLOW DYNAMICS

(variations in liquid holdup neglected so far)

A8. Constant molar flows

A4. Neglect vapor holdup   Vi-1 =Vi =Vi +1
Total material balance becomes

i+1
Li+1

Mi
i

Li

dM i
=Li+1 +Vi-1 +Li -Vi (1)
dt 39
 Tray Hydraulics
A9.: Assume simplified linear tray hydraulics

1
Li =Lio +  Mi -Mio  +λ  Vi -Vio  (2)
τL
 ΔM 
τL =  i  = hydraulic time constant
 ΔL V
 Li  = effect on change in Vi on Li (vapor may
λ=   push liquid off tray, λ>0.5: inverse
 Vi Mi
response)

Lio,Mio,Vio: Steady state values

Estimates
2/3
τ L :Francis weir formula M oi =k×Li
2 M oi 1 M i
τL =  (assume Moi  Mi/2)
3 L 3 L
Mi
(Packed:τ L =0.6 )
L
λ:Difficult. trays, λ>0usually
packed, λ  0 40
 Consider Deviations from Initial Steady-State (Δli=Li-Lio,…)

Then
dM i
each tray one tank =ΔLi+1 -ΔLi combine all trays
dt

Consider response in LB to change in L:

ΔL
N N tanks in series, each time constant τL

ΔL

0.5ΔL ΔLB

ΔV t
ΔLB θL=N·τL
(“almost” a dead time)

Laplace
1
ΔLB  S  = ΔL  s 
1 + τ L S 
N

41
Response in LB to change in V:
“Vapor pushes liquid off each tray”

ΔV

λ·ΔV
ΔLB

0
t
θL=N·τL

Laplace
 1 
ΔLB  S  = x 1   ΔV  s 
 1 + τ S N 
 L 

42
 1.6 Overall Dynamics

Composition Dynamics
Typical value
External flows , τ1 250 min
Internal flows , τ2 20 min

+ Liquid Flow Dynamics

“Dead time” from top to bottom, θL 3 minute
Essential for control !

+ Pressure dynamics Depend on tuning of

pressure and level
+ Top level – “ – controls. Typical time
+ BTM level – “ – constant: 0.5 minute
+ Valve dynamics,
0.2 minute
+ Heat transfer dynamics
(V and Vt indirectly controlled)
0.1 minute
+ Measurements Dynamics

43
 Simplified Model
Describe each effect independently
“Add” together to give overall dynamics
(Alternative: Linearize all equations)
NOTE: Exact value of τ1 not important for control!

44
 1.7 Nonlinearity

The dynamic response of distillation column is strongly

nonlinear. However, simple logarithmic transformations
counteract most of the nonlinearity.

dxi
Derivation.component balance: M i =L  x i +1 - x i  +V  y i-1 - y i 
dt
initial response to charge in L :
d
 x i+1 - x i 
Xi+1
M i xi =
dt Xi
dependsstrongly on operating point
"Trick": Divideby K :
dlnKi  x i +1 
Mi = - 1 DL
dt  x i  tray

45
 v
 α (bottom)- nearly indep.of operating point
L
In general:
Light key component on
X tray i
Use: Xi=ln Li
X NI Heavy component

May also be used for temperatures!

Temp. Top of Column

Ti -Ttop
Xi  ln Temp. on Tray i
TBTM -Ti
Temp. BTM. Of Column

More details: p.132-133 in NDF Notes 1991

46
 Initial Response to 10% ∆L: (V constant)
(Column A with Flow Dynamics)

nonlinear
Linear Extremely
model non linear

∆xB

∆yD

Log: Counteracts
nonlinear Nonlinearity

Linear
model -∆(n(1-yD)

∆ ln xB

47
1.8 Linearization
Need linear models for controller design
Obtain by

1) Put together simple models of individual

effects (previous page)
2) Linearize non-linear model

Given Tray

d α xi
 Mixi  =Li+1xi+1 +Vi-1yi-1  Li xi -Vi yi ; yi =
dt 1+  α-1 x i

Linearize, introduce deviation variables, simplification

here: assume: i) const., ii) const. molar flows
 Li = Li+1 = L
Vi-1= Vi = V

48
 Δx i
μi =Li1  Li +k i Vi  Δx i +Vi-1k i-1 +  x i+1 -x i  ΔL-  yi -yi-1  ΔV
dt
α dyi
Ki = =
1+  α=1 xi 
2
dKi

   
   
μi  1 dx i 1      
 dt =  L i  L i 1  K i 1Vi  K i Vi  x i 1  x  x  y i 1  y i   L 
 dx    +  i2 i 1
 L i  K i Vi   L 
μi i
  L K i 1Vi 1  x i   x i 1  x i y i  y i1    
 dt      
     
 
 

d Δx
Mi  A Δx + B ΔL
dt ΔV
inputs

State matrix Output matrix

(eigen values determine “states” (tray
speed of response) compositions)

Can derive transfer matrix G(12)

+ Equations for dMi/dt=…… 49
 (2) Distillation Control

Multivariable vs. Single-Loop Control…25

Choice of Control Configurations…27
LV
DV, L
L υ
DB
DB

1) Disturbance Sensitivity…………...31
2) Feedback Control Properties…….34
One-Point
Two-Point
3) Implementation, Level Control…..35

50
 Distillation Control

Manipulated inputs, μ1s: L,V,D,B,UT

Controlled outputs, Y1s: Mo, Mb, Mv
(P) ,YD, xB

L MD
V MB
D
COLUMN MV
YD
B
UT XB

51
5X5
 Controller Design

“Full” 5x5 multivariable controller? NO!

Instead use simpler scheme (decentralized control)

i) 3 single loops (PI, PID) for Md, MB, MV (levels and

pressure)
ii) There are now left two degrees of freedom for quality
controller (keep YD and xB at desired values).

Design as two single loops or 2x2 controller (e.g. decoupler)

PROBLEM 1: Choice of Control Configuration

(Structure). Which two degrees of freedom
should be left for equality control? (Same as
choice inputs for level control )
PROBLEM 2: Design of Controllers
i) Level controllers
ii) Composition controllers remaining 2x2
system
52
 Options Composition Control

a) “No control” that is, manual operation (e.g. reflux L and

boilup V are set manually by the operator
b) “One-point” control: One composition controlled
automatically. (e.g. yD controlled with L using PI-controller,
the other input is n manual, xB “floats”). Most common in
industry, often necessary because of constraints.
c) “Two-point” control: Both compositions unclear feedback
control.

$$! Potential for large economic savings.

(larger throughput, more products, less energy)

53
 Possible Controller (2x2)

(1) Single Loops  e.g.L  y 

 
D

 V  xB 

Problem: Interactions (performance)

Advantage: Robustness

(2) Mi variable, for example, decoupler

Problem: Often not robust
(Sensitive to errors)

54
 Problem 1.

1) LV-configuration (“conventional”, “energy balance”, “indirect

material-balance”)
Reflux L and boilup V used for composition control
Level control:
MD  D
MB  B
MV (P)  VT (cooling)
2) DV-configuration (“(direct) material balance”)
because D is used
Level control:
MD  L
MB  B
3) LB-configuration (“(direct) material balance”)
Level control:
MD  D
MB  V
4) L/D V/B-configuration (“double ratio”)
Level control:
MD  both L and D (with L/D constant)
MB  both V and B (with V/B constant) 55
 Top of Column

VT
LS LC LV-configuration
L+D D LB-configuration
L

VT
LC DS

L+D D DV-configuration
DB-configuration
L

56
 LV
VT - configuration
LC
DB
C

L+D D

L
DIV
(40)m Set manually or from
(L/D)s composition controller

Comment: Will usually use cascade control on inputs

for composition control LS
Lm P
LS

L L

No cascade (manipulated With cascade using flow measurements

valve position) (remove nonlinear valve characteristics

57
Open-loop response to flow
disturbance:
D  FV +V- L

58
D  O  FV +O- V

59
DB(!!) (Finco & Luyben, 1989)

ΔD = 0× ΔFV + 0× ΔV
(but = L =   FV  V)

BAD

Previously rejected from steady-

state considerations (Perry, 1973;
Shinskey, 1984; Skogestad &
Morari, 1987; Haggblom & Waller,
1988).
Works because mass may be
accumulated dynamically (change
in liquid level)

60
 D 1
D =  ΔFV +  ΔV
F L V
1+ +
D B

61
DIFFERENCE BETWEEN CONFIGURATIONS
1) Disturbance sensitivity (“self regulating” properties)
2) Interactions between loops, etc. (feedback control
properties)
3) Implementation level control

1. DISTURBANCE SENSITIVITY

Fact: Composition are mainly dependent on D/F

(external flows) should have D/F  ZF

High-purity columns: Composition extremely

sensitive to small changes in D/F:

dyO dx 1  D
x B  ×  dZF -d 
1-yO  x B BXB +0 1-yO   F
relative change in composition small for high-purity columns

Consequently: Disturbances which change D/F are “bad”

The effect of a given disturbance on D/F depends on the
configuration 62
 EXAMPLE
Configuration Disturbance

ΔFV ΔV
(increase in flow rate of (increase in boilup)
vapor in feed)

D DD  O
Optimal DD  DFV
F

LV DD =DFV ΔD = ΔV
DV DD =0 ΔD = 0
DB ΔD = 0
DD =0
LV 1
D ΔD = ΔV
DD = DFV L V
DB F 1+ +
D B

(see figures p. 29-30)

More detailed: See Table 3 in Skogestad, “Disturbance

rejection in distillation columns” CHEMDATA’88, Goteberg,
June 1988. LP. 65 Literature for 1989 NDF) or p. 40. 63
 1. Summary Disturbance Sensitivity

Disturbance in Disturbance in Feed forward

from Fm
V, L, qF F
Feed
boilup enthalpy
reflux

Poor Poor Difficult

LV
DV
Good Poor Easy  D F 
LB Good Poor Easy  B F 
LV Good Good Not needed
DB
Initially :Good Poor   D B
DB   Easy  
t   :Good Terrible L,V      F1 F 

Self-regulation ability

Note: The above analysis was based on open-loop. A

more careful analysis should take dynamics and
feedback (disturbance direction into account. (See
NDF Notes 1991, p. 205).
64
 (2) Feedback Control Properties

(2a) One-Point Control

(often necessary because of constraints)

Configuration Rating
LV Best

DV, LB Dangerous (if D or B fixed

L/D V/B Reasonable

DB Unworkable (material balance

locked)
(2a) Two-Point Control
* Use RGA as a function of frequency as a tool, want λRGA
(wB)≈1. See Facohsen et.al. (AIChEF, 1990 May)

65
 2x2 Controller (“Problem 2”)

Configuration Single loops (yD, Decoupler +

xB) (PI or PID) single loops

LV ←strong Poor if measure delay Hopeless (NO!)

interaction Bottom pure: OK; Top
pure: POOR GOOD, (but sensitive
DV to operating point)
Bottom pure: POOR; Top
LB pure: OK
Not worth it
L/D V/B Good (except very high
purity0
DB
Good Not worth it

66
 Implementation and Level Control

Implementation Level Control

(Constraints)

LV Easy Usually good, but

not for high reflux(*)
DV Easy OK
LB Easy Level btm: V
sometimes inverse
response
Level top: Poor high
reflux(*)
L/D V/B Difficult (need to measure Good
L,D,V,B!)
DB Easy OK, but level btm: V
may give inverse
response

67
 VT (large)

LC Almost impossible to
control level with small
flow (“no power”)

D (small)
LS (large)

 L LV 
 Have not mentioned V ("Ryskump") : Usually between and LV 
 D DB 

68
 Effect of Level Control tuning on
Composition Response

LV-configuration: Response time for level control has

almost no effect on composition response

LV
DV, DB,  Configuration, etc.: (as implemented on
D B p. 29-30): Need fast level control to
avoid undesired lags in composition
response.

e.g. DV – or DB – configuration (top)

Changing this only affects

Level controller compositions indirectly
sets L through change in L

LC DS

L
69
 Better solution if we do not want to have fast
level control:

Level controller Changing DS directly gives

sets L opposite change in L
(since Ltd is constant)

LC DS
FC

L
L D

L+D

LV
Conclusion (applies to DV, DB, ,-configurations, etc.)
DB
Top: Level controller always sets L+D
Bottom: Level controller always sets V+B
This avoids dependency on tuning of level loops.
70
3. Some Basic Control Theory

3.1 Transfer Functions | Matrices…………38

1. SISO Laplace, frequency
response
2. MIMO

3.2 Effect of Feedback Control ……………43

The sensitivity function
1
S
$ as a performance measure
1  gc

3.3 Analysis of Multivariable Processes…..45

Interactions
Relative Gain Array (RGA)

71
 Transfer Functions/Matrices
1. SISO (Single Input Single Output)
Steady-state

“Increasing L from 1.0 to 1.1 changes yD from 0.95

to 0.97” (for example, run Process)

yD = g o  L
0.97 - 0.95
stead- stategain = =0.2
1.1- 1.0
Dynamics
“The response has a dead time of 2 minutes and then
rises with a time constant of 50 minutes”

yD
0.97
Laplace: Δy D  s  = g  s  ΔL  s  63%

0.96

Transfer function:
0.2 -θs  =2min
g s  =
0.95
time
1+τs  =50min Θ=2 10 50

τ=50
72
 Alternative Method of Obtaining Transfer Function:
d
Take laplace of linearized model s
dt
Example
L (constant)
Material balance
Ki+1
d ln out
 inventory   -
dt sek sek
L d
xi  M i x i  =Lx i+1 -Lx i
Mi=constant dt
(Linearize) + Deviation variables
M i  s  x i  s   Lxi+1  Lx i  s 
Solve for x i
x i+1 M capacity
x i  , τ i =
1+τ  s L flow
1
Effect of x i+1 on x i :g  s   “gain”=1
1 τ s

Time constant
τ=Mi/L

Note: Gain is obtained by using s = o: “gain” = g(o) 73

Alternative Method of Obtaining Transfer Function:

d
Take laplace of linearized model s
dt
Example
L (constant)

Ki+1

L
xi
Mi=constant

Response:

Xi+1

Xi

τ time

74
 Material balance
d ln out
 inventory   -
dt sek sek
d
 M i x i  =Lx i+1 -Lx i
dt
(Linearize) + Deviation variables
M i  s  x i  s   Lxi+1  Lx i  s 
Solve for x i
x i+1 M capacity
x i  , τ i =
1+τ  s L flow

1
Effect of x i+1 on x i :g  s  
1 τ s “gain”=1

Time constant
τ=Mi/L

Note: Gain is obtained by using s = o: “gain” = g(o)

75
 2. MIMO (multivariable case)

Direct Generalization:
“Increasing L from 1.0 to 1.1 changes yD
from 0.95 to 0.97, and xB from 0.02 to 0.03”
“Increasing V from 1.5 to 1.6 changes yD
from 0.95 to 0.94, and xB from 0.02 to 0.01”

Steady-State Gain Matrix

 ΔYD   ΔL 
   G0  
 Δx B   ΔV 
 0.97  0.95 0.94  0.95 
g g12 0   1.1  1.0 1.6  1.5  0.2  0.1
G0    11    
g 21 g 22 0   0.03  0.02 0.01  0.02  0.1  0.1 
 1.1  1.0 1.6  1.5 
Effect of input 1ΔLon output 2 Δx B 

Can also include dynamics :

 
 0. 2 0.1  y D (Time constant 50 min for yD)

G s    1  50s 1  50s 
 0.1 0.1  x B
   (time constant 40 min for xB)
 1  40s 1  40s 
76
 Important Advantages With Transfer
Matrices:

1. G(s) is independent of input μ!

For given u(s) compute output y(s) as ys   Gs   μs 

Can therefore make block diagrams

μ(s) y(s)
G(s)

2. Frequency Response:
G(s), with s=jw (pure complex no.) gives directly
steady-state response to input sinwt!

μ y1
g11(s)
Let M1  t   A sin wt  y1  t   g11  jw   A  sin  wt  Lg11  jw  

77
 Response (as things settle)

A. |g11(jw)|
A y1(t)

time
μ1(t)

Lg11  jw 
Δt = -
w

78
 y1 Tells directly how much a
Note  g11  jw 
μ1 sine of frequency w is
amplified by process

2
Typical : g11(s) 
1  50 s

|g11(jw)|
(log scale)|
|g11|
2

1 w (log scale)
0.01 0.02 0.1
=1/50

0.1

Fast sinusoids are “filtered by

process (don’t come through

BODE PLOT (magnitude only) 79

 3.2 Effect of Feedback Control

d
ys μ
C(s) G(s) y

G(s): process (distillation column)

C(s): controller (multivariable or single-loop PI’s)
y: output , ys: setpoint for output
μ: input
d: effect of disturbance on output

80
Process : y  s   G  s   μ  s   d  s  (1)
Controller : μ  s   C  s   y  s  y  s
s (2)

Negative feedback

No Control μ s   0
y  s  d  s  No disturbance supression

With feedback control

Eliminate μ(s) in (1) and (2), ys=0

ys   I  Gs C(s) ds 

1

S(S)=“sensitivity function” = surpression disturbances

81
 y s     s   d  s 
Ideal:-Want  s   0
In practice:
Cannot do anything with fast changes, that is, ς(jw)= I at high frequency
S is often used as performance measure

 y1  S11 (s) S12 (s)  d1 (s) 

 y  S (s) . .  d (s) 
 2    21  2 
    
    
     
One Direction Small resonance peak
(small peak = large GM
S 21 jw  and PM)
Log- τ
scale

1 w (log)
wB
Bandwidth (Feedback no
help)
0.1

82
Slow disturbance d1: 90% effect on y2 removed
 “Summing up of Channels”

“maximum singular value” (worst

direction)

1 w (log)
wB

w (s)
“minimum singular value” (best
0.1
direction)

83
 3.3 Analysis of Multivariable processes
(Distillation Columns)

What is different with MIMO processes to SISO:

The concept of “directions” (components in u
and y have different magnitude”
Interaction between loops when single-loop
control is used

INTERACTIONS

Process Model

y1
μ1 g11
" Open  loop "
g12
y1  s   g11  s   g12  s  μ 2  s 
g21 y2  s   g 21  s   g 22  s  μ 2  s 
y2
G μ2 g22

84
Consider Effect of μ, on y1

1) “Open-loop” (C2 = 0): y1 = g11(s)·μ1

2) Closed-loop” (close loop 2, C2≠0)

 g g C 
y1   g11 s  12 21 2 μ1
 1  g22  C 2 
Change cause by
“interactions”

85
 Limiting Case C2→∞ (perfect control of y2)

 g g 
y1   g11 s  12 21 μ1
 g22 

How much has “gain” from u1 to y1 changed by

closing loop 2 with perfect control

y1/μ1 0L  g11 1 def

Relative Gain     λ RGA
y1/μ1  g11 
g12 g21
1
g12 g21 11

g22 g11 g22

The relative Gain Array (RGA) is formed by

considering all the relative gains

 y 1/u 1 OL y 1/μ2 OL 

 λ 11 λ 12   y 1/u 1 CL y 1/μ1 CL 
RGA  Λ   
 λ 21 λ 22   y 2 μ1 OL y 2 /μ2 OL 
 
 y 2 μ1 CL y 2 /μ2 CL 

Example from before

 0.2  0.1 1
G , λ 11  2
0.1  0.1  0.1 0.1
1
0.2 0.1
 
0.5

 2  1 86
RGA  
 1 2
Property of RGA:
Columns and rows always sum to 1
RGA independent of scaling (units) for μ and y.

Note: RGA as a function of frequency tell how

relative gain changes with frequency.

Sigurd’s Law!

87
 Use of RGA:

(1) Interactions
From derivation: Interactions are small if
relative gains are close to 1

Choose pairings corresponding to RGA

elements close to 1

Traditional: Consider Steady-state

Better: Consider frequency corresponding to closed-
loop time constant

But: Never choose pairing with negative steady-state

relative gain

88
Example:
0.2  0.1  y1 = 0.2  u1 -0.1  u1
G o   
0.1  0.1 y 2 = 0.1u1 -0.1μ1
2 - 1 
RGA =  
-1 2 
Only acceptablepairings :
μ1  y 1 With integral action :
μ2  y 2 Negative RGA  individual
Will not work : loop unstable or overall system unstable
μ1  y 2
μ 2  y1

89
 (2) Sensitivity measure

But RGA is not only an interaction measure:

Large RGA-elements signifies a process that is

very sensitive to small changes (errors) and
therefore fundamentally difficult to control

example Large (BAD!)

1 1  91  90 
G=   RGA   
0.9 0.91  90 91

1
  1.1%
90
1
Relativechange- makes matrix singular!
x 21
 1 
Then gˆ 12  g12 1    0.91
0.9
 90 

90
Singular Matrix: Cannot take inverse, that is,
decoupler hopeless.

Control difficult

91
92