You are on page 1of 133

CHEN90032

Process Dynamics and Control

Worksheet 1 – Automatic Control

The feedback controller is the cornerstone of automatic control within the


process industries. A feedback controller takes a measured value for some
important process variable (the controlled variable, c) and continually
compares it to the desired value (the set-point, r) to generate an error signal,
e. Some control algorithm within the controller then operates on the error
signal to arrive at a new value for the manipulated variable m that will bring
the controlled variable back to the desired value in a fast but stable way.

In order to completely define a control system we need to specify which


controlled variables are paired to which manipulated variables (the control
scheme), as well as details of the control algorithm. This worksheet presents
several processes for which we will develop control schemes based upon our
‘chemical engineering intuition’. The majority of this course will be devoted to
understanding how process variables change in time (process dynamics), and
how we can use that information to develop effective control algorithms within
a particular control scheme. Towards the end of the subject we will also see
how process dynamics can guide our pairing of controlled and manipulated
variables to arrive at the optimal control scheme.

W1-1
1) Develop a control scheme for the evaporator depicted below, which has
three controlled and manipulated variables, giving six possible configurations.
Identify (and label on the diagram) the pairs of controlled and manipulated
variables. Explain how you expect the controlled variables to interact with
each of the three manipulated variables.

Controlled variable 1 (c1):


Manipulated variable 1 (m1):

Controlled variable 2 (c2):


Manipulated variable 2 (m2):

Controlled variable 3 (c3):


Manipulated variable 3 (m3):

Vapour, Fv

P
Feed L

Steam, Fs
T

Liquid, Fl

W1-2
2) An absorber-stripper is used to remove CO2 from a mixed stream of H2 and
CO2, as shown below. Design a system to control the temperature, pressure,
and liquid levels in the absorber and stripper units.

CO2

H2

Cooling
Absorber Stripper
Water
Feed Steam

Solvent (MEA)

Cooling
Water

W1-3
3) The chemical plant depicted here converts reactants A and B to the products C and D in an exothermic reaction, then the
products are separated in a binary distillation column. Develop a control scheme including valves and controllers.


A process flow diagram for this chemical plant is provided online as a Microsoft Visio file. Use this file in developing a complete
control scheme.

Cooling
Feed Tank A Water

Distillation
Feed Heater Column Reflux Drum

Feed Tank B

Product D

Cooling Jacketed
Water Reactor
Reboiler Steam

Cooling
Product Cooler
Water

Product C

W1-4
CHEN90032
Process Dynamics and Control

Worksheet 1 – Automatic Control – SOLUTION

Question 1

Vapour

PC TC
LC

Feed

Steam

Liquid
Controlled variable 1: Liquid level in tank / volume (L)
Manipulated variable 1: Liquid flow rate (Fl)

Controlled variable 2: Temperature (T)


Manipulated variable 2: Steam flow rate (Fs)

Controlled variable 3: Pressure (P)


Manipulated variable 3: Vapour flow rate (Fv)

The steam flow rate has a significant effect on the manipulated variables
temperature and pressure. Increased steam flow directly increases temperature,
producing more vapour and therefore increasing pressure. There will be a small
effect on the liquid level, as liquid will be lost to evaporation. The liquid and
vapour flow rates will have relatively little impact on the system outside of their
controlled variables (liquid level and pressure, respectively). These manipulated
variables can therefore be tightly tuned for a fast response, with the steam flow
rate tuned loosely.

1
Question 2

Syngas operations convert hydrocarbons into H2 + CO2. The valuable H2 can


then be separated from the waste CO2 by absorption into a solvent. The CO2 is
released in a stripper unit, so that it can be sequestered.

The absorber unit operates by passing the feed gas stream through the solvent
MEA at relatively mild temperature and pressure conditions. The CO2 is
absorbed into the solvent, leaving H2 to be captured in the gas phase. The
loaded solvent is fed to the stripper, where it is heated to release CO2. The
solvent is recycled, and blended with a make-up stream. The solvent is cooled
before being returned to the absorber.

The manipulated variables are identified as the H2 and CO2 flow rates from the
absorber and the stripper, the steam flow to the stripper, the flow of water cooling
the solvent stream, the flow rate of solvent from the absorber, and the flow rate of
the solvent make-up stream. This results in six manipulated variables.

We need to control the pressures, temperatures, and liquid levels in the absorber
and stripper units (this gives us our six controlled variables). In both units the
pressure is controlled by the flow rate of the gas stream (H2 in the absorber, CO2
in the stripper). The absorber liquid level is controlled by the flow rate of liquid
out. Temperature in the stripper is maintained by the flow rate of heating steam.
In the absorber, temperature is maintained by the flow rate of water used to cool
the solvent feed to the absorber.

In the stripper unit, the liquid level is controlled by the solvent make-up stream.
This feedback loop is rarely used, and in some plants would be monitored
manually. Neglecting solvent degradation and vaporization, the total solvent
volume in the system is conserved. Given correct starting levels, and with the
absorber solvent level on negative feedback control, the solvent level in the
stripper will be correct. Occasional addition of MEA to replace lost solvent is
required.

2
CO2

H2 Cooling
Water
PC Absorber PC Stripper
TC
Feed TC

Steam
LC
LC

Solvent (MEA)

Cooling
Water

3
Question 3

Flow rates of A and B into the reactor are set equal to the design values. The
reactor temperature is maintained at the desired level by the flow of cooling
water; because the reaction is exothermic, heat needs to be removed. Vapour
produced in the reactor is vented to avoid pressure build-up. The liquid level in
the reactor is controlled by the flow of product (C + D) to the feed heater. This
product provides the feed for the distillation column.

Control of the distillation column is essentially as described in the notes to Topic


1. The distillate (product D) quality is controlled by the reflux rate, with the liquid
level in the reflux drum controlled by the distillate flow rate. The bottom product
(product C) quality is controlled by the flow of steam to the reboiler, and the
reboiler liquid level is controlled by the bottom product flow rate. The column
pressure is maintained by adjusting the flow of cooling water through the
condenser.

The feed to the distillation column is pre-heated in a heat exchanger,


simultaneously cooling the bottom product from the column. The feed
temperature is controlled by manipulating the amount of bottom product (i.e.,
product C) that bypasses the heat exchanger.

A relatively complete P&ID for this small chemical plant, based upon the above
control scheme, is shown on the following page.

4
PC
FC Cooling
AC Water
Feed Tank A

Reflux Drum LC
Distillation
FC Feed Heater Column

Feed Tank B LC
TC Product D

Reboiler
Cooling Water
Jacketed
Reactor LC Steam
TC

AC
Product Cooler

Cooling Water

TC
Product C

5
CHEN90032
Process Dynamics and Control

Worksheet 2 – Unsteady-State Models

One of the earliest techniques any chemical engineer learns is how to perform
steady-state material and energy balances. These steady-state balances are
one of the key tools in designing a process.

In order to design a complete control system we need to understand how the


important process variables change in time – in other words, we require a
dynamic model of the system. This dynamic model is often arrived at by
performing unsteady-state material and energy balances. Unsteady-state
models are more complex to develop and to solve than their steady-state
counterparts, as they constitute a system of differential equations; often non-
linear PDEs. Reasonable models can typically be obtained by approximating
the system as a set of linear ODEs that can be solved analytically or
numerically.

The problems in this worksheet illustrate unsteady-state material and/or


energy balances around some relatively simple processes involving mixing,
heating and chemical reaction. Each of these models constitutes a linear ODE
or set thereof. We will also consider the degrees of freedom in the system to
see if a solution can be obtained, and will attempt a numerical solution for one
case. In the final problem we will use the model to guide our thinking about
designing the control scheme, i.e., assigning variables as controlled or
manipulated and pairing them up appropriately.

W2-1
1) An enclosed stirred tank is used to heat a water stream flowing at F kg/min
from temperature Ti to T. Heat is supplied by a heating coil at a constant rate,
Qheater. Heat is lost to the surrounding ambient air at rate Qloss.

(a) Develop a dynamic model describing the temperature in the tank if the
incoming and ambient temperatures can both vary.

(b) Perform a degrees of freedom analysis to identify which variables need to


be assigned to solve for T(t).

Assume: The water density (ρ) and heat capacity (Cp) are constant. The inlet
temperature (Ti) is always larger than the ambient temperature (Ta). Heat loss
to the surroundings is described by Qloss = hAs∆T, where h is the overall heat
transfer coefficient (constant) and As is the tank surface area.

W2-2
2) A continuously stirred tank reactor is used to convert component A to B in
an irreversible first-order chemical reaction. A liquid stream containing A at
mass fraction xA,i = 0.02 in water enters the reactor with mass flow rate F = 50
kg/min.

(a) Develop a dynamic model to describe the mass fractions of A and B as a


function of time.

(b) Conduct a unit balance to demonstrate unit consistency for each of the
terms in the dynamic model.

The reactor is of constant volume V = 3 m3, with liquid of constant density


1000 kg/m3. The reaction A → B proceeds with rate constant k = 1×10-3 min-1.
The process is isothermal.


Solve the dynamic model numerically by adapting the MATLAB code for the
mixing tank without chemical reaction. Investigate the effect of k on the mass
fraction of A and B. Assume that the tank initially contains only water (i.e., xA =
xB = 0).

3) A stirred tank of variable volume is used to heat a process stream.

(a) Develop a dynamic model to describe both the volume (V) and
temperature (T) in the heated stirred-tank as a function of time. Liquid enters
the tank at flowrate Fi and leaves at flowrate F. Heat is supplied at rate Q.
Assume constant density and heat capacity.

(b) Perform a degrees of freedom analysis to determine if the model is


soluble. Design a control scheme to regulate the tank volume and
temperature, assigning variables as disturbance and manipulated variables.

W2-3
CHEN90032
Process Dynamics and Control

Worksheet 2 – Unsteady-State Models – SOLUTION

Question 1

(a) We begin by performing an energy balance:

dU
dt
( )
= −∆ FHˆ − Qloss + Qheater (1.1)

Where:

( )
−∆ FHˆ = FCP (Ti − T ) (1.2)

and

Qloss = hAs (T − Ta ) (1.3)

and

dU dT
= ρVCP (1.4)
dt dt

Substituting (1.2), (1.3), and (1.4) into (1.1) we get:

dT
ρVCP = FCP (Ti − T ) − hAs (T − Ta ) + Qheater (1.5)
dt

(b) NE = 1 (1.5). NV = 4 (T, Ti, Ta, F). Therefore, the number of degrees of
freedom (NF) is 3.

In order to solve the model for T(t) we need to specify the variables Ti, Ta, and F,
as a function of time.

1
Question 2

(a) We need to perform component balances for A and B.

d ( ρVx A )
= Fi x A,i − Fx A − k ρVx A (2.1)
dt

d ( ρVxB )
= k ρVx A − FxB (2.2)
dt

The change in mass of A per time due to the chemical reaction is the rate
constant (k) times the total mass of A (xAρV). The term due to chemical reaction
is negative in (2.1), as A is being consumed, and positive in (2.2), where B is
being produced.

(b) We will examine the units for each term in equation (2.1).

d ( ρVx A )
Term 1:
dt

ρ: kg/m3

V: m 3

xA: kg of A/kg

t: min

Therefore, this term is in units of kg of A/min

Term 2: Fi x A,i

Fi: kg/min

Given the units for xA listed above, we again find units of kg of A/min (the same
applies for Term 3).

Term 4: k ρVx A

k: min-1

2
Other units are as above. Therefore, we again have units of kg of A/min, and the
units are consistent.


The modified MATLAB code is available on
on-line.

Performing the simulations at the specified conditions produces the following


composition vs. time profiles:

This simulation essentially corresponds to process start-up (zero initial


concentration), and we find that the system takes several hours to reach steady
steady-
state. Because the reaction rate is slow relative to the residence time, conversion
of A to B is low.

As k is increased the conversion of A to B also increases, until xB approaches


0.02, the inlet mass fraction of A (i.e.,
( complete conversion):

3
Question 3

(a) First, perform a mass balance:

d ( ρV )
= Fi − F (3.1)
dt

dV
ρ = Fi − F (3.2)
dt

dV 1
= ( Fi − F ) (3.3)
dt ρ

Now we perform an energy balance:

d ( ρVCPT )
= FC
i PTi − FCPT + Q (3.4)
dt

This energy balance has been obtained using some of the same relationships as
in question 1. As in this earlier problem the density (ρ) and heat capacity (CP) are
assumed constant, but here volume (V) is not. So:

d (VT )
ρCP = FC
i PTi − FCPT + Q (3.5)
dt

d (VT ) FT
i i FT Q
= − + (3.6)
dt ρ ρ ρCP

Applying the chain rule:

d (VT ) d (T ) d (V )
=V +T (3.7)
dt dt dt

Therefore:

dT dV FT FT Q
V +T = i i − + (3.8)
dt dt ρ ρ ρCP

We already have a relationship for dV/dt from (3.3), and introducing this into (3.8)
gives:

4
dT 1 FT FT Q
V dt + T[ρ (F − F)] = ρ
− ρ
+ ρC (3.9)
p

dT 1 Q
V = (F T − FT − F T + FT) + (3.10)
dt ρ ρC p

dT 1 Q
V dt = ρ (F T − F T) + ρC (3.11)
p

dT 1 Q
V = (F T − F T) + (3.12)
dt ρ ρC p

dT F Q
dt
= ρV (T − T) + ρC (3.13)
p

(b) Degrees of freedom analysis: NE = 2 (3.3 and 3.12), NV = 6 (V, T, Fi, F, Q, Ti).
Therefore we have 4 degrees of freedom, and need to specify Q, Fi, F Ti.

In a control system the heating rate (Q) and the outlet flowrate (F) are likely to be
manipulated variables that we use to control T and V. The inlet temperature and
flowrate would be disturbance variables. An example control scheme is illustrated
below:

TC
LC

heating fluid

5
CHEN90032
Process Dynamics and Control

Worksheet 3 – Laplace Space

Laplace transforms are, in some respects, the language of control theory.


Laplace transforms provide a useful tool for analytically solving ordinary
differential equations, which are the basis for our dynamic process models.
The approach used to solve differential equations with Laplace transforms is
to transform our function from the time (t) domain to the frequency (s) domain
(Laplace space), rearrange the new function of s, then transform back to the
time domain.

Another reason that Laplace transforms are so widely-used in control is that


they transform problems of calculus into problems of algebra. The ability to
add, subtract, multiply and divide functions that correspond to differential
equations in the time domain is powerful, and we will find that it greatly
simplifies many complex control problems.

This worksheet contains several problems involving Laplace transforms.


Problems 1 and 2 involve converting to and from the time domain. The
remaining problems illustrate how we might solve a dynamic model for some
given input function. Problem 4 also demonstrates how the Laplace transform
of a differential equation can provide insight into the form of the time-domain
solution, even without completely solving the equation, which in itself is useful
in designing a control system.

W3-1
1) Evaluate L{f(t)} for:

(a) f(t) = sin2t

(b) f(t) = 3t – 5sin2t

s(s + 1)
2) Find x(t) = L -1{X(s)}, where X (s ) =
(s + 2)(s + 3)(s + 4)

dy 1
3) Using Laplace transforms solve the differential equation = (u − y )
dt τ
where u undergoes a step change from 0 to A at t = 0. y(0) = 0 and τ is a
constant.


Obtain a numerical solution to this differential equation in Excel using Euler’s
method and the Runge-Kutta method when τ = 5 and A = 2.

W3-2
4) Without solving for x(t), use Laplace transforms to identify the functions of t
in the following differential equations (e.g. t2, e-t, etc).

d 3x d 2x dx
a) 3
+ 2 2
+2 +x =3
dt dt dt

d 2x
b) 2
− x = 2e t
dt

d 3x
c) + x = sin t
dt 3

d 2 x dx
d) + =4
dt 2 dt

Identify the above problems where X(s) has complex factors. How will the
oscillatory term(s) in x(t) behave?

5) Using Laplace transforms, find x(t) for the following functions:

dx
a) − 12 x = sin 3t , x(0) = 0
dt

d 2x dx dx (0)
b) 2
+6 + 25 x = e −t , x(0) = 0, =0
dt dt dt

W3-3
CHEN90032
Process Dynamics and Control

Worksheet 3 – Laplace Space – SOLUTION

Question 1

(a) Evaluate L{sin2t}

ω
From the table of Laplace transforms L{sinωt} = . In our case we have ω =
s + ω2
2

2.

2
L {sin 2t } = (1a.1)
s + 22
2

2
L {sin 2t } = 2
(1a.2)
s +4

(b) Evaluate L{3t – 5sin2t}

By the linearity property of the Laplace transform:

L {3t − 5 sin 2t } = 3L {t } − 5L {sin 2t } (1b.1)

2
From (a) we know that L{sinωt} = 2
, so:
s +4

 2 
L {3t − 5 sin 2t } = 3 L {t } − 5 2  (1b.2)
s + 4

10
L {3t − 5 sin 2t } = 3 L {t } − 2
(1b.3)
s +4

1
From the Laplace transform table L{t} = , so:
s2
 1 10
L {3t − 5 sin 2t } = 3 2  − 2 (1b.4)
s  s + 4

3 10
L {3t − 5 sin 2t } = 2
− 2 (1b.5)
s s +4

1
L {3t − 5 sin 2t } =
( )
3 s 2 + 4 − 10 s 2 ( ) (1b.6)
( )(
s2 s2 + 4 )
3s 2 + 12 − 10s 2
L {3t − 5 sin 2t } = (1b.7)
( )(
s2 s2 + 4 )
12 − 7s 2
L {3t − 5 sin 2t } = (1b.8)
( )(
s2 s2 + 4 )

Question 2

x(t ) = L −1{X (s )} (2.1)

 s(s + 1) 
x(t ) = L −1   (2.2)
 (s + 2)(s + 3)(s + 4) 

Performing a partial fraction expansion:

s(s + 1) A B C
= + + (2.3)
(s + 2)(s + 3)(s + 4) s + 2 s + 3 s + 4

s(s + 1)
=
(s + 2)(s + 3)(s + 4)
(2.4)
A(s + 3)(s + 4) + B(s + 2)(s + 4) + C(s + 2)(s + 3)
(s + 2)(s + 3)(s + 4)

s(s + 1) = A(s + 3 )(s + 4) + B(s + 2)(s + 4) + C (s + 2)(s + 3) (2.5)

In the polynomial on the right hand side of (2.5) zeros occur for s = -2, -3, and-4.

At s = -2:

s(s + 1) = A(s + 3)(s + 4) (2.6)

− 2( −1) = A(1)(2) (2.7)

A =1 (2.8)

2
At s = -3:

s(s + 1) = B(s + 2)(s + 4) (2.9)

− 3( −2) = B( −1)(1) (2.10)

B = −6 (2.11)

At s = -4:

s(s + 1) = C(s + 2)(s + 3) (2.12)

− 4( −3) = C ( −2)( −1) (2.13)

C=6 (2.14)

Substituting A, B, and C back into (2.3) we get:

1 6 6
X (s ) = − + (2.15)
s+2 s +3 s+4

 1 
From the Laplace property table we see that L −1  −bt
 = e , so:
 s + b 

x(t ) = e −2t − 6e −3t + 6e −4t (2.16)

Question 3

We have the differential equation:

dy 1
= (u − y ) (3.1)
dt τ

Taking the Laplace transform of both sides:

 dy  1 
L   = L  (u − y ) (3.2)
 dt  τ 

3
 dy  1 1
L   = L {u} − L {y } (3.3)
 dt  τ τ

The input (u) undergoes a step change from 0 to A at t = 0, where the Laplace
transform of the step function u is A/s. Each of the terms in (3.3) become:

 dy 
L   = sY (s ) − y (0) (3.4)
 dt 

1 A
L {u} = (3.5)
τ τs

1 1
L {y } = Y (s ) (3.6)
τ τ

Substituting (3.4) to (3.6) into (3.3) (with y(0) = 0) gives:

A Y (s )
sY (s ) = − (3.7)
τs τ

Rearranging (3.7) to make Y(s) the subject:

Y (s ) A
sY (s ) + = (3.8)
τ τs

 1 A
Y (s ) s +  = (3.9)
 τ  τs

A
Y (s ) = τ (3.10)
 1
s s + 
 τ

Expanding (3.10) as a partial fraction, introducing the constants α and β :

α β
Y (s ) = + (3.11)
s  1
s + τ 
 

Applying the Heaviside method (with s = 0 and -1/τ):

4
A
α= τ (3.12)
 1
s + τ 
  s =0

A
α= τ (3.13)
1
τ

α=A (3.14)

A
β= τ (3.15)
s
s =−1 τ

A
β= τ (3.16)
s
s =−1 τ

A
β =− τ (3.17)
1
τ

β = −A (3.18)

(NB: You can use either this method or that used in Question 2 to obtain inverse
Laplace transforms of functions with unique real factors in the denominator
polynomial.)

Substituting α and β back into (3.11) gives:

A A
Y (s ) = − (3.19)
s  1
s + τ 
 

Applying inverse Laplace transforms to each component of (3.19), where L-1{


Y(s)} = y(t):

A
L −1   = A (3.20)
s 

5
   
 A   1 
−t τ
L −1  −1
 = AL   = Ae (3.21)

 s+ 1  
 s+ 1 
  τ     τ  

Accordingly, the solution


olution to the differential equation (3.1) is:

(
y (t ) = A 1 − e − t τ ) (3.22)


Numerical solutions to this differential equation (using A = 2 and τ = 5), in Excel,
are available on-line.

numerical Euler and 4th


The following figure compares the exact solution to the numerical
order Runge-Kutta Runge Kutta approach is seen to accurately
Kutta results. The Runge-Kutta
form solution using a relatively coarse time-step.
reproduce the closed-form time Similar
numerical protocols are employed for numerical simulation in MATLAB and
Simulink,, but with much finer time steps.

6
Question 4

To solve these problems we will convert the differential equations to Laplace


transforms, then rearrange them to give a fraction in which the denominator
contains the kind of factors that we have dealt with: (s + b) and (s2 + d1s + d0).

We can assume that all initial conditions are zero, as they will not affect the form
of the time-domain function.

d 3x d 2x dx
(a) 3
+ 2 2
+2 +x =3 (4a.1)
dt dt dt

The Laplace transform of this function is:

3
s 3 X (s ) + 2s 2 X (s ) + 2sX (s ) + X (s ) = (4a.2)
s

3
(s 3
)
+ 2s 2 + 2s + 1 X (s ) =
s
(4a.3)

3
X (s ) = (4a.4)
(
s s + 2s 2 + 2s + 1
3
)
3
X (s ) = (4a.5)
(
s(s + 1) s 2 + s + 1 )

The partial fraction expansion of (4a.5) is:

A B Cs + D
X (s ) = + + 2 (4a.6)
s s +1 s + s +1

We don’t need to know the constants A through D to determine the type of


functions in x(t). From (4a.6) we can see that the functions in x(t) will be e-t and
e-t(sint + cost). The A/s term will result in a constant.

d 2x
(b) 2
− x = 2e t (4b.1)
dt

7
2
s 2 X (s ) − X (s ) = (4b.2)
s −1

2
(s 2
)
− 1 X (s ) =
s −1
(4b.3)

2
X (s ) = (4b.4)
(
( s − 1) s 2 − 1 )

2
X (s ) = 2
(4b.5)
( s − 1) ( s + 1)

A B C
X (s ) = + + (4b.6)
( s − 1) ( s − 1) ( s + 1)
2

The resultant functions of t in x(t) will be e-t (twice) and te-t.

d 3x
(c) + x = sin t (4c.1)
dt 3

1
s 3 X ( s ) + X (s ) = 2
(4c.2)
s +1

1
(s 3
)
+ 1 X (s ) = 2
s +1
(4c.3)

1
X (s ) = (4c.4)
(s 2
)(
+ 1 s3 + 1 )

8
1
X (s ) = (4c.5)
( ) (
s + 1 ( s + 1) s 2 − s + 1
2
)

A Bs + C Ds + E
X (s ) = + 2 + 2 (4c.6)
(
( s + 1) s + 1 s − s + 1 ) ( )

The factor (s2 + 1) is in the form (s2 + d1s + d0), where d1 = 0. This gives a
function with terms e-t, e-t(sint + cost), and sint + cost.

d 2 x dx
(d) + =4 (4d.1)
dt 2 dt

4
s 2 X (s ) + sX (s ) = (4d.2)
s

4
(s 2 + s ) X (s ) = (4d.3)
s

4
X (s ) = 2
(4d.4)
s (s + s )

4
X (s ) = 2
(4d.5)
s (s + 1)

A B C
X (s ) = + 2+ (4d.5)
s s (s + 1)

We therefore find the functions t and e-t.

Problem (a) has the complex factor (s2 + s + 1). This factor results in a damped
oscillation (b > 0). Problem (c) has two complex factors, (s2 + 1) and (s2 - s + 1).
For the (s2 + 1) factor d1 = 0, so b = 0. This results in a stable, or repeating,
oscillation. With the (s2 - s + 1) factor b is negative, resulting in an exponentially
increasing oscillatory response in t.

9
Question 5

dx
a) − 12x = sin 3t (5a.1)
dt

The Laplace transform of (5a.1) with x(0) = 0 is:

3
sX (s ) − 12 X (s ) = 2
(5a.2)
s +9

3
X (s )(s − 12 ) = 2
(5a.3)
s +9

3
X (s ) = (5a.4)
(
(s − 12) s 2 + 9 )
3
X (s ) = (5a.5)
(
(s − 12) s 2 + 9 )
The partial fraction expansion of (5a.5) gives:

3 A Bs + C
X (s ) = = + 2 (5a.6)
( )
(s − 12 ) s + 9 (s − 12) s + 9
2
( )

Multiplying (5a.6) through by the denominator (s-12)(s2+9):

( )
3 = A s 2 + 9 + (Bs + C )(s − 12) (5a.7)

Solving (5a.7) at s = 12 gives A = 1/51. Equating coefficients of like powers


yields B = -1/51 and C = -12/51 (or -4/17).

1 1
51
− 51
s − 12
51
X (s ) = + (5a.8)
(s − 12) (s 2
+9 )
Taking the inverse Laplace transform of both side of (5a.8):

1 12t 1 -1 s + 12 
x (t ) = e − L   (5a.9)
51 51  s 2 + 9  ( )
10
1 12t 1 -1 s   12 
x (t ) = e − L  + 2  (5a.10)
51 51  s 2 + 9 ( ) (
  s +9 )


Completing the square in (s2 + 9) we get b = 0 and ω = 3. We therefore need the


numerator on the far right term to be equal to 3:

1 12t 1  -1 s  - 1 3 
x (t ) = e − L  2  + 4L  2  (5a.11)
51 51   s + 9  (
 s +9 ) ( )


1 12t 1
x (t ) = e − [cos 3t + 4 sin 3t ] (5a.12)
51 51

d 2x dx
b) 2
+6 + 25 x = e −t (5b.1)
dt dt


sଶ X(s)+6sX(s)+25X(s)= ௦ାଵ (5b.2)


ሺ‫ ݏ‬ଶ + 6‫ ݏ‬+ 25ሻX(s)= (5b.3)
௦ାଵ


X(s)= ሺ (5b.4)
ୱାଵሻሺୱమ ା଺ୱାଶହሻ

஺ ஻௦ା஼
X(s)= + (5b.5)
௦ାଵ ௦మ ା଺௦ାଶହ

ଵ/ଶ଴ ଵ/ଶ଴௦ାଵ/ସ
X(s)= ௦ାଵ
− ௦మ ା଺௦ାଶହ
(5b.6)

Completing the square gives (s2 + 6s + 25) = (s + 3)2 + 4, where b = 3 and ω = 4.

11
ଵ/ଶ଴ ଵ ௦ାହ
X(s)= ௦ାଵ
− ଶ଴ × ௦మ ା଺௦ାଶହ (5b.7)

ଵ/ଶ଴ ଵ ௦ାଷ ଵ ସ
X(s)= − × − × (5b.8)
௦ାଵ ଶ଴ ௦మ ା଺௦ାଶହ ସ଴ ௦మ ା଺௦ାଶହ

ଵ ଵ ଵ
X(s)= ݁ ି௧ − ݁ ିଷ௧ ቀ ܿ‫ݏ݋‬4‫ ݐ‬+ ‫݊݅ݏ‬4‫ݐ‬ቁ (5b.9)
ଶ଴ ଶ଴ ସ଴

12
CHEN90032
Process Dynamics and Control

Worksheet 4 – Transfer Function Modelling

In process control, dynamic models – either derived from first-principles or


otherwise – are conventionally expressed as transfer functions. The major
advantage of transfer functions is that they are independent of the input
function, and they can therefore be used to investigate process response to a
number of different inputs. Transfer functions have the disadvantage of
dealing in deviation form variables, and it is important to remember that their
output represents only the change in a variable from its initial value.

In this worksheet we will use transfer functions to study the response of a


process with known dynamics to various input functions, in what constitutes
an open-loop system. Later in this course we will see that control algorithms
can also be defined in terms of transfer functions, and by algebraically
manipulating the controller and process transfer functions we can obtain the
closed-loop response of the system being controlled.

Transfer function models can also be solved numerically, as explored in


problem 2 (here using Simulink). The advantage to this approach is that once
a simulation has been developed we can investigate the effect of changing
dynamic models and input functions without re-doing our calculations.

W4-1
1) A chemical plant generates a waste stream at a fixed rate q (kg s-1). Under
normal conditions the stream does not contain any toxic substances and can
be discharged directly into the municipal sewer system. However, when there
is a plant upset this stream may be contaminated by a toxic chemical, the
mass fraction of which varies with time (x1), and may for a brief period exceed
the maximum allowable discharge concentration (xmax). To cater for this
eventuality, the waste stream is discharged into the sewers via a buffer tank to
even out any unacceptably high concentration of the toxic chemical. The
buffer tank has constant volume V (m3) and can be assumed to be well-mixed.

(a) Obtain the transfer function between the outlet mass fraction x2 and
inlet mass fraction x1 in the buffer tank.

(b) If the input function x1 takes the form of a step of height A obtain the
outlet mass fraction x2(t).

(c) In the worst plant upset recorded the input function took the form of a
step of height 2xmax of duration W seconds. What volume tank is
required to cope with this disturbance? Would a 1 m3 tank be able to
cope with a disturbance of 1.7xmax for 30 s, if q = 40 kg s-1 and ρ =
1000 kg m-3?

W4-2
2) A composition sensor is used to continually monitor the contaminant level
in a liquid stream. The dynamic behaviour of the sensor can be described by
Cˆ m (s ) 1
the first-order transfer function = with time constant of 10 s,
Cˆ (s ) 10s + 1
where C is the actual contaminant concentration and Cm is the measured
value, both in deviation form. Both variables are initially 5 ppm.

An alarm sounds if the measured contaminant value exceeds the


environmental limit of 7 ppm. Suppose that the contaminant concentration c
gradually increases according to the expression c(t) = 5 + 0.2t (with t in
seconds). After the actual contaminant concentration exceeds the
environmental limit, what is the time interval until the alarm sounds?


Use MATLAB Simulink to numerically simulate the response of a first-order
process to a ramp input function. Solve the above problem in Simulink, and
compare with the exact result. Investigate the sensor response using different
process transfer functions, time constants and input functions.

2
3) The transfer function for a process is approximated by G(s ) = 2
.A
s + s +1
step change of magnitude 5 is introduced to the process. Calculate the
overshoot (fraction and percentage), the decay ratio, the new steady-state
output value, the maximum value of the output, the period of oscillation, and
the rise time.

W4-3
4) You are presented with the following plot of temperature (T, °C) versus time
(t, s) for an unknown heating tank under feedback control, in which the flow
rate of heating steam is known to have increased by 8 kg s-1 after operating
for a long period of time.

(a) Develop an approximate transfer function between the tank


temperature (output) and the steam flow rate (input), with numerical
values for all constants. What assumptions are you required to make?

(b) The tank temperature needs to converge rapidly to around the new set-
point, but the output deviation can not overshoot by more than 10%.
Suggest an appropriate value of ζ to use.

60

50
T ( C)

40
o

30

20
0 20 40 60 80 100 120 140

t (s)

W4-4
CHEN90032
Process Dynamics and Control

Worksheet 4 – Transfer Function Modelling – SOLUTION

1a) Perform a component mass balance across the tank (V and q are
constant):

dx 2
ρV = qx 1 − qx 2 (1a.1)
dt

dx 2
=
q
(x 1 − x 2 ) (1a.2)
dt ρV

dxˆ 2 qxˆ 2 qxˆ1


+ = (1a.3)
dt ρV ρV

The Laplace transform of (1a.3) gives:

sX 2 (s ) + X 2 (s ) = X 1 (s )
q q
(1a.4)
ρV ρV

X 2 (s )  q  q
 s +  = (1a.5)
X 1 (s )  ρV  ρV

Multiplying both sides by ρV/q:

X 2 (s )  ρV 
 s + 1 = 1 (1a.6)
X 1 (s )  q 

X 2 (s ) 1
= (1a.7)
X 1 (s )  ρV 
 s + 1
 q 

1
Introducing the time constant (which in this case is also the residence time) τ
= ρV/q:

X 2 (s ) 1
= (1a.8)
X 1 (s ) (τs + 1)

This is the transfer function between the outlet and inlet concentrations.

1b) The input step function x1(t) = A corresponds to the Laplace transform
X1(s) = A/s. Introducing this into our transfer function:

X 2 (s ) =
A
(1b.1)
s (τs + 1)

The inverse Laplace transform of (1b.1) gives:

 −t

x 2 (t ) = A1 − e τ  (1b.2)
 
 

1c) Using our analytical time-dependent model for x2(t) (1b.2) we can
calculate x2 as a function of time for a given residence time (ρV/q) and step
height (A). Recognising that the concentration x2 will peak at time w for a
pulse input of duration w seconds and height A kg m-3, we can find the volume
V at which x2 = xmax.

Letting x2 = xmax, A = 2xmax, and t = w:

 −w

x max = 2 x max 1 − e τ  (1c.1)
 
 

−w

1 = 2 − 2e τ
(1c.2)

2
−w
1
e τ
= (1c.3)
2

−w
1
e τ
= (1c.4)
2

−w  1
= ln  (1c.5)
τ 2

= ln(2)
w
(1c.6)
τ

w
τ = (1c.7)
ln(2)

Substituting in the definition of τ:

ρV w
= (1c.8)
q ln(2)

qw
V = (1c.9)
ρ ln(2)

Solving the above relationship for a given time (w), flowrate (q) and fluid
density (ρ) gives the minimum allowable tank volume.

For a tank volume of 1 m3, and A = 1.7xmax, w = 30 s, q = 40 kg s-1, and ρ =


100 kg m-3, we can again use the model for x2(t) (1b.2). Now, we want to see
if x2 exceeds xmax for the specified conditions. We will evaluate x2 at time t =
w, as this will correspond to the peak in concentration (assuming perfect
mixing).

 −30

x 2 = 1.7 x max 1 − e τ 
 (1c.10)
 
 

3
The residence time in this instance is 25 seconds (1000×1/40):

 −30


x 2 = 1.7 x max 1 − e 25  (1c.11)
 
 

x 2 = 1.19 x max (1c.12)

We see that x2 is greater than the maximum allowable concentration, and the
tank is therefore not large enough.

BONUS QUESTION: What volume of tank would we need to account for this
disturbance?

Letting x2 = xmax in equation (1b.2):

 −30

x max = 1.7 x max 1 − e τ 
 (1c.13)
 
 

Canceling xmax and rearranging:

−30
0 .7
e τ
= (1c.14)
1 .7

− 30  7 
= ln  = −0.887 (1c.15)
τ  17 

30
τ = = 33.8 (1c.16)
0.887

Where τ = 1000V/40 = 25V. Substituting into (1c.16):

25V = 33.8 (1c.17)

4
Solving (1c.17) we get V = 1.35 m3. So a tank of at least 1.35 m3 is needed to
handle the specified disturbance. With a 1.35 m3 surge tank a disturbance of
any larger magnitude or longer time would result in the output exceeding the
maximum allowable discharge concentration.

2) The sensor will achieve a reading of 7 ppm at some point later than when
this concentration is actually achieved, due to the time constant of the first
order process. Note that a smaller time constant will yield a faster response
time. The time at which c reaches 7 ppm can easily be calculated as c = 7 = 5
+ 0.2t, giving t = 10 seconds. To evaluate the time at which the sensor
registers a reading (cm) of 7 ppm we multiply the given transfer function by the
input Laplace transform, in deviation variable form.

c(t ) = 5 + 0.2t (2.1)

cˆ(t ) = c(t ) − c (t = 0) = 5 + 0.2t − 5 (2.2)

cˆ(t ) = 0.2t (2.3)

The Laplace transform of (2.3) is:

0 .2
C (s ) = (2.4)
s2

Introducing (2a.4) into the provided transfer function:

0 .2
C m (s ) = (2.5)
s (10s + 1)
2

The inverse Laplace transform of (2.5) is:

5
 −t 
cˆ m (t ) = 0.2t + 2 e 10 − 1 (2.6)
 
 

We have a function for cm(t) with terms in t0, t1 and e-t. This equation is not
easily rearranged to express t as a function of cm, so we solve it numerically
by trial and error.

We want to know t when cm in deviation form is 2 ppm (7 ppm absolute). We


now have to iteratively solve equation (2.6) to get deviation cm = 2 ppm. We
know that this time will occur after 10 seconds, and because the time constant
is 10 seconds the steady-state response in cm should be achieved on around
this time scale. Starting with an initial guess of t = 20 s we get cm = 2.27 ppm,
which is too large (therefore t is also too large). By trial and error we discover
that t = 18.4 s. Therefore, the alarm sounds 8.4 seconds after c reaches 7
ppm.

6

A Simulink model for Q2 is available on-line.

The figure below illustrates how the actual and measured concentrations, c
and cm, vary as a function of time. In accordance with the above solution, we
observe that the measured concentration indicates a deviation of 2 ppm
between 18 and 19 seconds, around 8.5 seconds after the alarm should have
sounded.

The dynamic behaviour of this system is characteristic of a 1st order process


responding to a ramp input function. The next figure illustrates the effect of
varying the time constant of the sensor, illustrating that a smaller time
constant results in a more rapid sensor response. In the final figure, white
noise is incorporated into the concentration signal. Observe that even though
the noise amplitude is significant compared to the slope of the ramp input
function, it is almost absent from the measured concentration. Here, the
sensor is effectively acting as a low-pass filter on the noise, because the time
constant (10 s) is much larger than the characteristic timescale of the noise.

7
8
3) Introducing the step change input function U(s) = 5/s into the transfer
function we obtain:

10
Y (s ) =
( )
(3.1)
s s + s +1
2

We have an underdamped second-order system, where τ = 1 and ζ = 0.5.

The analytical expression for overshoot (OS) is:

 − πζ 
OS = exp  (3.2)
 1− ζ 2 
 

 − π × 0 .5 
OS = exp  (3.3)
 
 1 − 0 .5 
2

Giving OS = 0.16, or 16%. The decay ratio (DR) is simply OS2, or 0.026. The
steady-state output is 10, and the maximum output value is 10 + 0.16×10
(+16% overshoot), or 11.63. The period of oscillation is:

2πτ
P= (3.4)
1− ζ 2

2π × 1
P= (3.5)
1 − 0 .5 2

We find that P = 7.2 s. The rise time (tr) is the most difficult property to
evaluate. The rise time is the earliest time at which y(t) is equal to the steady-
state value of 10. Finding tr requires taking the inverse Laplace transform of
equation (3.10), then evaluating this function at y = ys-s = 10 (don’t worry - I
won’t ask you to do this in either exam, but I do want you to see how it would
be done).

9
 −ζ t   1− ζ 2  ζ  1− ζ 2   

KA 1 − e τ cos 
yˆ (t ) = t+ sin  t    (3.6)
   τ  1− ζ 2  τ  
    

−ζt −ζt
1− ζ 2
ζ 1− ζ 2
1 = 1− e τ
cos t −e τ
sin t (3.7)
τ 1− ζ 2 τ

4a) The process response curve looks like an underdamped second order
process responding to a step change input. Assuming that this is the case,
the transfer function should look like:

K
G( s ) = (4a.1)
τ s + 2ζτs + 1
2 2

The process gain can be determined form the steady-state response to the
step input function, y = KA, where A is the step input height (8 kg s-1). From
the figure, estimating that KA = 46 – 20 = 26 °C, we evaluate K as 3.25 °C s
kg-1.

The damping coefficient can be determined solely form the overshoot or


decay ratio. Measurements of the overshoot are more accurate, so we will
use this approach. If Tmax = 58 °C and Ts-s = 46 °C, then OS = (58-46)/(46-20)
= 0.46 (46%). The overshoot is related to the damping coefficient by:

 − πζ 
OS = exp  (4a.2)
 1− ζ 2 
 

Rearranging:

ζ =
[ln(OS )]2 (4a.3)
π 2 + [ln(OS )]
2

10
ζ =
[ln(0.46)]2 (4a.4)
π 2 + [ln(0.46 )]
2

This gives ζ = 0.24. The time constant, τ, is obtained from ζ and the period,
P. More accurately, we can use tP instead of P (the time to peak is half of one
oscillation, or P/2). Estimating that the time to peak is 16 s, we get P = 32 s.

2πτ
P= (4a.5)
1− ζ 2

Rearranging:

P 1− ζ 2
τ = (4a.6)

1 − 0.24 2
τ = 32 (4a.7)

Finally, we obtain τ = 4.9 s. This gives us the process transfer function:

3.25
G( s ) = (4a.8)
24s 2 + 2.4s + 1

4b) We want to find the smallest value of ζ that results in no more than 10%
overshoot. This is independent of the time constant, and we can simply find ζ
= 0.59 from (4a.3), as shown below.

ζ =
[ln(0.1)]2 = 0.59 (4a.9)
π 2 + [ln(0.1)]
2

11
CHEN90032
Process Dynamics and Control

Worksheet 5 – Bode Plots

The Bode plot (or Bode diagram) is a widely used method for illustrating the
frequency response of a process, i.e., the amplitude ratio (AR) and phase
angle (φ) versus frequency (ω). The Bode plot of an open-loop process
provides information about the transfer function and therefore the underlying
process dynamics. We may be able to determine, for example, the presence
of time delays or high-order behaviour from a Bode plot. This information is
useful when thinking about how difficult a process will be to control. The Bode
plot for an open-loop process also contains information about that process in
a closed-loop feedback control configuration, and later in this course we will
revisit Bode plots and use this property to help arrive at fast-acting yet stable
control algorithms.

W5-1
1) The following frequency response data has been generated from a pulse
test.

(a) Generate a Bode diagram for the process (a blank template is


provided on the next page).

(b) Fit a transfer function to the data.

ω (rad/min) AR φ (deg)
0.0003 51.46 -0.09
0.0032 51.37 -0.81
0.041 51.21 -6.30
0.079 49.54 -15.10
0.13 46.69 -26.90
0.18 42.75 -34.60
0.24 38.71 -42.10
0.26 36.79 -43.90
0.79 16.63 -72.30
1.32 10.05 -83.40
1.85 7.24 -87.20
2.38 6.03 -88.70
2.64 5.09 -89.20
3.1 4.21 -89.70

W5-2
W5-3
10
2) A process is described by the transfer function G(s ) = 2
.
4s + 0.8s + 1
Sketch the Bode diagram that would result from a frequency response
analysis (time is in units of minutes), in the space below.

10

1
ARN

0.1

0.01

0.001
0.01 0.1 1 10 100

-60
φ (degrees)

-120

-180

-240

-300

-360
0.01 0.1 1 10 100

ω (rad/min)


Use the provided MATLAB code to generate the exact Bode plot for G(s).
Compare this plot to your approximate sketch. Investigate the effect of adding
a time delay that is significantly (i) smaller, and (ii) larger than the
characteristic time constant of the system.

W5-4
3) You are attempting to design a feedback controller to maintain reactant
concentration in a chemical reactor, and first need to understand the process
dynamics. Previous experimental testing has indicated that the outlet and
inlet concentration are adequately described by a first-order transfer function
with steady-state gain of 0.8 and time constant of 8 min. However, an
additional segment of piping has been placed at the reactor outlet, introducing
a transport delay (i.e., time delay) of 18 s.

(a) Sketch the Bode plot representing the original 1st order process
transfer function.

(b) Choose two frequencies at which you expect the phase angle to be
relatively unaffected by the 1st order process. Determine the phase
angle difference that would arise from introduction of the time delay.
Sketch the Bode plot for the new first order plus time delay
(FOPTD) transfer function.

(c) A Padé approximation is to be used to simplfy the FOPTD model in


designing the controller. Derive approximate overall transfer
functions using the 1/1 and 2/2 approximations.


(d) Plot all four transfer functions and compare.

W5-5
CHEN90032
Process Dynamics and Control

Worksheet 5 – Bode Plots – SOLUTION

1)

(a) Amplitude ratio and phase angle plots for the given process are depicted
below:

100
AR

10

1
0.001 0.01 0.1 1 10

0
φ (degrees)

-30

-60

-90
0.001 0.01 0.1 1 10

ω (rad/min)

(b) From the amplitude ratio diagram, at high frequencies the slope approaches
–1. This is indicative of a first-order process. On the phase angle diagram φ
varies between 0 and –90°, which is also characteristic of a first-order process.
We therefore conclude that the process transfer function looks like:

1
K
G( s ) =
τs +1

We need to determine K and τ in order to characterize the dynamic response.

At low frequencies AR approaches K, so K = 51.5.

The brake frequency, ωb = 1/τ, occurs on the amplitude ratio plot at the
intersection of the low and high frequency asymptotes. From the diagram below
we see that ωb = 0.26 rad/min, so τ = 3.8 minutes. Alternatively, the brake
frequency occurs at φ = –45°, and we estimate that this occurs on the phase
angle plot at 0.28 rad/min (τ = 3.6 minutes). Note: the small difference between
the two time constants reflects the variability in the experimental frequency
response data, and the uncertainty in reading these values off of the graph.

Using the pulse test data we have estimated the following transfer function:

51.5
G (s ) =
3.8s + 1

100

10
AR

1
0.001 0.01 0.1 1/τη
1/ 1 10

-30

-60
(degrees)
φ

-90
0.001 0.01 0.1 1/τη
1/ 1 10

ω (rad/min) 2
2) The transfer function is second-order, in the form:

K
G(s ) = 2 2
τ s + 2τζ s + 1

The process gain K is 10. The time constant is 2 minutes, and the damping
coefficient is 0.2.

The Bode diagram uses normalised amplitude ratio (ARN), so the low frequency
asymptote will approach ARN = 1. For a second-order process the high
frequency amplitude ratio asymptote approaches a slope of –2. The asymptotes
intersect at a brake frequency (1/τ) of 0.5 radians/minute. The process is
underdamped with a small damping coefficient, and the AR plot will also show a
peak at around ωb (this behaviour is typical for ζ of around 0.6 or below). Finally,
the phase angle plot will vary from 0° at low frequencies to -180° at high
frequencies. The brake frequency occurs at –90°.

The exact Bode diagram for this transfer function (generated in MATLAB) is
shown below, over similar ranges of ARN, φ, and ω. You should be able to
reproduce all of these major features by hand.

3
The following figures illustrate Bode plots for the same transfer function with a
time delay that is (i) an order of magnitude smaller (0.2 min) and (ii) an order of
magnitude larger (20 min) than the process time constant of 2 minutes.

Note that the introduction of a time delay has no effect on AR. In (i) we observe
that the small time delay has little influence on φ at low frequencies, but
dominates at high frequencies. Thus, it is still possible for time delays that are
much smaller than the characteristic time constant of a system to exert a
significant dynamic effect. This is particularly true for closed-loop (e.g., feedback
control) systems. From (ii) we notice that the phase angle is completely
dominated by the large time delay.

4
3)
.
(a) The first order Bode plot for   is provided below:


(b) At frequencies greater than about 1 min-1 the phase angle is close to 90°, and
in the FOPTD system any additional phase lag will be due to the time delay. The
time delay component of the phase angle (in radians) is given by the equation φ
= -ωθ (θ = 0.3 min), and at any given frequency the phase lag from the time delay
in the system can be determined. The total phase angle at high frequencies is
approximately -90 – 180ωθ/π degrees.

Evaluating the phase angle at 5 and 10 min-1:



φ5 min   90   5  0.3  176°



φ10 min   90   10  0.3  262°


.  .!"
The complete FOPTD Bode plot for    is shown below. Observe that
the phase angle values calculated above are close to the real values. Also note
that the AR plot is completely unaffected by the introduction of the time delay.

5
(c) The 1/1 Padé approximation is:
%

# $  &
%
&

. (
# .'  . (

Substituting into the transfer function:

. . (
  ) *) *
. (

. (
  0.8 ) . (
*


. (
  0.8 ) *
., & . (

6
Similar analysis with the 2/2 Padé approximation yields:

.-( & . (


  0.8 ).. ! .,-( & . (
*

Both approximate transfer functions provide the exact amplitude ratio.

The phase angle plot for the 1/1 and 2/2 approximants is illustrated below,
compared to the exact solutions with and without time delay. The 2/2
approximant provides a good fit up to relatively high frequencies. The 1/1
approximant diverges significantly at high ω, but does provide an adequate
description of the FOPTD system over a relevant frequency range.

7
CHEN90032
Process Dynamics and Control

Worksheet 6 – Empirical Modelling

Empirical dynamic models – derived from experimental data – are used widely
for control system design in the process industries. Empirical models are
advantageous when the unsteady-state balances are too complex to derive by
hand, or when the number of process variables that need transfer functions
are very large (complex, integrated plants may have hundreds or thousands of
such variables). Empirical models need to accurately describe the
experimental data (which includes reproducing all important features such as
dead time, oscillations) and must also be derived under conditions similar to
those under which the plant is expected to operate.

This worksheet illustrates several approaches available for fitting experimental


data to a dynamic (transfer function) model. In Problem 1 we will attempt to fit
two different transfer functions to the same data using graphical techniques.
For Problem 2 you are given a process transfer function and are asked to
develop an approximate model function of different form.

W6-1
1) The bottom product from a distillation column is cooled in a heat
exchanger. An increase of 2 m3/s in the coolant flow rate applied at time t =
10 s results in the temperature response plotted below.

Fit (a) a FOPTD and (b) a second-order transfer function to the output
response data. Obtain the time-domain response for the second-order model.

50

49
Temperature (°C)

48

47

46
0 20 40 60 80 100 120

Time (seconds)


Using Simulink, simulate the dynamic response of the open-loop system with
both of the developed models. Compare the simulation results to the raw data
(available on-line), and attempt to adjust the model parameters to obtain a
better fit.

W6-2
2) You are developing a control system for a natural gas furnace, and have
determined the following process transfer function G(s) that relates flue gas
temperature (K) to natural gas flowrate (L/s) (NB: time constant and time
delay in seconds):

4݁ ିଶ.଻ଽ௦
‫ܩ‬ሺ‫ݏ‬ሻ =
7.87‫ ݏ‬+ 1

In order to simplify the controller design, you have been asked to reduce this
model from FOPTD to second order. Using Smith’s method determine the
required parameters for the second order model. Determine the model error
for a step change input at t20 and t60. Is the approximate transfer function an
appropriate choice for designing the control system?

W6-3
W6-4
CHEN90032
Process Dynamics and Control

Worksheet 6 – Empirical Modelling – SOLUTION

1) First, we will calculate the process gain (K). For an input step change of 2
m3/s we get a steady-state output response of –3.8 °C. The steady-state output
response is the product of the gain and the step height, so K = -3.8/2 = -1.9
K·s/m3. Note that the gain is negative, indicating that an increase in the input
variable (coolant flow rate) results in a decrease in the output variable
(temperature).

(a) For a FOPTD (first-order plus time delay) model, we have the transfer
function:

Ke −θ s
G(s ) =
τs +1

We know the value of K, and need to obtain the time delay (θ) and time constant
(τ). Several methods exist for fitting FOPTD models; we will use the following
equations:

t 2 3 − t 13
τ=
0.7

θ = t − 0.4τ
1
3

From the output response we estimate that t1/3 and t2/3 occur at 48.7 and 47.5 °C,
respectively. From the diagram below we see that t1/3 = 8 s and t2/3 = 18 s. Note
that the time at which the step is applied is set to t = 0 s. The time constant (τ) is
calculated as 14.3 s and the time delay (θ) is 2.3 s. Accordingly, the FOPDT
transfer function is:

−1.9e −2.3s
G(s ) =
14.3s + 1

(b) The second-order model corresponds to a transfer function:

K
G(s ) = 2 2
τ s + 2ζτ s + 1

1
Again, we know that K = -1.9 K·s/m3, and we now need to determine τ and ζ.

To apply Smith’s graphical method, we require t20 and t60, the times
corresponding to 20% and 60% conversion, which respectively occur at 49.2 and
47.7 °C. On the attached figure we see that t20 = 5 s and t60 = 15 s. The ratio
t20/t60 is 0.33, and we use this with the attached figure to get ζ ~ 1.5 and τ ~ 4.3 s
(note that just from a visual inspection of the process response curve we would
expect a value of ζ greater than 1, with time constant on the order of several
seconds).

−1.9
G( s ) = 2
18.5s + 12.9s + 1

The time-domain response can be obtained by substituting K, τ, and ζ into the


equation derived in the notes.

2
50

49

48
Temperature (°C)

47

t 1/3 t 2/3
46
0 t0 20 40 60 80 100 120

Time (seconds)

50

49

48
Temperature (°C)

47

t 20 t 60
46
0 t0 20 40 60 80 100 120

Time (seconds)

3
4
As a test of the graphical method used to fit the second-order model, the actual
transfer function that we are trying to fit is:

−1.9
G( s ) = 2
25s + 16s + 1

The step response of the FOPTD and 2nd order model, and the actual process
transfer function, can be obtained from Simulink using a block diagram such as
the following:

The time-domain response of the actual transfer function and the two model fits
are shown in the figure below. Surprisingly, the FOPTD model provides the best
description of the actual data, apart from some small discrepancy at very early
times. The 2nd order model provides a slightly better account of the process
response initially, but deviates significantly at around 20 to 60 s. Note, however,
that the 2nd order model has the potential to exactly reproduce the actual process
data if we refine the fit, whereas the FOPTD model does not.

5
2) The approximate model will be in the following form, where K = 4 K s L-1:

‫ܭ‬
‫ ܩ‬ሺ‫ݏ‬ሻ =
߬ଶ‫ݏ‬ + 2ߦ߬‫ ݏ‬+ 1

In order to apply Smith’s method we need the FOPTD process output for a step
change input at 20 % and 60 % normalised response. We could plot the FOPTD
model response and determine t20 and t60 graphically, or alternatively we could
use the time-domain solution to the model.

For a first order transfer function responding to a step-change, the normalized


output response (yN) is given by:

‫ݕ‬ே ሺ‫ݐ‬ሻ = 1 − ݁ ି௧/ఛ

Rearranging to make t the subject:

݁ ି௧/ఛ = 1 − ‫ݕ‬ே ሺ‫ݐ‬ሻ

−‫ݐ‬/߬ = lnሾ1 − ‫ݕ‬ே ሺ‫ݐ‬ሻሿ

‫ = ݐ‬−߬ lnሾ1 − ‫ݕ‬ே ሺ‫ݐ‬ሻሿ

Substituting in τ = 7.87 s and solving for t at yN = 0.2 and 0.6 we get t20 = 1.76 s
and t60 = 7.21 s. Note however that we have neglected the time delay of 2.79 s,
and the actual values for t20 and t60 are therefore 4.55 and 10.00 s, respectively.

To apply Smith’s method we require t20/t60, which is 0.455. As shown on the


figure over page, this results in ζ ~ 0.7 and τ ~ 5.6 s. The approximate process
transfer function is therefore:

4
‫ ܩ‬ሺ‫ݏ‬ሻ =
31.36‫ ݏ‬+ 7.84‫ ݏ‬+ 1

Note that the damping coefficient of less than 1 suggests slightly underdamped
(oscillatory) behavior. This process response could never be achieved by a first
order plus time delay system, and the derived process transfer function is clearly
inadequate. We could, however, arrive at an improved approximation by setting
the damping coefficient to 1 (because this slows the process response an even
better fit would be obtained by decreasing the time constant to something like 5.2
s in order to compensate, but we will stick with 5.6 s). The process transfer
function is now:

4
‫ ܩ‬ሺ‫ݏ‬ሻ =
31.36‫ ݏ‬+ 11.2‫ ݏ‬+ 1

6
In order to determine the model error at t20 and t60 we need the time domain
solution for a critically damped 2nd order process responding to a step input:


‫ݕ‬ே ሺ‫ݐ‬ሻ = 1 − ቀ1 + ቁ ݁ ିഓ

With τ = 5.6 s we can calculate yN = 0.196 and 0.533 at t = 4.55 s and 10.00 s,
respectively. Compared to the actual response at these times (0.20 and 0.60) the
2nd order model under-predicts the FOPTD model by 2 % and 11 %.

Is the approximate transfer function an appropriate choice for designing the


control system? No. Despite some problems with using Smith’s method to fit the
2nd order model, it actually does provide a reasonable description of the FOPTD
behavior in the open-loop configuration. However, when designing controllers
we also need to consider the closed-loop system. In this case, the FOPTD
process will have some maximum controller gain at which the control system will
become unstable, whereas this will not be present in the 2nd order system (see
Topic 9). Accordingly, a fast-acting controller simulated on the 2nd order model
could result in an unstable response on the real FOTPD system.

7
8
CHEN90032
Process Dynamics and Control

Worksheet 7 – Closed Loop Systems

Feedback controllers constitute ‘closed loop’ systems, however it is still


possible to describe closed-loop dynamics using Laplace-space transfer
functions and block diagrams. This allows us to apply modeling techniques
developed earlier for open-loop systems. When dealing with closed-loop
systems we are interested in describing how process variables respond to
input changes (set-point and disturbances) and to different controller settings,
as well as understanding stability.

This worksheet presents several problems associated with block diagram and
transfer function representations of closed-loop system. For Question 1, a
block diagram needs to be developed from a process P&ID featuring a single
negative feedback loop. In Problems 2 and 3 you are provided with more
complex block diagrams and asked to derive closed-loop transfer functions
between input and output variables. In the final problem you are again asked
to derive a closed-loop transfer function, which is then used to simulate the
controller response.

W7-1
1) The temperature near the top of a distillation column is controlled by
manipulating the reflux rate, as depicted in Figure 1. Draw a block diagram
including transfer functions for the feedback control system, where the feed
flow rate (F) and composition (xF) are both considered to be disturbance
variables. The column temperature responds to changes in F, xF, and reflux
feed rate as an overdamped second-order process. The valve and
temperature analyser (TA) both behave as first-order processes. A PI
controller (TC) is used. [Numerical values for constants need not be included].

Coolant

TA TC

Feed
xF, F

Reflux Distillate

Bottom
Product

Figure 1. Feedback control system for distillate product quality in a distillation


column.

W7-2
2) A block diagram for internal model control (which we will be considering
later) is illustrated below (Figure 2). Derive the closed-loop transfer functions
for: (a) set-point changes, and (b) disturbances.

Ql Gl

+
+ E M +
R Km Gc Gp C
-

+
-
Gm

Figure 2. Internal model control block diagram.

3) Derive the closed-loop transfer function C/QL for the feedback control
system depicted in Figure 3.

Ql Gl

+
+ + +
R Gc G C
- -

+
-° s -
G* e

Figure 3. Feedback control system block diagram.

W7-3
4) The relationship between liquid level and inlet flowrate in a well-mixed tank
is described by a first order transfer function Gp with steady-state gain K = 1
and time constant τ = 2.8 min. A proportional controller with Kc = 4 is used to
control the liquid level, using the inlet flowrate as the manipulated variable.

(a) Derive both the laplace-domain and time-domain expressions for the
liquid level responding to a step change in the set-point, in terms of K,
τ, and Kc. Demonstrate that this closed-loop response corresponds to
a first-order process.

(b) The tank is at steady-state with liquid level of 2.8 m when the set-point
is changed to 3.8 m. What is the new liquid level at steady-state? How
long does it take for 95 % of the response to occur?


(c) Construct two Simulink models for the above problem, one using Kc
and Gp, the other using the transfer function derived in (a).
Demonstrate that the two systems exhibit the same dynamics. Confirm
that incorporating an integrator transfer function (1/τIs) into the negative
feedback loop eliminates steady-state offset.

W7-4
CHEN90032
Process Dynamics and Control

Worksheet 7 – Closed Loop Systems – SOLUTION

1) The process is described by the following block diagram. We have defined


transfer functions for the inlet flow rate and composition disturbances (Gq2 and
Gq1), the controller (Gc), the valve (Gv), the process (Gp), and the analyser
(Gm). We also need to include the measurement gain (Km) for the set-point.

Q2 Gq2

Q1 Gq1

+ +
+ + +
R Km Gc Gv Gp C
-

Gm

The transfer functions are:

 τ s + 1
Gc = K c  I 
 τIs 

Kv
Gv =
τ vs + 1

Kp
Gp = 2 2
(Similar to Gq1 and Gq2. ζ > 1)
τ s + 2ζτ ps + 1
p

Km
Gm = (Note that the measurement gain is the same as the set-point
τ ms + 1
gain)

1
2) The error function, E, for this form of feedback control is:

E = R − (C − MGm )

The manipulated variable, M, is given by:

M = Gc R − (C − MGm ) 

M = Gc R − Gc (C − MGm )

M = Gc R − GcC + MGcGm

M − MGcGm = Gc R − GcC

M (1 − GcGm ) = Gc R − GcC

Gc R − GcC
M=
1 − GcGm

The controlled variable, C, is the sum of MGp and QlGl:

C = QG
l l + MGp

Introducing our definition for M:

 Gc R − GcC 
C = QG
l l +  Gp
 1 − GcGm 

C (1 − GcGm ) = QG
l l (1 − GcGm ) + GcGp R − GcGpC

l l (1 − GcGm ) + GcGp R
C + GcGpC − GcGmC = QG

l l (1 − GcGm ) + GcGp R
C + GcGpC − GcGmC = QG

C (1 + GcGp − GcGm ) = QG
l l (1 − GcGm ) + GcGp R

2
For set-point changes (Ql = 0) we get:

C (1 + GcGp − GcGm ) = GcGpR

C
R
(1+ GcGp − GcGm ) = GcGp
C GcGp
Gsp = =
R 1 + GcGp − GcGm

For disturbance changes (R = 0) we get:

C (1 + GcGp − GcGm ) = QG
l l (1 − GcGm )

C
Ql
(1 + GcGp − GcGm ) = Gl (1 − GcGm )

C Gl (1 − GcGm )
Gdis. = =
Ql 1 + GcGp − GcGm

3
3) The block diagram can be re-drawn as the following equivalent form:

Ql Gl

+
+ + +
R Gc G C
- -

G* e-¬s

The internal negative feedback loop is comparing G* with G*e-θs, and can be
represented by the transfer function G*(1-e-θs):

Ql Gl

+
+ + +
R Gc G C
- -

G*(1-e-8S)

Introducing the transfer function Gc′ for the nested feedback loop:

Πf Gc
Gc′ = =
(
1 + Π e 1+GcG * 1 − e −θ s )

4
Ql Gl

+
+ +
R Gc’ G C
-

We want to derive the transfer function C/Ql, so we can assume that R = 0,


giving:

C = G QG − CGc′ 
 l l 

( )
C 1 + GGc′ = QGG
l l

C GGl
=
Ql 1 + GGc′

Substituting for Gc′ we get:

C GGl
=
Ql  Gc 
1+ G  
(
 1+GcG * 1 − e
−θ s
) 

5
4)

(a) The closed-loop transfer function between the a controlled variable (C) and
set-point (R) in a negative feedback control loop under proportional control is:
 
 =  
(4.1)

Where, for a first order process:



 = (4.2)


Substituting (4.2) in (4.1):

 

 = 
  (4.3)


This transfer function can be rearranged into the form of a first-order transfer
function, with gain K′ and time constant τ′:

 

 = 
  (4.4)


 
 =   
(4.5)

In (4.5) we want to place the denominator in the form τ′s + 1, and therefore
divide the numerator and denominator by 1 + KcKp:

 
 
 =   (4.6)
 

From (4.6) we find that:


 =  (4.7)

 
 = (4.8)
 

6
 = (4.9)
 

Indeed, the closed-loop system does resemble a first order process. Note
that this only holds for a proportional controller with first-order process transfer
function.

The Laplace- domain expression for liquid level (C) responding to a step-
change in the input (R = A/s) is:

 
 =   
(4.10)

The time-domain solution is:

"#
̂  =    1 − !  $ (4.11)

(b) The steady-state response in the closed-loop system can be obtained from
the overall process gain (K′) and step height (1.0 m):

̂ ∞ =    (4.12)

Introducing K = 1 and Kc = 4 into (4.8) yields:

&× &
 = &×
= ( = 0.8 (4.13)

Therefore the deviation at steady-state is 0.8 m, giving a final liquid level of


3.6 m. In this case the steady-state offset is 0.2 m (i.e., 3.8 – 3.6 m). Note
that as Kc approaches infinity the steady-state offset will approach zero (K′ →
1).

For the next part of the problem, we will rearrange (4.11) to give a normalized
response, cN(t):

7
"#

,  = = 1 − !  (4.14)
 

Introducing K = 1, Kc = 4, and τ = 2.8 min gives:

"#
,  = 1 − ! ../0 (4.15)

Evaluating (4.15) at cN = 0.95 yields t = 1.68 min.

(c) The following figure illustrates the equivalent closed- and open-loop block
diagrams in parallel. Running the model (available online), the two outputs
can be confirmed to be essentially the same.

1
4 C1
2.8s+1
Kc Process TF To Workspace1

Scope
Set Point, R

0.8
C2
0.56s+1
Process TF1 To Workspace2

Scope1
t
Clock To Workspace

The block diagram below incorporates integral action (1/8s) into the negative
feedback loop. The process response for a unit set-point change is also
illustrated, with and without the integrator block in place. Without integral
action (i.e., proportional control only) there is an off-set between the desired
(1) and actual (0.8) steady-state output. The transient response is, however,
rapid. Introducing integral action with τI = 8 min eliminates steady state offset,

8
although the controller dynamics are now more sluggish, with significant
overshoot.

1 1
4 C1
8s 2.8s+1
Set Point, R Kc Integrator Process T F To Workspace1

Scope

t
Clock To Workspace

9
CHEN90032
Process Dynamics and Control

Worksheet 8 – PID Controllers

Feedback controllers based on proportional (P), integral (I), and/or derivative


(D) action are the most commonly used form of automatic control in the
process industries. In order for these controllers to provide a rapid yet stable
control response, the constants associated with them need to be tuned. In
practice, tuning is often performed with the process on-line, where a
fundamental understanding of process dynamics can help guide the selection
of both the controller and the tuning constants. On-line tuning has several
downsides, and ideally a controller will be initially tuned off-line (with
subsequent fine-tuning performed on the operating process). In order to tune
a controller off-line we require an accurate dynamic process model.

This worksheet deals with PID-type controllers and their tuning constants. In
Question 1 the continuous cycling method of Ziegler-Nichols is used to
determine PID controller tuning constants based on the ultimate gain and
ultimate period for a closed-loop system. This technique can be used to tune
a controller on-line, but can also be performed off-line using an appropriate
process model. For Question 2 we look at how controller tuning constants
can be selected to provide a closed-loop controller response of a particular
desired form, while in Question 3 we examine a form of low-pass filter
commonly used with derivative control (which is particularly sensitive to noisy
input). Finally, Question 4 illustrates how frequency response techniques can
be used to guide the selection of a controller and its tuning constants.

W8-1
ଶ௘ షభೞ
1) A process is described by a first-order plus time delay model, ሺ‫ݏ‬ሻ = ହ௦ାଵ
.

Numerically, the ultimate gain and ultimate period are found to be 4.26 and
3.7, respectively. Obtain PID controller settings using the Ziegler-Nichols (Z-
N) and Tyreus-Luyben (T-L) tuning relations. Estimate the tuning constants
analytically by using a Padé approximation, and compare to the exact results.


Construct a block diagram representation of the closed-loop system in
Simulink. Simulate the control system performance, for both set-point tracking
and disturbance rejection, with the approximate and exact Z-N and T-L tuning
constants.

2) A first-order process with time constant τ and steady-state gain K is under


PI control (KC, τI). The transfer function between controlled variable and set-
point can be described as a second-order process, with an effective gain, time
constant and damping coefficient (Keff, τeff, and ζeff). Obtain these constants.

௔௦ାଵ
Hint: Rearrange the transfer function to place it in the form ‫ܩ‬ሺ‫ݏ‬ሻ = ௕௦మ ା௖௦ାଵ.

3) Proportional-derivative (PD) controllers are often formulated with the


ఛ ௦ାଵ
transfer function ‫ܭ‬௖ ቀఈఛವ ௦ାଵቁ, where a small value of α (typically 0.05 to 0.2)

acts as a low-pass filter on the derivative action (this form of transfer function
is known as a lead-lag unit). In practice, this type of controller can be realised
by placing a proportional controller (K2) in parallel with a first order transfer
௄భ
function ቀఛ ቁ. (a) Demonstrate that these two transfer functions in parallel
భ ௦ାଵ

are equivalent to the PD controller with derivative filter. (b) Determine


expressions that could be used to calculate required values of τ1, K1, and K2
given specified values of Kc, τD, and α.

W8-2
4) A continuous stirred tank reactor is being used to convert propylene oxide
into glycol, in an exothermic reaction. The reactor temperature is maintained
at the desired set-point of 50 °C by circulating cooling water around the
reactor. In order for the reactor products to be within specification, the reactor
temperature must be maintained within 5 °C of the set-point.

A Bode diagram has been generated for the transfer function that relates the
reactor temperature to cooling water flow rate, and is shown here in Figure 1.

(a) What are the values of the critical frequency (ωc), gain crossover
frequency (ωg), gain margin (GM), and phase margin (PM) for this process?

(b) When Kc = 1, how much additional time-delay must be added to or


removed from the system in order to make it critically stable?

(c) Recommend a type of controller and suggested controller settings to


maintain the reactor temperature at a set-point, despite process disturbances?
Provide a justification for your choice of controller and the controller settings.

W8-3
102

101
AR

100

10-1

10-2
0

-90
φ (degrees)

-180

-270

-360
0.0001 0.0010 0.0100 0.1000 1.0000 10.0000

ω (rad/min)

Figure 1. Bode plot for the open-loop transfer function (Question 4).

W8-4
CHEN90032
Process Dynamics and Control

Worksheet 8 – PID Controllers – SOLUTION

1) The relevant Z-N tuning relations are: Kc = 0.6Kcu, τI = 0.5Pu, τD = 0.125Pu.


The T-L relations are: Kc = 0.45Kcu, τI = 2.2Pu, τD = 0.159Pu.

The exact solution to the characteristic equation gives an ultimate gain (Kcu) of
4.26 and an ultimate period (Pu) of 3.7. The Z-N and T-L tuning constants are
thus found to be:

Kc τI τD
Z-N 2.56 1.85 0.46
T-L 1.92 8.14 0.59

The ultimate gain and period can be found without numerical simulation by
using a Padé approximation. We are interested in the characteristic equation:

1 + GOL = 0 (1.1)

The open-loop transfer function is the product of the process transfer function
and the controller transfer function (Gc = Kc; remember that the derivative and
integral actions are set to zero when finding the ultimate gain and period).

2K c e − s
1+ =0 (1.2)
5s + 1

We now introduce the Padé approximation e-s = (1 – 0.5s)/(1 + 0.5s):

 1 − 0.5s 
2K c  
1+  1 + 0.5s 
=0 (1.3)
5s + 1

1
2K c (1 − 0.5s )
1+ =0 (1.4)
(5s + 1)(1 + 0.5s )

2K c (1 − 0.5s )
1+ =0 (1.5)
5s + 2.5s 2 + 1 + 0.5s

2K c (1 − 0.5s )
= −1 (1.6)
2.5s 2 + 5.5s + 1

2K c − K c s = −2.5s 2 − 5.5s − 1 (1.7)

2.5s 2 + (5.5 − K c )s + 2K c + 1 = 0 (1.8)

Equation (2.8) can now be solved for the marginally stable system (where Kc
= Kcu). Marginal stability occurs when s = iω:

2.5(iω ) + (5.5 − K c )(iω ) + 2K cu + 1 = 0


2
(1.9)

− 2.5ω 2 + (5.5 − K c )iω + 2K cu + 1 = 0 (1.10)

Equating the imaginary components of (2.10):

(5.5 − K cu )ω = 0 (1.11)

K cu = 5.5 (1.12)

Equating the real components of (2.10):

− 2.5ω 2 + 2K cu + 1 = 0 (1.13)

ω = 0.8K cu + 0.4 (1.14)

2
Evaluating (2.14) at Kcu = 5.5:

ω = 0.8 × 5.5 + 0.4 = 2.19 (1.15)

Thus the ultimate gain is 5.5 and the ultimate frequency is 2.19. The ultimate
period is calculated as 2π/ω = 2.87. Using these values, the approximate Z-N
and T-L tuning constants are found to be:

Kc τI τD
Z-N 3.30 1.43 0.36
T-L 2.48 6.31 0.46

The approximate and exact tuning constants are in relatively good agreement.
It is important to note, however, that the approximated time delay has resulted
in a larger ultimate gain, and therefore a less robust controller. If the
approximation is sufficiently poor this could push the tuned controller to
become unstable.

The Simulink block diagram for the closed-loop system is depicted below.

1
t
5s+1
Clock
To Workspace Disturbance DisturbanceTF Scope

2
PID(s) C
5s+1
Set Point, R PID Controller Transport Process TF To Workspace1
Delay

The PID controller in Simulink is implemented as:

  =  + +  


3
Where P, I, and D are the proportional, integral, and derivative gains, and N is
the derivative filter coefficient (NB: N = ∞ provides the unfiltered response).
The following relationships relate these coefficients to the terms used above:

 =  = =  



Thus, the tuning constants for the four controllers become:

P I D
Z-N (exact) 2.56 1.38 1.18
Z-N (approx.) 3.30 2.31 1.19
T-L (exact) 1.92 0.24 1.13
T-L (approx.) 2.48 0.39 1.14

Note that the major difference between the two tuning relations is a larger
proportional gain with the Z-N settings and a much smaller integral gain with
the T-L settings.

The response of all four controllers to a unit set-point change applied at time t
= 0 and a unit disturbance (arbitrarily first order, τ = 5) at t = 25 are depicted
below (NB: the ‘choppy’ set-point tracking arises from the numerical
simulation, not the intrinsic dynamics).

4
The T-L controllers provide relatively good set-point tracking, but quite slow
disturbance rejection (a result of the small reset rate, 1/τI). The Z-N controllers
handle the disturbance better, but provide a lot of overshoot when responding
to the set-point change. For the Z-N controller settings obtained with the
approximate process model the overshoot and oscillations resulting from the
set-point change are particularly large, and the closed-loop system is unlikely
to be very robust (i.e., it could be very sensitive to process changes, which
could lead to control system instability). This highlights the need to select
both robust and fast-acting control settings. In cases where these two things
are unachievable, a more advanced form of single-loop control is probably
required.

5
2) The closed-loop transfer function for a set-point change is:

 =


(2.1)

 
 = 
 
(2.2)

The process and controller transfer functions are given by:

  = 


(2.3)

  =  1 +  

(2.4)


Substituting (3.3) and (3.4) into (3.2):

 % $ &

 =
!"#$ ! "
 $
 % &
(2.5)
!"#$ ! "

We will now attempt to rearrange the numerator into the form (as + 1).
Expanding terms:

 

 =
!"#$ ! "!"#$
$
  % &
(2.6)
!"#$ ! "

Giving the numerator terms a common denominator:

 '! "#$(

 =
! "!"#$
$
  % &
(2.7)
!"#$ ! "


  
Dividing the numerator and denominator by :

6
 =
 
 % $ &*+! "!"#$
) ,
(2.8)
!"#$ ! " 

The denominator can now be expanded into a second-order polynomial:

 = !"!"#$
 
 % $ &*+! "!"#$
) ,
(2.9)
 !"#$ ! " 

 = !"!"#$
 
$ $
) % &*-  .
(2.10)
 !"#$ ! "

 = !"!"#$
 
$
 % &
(2.11)
 ! "

 = !"!"#$
  
 
(2.12)


 = !! "/


 
! "
   
(2.13)
 

 = !! "/
 
 !
   
(2.14)
 

 = !! "/


 
$# 
  
(2.15)
 

We have now achieved a second-order transfer function with denominator in


the form τ2s2 + 2ζτs + 1. Note that this transfer function has a zero Keff(τIs+1)
in the numerator, with effective gain of 1. The presence of a zero alters the
process dynamics, but not the stability.

Equating the s2 terms:

011 = 
2 
(2.16)


7
011 = 3


(2.17)

Equating the s1 terms:

25011 011 = 


(2.18)

Substituting (2.17) into (2.18) for τeff:

25011 3 = 
 

(2.19)


$# 

5011 = 
!!
(2.20)
23


5011 = 6 8

7!!
(2.21)
2
7 

5011 = 9 :
 
2 
√
(2.22)
7 

5011 = 27 3


 /
(2.23)
 

5011 = 27 3 
 
(2.24)


8
and 2 in parallel we get the overall
$
$ 
3) (a) With transfer functions

controller transfer function (using the additive property of transfer functions):

 = 2 + 
$

(3a.1)
$

/ $  / $ $


 = + =
$
$   $ 
(3a.2)
$

/ $ $ / 
 =
$ 
(3a.3)

/ !$ " 
 =  + 2  $# /
$ 
(3a.4)

Note that 3a.4 is in the form of a lead-lag unit, like the filtered PD controller.

(b) Comparing the above result to the filtered PD controller   we get:


< 
=< 

> =  (3b.1)

 =
/ $
$ /
(3b.2)

 =  + 2 (3b.3)

We would like to obtain expressions for  ,  , and 2 in terms of  , > , and


 . From 3b.1 we get:

 = > (3b.4)

Substituting 3b.4 into 3b.2:

 =
/ =<
$ /
(3b.5)

9
1=
/ =
$ /
(3b.6)

 + 2 = 2 > (3b.7)

 = 2 > − 1 (3b.8)

Substituting 3b.3 into the above:

 =  −  > − 1 (3b.9)

 =  > − 1 −  > − 1 (3b.10)

 +  > − 1 =  > − 1 (3b.11)

> =  > − 1 (3b.12)

 =  
=@
=
(3b.13)

Finally, introducing 3b.13 into 3b.3 provides us with:

2 =  −  =  − 
=@
=
(3b.14)

2 =  %1 − &
=@
=
(3b.15)

2 =

=
(3b.16)

So, the required constans can be calculated as  = > ,  =  , and


=@
=

2 = . Note that the derivative time only appears in the calculation of  ,



=

whereas > appears in all three tuning constants.

10
4) (a) The figure below shows how we determine the gain margin (blue lines),
GM = 1/0.4 = 2.5, and the phase margin (red lines), PM = 55°. The gain
margin occurs at the critical frequency, ωc; the phase margin occurs at the
gain crossover frequency, ωg.

The critical frequency is around 0.5 rad/min; the gain crossover frequency is
ca. 0.2 rad/min.

102

101
AR

100

GM ~ 2.5
10-1

10-2
0

-90
φ (degrees)

PM ~ 55o
-180

-270

-360
0.0001 0.0010 0.0100 0.1000 1.0000 10.0000

ω (rad/min)

(b) Including a controller with Kc = 1 provides us with the same open-loop


transfer function as Gp alone, so the GM and PM determined above still apply.

The GM of 2.5 indicates that the open-loop system would be close-loop stable
with Kc = 1 (and critically stable when Kc = 2.5). As such, we would need to

11
add time delay to the system to make it critically stable; enough time delay to
give PM = 0°. We term this amount of phase lag ∆θmax. The equation to
calculate this is:

ΔBCDE = ×
FG K
HI LM

For PM = 55° and ωg = 0.2 rad/min we get ∆θmax = 4.8 min. The above
equation comes from rearranging the equation for time delay phase angle, π =
–ωθ. There are two properties that allow us to apply this: (i) In an nth order
plus time delay system the phase angle contributions are additive, and (ii) the
time delay has no effect on the amplitude ratio. Also, note that the term of π
radians per 180° is required to convert the phase margin from degrees to
radians, so that it cancels out units in the critical frequency.

(c) There are a number of viable controllers and tuning constants. First, we
need to select a type of controller. Integral action would be required for this
process, as it would be important to eliminate offset between the controlled
variable and the set-point. PID controllers are commonly employed for
temperature control, although PI or derivative-filtered PID might be used
instead if the temperature signal is particularly noisy. To fully specify the
controller, tuning constants are also required. These could come from (for
example) Z-N or T-L tuning relations, or from a dynamic simulation. It is also
important to consider whether or not simple feedback control is going to be
appropriate, or if an advanced control strategy should be implemented. In this
case the gain and phase margins are relatively wide, and it is unlikely that this
will be a particularly difficult control problem.

12
CHEN90032
Process Dynamics and Control

Worksheet 9 – Closed-Loop Stability

Systems that are open-loop stable can give unbounded responses when
under negative feedback control. For chemical processes, this can result in a
disturbance leading to potentially unsafe operating conditions. Furthermore,
the potential for instability places limits on the controller settings and response
time, significantly complicating the control problem. It is for these reasons that
we must be aware of the fundamental dynamic behavior that leads to closed-
loop instability, and familiar with techniques to determine stability limits

This set of problems deals with the stability of negative feedback control
loops. The first problem employs the Routh stability criterion to evaluate if a
particular closed-loop transfer function would yield a stable control response.
In the second problem the Routh stability criterion is again implemented,
although in this case the controller gain (Kc) is unspecified, and the stability
limits need to be determined. The final problem makes use of the Bode
stability criterion, which allows us to obtain information about closed-loop
stability from the open-loop frequency response characteristics for a system.

W9-1
1) Using the Routh stability criterion, determine if the following closed-loop
transfer function is stable:

s 2 + 100s + 1
G( s ) =
s 4 + 10s 3 + 100s 2 + 10s + 1

 Simulate the dynamic behavior of the above transfer function to confirm


that it is or is not stable.

2) A process is described by the following open-loop transfer function (GOL).


Determine the stability limits on Kc using a Routh array.

5K c
GOL (s ) =
( s + 3 )( s + 4 )( s + 6 )

 Simulate the control system response using the upper and lower bounds
on Kc, as well as intermediate values. What is the significance of the lower
stability limit?

3) The feedback control system in Figure 1 has the following transfer


functions. Determine the open-loop transfer function, and make a comparison
to the Bode diagram provided (Figure 2). (a) Calculate the value of Kc that
provides a phase margin of 30°. (b) What is the gain margin when Kc = 10?

 2s + 1  2
G c (s ) = K c   Gv (s ) =
 0.1s + 1 0.5s + 1

0. 4 3
G p (s ) = G l (s ) =
s (5s + 1) 5s + 1

W9-2
Ql(s) Gl(s)

+
+ +
R(s) Gc(s) Gv(s) Gp(s) C(s)
-

Figure 1: Process block diagram for Question 3.

Figure 2: Open-loop transfer function Bode diagram.

W9-3
CHEN90032
Process Dynamics and Control

Worksheet 9 – Closed-Loop Stability – SOLUTION

1) The transfer function denominator determines stability, so we are


interested in the polynomial:

GOL + 1 = s 4 + 10s 3 + 100s 2 +10s + 1 = 0

The Routh array will have five rows (fourth-order), and will look like:

1 a4 a2 a0
2 a3 a1 0
3 b1 b2 0
4 c1 0 0
5 d1 0 0

We have a fourth-order polynomial with a4 = 1, a3 = 10, a2 = 100, a1 = 10, a0 =


1.

The remaining constants are given by:

a3a2 − a4a1 1000 − 10


b1 = = = 99
a3 10

b2 = a0 = 1

b1a1 − a3 b2 990 − 1
c1 = = ≅ 9.9
b1 99

d 1 = a0 = 1

The Routh array becomes that shown below. The first column is all positive,
and the system is therefore stable.

1
1 1 100 1
2 10 10 0
3 99 1 0
4 9.9 0 0
5 1 0 0

The following diagram shows a Simulink block diagram for this transfer
function responding to a unit set-point change:

t
Clock To Workspace Scope

s2 +100s+1
C
den(s)
Set Point, R Process TF To Workspace1

The process response is shown below. Despite a large overshoot, the


response is still closed-loop stable.

2
2) For the characteristic equation 1 + GOL = 0:

5K c
1 + GOL (s ) = 1 + =0
( s + 3 )( s + 4 )( s + 6 )

5K c
= −1
( s + 3 )( s + 4 )( s + 6 )

5Kc = − ( s + 3 )( s + 4 )( s + 6 )

( s + 3 )( s + 4 )( s + 6 ) + 5Kc = 0

Expanding:

(s 2
)
+ 7s + 12 ( s + 6 ) + 5K c = 0

(s 2
)
+ 7s + 12 ( s + 6 ) + 5K c = 0

s 3 + 13s 2 + 54s + 72 + 5K c = 0

For a third-order polynomial the Routh array is:

1 a3 a1
2 a2 a0
3 b1 0
4 c1 0

The constants are:

a2 a1 − a3 a0 13 × 54 − (72 + 5K c )
b1 = = = 48.46 − 0.38K c
a2 13

c1 = a0 = 72 + 5K c

3
The Routh array becomes:

1 1 53
2 13 72 + 5K c
3 48.46 − 0.38K c 0
4 72 + 5Kc 0

The system will become unstable for (i) 48.46 - 0.39Kc < 0, or Kc < 124.3, and
(ii) 5Kc + 72 < 0, or Kc > -14.4. So, the system is stable for controller gain
values of:

-14.4 < Kc < 124.3

(Actually, ultimate gain is 126 when the calculations aren’t rounded…)

The Simulink block diagram for this system is depicted below. A unit
disturbance is applied at time t = 1.

t
Clock To Workspace Disturbance, Ql Scope

5
-K- C
s3 +13s2 +54s+72
Set Point, R Kc Process TF To Workspace1

The process response obtained with Kc = +126, +63, -7.2, and -14.4 is plotted
in the diagram below. Note that the dotted line depicts the open-loop
response.

4
With the largest gain (+126) the process is indeed seen to be critically stable.
When a value of +63 is used instead (i.e., GM = 2) the response is stable,
although still quite oscillatory.

Using the largest negative controller gain that provides a critically stable
response, -14.4, we see that the system responds as a ramp, with the initial
disturbance ever-growing. With a smaller negative controller gain (-7.2) the
output eventually reaches a steady-state value, but the initial disturbance has
still been amplified. With a negative controller gain the controller is forcing the
manipulated variable to act in the opposite direction to what is actually
required, and even though the response might be stable it is still undesirable.
For a reverse-acting process, where an increase in the manipulated variable
results in a decrease in the controlled variable, we would actually need to
select a negative value for the controller gain to provide the desired response.
In short, before performing the stability analysis we should be aware of
whether we are after a positive or negative value of Kc.

5
3)
(a) The phase margin is related to φg, the phase angle at the gain crossover
frequency (ωg), by the following equation:

PM = 180 + φ g (3.1)

When PM = 30°, φg = -150°. From the Bode diagram at this phase angle the
frequency ωg is 1.72 rad/sec. On the Bode diagram the amplitude ratio
divided by the controller gain (AR/Kc) at ωg = 1.72 rad/sec is 0.144. Because
AROL = 1 at ωg, we find that the controller gain Kc that corresponds to a phase
margin of 30° is 6.94.

6
(b) For the gain margin, we are interested in the critical frequency (ωc), which
occurs at a phase angle of -180°. From the Bode diagram we determine ωc to
be 4.05 rad/s. The amplitude ratio at ωc is 0.0326. When Kc = 10 the critical
amplitude (ARc) is 0.326. The gain margin, 1/ARc, is therefore 3.07.

7
CHEN90032
Process Dynamics and Control

Worksheet 10 – Advanced Single-Loop Control

We have seen that for single control loops, in which one controlled variables is
paired with one unique manipulated variable, conventional feedback control
(often using PID control algorithms) provides a satisfactory closed-loop
response in many cases. There are, however, a number of circumstances
where feedback control is not sufficient, and here we need to look to
advanced single-loop control strategies. Cases where this might apply
include processes with slow dynamics and/or large time delays, or systems
that are particularly prone to disturbances, either known or unknown.

This worksheet presents several problems associated with different advanced


forms of control. The first problem examines the block diagram representation
and transfer function for a cascade controller, commonly used to handle
disturbances in manipulated variables. The second problem presents a form
of controller designed to reduce the effects of time delays on PI control. In
question 3 you are asked to derive a feedforward controller, which can be
used to handle a specific process disturbance when an accurate process
model is known. The final two problems are concerned with the direct
synthesis approach to controller design, which uses a process model and a
control objective to define the control algorithm.

W10-1
1) Derive the transfer function relating the controlled variable C(s) to changes
in the set-point R(s) for the following cascade control block diagram:

Ql2 Ql1
Gl2 Gl1

+ +
R + E + + + C
Km1 Gc1 Gc2 Gv Gp2 Gp1
- -

Gm2

Gm1

 τIs + 1 
2) A feedback controller of the form Gc = K c  −θ s 
has been
 τ I s + 1− e 
proposed as a time-delay compensator. Show that this controller eliminates
the time delay term from the characteristic equation for the FOPTD process
 e −θ s 
Gp =   when Kc = 1/Kp and τ = τI.
 τ s + 1

3) Design the ideal feedforward controller for a process with transfer function
−θ p s
K pe K l e −θ l s
Gp = and disturbance transfer function Gl = . Comment on
τ ps + 1 τls +1

the feedforward controller when θp > θl and θp < θl.

W10-2
4) Design a controller for the third-order process
K 1
G( s ) = , where the control objective is Gobj (s ) = .
(τ 1s + 1)(τ 2 s + 1)(τ 3 s + 1) τ mc s + 1
What is the effective controller gain? How would this controller be
implemented?

2e −0.2s
5) For a process described by the transfer function G(s ) = use the
s +1
1 e −θ s
direct synthesis method with Gmc (s ) = to design a controller when
G (τ mc + θ ) s

τmc = 0.2 and 1.0. What are the tuning constants for the resultant PI
controller?

 Simulate the control system response using a PI controller and the


settings derived above. How does the set-point response compare to the
implied control objective?

W10-3
CHEN90032
Process Dynamics and Control

Worksheet 10 – Advanced Single-Loop Control – SOLUTION

1) Because we are interested in the set-point transfer function, we can set Ql1
and Ql2 to zero:

R + E + C
Km1 Gc1 Gc2 Gv Gp2 Gp1
- -

Gm2

Gm1

The transfer function for the internal feedback loop is:

Gc 2Gv G p 2
(1.1)
1 + Gc 2Gv G p 2Gm 2

We can now reduce the block diagram to:

R + E Gc1Gc 2Gv G p1G p 2 C


Km1
1 + G c 2G v G p 2G m 2
-

Gm1

1
The controlled variable can be expressed as:

Gc1Gc 2GvG p 1G p 2
C=E (1.2)
1 + Gc 2Gv G p 2Gm 2

The error signal E is:

E = RK m1 − CGm1 (1.3)

Substituting (1.3) into (1.2):

K m1Gc1Gc 2GvG p 1G p 2 Gm1Gc1Gc 2GvG p 1G p 2


C =R −C (1.4)
1 + Gc 2Gv G p 2Gm 2 1 + Gc 2Gv G p 2Gm 2

Grouping the C and R terms we can obtain the transfer function C/R:

K m1Gc1Gc 2GvG p 1G p 2
C 1 + Gc 2Gv G p 2Gm 2
= (1.5)
R Gm1Gc1Gc 2GvG p 1G p 2
1+
1 + Gc 2Gv G p 2Gm 2

K m1Gc1Gc 2GvG p 1G p 2
C 1 + Gc 2Gv G p 2Gm 2
= (1.6)
R 1 + Gc 2Gv G p 2Gm 2 + Gm1Gc1Gc 2GvG p 1G p 2
1 + Gc 2Gv G p 2Gm 2

C K m1Gc1Gc 2GvG p 1G p 2
= (1.7)
R 1 + Gc 2Gv G p 2Gm 2 + Gm1Gc1Gc 2GvG p 1G p 2

This is the set-point transfer function for cascade control.

2
2) The transfer function for set-point changes (for example) is:

C GcGp
G= = (2.1)
R 1 + GcGp

Where:

 e −θ s 
Gp =   (2.2)
 τ s + 1

1  τs +1 
Gc = (2.3)
K p  τ s + 1 − e −θ s 

The term GcGp in (2.1) becomes:

 e 
−θ s
K  τs +1
GcGp = p    τ s + 1 (2.4)
K p  τ s + 1 − e −θ s  

 e −θ s 
GcGp =  −θ s 
(2.5)
 τ s + 1− e 

Substituting (2.5) into (2.1):

 e −θ s 
 −θ s 
τ s + 1− e 
= 
C
(2.6)
R  e −θ s 
1+  −θ s 
 τ s + 1− e 

 e −θ s 
 −θ s 
C
=  τ s + 1− e  (2.7)
R  τ s + 1 − e −θ s + e −θ s 
 −θ s 
 τ s + 1− e 

3
C e −θ s
= (2.8)
R τs +1

Thus, the time-delay term e-θs has been eliminated from the characteristic
equation for the closed-loop transfer function.

3) An ideal feedforward controller is given by the transfer function:

Gl
Gffc = − (3.1)
Gp

In this problem, the ideal feedforward controller is:

 K l e −θl s 
 
τ s +1
Gffc = −  l −θ s  (3.2)
 K pe p 
 
 τ ps + 1 
 

 e −θl s 
 K l   τ l s + 1 
Gffc = − 
 K   K e −θps 
(3.3)
 p p
 
 τ ps + 1 
 

 K   τ s + 1  e −θl s 
Gffc = −  l   p (3.4)
 K p  τ l s + 1  e −θps 
  

 K   τ s + 1  (θ p −θl )s
Gffc = −  l   p e (3.5)
 Kp  τls + 1 
   

When θp > θl the time delay term in (3.5) is e+θs, and the controller cannot be
physically realised. When θl > θp the term becomes e-θs and the feedforward
controller can be achieved.

4
4) In a negative feedback control loop the model based controller Gmc is
related to the process transfer function G and the control objective Gobj by:

1  Gobj (s ) 
Gmc (s ) = (4.1)
G(s )  1 − Gobj (s ) 

Introducing the process and control objective transfer functions into (4.1), then
rearranging:

 1 
 
Gmc (s ) =
(τ 1s + 1)(τ 2 s + 1)(τ 3 s + 1)  τ mc s + 1 
(4.2)
K  1 
1− 
 τ mc s + 1 

 1 
 
Gmc (s ) =
(τ 1s + 1)(τ 2 s + 1)(τ 3 s + 1)  τ mc s + 1 
(4.3)
K  τ mc s + 1 − 1 
 
 τ mc s + 1 

(τ 1s + 1)(τ 2s + 1)(τ 3 s + 1)  1 
Gmc (s ) =
K  τ s  (4.4)
 mc 

Gmc (s ) =
(τ 1s + 1)(τ 2s + 1)(τ 3 s + 1) (4.5)
Kτ mc s

Gmc (s ) =
(τ τ s
1 2
2
)
+ τ 1s + τ 2 s + 1 (τ 3 s + 1)
(4.6)
Kτ mc s

τ 1τ 2τ 3 s 3 + τ 1τ 2 s 2 + τ 1τ 3 s 2 + τ 2τ 3 s 2 + τ 1s + τ 2 s + τ 3 s + 1
Gmc (s ) =
Kτ mc s
(4.7)

Grouping together the different sn terms:

5
Gmc (s ) =
(τ 1τ 2τ 3 )s 3 (τ 1τ 2 + τ 1τ 3 + τ 2τ 3 )s 2 (τ 1 + τ 2 + τ 3 )s
+ + +
1
(Kτ mc )s (Kτ mc )s (Kτ mc )s (Kτ mc )s
(4.8)

τ 1τ 2τ 3 2 τ 1τ 2 + τ 1τ 3 + τ 2τ 3 τ +τ2 +τ3 1
Gmc (s ) = s + s+ 1 +
Kτ mc Kτ mc Kτ mc (Kτ mc )s
(4.9)

We find that there are terms s-1, s0, s1, and s2. These correspond to the
standard PID elements, but with s2 also included – this is a second derivative
in the time domain. This indicates that we need to consider the second
derivative of the error signal if we want to control a third order process with a
response that looks like a first order process. This kind of control system
performance is therefore unattainable using a standard PID controller. More
advanced hardware would be required to implement this control system [I
have never heard of anyone implementing a second-derivative controller in
real life, but I imagine it would be unwise].

The effective controller gain is the constant (i.e., P) term in (4.9): Kc =


(τ1+τ2+τ3)/Kτmc.

5) Substituting the process transfer function and θ = 0.2 into Gmc(s) we get:

s +1 e −0.2s
Gmc (s ) = (5.1)
2e −0.2s (τ mc + 0.2 ) s

0.5 ( s + 1)
Gmc (s ) = (5.2)
(τ mc + 0.2 ) s

0.5s 0.5
Gmc (s ) = + (5.3)
(τ mc + 0.2) s (τ mc + 0.2 ) s

6
0.5 0.5
Gmc (s ) = + (5.4)
(τ mc + 0.2 ) (τ mc + 0.2 ) s

0.5  1
Gmc (s ) = 1+ (5.5)
(τ mc + 0.2)  s 

0.5
Equation (5.5) is in the form of a PI controller, with K mc = and τI =
(τ mc + 0.2 )
1.

When τmc = 0.2 we obtain the tuning constants Kmc = 1.25 and τI = 1. When
τmc = 1 the constants are Kmc = 0.42 and τI = 1.

The Simulink block diagram developed to simulate the direct synthesis PI


controller is depicted below (an arbitrary transfer function has been used to
describe the disturbance dynamics):

1
s+1
Load Disturbance TF C

2
PI(s) C
s+1
Set Point DS Controller Time Delay Process TF To Workspace

Simulations were performed using both sets of tuning constants derived


above, with a unit set-point change applied at time zero and a unit disturbance
applied at t = 10. Results obtained with both controllers are illustrated in the
following figure.

7
Note that the smaller value of τmc provides the faster control response, which
is consistent with the larger controller gain calculated above.

In this problem the control objective for set-point tracking implied by the form
of the model-based controller is first-order plus time delay, with the first term
of a Taylor series expansion used for e-θs in the denominator (i.e., τs + 1 - e-θs
becomes [τ + θ]s). From the set-point response with the less-aggressive
control settings the control objective is essentially met – this looks like a
classic first order plus time delay response (if it was second order or higher it
would be S-shaped). The time delay of 0.2 minutes is also observed. The
time constant for the process response, which should be 1.2 (τ + θ), looks to
be more like 2, and this is a result of the pure FOPTD objective being
physically unobtainable (which is why we required an approximation for time
delay in the controller denominator). At the more aggressive setting of τmc =
0.2 we see that there is some overshoot when tracking a set-point change.
This clearly violates the control objective, and results from the system
approaching its stability limit (NB: Kcu ~ 4.3).

8
CHEN90032
Process Dynamics and Control

Worksheet 11 – Multiloop Control Systems

In practice, most unit operations and chemical processes represent multiloop


control problems, where there are multiple controlled variables, which are
influenced by numerous manipulated variables. In a multiple-input multiple-
output (MIMO) control system, as well as the direct effect that a manipulated
variable has on its paired control variable, we also need to consider the
impact of process interactions – the unwanted effects of the remaining
manipulated variables on our controlled variable of interest.

When dealing with a multiloop control system, there are several issues to deal
with that we do not find in a single-loop control configuration. Process
interactions can compromise control system stability, degrade the controller
response, and complicate controller tuning. We also need techniques to help
us pair controlled and manipulated variables, now that there are multiple
options available. The first question here considers the 1-2/2-1 C-M pairing in
a 2×2 MIMO control system, looking at the characteristic equation, which is
needed in order to analyse the closed-loop stability. In the second and third
problems we use steady-state process information to help us decide on C-M
pairings, and in question 4 we examine an advanced control strategy to
reduce process interactions in MIMO systems.

W11-1
1) Derive the characteristic equation for the 1-2/2-1
1 1 2×2 MIMO control
configuration (shown below), and compare it to the 1-1/2-2
1 2 characteristic
equation.

2) A distillation column uses the temperatures at trays 17, 24, 30 (analogs of


composition) as controlled variables, where the steam pressure in the reboiler
(m1) and the reflux ratio (m
( 2) are the manipulated variables. The following
steady-state relationships
hips have been derived between the tray temperatures
and the manipulated variables (in deviation form):

T17 = 1.5m1 + 0.5m2

T24 = 2.0m1 + 1.7m2

T30 = 3.4m1 + 2.9m2

Use a relative gain array to select pairings of controlled and manipulated


variables in a 2×2 control system (note: there are three possible combinations
of controlled and manipulated variables). Use the Niederlinski index to test for
control system instability.
bility.

W11-2
3)
(a) A two-tank pH neutralisation process can be described using the following
4 × 4 relative gain array, where hn and pHn refers to the liquid level and pH in
tank n. The manipulated variables q1 – q4 are liquid flowrates into the two
neutralisation tanks.

Suggest pairings of controlled and manipulated variables for a 4 × 4 MIMO


control system.
q1 q2 q3 q4
h1  0.64 0.72 −0.20 −0.20 
pH1  0.87 0.85 −0.35 −0.35 
h2  −0.18 −0.21 0.70 0.70 
 
pH2  −0.36 −0.37 0.85 0.88 

(b) pH2 and h2 are to be controlled solely by q3 and q4. Develop a relative
gain array for this 2 × 2 MIMO process. Recommend pairings of controlled
and manipulated variables based on steady-state and dynamic
considerations. The relevant process transfer functions are:
h2 0.42e −0.4s h2 0.41e −0.1s
= =
q3 3.32s + 1 q4 2.07s + 1

pH2 −0.32e −0.8s pH2 0.32e −0.4s


= =
q3 2.36s + 1 q4 2.03s + 1

 Develop block diagram representations for both controller configurations.


Simulate the control system response to set-point changes in both
manipulated variables, and using this numerical model attempt to arrive at
effective tuning constants for both configurations. Compare your findings to
the recommended pairings from the RGA. Suggest ways to improve the
control system performance.
[Hint: First tune each controller in a single-loop configuration, then de-tune the
controllers in the MIMO setups to arrive at a satisfactory 2×2 closed-loop
response.

W11-3
4) In a blending process, liquid level h is controlled by outlet flow-rate q3, and
outlet concentration c3 is controlled by inlet flow-rate q1. Derive expressions
for the two ideal decouplers, T21 and T12. The transfer function array in this 2
× 2 MIMO system is:

 −0.0212 0.0212 
 s ( 0.167s + 1) s ( 0.167s + 1) 
 
 0.0011 
 0
 (1.06s + 1)( 0.167s + 1) 

W11-4
CHEN90032
Process Dynamics and Control

Worksheet 11 – Multiloop Control Systems – SOLUTION

1) The characteristic equation is the denominator of any of the four transfer


functions C1/R1, C1/R2, C2/R1, or C2/R2. We will look at the case of C1/R1, so
we can set R2 = 0.

From the block diagram we can derive the following relationships for C1 and
M2:

C1 = M1Gp11 + M 2Gp12 (1.1)

M2 = E1Gc1 = ( R1 − C1 ) Gc1 (1.2)

Substituting (1.2) into (1.1) gives:

C1 = M1Gp11 + ( R1 − C1 ) G c1Gp12 (1.3)

The following relationships are obtained for M1 and C2:

M1 = E2Gc 2 = −C2Gc 2 (1.4)

C2 = M2Gp 22 + M1Gp 21 (1.5)

Substituting (1.5) into (1.4):

M1 = −M2Gp 22Gc 2 − M1Gp 21Gc 2 (1.6)

Introducing (1.2) for M2:

M1 = −Gp 22Gc1Gc 2 ( R1 − C1 ) − M1Gp 21Gc 2 (1.7)

1
M1 + M1Gp 21Gc 2 = −Gp 22Gc1Gc 2 ( R1 − C1 ) (1.8)

M1 (1 + Gp 21Gc 2 ) = −Gp 22Gc 1Gc 2 ( R1 − C1 ) (1.9)

Gp 22Gc1Gc 2
M1 = − ( R1 − C1 ) (1.10)
1 + Gp 21Gc 2

Introducing (1.10) into (1.3):

Gp11Gp 22Gc 1Gc 2


C1 = − ( R1 − C1 ) + G c1Gp12 ( R1 − C1 ) (1.11)
1 + Gp 21Gc 2

Gp11Gp 22Gc1Gc 2 Gp11Gp 22Gc 1Gc 2


C1 + G c 1Gp12C1 − C1 = − R1 + G c1Gp12R1
1 + Gp 21Gc 2 1 + Gp 21Gc 2

(1.12)

 G G G G   G G G G 
C1  1 + G c1Gp12 − p11 p 22 c1 c 2  = R1  G c 1Gp12 − p11 p 22 c1 c 2 
 1 + Gp 21Gc 2   1 + Gp 21Gc 2 
 
(1.13)

 Gp11Gp 22Gc1Gc 2 
 G c 1Gp12 − 
C1  1 + Gp 21Gc 2 
= (1.14)
R1  Gp11Gp 22Gc1Gc 2 
 1 + G c1Gp12 − 
 1 + Gp 21Gc 2 

 G c1Gp12 (1 + Gp 21Gc 2 ) − Gp11Gp 22Gc1Gc 2 


 
C1  1 + Gp 21Gc 2 
=   (1.15)
R1  (1 + G c1Gp12 )(1 + Gp 21Gc 2 ) − Gp11Gp 22Gc1Gc 2 
 
 1 + Gp 21Gc 2 
 

2
C1 G c1Gp12 (1 + Gp 21Gc 2 ) − Gp11Gp 22Gc 1Gc 2
= (1.16)
R1 (1 + G c1Gp12 )(1 + Gp 21Gc 2 ) − Gp11Gp 22Gc 1Gc 2

From equation (1.16) we get the characteristic equation for the 1-2/2-1
configuration of:

(1 + G Gp12 )(1 + Gc 2Gp 21 ) − Gp11Gp 22Gc 1Gc 2 = 0


c1 (1.17)

For comparison, in the 1-1/2-2 configuration the characteristic equation is:

(1 + G Gp11 )(1 + Gc 2Gp 22 ) − Gp12Gp 21Gc 1Gc 2 = 0


c1 (1.18)

2) The three possible selections of controlled and manipulated variables are:


(i) T17 and T24, m1 and m2, (ii) T17 and T30, m1 and m2, and (iii) T24 and T30, m1
and m2.

Because the RGA relies only on steady-state information, we can select


controlled and manipulated variables based solely on steady-state gains. In a
2×2 RGA, the first element is given by:

1
λ11 =
K K
1 − 12 21
K11K 22

(i) T17 (c1) and T24 (c2), m1 and m2

The gain array for case (i) is:

 1.5 0.5 
 
 2.0 1.7 

3
Using the above equation we obtain λ11 = 1/[1-(0.5×2.0)/(1.5×1.7)] = 1.65.
The RGA is therefore:

 1.65 −0.65 
 
 −0.65 1.65 

The pairing of T17 with m1 and T24 with m2 (i.e., 1-1/2-2) is the best choice
here.

(ii) T17 (c1) and T30 (c2), m1 and m2

For case (ii) we get the following RGA:

 1.64 −0.64 
 
 −0.64 1.64 

The 1-1/2-2 pairing of T17 with m1 and T30 with m2 is the best choice, and is
essentially indistinguishable from the 1-1/2-2 pairing in case (i).

(iii) T24 (c1) and T30 (c2), m1 and m2

This selection of controlled and manipulated variables results in the RGA:

 290 −289 
 
 −289 290 

The best pairing in this instance is T24 with m1 and T30 with m2, but this is a
poor selection due to the very large relative gains, and therefore the large
controller gains that would be required to elicit an effective control response.

4
Finally, we will use the Niederlinski index (NI) to determine if the controller-
manipulated variable pairings from cases (i) – (iii) are potentially unstable.
The NI for a 2 × 2 system is:

K11K 22 − K12K 21
NI =
K11K 22

For (i) we find that NI = 0.61, for (ii) NI = 0.61, and for (iii) NI = 0.003. All three
cases pass the NI test, although NI for case (iii) is very close to becoming
negative, and small changes in any of the process gains could make the
control system unstable.

3)
(a) Considering the RGA, we find the best pairings to be: h1-q2, pH1-q1, h2-q3,
pH2-q4.

(b) When the two controlled variables are h2 and pH2, and the two
manipulated variables are q3 and q4, we get the following gain array:

 0.42 0.41
 
 −0.32 0.32 

1
Using λ11 = the RGA is:
K12K 21
1−
K11K 22

 0.506 0.494 
 
 0.494 0.506 

Based on steady-state considerations the optimum pairing is 1-1/2-2, or h2


with q3 and pH2 with q4. Considering process dynamics, the pairing of pH2
with q3 is unwanted because of the relatively large time delay (time constants
are all similar); this also supports a 1-1/2-2 pairing.

5
In this example the RGA suggests that there is relatively little difference
between the two controller configurations. There is not much difference
between the process transfer functions in terms of dynamics either. In this
instance a numerical simulation of the system can be very instructive, and
may avoid the expense of implementing the two controller configurations to
see which is best (or picking one and hoping). It also provides a good starting
point for tuning constant selection.

We begin by examining the dynamics of each control loop individually. Four


single feedback loops can be constructed, each with their own ultimate gain
(Kcu) and ultimate period (Pu). Using T-L PID controllers (or whatever you
prefer…), SISO controller settings can be calculated, which are seen to
provide good set-point tracking and disturbance rejection (response times on
the order of a minute – see Simulink file available online). Note here that the
2-1 control pairing requires a reverse acting controller, with negative controller
gain.

Kcu Pu Kc (T-L) τI (T-L) τD (T-L)


[-] [min] [-] [min] [min]
1-1 33 1.4 14.9 3.1 0.2
1-2 82 0.35 36.9 0.8 0.1
2-2 27 1.3 12.2 2.9 0.2
2-1 -16.5 2.4 -7.4 5.3 0.4

When the tuned single-loop controllers are used in the multiloop control
configuration the control response becomes highly oscillatory, and is
operating close to the stability limits (in both 1-1/2-2 and 1-2/2-1). The
controllers in the MIMO setup need to be de-tuned.

So that we don’t have 6 independent parameters to vary for each


configuration, we will de-tune the controllers by multiplying all constants (P, I,
D in the Simulink parlance) by the same de-tuning factor, FT (note that if you
are using integral time, τI, you would need to divide by FT).

6
In both cases, FT = 0.5 seems to provide a good compromise between
response time and robustness. The following figure illustrates the effect of
unit set-point changes in C1 at 0 min and in C2 at 25 min, for both the 1-1/2-2
and 1-2/2-1 configurations (Simulink file available online). We see that the
time required to reach steady-state is now around 5 to 10 minutes, which may
be reasonable but is certainly much slower than in the SISO set-ups. It is hard
to pick between 1-1/2-2 and 1-2/2-1, but the former probably performs better
for C2, the latter for C1.

Finally, if the tuned MIMO controllers were not acting fast enough there are
several options to improve performance. We could investigate decouplers, or
perhaps some form of time delay compensation. If one controlled variable
was more important than the other we could selectively de-tune only one of
the controllers – remembering that in this example C1 is liquid level and C2 is
pH, we might be able to allow level to ‘float’ between acceptable values, and
more tightly regulate the tank pH.

7
4) The ideal decoupler T12 = -Gp12/Gp11. Substituting in the transfer functions
we get:

0.0212
s ( 0.167s + 1)
T12 = −
0.0212

s ( 0.167s + 1)

Similarly, for T21:

0
T21 = −
0.0011
(1.06s + 1)( 0.167s + 1)

The decouplers T12 and T21 are +1 and 0, respectively.

You might also like