PC Lessons On Process Control

10.
450 Process Dynamics, Operations, and Control

Lecture Notes - 1
Lesson 1. Design of a surge tank to smooth out fluctuations in flow. Definition of
important process control terms.
1.0 Context
Much of the chemical engineering curriculum concerns continuous
processes operating at steady state. Well and good, but there's more to it:
continuous processes may be disturbed in a variety of ways, and the
effects propagate through the process as a function of time – throughout
the process, temperature, pressure, flow, and composition may rise or fall.
Process Control is about managing disturbances, for product quality, for
economics, for safety. We begin with a simple example:
1.1 Surge tank

Envision two continuous processes operating in series. Process 1 feeds a
stream wi to process 2.
from other processes
wi
process 1 process 2
to other processes
Stream wi has some steady design flowrate, but in practice it varies,

causing problems in process 2. We might attack the problem by reducing
the cause of variation in process 1; we might also attempt to mitigate the
effect of the variation on process 2. Propagation of disturbances between
processes is a common problem, and a common solution is the surge tank;
its job is to damp out changes in wi from the upstream process and thus
deliver a steadier wo to the downstream process.
from process 1
wi
to process 2
wo
1
10.450 Process Dynamics, Operations, and Control
Lecture Notes - 1
Notice that the inlet flow is unconstrained, and the outlet flow is pumped.
Because the surge tank itself is not a steady-state process, IN does not
equal OUT, and thus the liquid level will vary with time.
1.2 Designing the tank

We might approach the design by
(1) characterizing the inlet stream
(2) specifying limits on the outlet stream
(3) making a suitable process model to connect inlet with outlet
Stated in this way, it's not so different from approaching a steady state
process design.
Inlet: suppose that the flow swings by ±40% over a 20-minute period.
2πt
wi = 10,000 + 4000sin (1.2.1)
20
where flow is in kg h-1 and time in minutes. Of course, real data would be
more messy, but that's just a matter of detail. One can learn a lot about the
system by simplifying to the essential features.
Outlet: the desired flow wo is 10,000 kg h-1. Let's be hard-nosed and insist
on no variation.
Process model: by instinct a chemical engineer writes a material balance.
dh
ρA = wi − wo h(0) = ho
dt
t (1.2.2)
1
(wi (t) − wo (t) )dt
ρA ∫0
h(t) = ho +
This is the process model; it describes how level varies with time as wi
changes. Substituting (1.2.1) into (1.2.2) and integrating, we find
212  2πt 
h(t) = ho + 1 − cos  (1.2.3)
ρA  20 
where height is in m, area in m2, time in minutes, and density in kg m-3.

To complete the design, we must choose the tank cross-sectional area A;
spending more money for a larger tank will reduce the amplitude of the
level variations.
2
Lecture Notes - 1
This tank has a free inlet and a pumped outlet; intuitively it seems possible
that the tank may overflow or run dry. We can confirm this sobering
thought by applying our tank model (1.2.2) to a persistent imbalance
between wi and wo. Suppose the simple case of
wi − wo = C (1.2.4)
Substituting (1.2.4) into (1.2.2), we find
C
h(t) = ho + t (1.2.5)
ρA
Therefore, we should protect our tank with some automatic process

control – something that will measure the level and take corrective action
should the level become too high or low.
1.3 Definitions to get us started

process: the equipment within some boundary, along with the streams of
matter and energy that cross that boundary -- what we usually
mean when we think of 'chemical process'. In this example, it's the
tank, pump, piping, and fluid.
disturbance: a change imposed on the process. In this example, the input
stream wi varies with time.
controlled variable: some feature of the process that we would like to
control. It may be a stream crossing the boundary or some
quantity within. We want to control it because the disturbance
makes it change with time, in a way that we don't like. In this
example, the controlled variable is the liquid level.
set point: the desired value of the controlled variable. In this example, we
have no set point, but we do want to confine the controlled variable
between high and low limits.
manipulated variable: some feature of the process that we adjust so that
we can exert influence on the controlled variable. In most
chemical processes, the manipulated variable will be the flowrate
of a stream. In this example, it seems reasonable to manipulate the
outlet flow.
final control element: a device that adjusts the manipulated variable. If the
manipulated variable is usually a stream, the final control element
is usually a valve, referred to as a control valve.
measured variable: most often, synonymous with the controlled variable –
we measure it so that we can tell how well our control scheme is
working. Of course, we may also measure the manipulated
variable and other variables, as well.
sensor: a measuring instrument. For chemical processes, the most
common measurements are of flow (F), temperature (T), pressure
3
Lecture Notes - 1
(P), level (L), and composition (A, for analyzer). The sensor will
detect the value of the measured variable as a function of time.
controller: the device that detects the output of the sensor, decides how
seriously the controlled variable deviates from the set point, and
directs the final control element in response. The controller
performs calculations based on its control algorithm.
transducer: this may be more than you wanted to know. The controller
must be able to communicate with sensor and final element.
Transducers convert and transmit signals to make this possible.
There are more details, of course, but they can wait. We install a level
sensor on the tank, put a control valve at the pump discharge, and connect
the two with a controller. Notice the symbols: the circle containing L
represents the sensor, and LC represents the controller. The control valve
has a mushroom on it for reasons we'll cover later. In the schematic, the
sensor communicates with the process by a solid line, and with the
controller and valve by a dashed line. We call this control structure
feedback control - the value of the controlled variable is fed back to a
controller, which adjusts the manipulated variable in response.
wi
L = level sensor
C = calculation or controller
final control element

L LC (control valve)
h
wo
When the level sensor indicates approach to high or low limits, the
controller computes a response by its algorithm and directs the control
valve to open or close appropriately. The outlet flow wo may not be
constant, as we wanted, but by suitable choice of tank size and control
algorithm we can significantly reduce its variability, and hence the effects
on downstream processes.
1.4 That’s process control?

Enough for now. We've modeled a simple process, defined our terms, and
sketched out a control scheme. Before we attempt to specify more about a
controller, however, we must learn more about the ways in which
processes might be disturbed.
4
Lecture Notes 2
Chapter 2. Dynamic system
2.0 Context
In this chapter, we define the term 'system' and how it relates to 'process'
and 'control'. We will also show how a simple dynamic system responds
to several disturbances.
2.1 System
In Chapter 1, we introduced a process - a surge tank with pumped outlet -
that was subject to disturbances in time. We thought of the process as a
collection of equipment and other material, marked off by a boundary in
space, communicating with its environment by energy and material
streams.
wi
wo
'Process' is a good notion; another useful notion is that of 'system'. A

system is some collection of equipment and operations, usually with a
boundary, communicating with its environment by a set of inputs and
outputs. By these definitions, a process is a type of system, but system is
more abstract and general. For example, the system boundary is often
tenuous: suppose that our system comprises the equipment in the plant and
the controller in the central control room, with radio communication
between the two. A physical boundary would be in two pieces, at least;
perhaps we should regard this system boundary as partly physical (around
the chemical process) and partly conceptual (around the controller).
Furthermore, the inputs and outputs of a system need not be material and
energy streams, as they are for a process. System inputs are "things that
cause" and outputs are "things that respond".
inputs system outputs

(causes) (responses)
1
Lecture Notes 2
To approach the problem of controlling our surge tank process, let’s think
of it in system terms: the primary output is the liquid level h -- not a
stream, certainly, but an important response variable of the system.
Disturbances are of course inputs, and so the stream wi is an input. And
peculiar as it first seems, the outlet flow wo is also an input, because it
influences the liquid level, just as does wi.
The point of all this is to look at a single schematic and know how to view
it as a process, and as a system. View it as a process (wo as an output) to
write the material balance and make fluid mechanics calculations. View it
as a system (wo as an input) to analyze the dynamic behavior implied by
that material balance and make control calculations.
2.2 Systems within systems

We call something a system and identify its inputs and outputs as a first
step toward understanding, predicting, and influencing its behavior. We
recognize that our understanding may improve if we determine some of
the structure within the system boundaries; that is, if we identify some
component systems. Each of these, of course, would have inputs and
outputs, too.
system
inputs outputs
1
2
Considering the relationship of these component systems, we recognize

the existence of intermediate variables within a system. Neither inputs
nor outputs of the main system, they connect the component systems.
Intermediate variables may be useful in understanding and influencing
overall system behavior.
When we add a controller to a process, we create a single time-varying

system; however, it is useful to keep process and controller conceptually
distinct as component systems. This is because relatively few control
schemes (relationships between process and controller) suffice for myriad
process applications. Using the terms we defined in Chapter 1, we
represent a control scheme called single-loop feedback control in this
fashion:
2
Lecture Notes 2
system
other inputs
other outputs
process
manipulated variable (intermediate

process controlled
set point controller final element variables) variable
sensor
Inside the block called "process" is the physical process, whatever it might
be, and the block is the boundary we would draw if we were doing an
overall material or energy balance. HOWEVER, we remember that the
inputs and outputs for the block are NOT necessarily the same as the
material and energy streams that cross the process boundary. From among
the outputs, we may select a controlled variable (VC), and provide a
suitable sensor to measure it. From the inputs, we choose a manipulated
variable (VM) and install an appropriate final control element. The
measurement is fed to the controller, which decides how to adjust VM to
keep VC at the set point. Other inputs are disturbances that affect VC, and
so require action by the controller.
We keep in mind this feedback control scheme, and how it relates the
controller to the process, when we represent the equipment in schematic
form, as with the surge tank of Chapter 1.
wi
L = level sensor
C = calculation or controller
final control element

L LC (control valve)
h
wo
We'll have much more to say about feedback control later. For now, it's
important to think of a chemical process as a dynamic system that
responds in particular ways to its inputs. We attach other dynamic
systems (sensor, controller, etc.) to that process in single-loop feedback
scheme and arrive at a new dynamic system that responds in different
3
Lecture Notes 2
ways to the inputs. If we do our job well, it responds in better ways, so to
justify all the trouble.
To do our job well, we must understand more about system dynamics --

how systems behave in time. That is, we must be able to describe how
important output variables react to arbitrary disturbances.
2.3 Dynamics of a tank, without any control

From Chapter 1, our process model was
dh
ρA = wi − wo h(0) = ho
dt
t (2.3.1)
1
(wi (t) − wo (t) )dt
ρA ∫0
h(t) = ho +
Now mindful of our system concepts, we recognize h(t) as the output and
wi(t)-wo(t) as the input. Indeed the flow rates are separate inputs, but our
model of the process indicates that they always influence the output liquid
level by their difference. For convenience, let us represent this difference
as x(t). Our model (2.3.1) captures the system dynamics; it tells us how
the output h(t) responds in time to input disturbances x(t). We now
integrate (2.3.1) for several specific cases of x(t).
2.4 Response to rectangular pulse at time td

Let the tank be operating at steady state, so that the flows are initially
balanced, and x is zero. Suppose that at time td, extra liquid is injected
into the feed stream: mass M is added over time interval ∆t before the inlet
flow returns to normal. We can idealize this as a rectangular pulse.
x(t) = 0, 0 ≤ t < td
M
, t d ≤ t ≤ t d + ∆t (2.4.1)
∆t
0, t d + ∆t < t
Inserting the disturbance (2.4.1) into process model (2.3.1), we compute

the response.
h(t) = ho , 0 ≤ t < td
M
ho + (t − t d ), t d ≤ t ≤ t d + ∆t (2.4.2)
ρA∆t
M
ho + , t d + ∆t < t
ρA
4
Lecture Notes 2
As shown in Figure 2-1, the level never recovers its former value ho; the
temporary disturbance has had a permanent effect on the output.
1.2
output
1
input
x∆t
(h − ho )ρA
normalized input and output
M
0.8 M
0.6
0.4
0.2
0
0 0.5 1 1.5 2 2.5 3 3.5
t/td
Figure 2-1. The pulse width ∆t has arbitrarily been set equal to the onset time td.
2.5 Impulse at time td

If in (2.4.1) we decrease the time interval of fluid injection, we finally
arrive at an infinite flow rate in an infinitesimal time interval, delivering
extra mass M to the fluid. Thus we introduce the mathematical delta
function -- a singularity at time td. Its value is infinite there and zero
elsewhere. Its integral over all time is 1. It has the dimension of
reciprocal time, although it’s not clear whether the actual units of a single
time point are at all significant.
x(t) = Mδ (t − t d )
h(t) = ho , 0 ≤ t < td (2.5.1)
M
ho + , td ≤ t
ρA
Again, the level never recovers from the brief disturbance.
5
Lecture Notes 2
input
xt d
output
1
(h − ho )ρA
M
0
0 1 2 3 4
t/td
Figure 2-2. The impulse is shown as a vertical arrow at the time of its action. The input
flow x(t) has been made non-dimensional by multiplying by td.; however, the sense of the
plot would be unchanged if this were not done.
2.6 Step at time td

The dimensionless unit step function is zero before td and one thereafter.
We use it to represent a sudden, permanent change in the inlet flow rate.
x(t) = ∆wu(t − t d )
h(t) = ho , 0 ≤ t < td (2.6.1)
∆w
ho + (t − t d ), t d ≤ t
ρA
The permanent change in the input has caused the output to rise without
limit. This is certainly reason to consider adding a control system.
6
Lecture Notes 2
normalized input and output 3
output
2
(h − ho )ρA
t d ∆w
x
input
∆w
1
0
0 1 2 3 4
t/td
Figure 2-3. Of course, the model ceases to be applicable when the liquid level reaches
the top of the tank.
2.7 Sine
At time td, the inlet flow begins to oscillate with radian frequency ω.
x(t) = ∆wsin(ωt − ωt d )
∆w (2.7.1)
h(t) = ho + (1− cos(ωt − ωtd ) )
ρAω
The liquid level oscillates at the frequency of the disturbance. Its response
is delayed, in that the level reaches its peak some time after the inlet flow
has peaked. Notice that the amplitude of oscillation decreases as the
frequency increases. This indicates that the tank cannot follow fast
changes.
7
Lecture Notes 2
2.5
liquid level
2
(h − ho )ρAω
∆w
1.5
0.5
-0.5
x
inlet flow ∆w
-1
-1.5
0 2 4 6 8 10 12 14
ωt
Figure 2-4. The onset of disturbance ωtd has arbitrarily been set to 1.
More detail
Response to a sine disturbance has two parts – the initial transient,
and a recurring oscillation. We can recast the response in this form
by using the sum-of-angles formula to write the cosine as a sine
that includes a phase angle.
∆w ∆w  π
h(t) = ho + + sin  ω (t − t d ) −  (2.7.2)
ρAω ρAω  2
Equation (2.7.2) shows that the output lags the input by π/2
radians, or 90°. The liquid level differs at most from its initial
value by twice the amplitude. It either exceeds or stays below the
initial level according to the sign of ∆w; that is, whether the flow
initially increased or decreased.
2.8 Typical disturbances

Knowing how a system responds to disturbances is a prerequisite for
controller design. We will use the impulse, step, and sine disturbances
repeatedly to test various dynamic systems. While we can never test our
control designs against every conceivable disturbance, testing against
8
Lecture Notes 2
these standard ideal disturbances will usually tell us what we need to
know.
9
Lecture Notes - 3
Lesson 3. Math review.
3.0 Context
In the previous chapters, we solved a differential equation for different
forcing functions. Here we will review this and other mathematical topics
that we will need.
3.1 Quadratic equation

The roots of
αs 2 + βs +1 = 0 (3.1.1)
are
− β ± β 2 − 4α
s= (3.1.2)
2α
If α is less than zero, the roots s will be real, of opposite sign, and of
unequal magnitude. For α greater than zero, the roots may be real or
complex; the real parts will have the same sign. The term under the
radical, aptly called the discriminant, determines whether the roots are
complex.
3.2 complex numbers

Consider the complex number
z = a + jb, where j = − 1 , Re(z) = a, Im(z) = b (3.2.1)
The same number may be written in polar form
b
z = z e jϕ , where z = a 2 + b 2 and ϕ = tan −1 (3.2.2)
a
The diagram shows the complex plane, in which real numbers are
confined to the horizontal axis. A complex number appears as a vector
from the origin. The diagram relates the Cartesian and polar forms of the
complex number. The phase angle ϕ is measured counterclockwise from
the positive real axis.
1
Lecture Notes - 3
imaginary
a 2 + b2
ϕ b
a real
If we let the magnitude |z| equal 1, we obtain Euler's formula relating the
exponential and trigonometric functions.
e ± jϕ = cos ϕ ± j sin ϕ (3.2.3)
Each complex number z has a conjugate z*
z = a + jb = z e jϕ (3.2.4)
z* = a – jb = z * e − jϕ (3.2.5)
The conjugates have the same magnitude.
zz* = |z|2 =|z*|2 = a2 + b2 (3.2.6)
Use the conjugate to eliminate j in the denominator of a ratio.
c a − jb ca − cb
= 2 +j 2 (3.2.7)
a + jb a − jb a + b 2
a + b2
3.3 Oscillatory behavior

Suppose a variable y (a liquid level, a temperature, a chemical
concentration, a flow rate) oscillates regularly in time:
 t 
y (t ) = A sin 2π + ϕ  = A sin (2πft + ϕ ) = A sin(ωt + ϕ ) (3.3.1)
 p 
where
A is the amplitude (dimension of y)
p is the period (dimension of time)
f is the cyclical frequency (dimension of time-1)
ω is the radian frequency (dimension of radians time-1)
ϕ is the phase angle (dimension of radians)
Inserting 2π into the argument is necessary because we are attempting to

describe physical behavior (something varying in time) by an abstract
2
Lecture Notes - 3
math function that doesn't care what our time units are. View ω = 2π/p =
2πf as the conversion factor between time units and radians.
The phase angle ϕ represents an advance in the signal y(t) with respect to
some other signal. That is, if
x(t) = B sin (ωt ) and y(t) = C sin (ωt + ϕ ) (3.3.2)
the oscillation y(t) is ahead of that of x(t) at any time t. However, we will
most often encounter phase lags, so that the phase angle ϕ will have a
negative value. If ϕ = 0, x(t) and y(t) are said to be in phase.
Representing an oscillation with a phase angle is quite useful, but on

occasion it is helpful express the same signal in another fashion. This
phase angle identity, derived from the trigonometric sum-of-angles
formula, shows that the signal can be expressed as a combination of sine
and cosine functions:
  C 
S 2 + C 2 sin ωt + tan −1    = S sin ωt + C cos ωt (3.3.3)
  S 
3.4 Arctangent
The arctangent is a multi-valued function, and thus must be treated
carefully in calculations. For example, suppose we wish to express the
complex number -1-j in polar form. From the figure we see that the phase
angle ϕ should be designated as either 225° (3.927 radians) or -135° (-
2.356 radians).
imaginary
real
ϕ
However, most calculators and spreadsheets will process (3.2.2) to give

45° (0.785 radians).
 −1
ϕ calculator = tan −1  −1
 = tan (1) = 45
o
(3.4.1)
−
  1
Calculators and spreadsheets tend to work between -90° and +90°.

Because we will be considering phase lags, we will instead tend to apply
3
Lecture Notes - 3
the arctangent between -180° and 0°. Hence angles in the upper right
quadrant should be corrected.
ϕ = ϕ calculator − 180o , 0 ≤ ϕ calculator ≤ 90o (3.4.2)
Arctangent Function
0
arctan (radians)
-1 as calculated
-2 180 deg offset
-3
-4
-5
-30 -20 -10 0 10 20 30
argument
By the way, the conversion factor between degrees and radians is
o
180
1= (3.4.3)
π radians
3.5 First-order, linear, variable-coefficient ODE

We are addressing systems that vary in time, so our independent variable
is always t.
dy
a (t ) + y (t ) = Kx(t ) y (t0 ) = known (3.5.1)
dt
In writing (3.5.1) we have arranged the coefficient functions to isolate the

dependent variable y(t). By this means, a(t) must have dimensions of
time t, and K has dimensions of y/x. We solve this equation by defining
the integrating factor p(t)
dt
p (t ) = exp ∫ (3.5.2)
a(t )
4
Lecture Notes - 3
and integrating to find
t
K p (t ) x(t ) p (t0 ) y (t0 )
y (t ) = ∫
p(t ) t0 a(t )
dt +
p(t )
(3.5.3)
The solution y(t) comprises contributions from the forcing function Kx(t)
and the initial condition y(t0). These are known as the particular (depends
on the right-hand side) and homogeneous (as if the right-hand side were
zero) solutions. In the language of dynamic systems, we can think of y(t)
as the response of the system to input disturbances Kx(t) and y(t0).
3.6 First-order ODE, special case for process control applications

If a(t) is constant in (3.5.1), we call it the time constant τ. In this context,
K is called the gain. By its magnitude and sign, the gain influences how
severely y responds to x. The solution (3.5.3) becomes
K −t
t
t − (t − t 0 )
y (t ) =
τ
e τ
∫ e τ x(t )dt + y(t0 )e
t0
τ
(3.6.1)
or for initial time at zero,
t
K −t t −t
y (t ) =
τ
e τ
∫ e τ x(t )dt + y(0)e
0
τ
(3.6.2)
3.7 First-order ODE, special case for no disturbance (forcing function)

If K = 0, the system response depends only on the initial conditions. This
can be obtained from the general solution (3.5.3), of course, but can also
be obtained by directly integrating equation (3.5.1), which has become
separable.
dy − dt
=
y a(t )
t
(3.7.1)
y dt
ln = −∫
y (0) 0
a(t )
3.8 First-order ODE, special case for missing dependent variable

If the y(t) term is removed (for example, if K and a(t) are both very large),
another separable equation results.
5
Lecture Notes - 3
Kx(t )
dy = dt
a(t )
t
(3.8.1)
x(t )
y = y (0) + K ∫ dt
0
a (t )
Separable equations are convenient to solve by straightforward function
integration. The surge tank of Chapters 1 and 2 was described by a
separable equation of form (3.8.1).
3.9 First-order ODE, delayed disturbance
Let the forcing function be delayed; suppose x(t) is a unit step at time td >
0. Then from (3.5.3)
K  d p(t)(0) p(t)(1)  p(0) y(0)

t t
y(t) = ∫ dt + ∫ dt  +
p(t)  0 a(t) td
a(t)  p(t)
(3.9.1)
t
K p(t ) p(0) y (0)
y(t) = ∫
p(t) td a (t )
dt +
p(t)
3.10 Second-order, linear, constant-coefficient ODE

d2y dy dy
α 2 +β + y(t) = Kx(t) y(0), = known (3.10.1)
dt dt dt 0
The coefficient a2 has the dimension of time squared. The solution to

(3.10.1) is the sum of two terms:
y(t) = yh (t) + y p (t) (3.10.2)
The homogeneous solution yh, which depends only on the left-hand-side

of (3.10.1), is itself the sum of two linearly independent exponential
functions
yh (t) = C1e r1t + C2 e r2t (3.10.3)
where r1 and r2 are the roots of the characteristic equation of (3.10.1).
− β ± β 2 − 4α
r1,2 = (3.10.4)
2α
The value of the discriminant determines three distinct forms of the

solution.
6
Lecture Notes - 3
real, unequal roots for r
− βt  β 2 − 4α t − β 2 − 4α t 
yh (t) = e 2α  C1e 2α
+ C2 e 2α 
 (3.10.5)
 
In process control, we prefer stable systems, those in which disturbances

do not grow with time. We observe that (3.10.5) decays if α and β have
the same sign.
real, equal roots for r

− βt
yh (t) = e 2α
(C1 + C2t ) (3.10.6)
The solution will decay if α and β have the same sign.
complex roots for r

 
 C1 cos 4α − β t + C2 sin 4α − β
− βt 2 2
yh (t) = e 2α
t (3.10.7)
 2α 2α 
 
Once again, the solution will decay if α and β have the same sign. The
coefficient of t in the trigonometric functions is the radian frequency of the
oscillation.
The particular solution yp for any disturbance x(t) may be determined by

the 'method of undetermined coefficients', or the 'method of variation of
parameters'. The initial conditions are then applied to the solution y(t) to
determine coefficients C1 and C2.
The response of the system (3.10.1) then depends on

• the character of the system itself (through the left-hand-side coefficients, affecting the
exponential and trigonometric terms in the homogeneous solution)
• the initial conditions (affecting coefficients C1 and C2 in the homogeneous solution)
• the nature of the disturbance (through the particular solution, as well as C1 and C2, if the
disturbance is initially non-zero)
In a later lesson, we will introduce Laplace transforms as an alternative

method for determining the solution.
3.11 Representing functions by Taylor series

We specify some reference value of the independent variable, and
represent the function in the neighborhood of that reference as a series of
terms. For a function of one variable:
f (x) = f (xs ) +
df
(
(x − xs ) + O (x − xs ) 2 ) (3.11.1)
dx xs
7
Lecture Notes - 3
For a function of more than one variable:
f ( x, y,...) = f ( xs , y s ,...) + ∂f ( x − xs )
∂x x s , y s ,...
(3.11.2)
+ ∂f ( y − ys ) + ... + O (( x − xs ) ,( y − ys ) ,...)
2 2
∂y xs , y s ,...
By retaining only linear terms, we obtain a linear approximation. The

derivatives are evaluated at the reference point. Of course, the
approximation is exact at the reference, and it is often satisfactory in some
region about the reference value. As the figure indicates, however,
extrapolation to x = 0 would be erroneous.
Taylor Series Linearization about x = 2
5
4.5
4
3.5
3
nonlinear
2.5
y
linear
2
1.5
1
0.5
0
0 0.5 1 1.5 2 2.5 3
x
3.12 Chain rule for differentiation

d dg df
g( f (s)) = (3.12.1)
ds df ds
The functions g and f are said to be nested. For example, let g be the
exponential and f the square root.
8
Lecture Notes - 3
d
ds
e s
( ) 12 s
= e s
−1
2 


(3.12.2)
The chain rule applies to the special cases of a product
d dg df
f (s)g(s) = f (s) + g(s) (3.12.3)
ds ds ds
or a quotient
dN dD
D(s) − N (s)
d N (s) ds ds
= (3.12.4)
ds D(s) D(s) 2
3.13 Must we?

The math topics collected here will be used during the course. Please
review any that seem unfamiliar.
9
Lecture Notes - 4
Lesson 4. Dynamic behavior of 'first-order' processes.
4.0 Context
System dynamics is an engineering science useful to mechanical,
electrical, and chemical engineers, as well as others. This is because
transient behavior, for all the variety of systems in nature and technology,
can be described by a very few elements. This lesson concerns one of
those elements.
4.1 The first order lag: mixed tank

Before we explain the term "first-order lag", we will work with an
example of one: consider a tank equipped with a stirrer to mix the inlet
stream into the contents of the tank. The composition of the liquid in the
tank is uniform, and it is equal to the composition of the outlet stream.
The tank is arranged with large overflow; thus the inlet and outlet flows
are essentially equal at any time, and the volume is constant. This is a
surge tank for smoothing concentration changes; contrast it with the flow
surge tank of Lesson 1.
For simplicity, we will consider the flow to be constant, but the inlet
concentration may vary with time; we wish to determine the effect on the
outlet concentration.
Let's outline the procedure we will be following:

• write material and energy balances and other equations required to
describe the process
• (for most process control applications) identify a steady state operating
condition to serve as a reference
• substitute Taylor series approximations about the reference condition
for nonlinear terms in the model
• solve the model for any required operating parameters at the reference
condition
• subtract the steady state condition from the model equations to express
all variables in deviation form
• arrange the equations in standard form, identifying dynamic
parameters of known significance
1
Lecture Notes - 4
• solve for the output variables as functions of the inputs
• introduce particular disturbances and calculate the responses
First, we write a component material balance on the solute.
d
VCo = FCi (t) − FCo (t) Co (0) = Cs (4.1.1)
dt
Because the flow F and volume V are constant, there are no nonlinear
terms in the equation. We write (4.1.1) at steady state with reference inlet
concentration Cs.
dVCo
= 0 = FCs − FCos (4.1.2)
dt s
From (4.1.2) we see that the outlet concentration at the reference condition
is also Cs. Subtracting (4.1.2) from (4.1.1), we obtain the process model
in terms of deviation variables, indicated by an asterisk superscript. These
variables are zero when the process is at the reference condition; nonzero
values indicate deviation from the reference.
d
V (Co (t) − Cs ) = F (Ci (t) − Cs ) − F (Co (t) − Cs )
dt
(4.1.3)
d
VCo* = FCi* (t) − FCo* (t) Co* (0) = 0
dt
This is a first-order ODE with constant coefficients. We rearrange it to

standard form.
V dCo*
+ Co* (t ) = Ci* (t ) (4.1.4)
F dt
In mathematical nomenclature, Ci* is the forcing function and Co* the

dependent variable. In our system nomenclature, Ci* is the input and Co*
the output. In standard form, the ratio of tank volume to flow rate clearly
takes on the significance of a characteristic time, the time constant τ.
dCo*
τ + Co* = Ci* (t ) Co* (0) = 0 (4.1.5)
dt
The solution (by (3.6.1)) is
2
Lecture Notes - 4
−t
τ t
e t
C (t) = ∫ e τ Ci (t)dt
* *
(4.1.6)
τ
o
0
Equation (4.1.6) is the process model for the mixing tank, showing how
the outlet concentration behaves for arbitrary disturbances in the inlet. For
example, if inlet concentration undergoes a step change ∆C at td,
 −(t −t d )
τ 
C o* = ∆C 1 − e  (4.1.7)
 
As the disturbance is introduced, the outlet concentration begins to

change; it gradually becomes equal to the inlet concentration. Notice that
the tangent to the initial response reaches the final value in one time
constant.
1.2
inlet concentration
1
outlet concentration
0.8
0.6
0.4
0.2
0
0 1 2 3 4 5 6 7
(t - τ )/ τ
Fig 4-1. The ordinate has been normalized by the magnitude of the step
change, and the abscissa by the time constant. Thus this non-dimensional
plot is characteristic of all first order lag step responses. The time td at
which the input occurs has been set for convenience to equal the time
constant.
3
Lecture Notes - 4
Systems described by (4.1.5), with solution (4.1.6), are called first-order

lags. "First order" refers to the order of the governing differential
equation (4.1.5). "Lag" refers to the way in which the output lags behind
the input. The lag occurs because the system has storage capacity, and
that capacity takes time to fill or deplete when conditions change. In this
problem, the system stores the dissolved component.
First order lags always feature a time constant τ that indicates the speed of
response, because time is normalized by, or scaled to, the time constant.
From the properties of the exponential function, we see that the step is
95% complete when time equal to three time constants has elapsed. If the
tank time constant is large (large volume, low flow) this time will be large.
If the time constant is smaller (small volume, large flow) the outlet
concentration will respond more quickly. This is consistent with intuition
and experience.
In addition to speed of response, we are also interested in the degree to

which a dynamic system amplifies or attenuates the input signal. This is
often expressed by the steady-state gain, which is the ratio of steady
output change to input change following a step disturbance.
C o* (∞) − C o* (0) ∆C
gain = = =1 (4.1.8)
C i* (∞) − C i* (0) ∆C
For the mixing tank, the gain is 1, showing that permanent disturbances
are merely passed through the system. Both time constant and gain are
independent of the size of the disturbance ∆C.
4.2 Integrator: pumped outlet tank

The pumped outlet tank of Lessons 1 and 2 is an example of a first order
integrator.
dh
ρA = wi − wo h(0) = hs (4.2.1)
dt
All terms in the equation are linear. We define a steady state reference
condition in which the liquid level is hs, and the inlet and outlet flows are
equal to ws. In deviation variables, (4.2.1) becomes
dh*
ρA *
= wi − wo
*
h* (0) = 0 (4.2.2)
dt
To be strict about placing (4.2.2) in standard form, we should define a gain

and a time constant. Gain always has dimensions of output/input, or in
4
Lecture Notes - 4
this case, length divided by mass flow. Hence we multiply (4.2.2) by the
height of the tank, and divide by ws.
hT dh* hT
ρA =
* *
(wi − wo ) h* (0) = 0 (4.2.3)
ws dt ws
Collecting terms, we find
dh*
τ * *
= K (wi − wo ) h* (0) = 0 (4.2.4)
dt
Equation (4.2.4) is separable; its solution is
* * t
h * = K (wi − wi ) (4.2.5)
τ
and the response to a step change of magnitude ∆w in inlet flow at time td

is
(t − td )
h* = K∆w (4.2.6)
τ
The integrator has no steady state response to a step disturbance, so K

cannot be viewed as a steady-state gain. The time constant is the
residence time for a full tank.
5
Lecture Notes - 4
3
liquid level response
change in inlet flow
0
0 1 2 3 4
(t - τ)/ τ
Fig 4-2. The ordinate has been normalized by the product of the gain and
step disturbance, and the abscissa by the time constant. The time td at
which the input occurs has been set for convenience to equal the time
constant.
4.3 Summary of disturbance responses

Initial condition is zero; disturbance is introduced at time td.
system first order lag first order integrator

model dy dy
τ + y (t ) = Kx(t ) τ = Kx(t )
dt dt
step: − ( t − td ) AK (t − td )
AK 1 − e τ 

x = Au(t - td)
  τ
steady-state gain = K
impulse: AK − (t − td ) AK
τ
e
x = Aδ(t - td) τ τ
sine: − ( t − td )
AK sin(ω (t − td ) + φ ) AK   π 
AKωτe 1 + sin ω (t − td ) −  
τ
x = Asin(ωt - ωtd) + τω   2 
1 + ω 2τ 2 1 + ω 2τ 2
φ = tan −1 ( −ωτ )
6
Lecture Notes - 4
4.4 Multiple system inputs from multiple inlet streams

The first order lag equation in Section 4.3 is the mathematical form that
results from applying material and energy balances to well-mixed
volumes, as illustrated by the mixing tank in Section 4.1. Of course, a
tank may have more than one inlet stream. If so, it will usually be added
to the right-hand side of the equation. In illustration, consider a well-
stirred tank heated by an electric resistance element of output power Q.
Our previous first-order systems have stored mass; this one stores energy.
The energy balance is
d
(ρC pV (To − Tref )) = ρC p F (Ti − Tref ) − ρC p F (To − Tref ) + Q (4.4.1)
dt
where Tref is a reference temperature for computing the enthalpy of the

flowing stream. Presuming that flow F is constant and the physical
properties are not a function of temperature, we see that (4.4.1) is a first-
order linear ODE with constant coefficients. The energy balance at steady
conditions is
d
(ρC pV (Tos − Tref )) = 0 = ρC p F (Tis − Tref ) − ρC p F (Tos − Tref ) + Qs (4.4.2)
dt
Subtracting (4.4.2) from (4.4.1) and introducing deviation variables
d
(ρC pV (To − Tos )) = ρC p F (Ti − Tis ) − ρC p F (To − Tos ) + Q − Qs
dt
(4.4.3)
d
dt
( *
)
ρC pVTo = ρC p FTi − ρC p FTo + Q
* * *
We rearrange (4.4.3) to standard form and consider the case of initial

steady state.
7
Lecture Notes - 4
*
V dTo * * 1 *
+ To = Ti + Q* To (0) = 0 (4.4.4)
F dt ρC p F
The time constant for temperature change is the tank residence time, equal
to the tank volume divided by the volumetric flow rate. We find that the
outlet temperature response depends on two inputs: the inlet temperature
and the heater power; either can act as a disturbance to the first-order
system. The gain for inlet temperature disturbances is unity; thus a step
change in temperature would ultimately propagate through the tank. The
gain for power disturbances converts dimensions of power to dimensions
of temperature. This gain is a function of the flow rate, so that, for
example, power disturbances have less effect on To when the flow F is
large.
Equation (4.4.4) is linear first-order, and can be solved by (3.6.1), just as

we did in Section 4.1.
−t
e t  *
τ t
Q* 
τ
τ ∫0 
T (t ) =
*
e T + dt (4.4.5)
ρC p F 
o i
In a linear model, the effect of the disturbances is additive. That is, each
affects the response independently of the other, and the effects are simply
added. Consider a step increase in inlet temperature at time τ/2, followed
at 2τ by a compensating step decrease in heater power. The outlet
temperature first rises and then falls in response.
8
Lecture Notes - 4
1.5
change in inlet temperature

1
0.5
outlet temperature response
-0.5
change in heater power

-1
-1.5
0 1 2 3 4 5 6
(t - τ)/ τ
Figure 4-3 The disturbances to linear equations produce additive responses.
4.5 Multiple Outlet Streams

A tank may have multiple outlet streams, as well. In modeling first order
systems, we often find that these streams depend on the response variable.
In this case, the effect of additional outlet streams is to alter the time
constant and gain of the system. For example, suppose that a mixed
overflow tank is cooled by convective heat transfer to a condensing vapor.
Thus there are two outlet streams: the enthalpy carried out with the outlet
flow, and the heat transferred to the cooling coil. The energy balance is
d
(ρC pV (To − Tref )) = ρC p F (Ti − Tref ) − ρC p F (To − Tref ) − UA(To − Tc )
dt
(4.5.1)
where the overall heat transfer coefficient is U and the coolant condenses
at temperature Tc. Writing (4.5.1) at steady state and subtracting this
result from (4.5.1) gives
d
dt
( )
ρC pVTo* = ρC p FTi * − ρC p FTo* − UATo* + UATc* (4.5.2)
9
Lecture Notes - 4
where we have two terms that depend on the response variable To*.
Writing (4.5.2) in standard form gives
ρC pV dTo
*
* ρC p F * UA *
+ To = Ti + Tc (4.5.3)
ρC p F + UA dt ρC p F + UA ρC p F + UA
Because there are two paths out for enthalpy - flow and heat transfer - the
time constant for temperature change is less than the tank residence time,
in contrast to (4.4.4). Equation (4.5.3) also features two enthalpy inputs:
inlet flow and the coolant. In fact, if the coolant temperature Tc exceeds
Ti, (4.5.3) will describe heating of the tank. Notice that the gain for an
inlet temperature disturbance is less than unity: because there are two
paths out, the outlet response will not grow to equal a permanent inlet
disturbance. Equation (4.5.3) is solved as before, and the response is
shown in the figure for a gain of 0.5 (that is, UA = ρCpF) and a step
decrease in Ti at time τ.
-2
input and output (deg C)
-4
outlet temperature response

-6
-8
-10
change in inlet temperature
-12
0 1 2 3 4 5
(t - τ )/ τ
4.6 Summing up
First order systems arise from material and energy balances on perfectly
mixed volumes. The system output or response variable is a measure of
the storage in the system - for example, liquid level or concentration for
10
Lecture Notes - 4
mass and temperature for energy. Two parameters, the time constant and
the gain, characterize the response of the output variable to disturbances.
11
Lecture Notes – Lesson 5
Laplace transforms
5.0 Context
Remember how one multiplies with logarithms:
Given 3.5×4.3
- transform the problem into "log space"
log (3.5×4.3) = log 3.5 + log 4.3
- perform the logarithm addition
log (3.5×4.3) = 0.5441 + 0.6335 = 1.1775
- transform the problem back into original terms by finding the antilog of the sum
3.5×4.3 = log-1(1.1775) = 15.05
We exchange one multiplication for two transformations and an addition.

In an analogous way, the Laplace transform can be used to solve linear
ordinary differential equations with constant coefficients.
5.1 The Laplace transform

The Laplace transform changes a function of time t into a function of a
new independent variable s. This new variable may take complex values.
We will go through the trouble for the same reason we use logarithms – it
may make the original problem easier to solve.
∞
y(s) = L[ y(t)] = ∫ y(t)e −st dt (5.1.1)
0
We will indicate a Laplace transform of y(t) or y by writing y(s). This is

not to be interpreted as a simple variable substitution of s for t in the
function y.
5.2 Illustration
Consider the unit step function u(t). From (5.1.1) we find the Laplace
transform of u(t) as
∞
u(s) = L[u(t)] = ∫ u(t)e −st dt (5.2.1)
0
Before integrating, however, let’s get a sense of what this looks like: plot
the product u(t)e-st versus t for different values of s.
5-1
1.2
0.8
s=
0.1
u(t) exp(-st)
0.5
0.6
1
5
0.4
0.2
0
0 0.5 1 1.5 2 2.5 3 3.5
t
Integrating under each of these curves gives an area associated with a

value of s. Plotting these versus s, we see the behavior of the Laplace
transform of u(t), at least along the real axis.
12
10
8
integral of u(t) exp(-st)
0
0 1 2 3 4 5 6
s
5-2
At each value of s, y(s) is based on the entire t-dependence of y. The
magnitude of s acts as a weighting parameter to emphasize different
regions of the t domain. At large values of s, the exponential function will
be appreciably different from zero only for small values of t; hence large s
emphasizes the initial region of y. By contrast, small values of s allow the
influence of longer times on the value of y(s). The plot of y(s) looks like
s-1; performing the integration of (5.1.1) verifies this guess.
5.3 Functional and operational transforms

There are two categories of Laplace transforms we'll consider: transforms
of particular functions and transforms of general operations.
Example functional transforms:

unit step function u(t - td).
1
∞
L(Cu(t − t d ) ) = ∫ Cu(t − t d )e −st dt
0 td
∞
= C ∫ e −st dt
td (5.3.1)
−C −st ∞
= e
s td
Ce −std
=
s
exponential decay
∞
(
L Ce −at
) = ∫ Ce −at −st
e dt
0
− C −(s+a)t ∞
= e (5.3.2)
s+a 0
C
=
s+a
Example operational transforms:

the first derivative.
Using integration by parts,
5-3
∞
 df (t)  df (t) −st
L =∫ e dt
 dt  0 dt
∞
∞
= s ∫ f (t)e −st dt + f (t)e −st (5.3.3)
0
0
= sL( f (t) ) − f (0)
Further detail
∞
 df (t)  df (t) −st
L  ∫
= e dt
 dt  0 dt
uv = ∫ udv + ∫ vdu
df
u= f du = dt
dt
v = e −st dv = −se −st dt
∞ ∞
−st ∞ df (t) −st
f (t)e = −s ∫ f (t)e dt + ∫
−st
e dt
0
0 0
dt
∞
 df (t)  ∞
 = s ∫ f (t)e dt + f (t)e −st 0
−st
∴ L
 dt  0
= sL f (t) ) − f (0)
(
the second derivative

Using (5.3.3),
 d 2 f (t )  ∞ d 2 f (t ) −st
L 2
 = ∫ 2
e dt
 dt  0 dt
∞
df −st df
= s∫ e dt −
0
dt dt 0
(5.3.4)
 ∞  df
= s  s ∫ f (t )e −st dt − f (0) −
 0  dt 0
df
= s 2 L( f (t ) ) − sf (0) −
dt 0
the integral
Using integration by parts,
5-4
t  ∞ t 
L ∫ f (ξ )dξ  = ∫  ∫ f (ξ )dξ e −st dt
0  00 
(∫ )
∞
∞ t
1 1
=∫ f (t) e −st dt − f (ξ )dξ e −st (5.3.5)
0
s s 0 0
1
= L( f (t) )
s
Further detail
t  ∞ t 
L ∫ f (ξ )dξ  = ∫  ∫ f (ξ )dξ e −st dt

0  00 
uv = ∫ udv + ∫ vdu
u = ∫ f (ξ )dξ du = fdt
v = e −st dv = −se −st dt
(∫t
0
)
f (ξ )dξ e
−st
∞
0
∞ t
  ∞
= ∫  ∫ f (ξ )dξ (− s )e −st dt + ∫ f (t)e −st dt
00  0
(∫ )
∞
t  1∞ 1
t
∴ L ∫ f (ξ )dξ  = ∫ f (t)e −st dt −
 f (ξ )dξ e −st
0  s0 s 0 0
1
= L( f (t) )
s
the time delay
Suppose that some function f(t) is delayed by an interval of time θ. Thus
the delayed function g(t) may be written
f(t) g(t)
1
0 0 < t < θ
g(t) =  (5.3.6)
 f (t − θ ) θ ≤ t θ
The Laplace transform of g(t) is
θ ∞
L(g (t) ) = ∫ 0e −st dt + ∫ f (t − θ )e −st dt (5.3.7)
0 θ
Make a substitution of independent variable, ζ = t – θ, or t = ζ + θ.
5-5
∞
L(g(t) ) = ∫ f (ζ )e −s(ζ +θ ) dζ
0
∞
= ∫ f (ζ )e −sζ e −sθ dζ
0 (5.3.8)
∞
= e −sθ ∫ f (ζ )e −sζ dζ
0
=e −sθ
L( f (t) )
The last step happens because there is no reason not to rename the
independent variable ζ as t. Equation (5.3.1) is an example of the time
delay: the transform of the unit step at td may be viewed as the transform
of the unit step at time 0 multiplied by the exponential in td.
5.4 Solving the first-order lag with Laplace transforms

Recall the mixing tank of Section 4.1.
dCo*
τ + Co* (t) = Ci* (t) Co* (0) = 0 (5.4.1)
dt
Take Laplace transforms of (5.4.1)
τ [sC o* (s) − C o* (0)] + C o* (s) = C i* (s) (5.4.2)
Solve for the outlet concentration transform
1
Co* (s) = Ci* (s) (5.4.3)
τs + 1
As before, the inlet concentration undergoes a step change ∆C at td.

Transforming the step change by (5.3.1) and inserting, we get
1 ∆Ce −td s
Co* (s) = (5.4.4)
(τs + 1) s
To perform the reverse transform, it is convenient to use a table of

transform pairs, such as that provided by Marlin (2000). The reverse
transform of everything except the exponential factor is
 ∆C   −t

L−1   = ∆C 1− e τ  (5.4.5)
 s(τs +1)   
5-6
The effect of the exponential factor is to introduce a time delay into this
function. By (5.3.8)
Co* (t ) = 0 0 ≤ t < td
 (5.4.6)
τ 
− ( t −t d )
= ∆C 1 − e  td ≤ t
 
5.5 Example - first-order lag driven by its initial conditions

dy
τ +y=0 y (0) = A (5.5.1)
dt
The constant A has the dimensions of variable y. Take the Laplace

transform of the equation
τ [sy ( s) − y (0)] + y ( s) = 0 (5.5.2)
Solve for y(s), substituting the known initial condition.
Aτ
y ( s) = (5.5.3)
τs + 1
Now use a table of transform pairs to invert this function
1 −t −t
y (t ) = Aτ e τ
= Ae τ
(5.5.4)
τ
The solution (5.5.4) decays from the initial condition to a steady value of
zero.
5.6 Example - first-order lag driven by impulse disturbance at time zero

dy
τ + y = Aδ (t − 0) y(0) = 0 (5.6.1)
dt
Because the delta function has the dimension of reciprocal time, the
constant A has the dimensions of variable y multiplied by time. Take the
Laplace transform of the equation
τ [sy(s) − y(0)] + y(s) = A (5.6.2)
A
y(s) = (5.6.3)
τs +1
5-7
1 −t
y(t) = A e τ (5.6.4)
τ
The solution decays from the impulse to a steady value of zero. Certainly
(5.6.4) has the same form as (5.5.4), but one should not confuse the initial
condition of (5.5.1) with the delta function disturbance in (5.6.1).
5.7 Example - first-order lag driven by impulse disturbance at time td

dy
τ + y = Aδ (t − t d ) y(0) = 0 (5.7.1)
dt
Transforming,
τ [sy(s) − y(0)] + y(s) = Ae −t ds

(5.7.2)
Rearranging and substituting,
A −t d s
y(s) = e (5.7.3)
τs +1
The exponential factor contributes a time delay in the inverse transform.
y(t) = 0 0 ≤ t < td
1 −(t −td )τ (5.7.4)
=A e td ≤ t
τ
Solution (5.7.4) is just (5.6.4) shifted to a later time of occurrence.
5.8 Example - first-order lag driven by step disturbance at time td

dy
τ + y = Au(t − t d ) y(0) = 0 (5.8.1)
dt
Transforming, substituting, and rearranging,
1 A −td s
y(s) = e (5.8.2)
τs +1 s
Inverting with the aid of a transform pair table,
y(t) = 0 0 ≤ t < td
 −(t −t d )
τ 
(5.8.3)
= A1− e  td ≤ t
 
5-8
Solution (5.8.3) is just the general form of (5.4.6). Under a continuing
steady disturbance, a first-order lag approaches a new steady value.
5.9 Example - first-order lag driven by initial conditions and step disturbance at time td
dy
τ + y = Au(t − t d ) y(0) = B (5.9.1)
dt
Transforming
A −td s
τ (sy(s) − y(0) ) + y(s) = e (5.9.2)
s
Substituting and rearranging
τB A
y(s) = + e −td s (5.9.3)
τs + 1 s(τs + 1)
The two terms on the RHS are inverted separately; the time delay applies
only to the second term.
−t  −(t −t d )
τ 
y(t) = Be τ
+ A1 − e  (5.9.4)
 
The solution starts at the initial condition B and decays; at time td, it
begins to respond to the step input. At long times, the influence of the
initial conditions is negligible, and the solution approaches the value of the
step, A, as in (5.8.3). In interpreting (5.9.4), the second term is understood
to be zero at times less than td.
5.10 Doing it wrong: forgetting coefficients when substituting the derivative transforms
Let’s return to the problem of (5.5.1):
dy
τ +y=0 y(0) = A (5.10.1)
dt
The constant A has the dimensions of variable y. Take the Laplace

transform of the equation
τ [sy(s) − y(0)] + y(s) = 0 (5.10.2)
It’s simple algebra, but it’s possible to neglect to distribute the time
constant across the transform
τ [sy(s)] − y(0) + y(s) ≠ 0 (5.10.3)
5-9
A
y(s) ≠ (5.10.4)
τs +1
1 −t
y(t) ≠ A e τ (5.10.5)
τ
Compared to (5.5.4), the solution has the right form, but the dimensions
are wrong: y is set equal to a quantity with dimensions of y/time.
5.11 Doing it wrong: inverting before applying disturbance

The Laplace transform gives a solution in the Laplace domain; apply the
disturbance before inverting to find the time-domain solution. Here’s the
wrong way: consider a first-order lag.
dy
τ + y = x(t) y(0) = 0 (5.11.1)
dt
Transforming and rearranging,
1
y(s) = x(s) (5.11.2)
τs +1
If, for example, x is a step disturbance Au(t), the correct procedure is to

make the functional transform of the step, substitute that into (5.11.2) for
x(s), and then invert the right-hand-side. Don’t invert as if x(s) is a
constant, and then apply the disturbance!
−t
y(t) ≠ e τ
Au(t) (5.11.3)
This solution will not satisfy the initial conditions on y; neither will it
show a long-term change from the step input.
5.12 What’s missing

Lots. We are skipping the mathematical questions of existence and
uniqueness, extensions to the related Fourier transforms, and the whole
question of how to invert an arbitrary Laplace transform.
5.13 References
Marlin. Process Control. 2nd ed. McGraw-Hill, 2000, p.100.
5-10
Lecture Notes - 6
Lesson 6. Further topics on Laplace transforms and system models
6.0 Context
In Lesson 4 we saw how first-order models could describe several
physical systems of interest. We derived the models from conservation
equations, and solved them by the integrating factor method. In Lesson 5
we examined the Laplace transform as an alternative method of solving
the models. Here we emphasize several key topics on which the making
and solving of models depends: making equations linear, using deviation
variables, and inverting Laplace transforms.
6.1 Linearization
Laplace transforms are particularly useful for linear equations with
constant coefficients. If the process models include products or non-linear
functions of any of the dependent variables, it is common to make linear
approximations by Taylor series so that they can be more easily solved.
Because control is intended to keep a process near a set point, and because
a linear approximation is usually accurate around a fixed point, the linear
model is often perfectly satisfactory for control design.
6.2 Approximating a nonlinear function of the dependent variable

For turbulent flow through an obstruction, the volumetric flow is
proportional to the square root of the head loss.
Fo = kh 0.5 (6.2.1)
Fo is a function of one variable, h; it may be approximated by Taylor

series.
Fo = Fo (h)
(6.2.2)
Fo ≈ khs0.5 + 0.5khs−0.5 (h − hs )
where hs is the reference head loss. The approximation is exact at h = hs.
6.3 Approximating a product of variables

Even if the governing equation is linear, it may have variable coefficients
that include products of variables. For example, the flow of solute A
depends on its concentration and the total volumetric flow rate.
wA = FC A (6.3.1)
A two-variable Taylor expansion is required.
wA = wA (F , C A )
(6.3.2)
wA ≈ Fs C As + Fs (C A − C As ) + C As (F − Fs )
1
Lecture Notes - 6
where the subscript s denotes reference values of flow and concentration.
6.4 Approximating a product including a nonlinear function

As a further example, consider a reaction rate term that might appear in a
material or energy balance:
−E
ko e RT
C A = f (T ,C A ) (6.4.1)
Both temperature T and concentration CA might be variables in a dynamic

system; a two-variable Taylor expansion is required.
∂f E − E RT
= ko e CA
dT RT 2
∂f −E
= ko e RT
dC A
−E −E E − E RTs −E
ko e RT
C A ≈ ko e RTs
C As + ko 2
e C As (T − Ts ) + k o e RTs
(C A − C As )
RTs
(6.4.2)
6.5 Deviation variables

In process control, we try to keep a controlled variable at a set point; thus,
we are concerned to reduce deviation from the set point. This suggests
that we express our models in deviation variables, so that any non-zero
value of the deviation indicates a problem.
We used deviation variables when we analyzed the mixing tank in Lesson

4. Let's examine the procedure for the first-order lag in general.
dy
τ + y = Kx(t) y(0) = A (6.5.1)
dt
where y(t) is the system output, x(t) the input, τ a characteristic time
constant, K the steady state gain, and A the initial value of y. Let's
consider the deviation of the input and output variables from a constant
reference state xref, yref.
y * (t) = y(t) − y ref

(6.5.2)
x * (t) = x(t) − x ref
Substituting into (6.5.1), we obtain
2
Lecture Notes - 6
d ( y * + y ref )
τ + ( y * + y ref ) = K ( x * + x ref ) y * (0) + y ref = A
dt
(6.5.3)
and rearranging,
dy *
τ + y * = Kx * + (Kx ref − y ref ) y * (0) = A − y ref (6.5.4)
dt
Equation (6.5.4) is a first order lag with output y* determined by input

Kx* + (Kxref – yref) and initial condition A – yref. The fundamental
behavior of the solution will not differ from that of (6.5.1); we merely may
choose the reference condition for our convenience. We consider several
cases:
reference at initial conditions: xref = x(0); yref = A

Substituting into (6.5.4)
dy *
τ + y * = Kx * + (Kx(0) − A) y * (0) = 0 (6.5.5)
dt
The deviation variable y* will start at zero and respond to the right-hand
forcing function. The deviation input Kx* may begin at some arbitrary
time, but the second term affects the solution from time zero. At long
time,
y * (∞) = Kx* (∞) + Kx(0) − A (6.5.6)
which may or may not be zero.
reference at initial steady state: the classic process control problem

For much of our development of process control theory, we will conceive
of a system operating at steady state that is then disturbed at some
arbitrary time. In (6.5.4), then,
yref = A = Kxref (6.5.7)
so that (6.5.4) becomes
dy *
τ + y * = Kx* y * (0) = 0 (6.5.8)
dt
The deviation variable y* will have no contribution from the initial

condition and will become nonzero only when disturbance x* becomes
3
Lecture Notes - 6
nonzero. The difference between (6.5.5) and (6.5.8) is that the initial
condition is a steady state, as specified in (6.5.7). At long time,
y * ( ∞) = Kx * ( ∞) (6.5.9)
If the disturbance returns to zero, as for an impulse, so will the response.

A persistent disturbance, such as a step, gives a persistent response.
reference at long-term conditions: xref = x(∞); yref = Kxref

Other reference conditions are possible. If a long-term condition is
identified, (6.5.4) becomes
dy *
τ + y * = Kx* y * (0) = A − Kx( ∞) (6.5.10)
dt
The deviation variable y* will start at A - Kxs, which may or may not be
zero, and respond to the right-hand forcing function. Because the
reference has been selected as the long-term state, y* will necessarily tend
to zero.
reference at some arbitrary condition: xref, yref

The deviation variable will in general not start at zero, nor ultimately tend
to zero. The solution will mark the deviation from the reference state, as
driven by the initial conditions and forcing function.
6.6 An example with linearization and deviation variables

Consider a tank in which the outlet flow occurs by gravity.
Fi
ho Fo
Following the modeling procedure in 4.1, we begin with a material

balance.
d
ρAh = ρFi − ρFo (6.6.1)
dt
4
Lecture Notes - 6
A mechanical energy balance relates the outlet flow to the liquid level in
the tank through the frictional loss in the exit valve. For the normal
condition of turbulent flow through the valve,
dh
= Fi − α (h + ho )
0.5
A (6.6.2)
dt
where ho is the vertical distance between the tank bottom and the exit
nozzle.
Further detail
We write the mechanical energy balance between the liquid
surface (point 1) and the outlet nozzle (point 2).
2 2
v2 v
z1 − z2 = + KL 2 (6.6.3)
2g 2g
where KL is the loss coefficient. We represent the distance z1 – z2

by the sum of the liquid level h and the distance between tank
bottom and outlet nozzle, ho. In addition, we express the outlet
velocity in terms of the volumetric flow and the cross-sectional
area Ao of the exit pipe.
2
1+ K L  Fo 
h + ho =   (6.6.4)
2g  Ao 
Solving for flow
0.5
 2gAo2 
Fo =  (h + ho ) = α (h + ho )
0.5
(6.6.5)
 1+ K L 
and substituting (6.6.5) into (6.6.1), we obtain (6.6.2).
The nonlinear term may be approximated by Taylor series, referred to the

steady level hs.
1
(h + ho )0.5 ≈ (hs + ho )0.5 + (h − hs ) (6.6.6)
2(hs + ho )
0.5
On substituting into (6.6.2), we obtain the linearized model
dh α
= Fi − α (hs + ho ) − (h − hs )
0.5
A (6.6.7)
2(hs + ho )
0.5
dt
5
Lecture Notes - 6
Writing (6.6.7) at the steady reference condition,
dhs α
= 0 = Fis − α (hs + ho ) − (hs − hs )
0.5
A (6.6.8)
2(hs + ho )
0.5
dt
From (6.6.8) we relate the steady inflow and liquid level.
2
F 
hs =  is  − ho (6.6.9)
α 
Subtracting reference (6.6.8) from model (6.6.7), we obtain deviation

variables.
d (h − hs ) α
A = (Fi − Fis ) − (h − hs ) (6.6.10)
2(hs + ho )
0.5
dt
We indicate the deviation variables by * and rearrange into standard form,

thus identifying the characteristic time constant and steady-state gain.
2 A(hs + ho ) 2(hs + ho )
0.5 0.5
dh* *
+ h* = Fi (6.6.11)
α dt α
The time constant can be rewritten using (6.6.9) to show that it is more
than twice the residence time of the volume of liquid at the steady
reference condition.
2 A(hs + ho )
τ= (6.6.12)
Fis
Equation (6.6.11) may be solved by Laplace transforms for particular

disturbances in the inlet flow rate Fi.
6.7 Deviation variables are best in linearized equations

Consider a first-order lag in y disturbed by the ratio of independent inputs
x1 and x2.
dy x (t )
τ +y=K 1 y (0) = y s (6.7.1)
dt x 2 (t )
We write (6.7.1) at steady state and subtract the result from (6.7.1).
6
Lecture Notes - 6
τ
d
( y − y s ) + ( y − y s ) = K x1 − K x1s
dt x2 x2s
*
(6.7.2)
dy * x +x x
τ + y * = K 1* 1s − K 1s y (0) = 0
*
dt x2 + x2s x2 s
Deviation variables may be used, of course, but they are best suited to the
linear terms on the left-hand side. For process control purposes, we would
usually linearize (6.7.1).
6.8 The partial fraction expansion

If the function to be inverted is complicated, it may not be found in a table
of transform pairs. One method to simplify the polynomial functions often
found in control engineering is called partial fraction expansion.
N ( s) N ( s) C1 C2 C3
= = + + ... + + ...
D( s) ( s − α1 ) ( s − α 2 ) ... ( s − α1 )
n m n
( s − α1 ) n −1
(s − α 2 )m
(6.8.1)
The complicated ratio in (6.8.1) can be inverted if is expanded into a series

of simpler fractions. The inverse transform of each term will involve an
exponential function of the root αi.
C Ct n−1eαt
f (s) = ⇒ f (t) = (6.8.2)
(s − α ) n (n −1)!
Variety in the values of the coefficients Ci comes from the numerator
function N(s).
how to write the expansion
Arrange the denominator so that the coefficient of each s is 1. If there are
no repeated roots, each root appears in one term.
N ( s) C1 C2 C3
= + + (6.8.3)
( s − α1 )( s − α 2 )( s − α 3 ) ( s − α1 ) ( s − α 2 ) ( s − α 3 )
If a root is repeated, it requires a term for each repetition.
N (s) C1 C2 C3
= + + (6.8.4)
( s − α1 )( s − α 2 ) 2
( s − α1 ) ( s − α 2 ) 2
(s − α 2 )
The roots in (6.8.3) may appear as complex conjugate pairs. As an

alternative, they may be left in quadratic form.
7
Lecture Notes - 6
N ( s) C1 Bs + C
= + 2 (6.8.5)
( s − α1 )( s + bs + c ) ( s − α1 ) ( s + bs + c ) 2
2
how to solve for the coefficients

(1) For each of the real, distinct roots, multiply the expansion by each RH
denominator and substitute the value of the root for s to isolate the
coefficient. This also works for the highest power of a repeated root.
(2) With some roots determined, it may be easiest to substitute arbitrary
values for s to get equations in the unknown coefficients.
(3) For repeated roots, either
(3a) multiply the expansion by s and take the limit as s → ∞.
However, this will not isolate coefficients associated with
repeated complex roots.
(3b) multiply the expansion by the RH denominator of highest power.
Differentiate this equation with respect to s, and substitute the
value of the root for s. Continue differentiating in this manner to
isolate successive coefficients.
(4) For complex roots written as separate roots in (6.8.3), solving for one
coefficient is enough. The other will be the complex conjugate.
6.9 Partial fraction examples to make sense of those rules

the first order lag step response
From (5.4.4),
∆Ce −td s ∆Ce −td s 1

Co* ( s ) = = (6.9.1)
s (τs + 1) τ  1
s  s + 
 τ
Partial fraction expansion is applied only to the polynomial part of this

function. The denominator has already been factored, and we have
divided top and bottom by τ to put it in the form of (6.8.1). We write the
expansion as in (6.8.3); there are no repeated factors to consider.
1 C1 C2
= + (6.9.2)
 1 s  1
s  s +  s + 
 τ  τ
Using step (1), we solve (6.9.2) first for C1, and evaluate the equation at s
= 0.
8
Lecture Notes - 6
1 C2 s
= C1 +
 1  1
s +  s + 
 τ  τ
(6.9.3)
1
= C1
1
 
τ 
We proceed similarly for C2, with the other root.
1 C1  1
=  s +  + C2
s s  τ
(6.9.4)
1
= C2
−1
τ
The completed partial fraction expansion
1 τ −τ
= + (6.9.5)
 1 s  1
s  s +  s + 
 τ  τ
is substituted into (6.9.1), which is now easily inverted to the sum of a step
function and an exponential, both delayed by time td.
   
−t d s   1 
∆Ce τ − τ −1
Co* (s) =  +  = ∆C  + e −td s (6.9.6)
τ s s+  1 1
s s+ 
   
 τ  τ
The result is given in Lesson 5 as Equation (5.4.6).
complex roots
s s C1 C2
= = + (6.9.7)
s +ω
2 2
(s + jω )(s − jω ) (s + jω ) (s − jω )
Using step (1), we solve for each coefficient in turn.
9
Lecture Notes - 6
s C (s + jω )
= C1 + 2
(s − jω ) (s − jω )
− jω 1
= C1 =
(− jω − jω ) 2
(6.9.8)
s C (s − jω )
= 1 + C2
(s + jω ) (s + jω )
jω 1
= C2 =
( jω + jω ) 2
Notice that C1 and C2 indeed qualify as complex conjugates, as indicated

in step (4), although the absence of an imaginary term makes it an
incomplete test of the rule. The inverse transform is a sum of exponential
functions, which may be rearranged by Euler's equation to the cosine
function.
1 1 1 
f (s) =  + 
2  s + jω s − jω 
1
(
f (t) = e − jωt + e jωt
2
) (6.9.9)
1
= (cos ωt − j sin ωt + cos ωt + j sin ωt )
2
= cos ωt
repeated roots
Using (6.8.4),
s 1 s C1 C2
= 2 = + (6.9.10)
(τs +1) 2
τ  1  1  1
2 2
s +  s +  s +τ 
 τ  τ  
Isolate C1 according to step (1).
s  1
= C1 + C2  s + 
τ2
 τ
(6.9.11)
−1
τ = C = −1
τ τ
2 1 3
Using step (2), we find it easy to evaluate (6.9.10) at s = 0.
10
Lecture Notes - 6
−1
1 0 τ3 + C2
= (6.9.12)
τ 
2
1
2
 1
2
 1
0 +  0 +  0 + 
 τ  τ  τ
which gives C2 = τ-2.
Alternatively, we proceed by step (3a).
−s
1 s2 τ 3 + C2 s
= (6.9.13)
τ2  1
2
 1
2
 1
s +  s +  s + 
 τ  τ  τ
Taking the limit as s → ∞, we find, as before,
1
= 0 + C2 (6.9.14)
τ2
By step (3b), it is necessary to differentiate (6.9.11) with respect to s. This

leads quickly to the same result for C2. Thus, substituting into (6.9.10),
−1 3 1 2
s τ τ
= + (6.9.15)
(τs + 1) 2
 1   s + 1 
2
s + 
 τ  τ
and the inverse transform is
−t −t
−1 te τ 1 e τ
τ 1
+ τ
3 2 −t
= (τ − t )e τ
(6.9.16)
(2 − 1)! (1 − 1)! τ 3
repeated complex roots
8
(6.9.17)
s ( s + 2 s + 2) 2
2
Write the partial fraction expansion. Each repeated root accounts for 2
terms.
8 C C2 C3 C4 C5
= 1+ + + +
s( s + 1 − j ) ( s + 1 + j )
2 2
s (s + 1 − j) (s + 1 − j) (s + 1 + j) (s + 1 + j)
2 2
(6.9.18)
11
Lecture Notes - 6
Solve for C1, associated with the single root s = 0, by step (1).
8
= C1 + s{others} (6.9.19)
(s + 1 − j) (s + 1 + j)2
2
Evaluate the expression at the single root. C1 is found to be 2.
By step (1), solve for C2, associated with the highest power of the repeated
root s = -1 + j.
8
= C2 + ( s + 1 − j ) 2 {others} (6.9.20)
s( s + 1 + j ) 2
Evaluate the expression at the root, remembering to multiply by complex

conjugates, if needed, to remove j from the denominator. C2 is found to be
1 + j.
For repeated roots, step (3a) will not help, so we use step (3b). Rewrite
(6.9.19) to reveal C3, the coefficient associated with the next lower power
of the repeated root s = -1 + j.
8
= C2 + C3 ( s + 1 − j ) + ( s + 1 − j ) 2 {others} (6.9.21)
s( s + 1 + j ) 2
Now differentiate (6.9.21) with respect to s; the left-hand-side requires use

of the chain rule for differentiating a quotient.
[
0 − 8 (s +1+ j) 2 + 2s(s +1+ j)
= C3 +
]d
[
(s +1− j) 2 {others} ]
s (s +1+ j)
2 4
ds
(6.9.22)
When the product differentiation is performed on the last term, all its
component terms will feature a factor of (s + 1 – j) to some power. Thus
when the equation is evaluated at the repeated root, they will vanish. C3 is
found to be –1 + 2j. Step (2) could have given the same result, but would
require some lengthy algebra.
Because C4 and C5 are associated with the complex conjugate root, they
will be complex conjugates of C2 and C3. Thus by step (4) we
immediately find that C4 is 1 – j, and C5 is –1 - 2j. Then the complete
partial fraction expansion is
12
Lecture Notes - 6
8 2 1+ j −1+ 2 j 1− j −1 − 2 j
= + + + +
s ( s + 2 s + 2)
2 2
s (s + 1 − j) (s + 1 − j) (s + 1 + j) (s + 1 + j)
2 2
(6.9.23)
As a check on our manipulations, we test the expansion for particular

values of s. For s = 1, we find each side equal to 0.32. For s = 1 - j, we
find each side equal to –0.125 + 0.125j.
6.10 Short cuts

Partial fraction expansion leads to the following short cuts for inversion to
time domain:
terms in a partial fraction type of root corresponding time domain

expansion inversion
C single real root Cebt
s −b
C repeated real root Ctebt
( s − b) 2
C C* pair of complex roots 2 C eαt cos( βt + φ )
+
s − b s − b*
C C* repeated pairs of complex 2 C teαt cos( βt + φ )
+ roots
( s − b) 2 ( s − b* ) 2
In this table,
• C* and b* are conjugates of C and b
• C = |C|ejφ.
• b = α + jβ
6.11 Summing up
By using linearization, deviation variables, and Laplace transforms, we
can solve a great variety of process control model equations.
13
Lecture Notes - 7
Lesson 7. Transfer functions and block diagrams
7.0 Context
In Lesson 5, we introduced Laplace transforms as a method of solving the
linearized equations of system dynamics. However, a more important
reason is that so many of the concepts of dynamics and control theory are
expressed in the LT language, even if the methods are not required for
solution. In this lesson we present the transfer function, a concise
description of a dynamic system that is based on Laplace transforms. We
also present block diagrams, a convenient way to represent the structure of
dynamic systems.
7.1 Dynamics of systems

Process control deals with systems that change in time. In Lesson 2, we
asserted that systems are characterized by input disturbances (causes), and
output responses (effects).
We have claimed that a variety of physical systems can be satisfactorily

described by relatively few mathematical models. We have dwelt on the
first-order lag as a prime example:
dy *
τ + y * = Kx * (t ) y * (0) = 0 (7.1.1)
dt
The system model is the ordinary differential equation (7.1.1), relating

input x* and output y*as they vary in time. In (7.1.1), x*(t) is the
mathematical forcing function, and y*(t) the dependent variable. After
taking Laplace transforms, we can relate input and output by an algebraic
equation:
K *
y * (s) = x (s) (7.1.2)
τs + 1
The ratio in (7.1.2) contains all the information about the ODE (7.1.1).
When multiplying x*(s), the transform of disturbance x*(t), the ratio
converts it into y*(s), the transform of the response y*(t). We call this ratio
the transfer function.
1
Lecture Notes - 7
7.2 Transfer functions
Let's take a larger view of getting transfer functions from differential
equations: we have a lot more to learn about dynamic systems, but it’s not
too early to speculate a bit from what we already know. Perhaps systems
more complicated than our first-order lag may be described by higher-
order equations. If we linearize such equations, and express them in
deviation variables, they must look like
d n y* d n−1 y * dy *
an n
+ a n−1 n−1
+ K + a1 + y * = f (x * ) (7.2.1)
dt dt dt
It’s not outlandish to speculate that a complicated dynamic system might

depend not only on the disturbance x*, but its rate of change, as well. For
that matter, it may depend on higher derivatives of x*, leading us to write
(7.2.1) as
d n y* d n−1 y * dy * d l x* dx*
an n
+ a n−1 n−1
+ K + a1 + y *
= bl l
+ K + b1 + b0 x* (t)
dt dt dt dt dt
(7.2.2)
We have already encountered systems with multiple disturbances in

Section 4.4. Hence we may expand our speculative model further.
* *
d n y* d n−1 y * dy * d l x1 dx *
an n
+ a n−1 n−1
+ K + a1 + y *
= bl l
+ K + b1 1 + b0 x1 (t)
dt dt dt dt dt
m * *
d x2 dx *
+ cm m
+ K + c1 2 + c0 x 2 (t)
dt dt
+K
(7.2.3)
As is usual in process control, we presume all initial conditions are zero,

which describes a system expressed in deviation variables initially at
steady state. Taking Laplace transforms of (7.2.3) leads to
y * (s) =
(b s + b s
l
l
l −1
l −1
+ K + b1 s + b0 ) *
x1 (s)
(a s + a s
n
n
n−1
n−1
+ K + a1 s + 1)
+
(c s + c
m
m
m−1s m−1 + K + c1 s + c0 ) *
x2 (s) (7.2.4)
(a s + a
n
n
n−1 s
n −1
+ K + a1 s + 1)
+K
If we set all the (7.2.4) coefficients except b0 and a1 to zero, we recover

the particular example of (7.1.2). The ratios of polynomials in (7.2.4), like
2
Lecture Notes - 7
the ratio in (7.1.2), are transfer functions. We will represent a general
transfer function by G(s). Thus (7.2.4) becomes
* *
y * (s) = G1 (s)x1 (s) + G2 (s)x 2 (s) + K (7.2.5)
G1(s) is the transfer function that relates y*(s) to x1*(s). G2(s) similarly
relates y*(s) to x2*(s). Notice that
• The Laplace transforms of the disturbances, when substituted for the

xi(s) variables, will not change the polynomial nature of the Gi(s)
terms in (7.2.4). Thus polynomial ratios in the Laplace variable s will
always result from the linear, constant-coefficient ordinary differential
equations of process control.
• It is the nature of the linear ODE that the effects of the inputs are
additive. Each disturbance xi*(s), when processed through its
particular transfer function, contributes to the overall response of y*(s).
• As we learned from the partial fraction expansion of Lesson 6, the
time-domain response will finally be a sum of exponential and
trigonometric terms. The various time constants, frequencies, and
phase lags in these terms are determined by the coefficients in the
transfer functions, and thus the original differential equation.
The dynamic response calculated from Equation (7.2.5) may be

complicated indeed, but the essential concept - a dynamic system
approximated by a linear equation and expressed in terms of transfer
functions - is no different from what we have already studied in the
simpler first order system of (7.1.2).
7.3 Using the transfer function

A transfer function represents a differential equation. Just as we classify
differential equations into recognizable types, we will classify transfer
functions, learn their characteristics, and use them as a concise
representation of particular behaviors. For example, the transfer function
in (7.1.2) represents a first-order lag; it contains the same information as
the ODE of (7.1.1).
Given the transfer function for a system, therefore, we can predict some
features of its behavior without actually calculating its response to
particular disturbances. Consider these terms:
order – the highest power of s in the denominator. Equivalent to the order

of the differential equation describing the system.
The first-order lag is described by a first-order differential
equation; its transfer function has a single s in the denominator.
pole – root of the denominator. In later lessons, we will learn that poles
with negative real parts result in output signals that decay in time,
3
Lecture Notes - 7
so that the system will be stable. If there exist poles with
imaginary parts, the system may oscillate, even without oscillatory
disturbances.
The first order system has a pole at -τ-1; this negative, real value
indicates a stable response with no oscillation.
zero – root of the numerator. These generally have no influence on

stability, but can influence the rate and character of the dynamic
response.
The first-order system has no zeroes.
steady state gain – the ratio of long-term output change to input step
change. The gain is a measure of how sensitive the system is to
disturbances. If the system is a chemical process, we would like a
low value of gain, so that disturbances would have little effect on
the output variable. In a sound system, we would like a large gain,
so that tiny input signals from the source (tape, vinyl, CD) are
amplified to audibility. The gain is found by setting s = 0 in the
transfer function.
Here we summarize the first-order lag and integrator with respect to these
properties
type equation transfer poles steady state gain

function
lag dy K -τ-1 K
τ + y (t ) = Kx(t )
dt τs + 1
integrator dy K 0 none; increases
τ = Kx(t ) without bound
dt τs
7.4 Transfer function for the stirred reactor

Let’s combine our knowledge of modeling first-order systems from
Lesson 4, Laplace transforms from Lessons 5 and 6, and notions of the
transfer function. In Section 4.1 we modeled a stirred overflow tank
containing a dissolved substance A. Let’s now assume that A disappears
by first order chemical reaction.
1 dN A
rA ≡ = −kC A (7.4.1)
V dt
where the negative sign shows that A is consumed in the reaction. The
component material balance is written assuming that the volumetric flow
rate F is constant.
dC Ao
V = F (C Ai − C Ao ) −VkC Ao C Ao (0) = C As (7.4.2)
dt
4
Lecture Notes - 7
There are no nonlinear terms; subtracting the steady state from (7.4.2)
leaves deviation variables.
*
dC Ao * * * *
V = F (C Ai − C Ao ) − VkC Ao C Ao (0) = 0 (7.4.3)
dt
Equation (7.4.3) is then put into standard form:
*
V dC Ao * F *
+ C Ao = C Ai
F + Vk dt F + Vk
*
(7.4.4)
dC Ao
τ *
+ C Ao = KC Ai
* *
C Ao (0) = 0
dt
The time constant is smaller than that for the mixing tank in Equation
(4.1.4) -- in a manner analogous to the multiple outlet streams in Section
4.5, the combination of outflow and chemical consumption in (7.4.4)
reduces the time response of the outlet concentration. Similarly the gain is
less than unity -- a disturbance in inlet concentration is only partly
transmitted to the outlet stream
Taking Laplace transforms of (7.4.4) gives
K
C *Ao (s) = *
C Ai (s) (7.4.5)
τs + 1
Guided by (7.1.2), we identify the transfer function of the mixing tank.
K  C Ao
*
(s) 
G(s) =  = *  (7.4.6)
τs +1  C Ai (s) 
As we remarked in Section 7.2, the transfer function depends only on the

geometry and operating conditions of the tank itself, not on the
disturbance. The particular nature of the inlet disturbance CAi*(s), when
worked through the transfer function G(s), gives the particular nature of
the output response CAo*(s).
Detail
In process control, we think of the gain K as a measure of how a
permanent change in input CAi* affects the output CAo* in the long
term. However, the steady-state performance of the reactor -- that
is, how well it converts CAis to CAos -- is also indicated by the gain.
From material balance (7.4.2), written at steady state, we find that
5
Lecture Notes - 7
the gain depends on the reaction rate constant k and the reactor
residence time τR.
C Aos F 1
=K= = (7.4.7)
C Ais F + Vk 1+ kτ R
Low gain means good conversion of reactant A.
By placing the equations in standard form, we found that the time

constant τ depends on the tank volume, the flow, and the reaction
rate constant.
V τR
τ= = (7.4.8)
F + Vk 1+ kτ R
These parameters are important individually, of course, but when

we are concerned with the dynamic response, it is important to
identify how they interact to affect the time constant. We see that
the time constant is related to the reactor residence time τR, which
we use in designing a stirred reactor to produce a desired outlet
concentration of reactant. For understanding how that outlet
concentration varies in time, however, the time constant is more
significant. Notice that for no reaction, the time constant reduces
to the residence time. As the reaction rate increases, the time
constant decreases, indicating that the outlet concentration
responds more quickly to disturbances
7.5 Block diagrams
The block diagram is a graphical display of the process model in the
Laplace domain. It comprises blocks and arrows, and thus resembles
many other types of flow diagram. In our use with control systems,
however, the arrows represent signals, variables that change in time, which
are not necessarily actual flow streams. The block contains the transfer
function, which may be as simple as a units conversion between x and y,
or as complicated as a full chemical process. Remember that the transfer
function incorporates all the dynamic information in the process model.
x(s) y(s)
G(s)
This diagram means
y(s) = G(s)x(s) (7.5.1)
6
Lecture Notes - 7
Returning to the mixed reactor in Section 7.4, we can represent the

dynamic behavior of the reactor by a block diagram that is equivalent to
Equation (7.4.5):
CAi*(s) K CAo*(s)
τs + 1
7.6 Block diagram structures

The real value of block diagrams is to represent the flow of signals among
multiple blocks.
The Block Diagram Rules:

• only one input and output to a block. The figure in Section 7.1 was fine for its purpose, but
does not qualify as a process control block diagram.
• two signals may be summed at an explicit summing junction. The

algebraic sign is indicated at the junction.
x1(s) y1(s)
G1(s)
+ x3(s) y3(s)
G3(s)
-
x2(s)
G2(s)
y2(s)
• a single signal may feed its value to multiple blocks. This does NOT
indicate that the signal is split among the blocks.
y1(s)
x1(s) x2(s)
G1(s) G2(s)
y2(s)
x2(s)
G3(s)
Block diagrams may be turned into equations by simple algebra. It is

usually most convenient to start with an output and work backwards by
substitution. In the summing diagram
7
Lecture Notes - 7
y3 ( s ) = G3 ( s ) x3 ( s )
= G3 ( s )( y1 ( s ) − y2 ( s ) )
(7.6.1)
= G3 ( s )(G1 ( s ) x1 ( s ) − G2 ( s ) x2 ( s ) )
= G3 ( s )G1 ( s ) x1 ( s ) − G3 ( s )G2 ( s ) x2 ( s )
In the multiple assignment diagram
y1 ( s ) = G2 ( s ) x2 ( s )
= G2 ( s )G1 ( s ) x1 ( s )
(7.6.2)
y2 ( s ) = G3 ( s ) x2 ( s )
= G3 ( s )G1 ( s ) x1 ( s )
Similarly, equations may be turned into block diagrams. System (7.6.1)

has two inputs and thus requires at least 2 blocks.
x1(s)
G3(s)G1(s)
+ y3(s)
-
x2(s)
G3(s) G2(s)
System (7.6.2) has two outputs for one input. Input x1 is not split – its full
value is sent to each of two blocks.
x1(s) y1(s)
G2(s)G1(s)
y2(s)
G3(s)G1(s)
This pair of block diagrams is equivalent to the pair from which they were
derived.
7.7 Splitting a signal

The block diagram rules contain a summing junction, but no explicit
provision for splitting a signal. And yet, chemical processes frequently
contain junctions at which a flowing stream is divided -- certainly the full
value of the stream is not fed to each destination! To represent such splits
on a block diagram, we feed the input signal into separate blocks. The
transfer functions in these blocks express the manner in which the flow
was split.
8
Lecture Notes - 7
For example, suppose that an inlet flow always divides so that one branch
receives fraction ε of the flow and a second receives the remainder. Then
the material balance is
wi = wo1 + wo 2
(7.7.1)
= εwi + (1 − ε )wi
and the block diagram is drawn
wi ε wo1
1-ε wo2
The gain of each transfer function is less than unity, showing that each
output signal is diminished from the input value.
Detail
In fact, the transfer functions are nothing but gain, which implies
that the response of the outlet to inlet disturbances is instantaneous.
Of course, we assumed this in writing the material balance (7.7.1).
For an incompressible flow, this is a good description.
The flow junction might behave so that the flow fraction ε depends
on the magnitude of the inlet flow wi. For a compressible flow, the
dynamic response of the branches might differ, so that the transfer
functions would depend on Laplace variable s. The complexity of
the transfer functions depends on the detail of our modeling;
however, the principle of splitting a signal on a block diagram is
the same as in our simple incompressible case (7.7.1).
7.8 Describing systems

We began our study of dynamic systems by writing differential equations.
Then we adopted the Laplace transformation of these equations as an
equivalent description. Now we have introduced block diagrams as yet
another description. That does it for descriptions -- we will now apply
them to increasingly complicated systems. Whatever the means, our
purpose is to calculate an output response for an input disturbance, as we
have in many examples of first-order systems.
9
Lecture Notes - 8
Lesson 8. Arrangements of first-order systems -- series, parallel, and recycle
8.0 Context
We have studied varieties of the first-order lag in Lessons 4 - 7. Here we
take two stirred reactors and arrange them in different ways. This will
require us to use more than one differential equation in modeling, and thus
lead us to higher-order systems. We shall also gain more experience in
using Laplace transform and block diagram methods.
8.1 Reactors in series

Combine identical well-mixed overflow reactors in series.
F, CA0
F, CA1
F, CA2
V
Recalling Section 7.5, the component material balance is written for each
tank, assuming that the volumetric flow rate F is constant.
dC A1
V = F (C A0 − C A1 ) − VkC A1
dt
(8.1.1)
dC A2
V = F (C A1 − C A2 ) − VkC A2
dt
The balances are evaluated at a steady state reference condition.
0 = F (C A0s − C A1s ) −VkC A1s

(8.1.2)
0 = F (C A1s − C A2s ) −VkC A2s
The steady state balances are subtracted from the transient balances so that
the concentrations may be expressed as deviation variables. We choose
the initial condition as the reference steady state, so that we may examine
disturbances to that condition. Rearranging to standard form, we
recognize first-order systems in which we can group the physical
parameters of flow, volume, and rate constant to identify a time constant
and gain.
1
Lecture Notes - 8
*
V dC A1 * F * *
+ C A1 = C A0 C A1 (0) = 0
F + Vk dt F + Vk
*
(8.1.3)
V dC A 2 * F * *
+ C A2 = C A1 C A 2 ( 0) = 0
F + Vk dt F + Vk
Taking the two equations together to represent a single system, we identify

the input disturbance as CA0*, the output response as CA2*, and an
intermediate variable as CA1*.
8.2 In series: time domain solution by successive integration

These first-order equations are sequential: we can integrate the first
equation to determine CA1* for CA0* as a forcing function. Then we
substitute CA1* as the forcing function in the second equation. Two
equations, and two integrals in the solution, indicate a 2nd order system.
−t
τ t
e t
∫ e τ C A0 (t )dt
* *
C =K
τ
A1
0
−t
(8.2.1)
e τ t  t
t 
∫  ∫ e τ C (t )dt dt
* *
C A2 = K 2
τ 2 A0
0 0 
where
V F
τ= K= (8.2.2)
F + Vk F + Vk
Upon substituting a particular disturbance for the input CA0*, we can

perform the double integration to obtain the response CA2*.
8.3 In series: Laplace transform solution

Transforming (8.1.3),
( )
τ sC A1* ( s) − C A1* (0) + C A1* ( s) = KC A0* ( s )
(8.3.1)
τ (sC A2
*
( s) − C A2
*
(0) ) + C A2
* *
( s ) = KC A1 ( s )
Eliminating intermediate variable CA1*, we relate CA2* to disturbance CA0*

through the transfer function for the system.
* K2 *
C A2 ( s) = C A0 ( s ) (8.3.2)
(τs + 1) 2
Upon substituting a particular disturbance for the input CA0*, we can

perform the inverse transform to obtain the response CA2*(t). Even
2
Lecture Notes - 8
without doing any calculations, we notice that the second order transfer
function has a repeated pole at –τ-1. Thus all poles are real and negative,
indicating a stable, non-oscillating response to step and pulse disturbances.
8.4 In series: block diagram solution

It is instructive to express Equations (8.3.1) as transfer functions, as shown
in the block diagram.
CA0*(s) CA1*(s) CA2*(s)

K K
τs + 1 τs + 1
By the rules of block diagrams, we relate output CA2* to input CA0*:
* K *
C A 2 (s) = C A1 (s)
(τs +1)
(8.4.1)
K  K * 
=  C A0 (s)
(τs +1)  (τs +1) 
which leads again to (8.3.2).
8.5 In series: step response

*
If CA0 is a step disturbance ∆C at time td, the response from (8.3.2) is
* K 2 ∆Ce −td s
C A2 (s) = (8.5.1)
(τs +1) 2 s
Inverting via partial fractions or a table of transform pairs,
*   (t − t d )  −(t −td )τ 
C A2 (t) = ∆CK 2 1− 1+ e  (8.5.2)
  τ  
where the response is understood to be zero before time td. Equation
(8.5.2) could also have been obtained by integrating the time domain
solution (8.2.1) for a step disturbance.
step response example
For example, the step in input concentration is 1 mol m-3, the time
constant is 8 minutes, and the process gain is 0.5, where both gain and
time constant depend on flow and kinetics by (8.2.2). If the step
disturbance occurs at td = 5 minutes, the outlet concentration responds as
shown in Figure 8-1.
3
Lecture Notes - 8
1.2
disturbance
1
disturbance and response (mol min -1)
0.8
0.6
intermediate C1
0.4
equivalent single tank
0.2
output C2
0
0 5 10 15 20 25 30 35
time (min)
Figure 8-1. Series response compared to first order
Notice that outlet concentration C2 reacts immediately to the disturbance

(a consequence of the complete mixing assumption). However, the rise is
gradual and not sharp as for a first-order system (such as intermediate
concentration C1 leaving the first reactor). Also shown is the response of a
single tank. Its volume is equal to the combined volume of the tanks in
series, so that the reactor residence times are identical. At 32 minutes, the
equivalent tank has reached 92% of its ultimate concentration, compared
to 85% for the tanks in series. The equivalent tank has a larger gain
(lower conversion).
8.6 In series: unequal tanks

Let’s generalize our results slightly; the two tanks have different volumes.
CA0
V1
V2
4
Lecture Notes - 8
From (8.3.1) the Laplace domain material balances for the two tanks are
K1 K2
*
C A1 (s) = *
C A0 *
(s) and C A2 (s) = *
C A1 (s) (8.6.1)
τ 1s +1 τ 2 s +1
We can represent these equations in a block diagram, which resembles the

process sketch.
CA0 CA1 CA2

G1(s) G2(s)
Eliminating C*A1(s), either algebraically or by applying block diagram

rules, we find the series transfer function.
*
C A2 (s) K1 K 2
= G(s) = (8.6.2)
*
C A0 (s) (τ 1 s + 1)(τ 2 s + 1)
The response to a step of ∆C at time td is
 τ1 −(t −t d )
τ1 τ2 −(t −t d )
τ2 
*
C A2 = ∆CK1 K 2 1+ e − e  τ 1 ≠ τ 2 (8.6.3)
 τ 2 −τ1 τ 2 −τ1 
In contrast to (8.5.2), (8.6.3) combines two separate exponential terms.

Equation (8.5.2) is the limiting case when the two time constants are
identical.
8.7 Reactors in parallel arrangement
(1-b)F CA1
1
CAO CA
bF
2
CA2
The stream is split so that concentration CA0 is fed to each reactor. The
resultant concentration CA is NOT the sum of the outlet concentrations,
but is given by material balance around the mixing point.
C A = (1 − b)C A1 + bC A2 (8.7.1)
Combining (8.7.1) with first-order transfer functions yields the parallel

structure transfer function.
5
Lecture Notes - 8
C A* (s) K (1− b) K 2b
= G(s) = 1 + (8.7.2)
*
C A0 (s) (τ 1s +1) (τ 2 s +1)
In expressing this as a block diagram, we use in each path the block for the
individual reactor, followed by another to apply the material balance (flow
fraction). These material balance blocks are easy to neglect if the block
diagram is drawn too hastily from the process diagram. Remember that
the arrows in the block diagram do not represent material flows, as much
as they seem to!
1 (1-b)
C*AO(s) C*A(s)
2 b
Rearranging to one term
(K1 (1− b)τ 2 + K 2bτ 1 )s + K1 (1− b) + K 2b

G(s) = (8.7.3)
(τ 1s +1)(τ 2 s +1)
This parallel flow transfer function shows negative, real poles, so it will be
stable. However, it has a numerator zero, so that its dynamic response will
differ from that of the series structure. Indeed, the step response is the
sum of two distinct first-order responses, weighted by the flow split.
C A*  − ( t −t d )
τ1   − ( t −t d )
τ2 
= K1 (1− b)1− e  + K 2b1− e  (8.7.4)
∆C    
8.8 Reactors in recycle arrangement
CAO (1+b)F CA
Cm 1
bF
2
CA2
Letting Cm represent the concentration after the mixing point, the transfer
functions are
6
Lecture Notes - 8
C A* (s) K1 F (1+ b) V1
= G1 (s) = K1 = τ1 =
*
Cm (s) τ 1s +1 F (1+ b) + kV1 F (1+ b) + kV1
C A* 2 (s) K2 Fb V2
= G2 (s) = K2 = τ2 =
*
C A (s) τ 2 s +1 Fb + kV2 Fb + kV2
C A* 0 (s) + bC A* 2 (s) = (1+ b)Cm* (s)
(8.8.1)
Combining, the recycle structure transfer function is
1
K1 (τ 2 s +1)
G(s) = 1+ b (8.8.2)
(τ 1s +1)(τ 2 s +1) − b K1K 2
1+ b
This is consistent with the block diagram
C*AO(s) C*A(s)
1/(1+b)
1
b/(1+b) 2
Once again, the material balance blocks modify the input signals to the
summing junction. The block diagram resembles the process, but the
arrows that represent its signals (change in concentration) cannot be
summed by direct analogy to the arrows representing flow diagram
streams.
In the transfer function, the poles are no longer those of the individual
reactors. Rewriting to allow later expansion by partial fractions
1 K1
(τ 2 s +1)
1+ b τ 1τ 2
G(s) = (8.8.3)
(s − α1 )(s − α 2 )
where
b
− (τ 1 + τ 2 ) ± (τ 1 − τ 2 ) 2 + 4 τ 1τ 2 K1 K 2
α1,2 = 1+ b (8.8.4)
2τ 1τ 2
7
Lecture Notes - 8
For CSTR, 0 < K < 1, and so these roots will be real and negative.
However, for more general first order systems, this equation form admits
the possibility of positive roots, so that arranging two stable systems in a
recycle structure could make them unstable. The step response comprises
three terms from the partial fraction expansion
C A* 1 K1  1 (τ α +1) α1t (τ α +1) α 2t 

=  + 2 1 e + 2 2 e  (8.8.5)
∆C 1+ b τ 1τ 2  α1α 2 α1 (α1 − α 2 ) α 2 (α 2 − α1 ) 
Figure 8-2 shows step responses for two reactors (1 and 0.6 m3) arranged
alternately in series, in parallel with 40% of the flow going to the smaller
vessel, and in recycle with 40% of the throughflow returned through the
smaller vessel. Flow is 0.01 m3 s-1 and rate constant 0.01 s-1 in all cases.
The latter arrangements show faster response but less satisfactory
mitigation of the disturbance (that is, higher gain) than the series
arrangement.
0.5
0.45 recycle
parallel
0.4
0.35
0.3
CA /step
series
0.25
0.2
0.15
0.1
0.05
0
0 100 200 300 400 500
time (s )
Figure 8-2. Response of series, parallel, and recycle reactor

arrangements to a step in inlet concentration. The change in outlet
concentration is normalized by the magnitude of the step.
8.9 Higher-order systems, and dead time

Connecting further tanks makes the system correspondingly higher order.
Thus by a series of first order systems, we can get an infinite number of
higher order systems. Suppose we have a well-mixed overflow tank of
time constant τ. If we introduce a step increase in the inlet temperature or
concentration, we will (by the well-mixed assumption) immediately detect
a rise in the outlet stream – the familiar first-order lag response. If we
8
Lecture Notes - 8
have instead two tanks in series, each half the volume of the original, we
will detect a second-order, sigmoid response at the outlet.
If we continue to increase the number of tanks in the series, always

maintaining the total volume, we observe a slower initial response with a
faster rise around the time constant. This behavior is shown in Figure 8-3.
1.2
0.8
step response
0.6
0.4
1
2
6
12
0.2
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
time/tau
Figure 8-3. Step response for series of tanks of total time constant τ.
9
Lecture Notes - 8
If taken to the limit of an infinite number of tanks, we finally obtain a pure

delay, in which the full step disturbance is not seen at the outlet until time
τ has passed. This is the dead time, or transmission delay; it is familiar to
anyone who has waited at the faucet for the hot water to arrive.
Now, imagine a pipe of volume V carrying a volumetric flow F. The

residence time of the pipe is
V L
θ= = (8.9.1)
F v
An observer at the entrance of the pipe injects a short pulse of a tracer

material, also measuring the concentration. An observer at the exit also
measures the concentration. Figure 8-4 shows the data. Of course,
nothing can be observed at the exit until sufficient time has passed to
convey the pulse through the pipe. This is the dead time.
x(t) seen by observer

at entrance of pipe y(t) seen by observer
at exit of pipe
signal
time
td td + θ
Figure 8-4. Pulse disturbance in a pipe of dead time θ.
The dead time can be expressed as
y(t) = x(t − θ ) (8.9.2)
where the response y at time t is identical to the disturbance x as it was

earlier at a time θ before t. The Laplace transform of dead time is
L(x(t − θ ) ) = x(s)e=−sθ (8.9.3)
where x is any function.
Imagine a simple process consisting of a mixed tank in series with a pipe.
10
Lecture Notes - 8
As always, we wish to know the effect of input disturbances on the output.

By a component material balance, we find that
*
dCo
τ +=Co = Ci (t − θ ) =
* *
(8.9.4)
dt
where τ is the time constant of the tank, and θ the dead time due to the
pipe, from (8.9.1). Taking Laplace transforms, we find the transfer
function to be
*
Co (s) K −sθ=
= e (8.9.5)
Ci (s) τs +1
*
where K is unity. Now imagine a step disturbance at time td. Thus
* 1 −sθ= A −st d
Co ( s ) = e e (8.9.6)
τs + 1 s
Inverting, we find the response combines first order and dead time
behaviors.
* θ)
Co (t )  −(t −t d −=
τ= 
= 1 −=e  (8.9.7)
A = =
The results are plotted in Figure 8-5. The onset of disturbance td has been
set equal to τ, and the dead time to τ /2. With these arbitrary choices, the
disturbance occurs at dimensionless time 1. At 1.5, the output shows a
first-order response. The results do not depend on the physical order of
the lag and the dead time element.
Notice the resemblance of the output trace in 8-5 to the higher-order traces
in Figure 8-3. We will find that the first-order-dead-time (FODT) process
of (8.9.5) is frequently used to represent high-order processes whose
transfer functions are not known.
11
Lecture Notes - 8
1.2
1
disturbance and response
0.8
0.6
0.4
0.2
0
0 1 2 3 4 5 6 7
time/tau
Figure 8-5. FODT (first order plus dead time) step response.
Detail
By the way, here are step responses for series of identical first
order lags.
1 1 −t
⇒ 1− e τ
τs + 1 s
1 1  t  −t
⇒ 1 −  1 + e τ
(τs + 1) s
2
 τ
(8.9.8)
1 1  t 1 t 2  −t τ
⇒ 1 − 1 + + e
2 
(τs + 1)3 s  τ 2τ 
1 1  t 1 t2 1 t n  −t τ

⇒ 1 − 1 + + +L+ e
n 
(τs + 1)n+1 s  τ 2 τ 2
n! τ 
12
Lecture Notes - 9
Lesson 9. Second-order systems
(I apologize; these notes are incomplete.)
9.0 Context
In Lesson 8, combining two first-order lags gave us a variety of second-
order systems. In this lesson we will find that second-order systems are
even more versatile than we have seen.
9.1 The second-order system

In Lesson 4 we introduced the first-order system by writing a material
balance on a mixed tank and obtaining a first-order differential equation.
Here we will introduce the second-order system directly from its equation.
d2y dy dx
α 2
+β + y(t) = Kx(t) + M
dt dt dt
(9.1.1)
dy
y(0) = =0
dt t=0
As before, we will assume that all initial conditions on response variable

y(t) are zero, so that the system is initially at steady state, and will be
driven only by disturbance x(t). Parameter K has dimensions of y divided
by x; it is the steady-state gain. Parameter M measures the sensitivity of
the system to the rate of change of the disturbance x(t); it has the
dimensions of K multiplied by time. Both K and M may be positive or
negative, as may α and β.
We solve (9.1.1) by Laplace transform.
y(s) Ms + K
= 2 (9.1.2)
x(s) αs + βs + 1
The behavior of the solution depends fundamentally on the poles of

(9.1.2). A map relating system stability to coefficients α and β is given in
Figure 9-1.
1
Lecture Notes - 9
3
unequal, negative
poles: stable, non-
oscillating
2 equal, negative
poles: stable, non-
oscillating
1
complex negative
poles: stable,
unequal poles of
oscillating
opposite sign:
beta
0 excursively
unstable complex positive imaginary poles:
poles: unstable, persistent
-1 oscillating oscillation
equal, positive
poles: excursively
-2 unstable
unequal, positive
poles: excursively
unstable
-3
-1.6 -1.2 -0.8 -0.4 0 0.4 0.8 1.2 1.6 2
alpha
Figure 9-1. Second-order system stability related to equation

coefficients
From the properties of the quadratic equation (3.1.1) we recall that the
poles are real for
β2
α< (9.1.3)
4
With real poles, it is convenient to factor the denominator, and thus

express the second-order system in terms of two first-order systems in
which the characteristic times τ1 and τ2 may be positive or negative. We
compare side-by-side second-order and first-order systems:
second-order first-order
equation 2
d y dy dx dy
τ 1τ 2 2 + (τ 1 + τ 2 ) + y (t ) = Kx(t ) + M τ + y (t ) = Kx (t )
dt dt dt dt
dy y ( 0) = 0
y ( 0) = =0
dt t =0
transfer y( s) K + Ms y ( s) K
= =
function x ( s ) (τ 1s + 1)(τ 2 s + 1) x ( s ) τs + 1
2
Lecture Notes - 9
Poles 1 1 1
s=− ,− s=−
τ1 τ2 τ
Zeroes −K none
s=
M
The system will be stable if both τ1 and τ2 are positive, as they were for the
mixing tanks of Lesson 8. However, we will see an example of a negative
characteristic time in the energy balance for an exothermic chemical
reaction. Second-order responses to common disturbances are given in
Table 9.1. Disturbances occur at arbitrary time td.
Step: x (t ) = Au (t − t d )
 − (t − td ) − (t − td ) − (t − t d )
 M  τ1  M  τ2 τ2   τ 
y (t ) = AK u (t − td ) − 1 −  e τ1
+ 1 −  e  y (t ) = AK u (t − t d ) − e 
  K τ 1 1τ − τ 2  K τ 2  1τ − τ 2  
Impulse: x (t ) = Aδ (t − td )
 M   M  
 1 −  −(t −t ) 1 −  −(t −t ) 
Kτ 1  Kτ 2 
y (t ) = AK   +
d d
τ1 τ2 
e e
 τ1 − τ 2 τ 2 − τ1 
 
 
Sine:
When the coefficient of the second derivative is positive (right-hand-side

of Figure 9-1), it is customary to express the coefficients α and β in an
alternative way:
equation d y 2
dy dx dy
τ 2 2 + 2ξτ + y (t ) = M + Kx(t ) τ + y (t ) = Kx (t )
dt dt dt dt
dy y ( 0) = 0
y ( 0) = =0
dt t =0
transfer y ( s) Ms + K y ( s) K
function = 2 2 =
x( s ) τ s + 2ξτs + 1 x ( s ) τs + 1
poles 1
s = − ξ ± ξ 2 −1
τ
( ) s=−
1
τ
zeroes −K none
s=
M
3
Lecture Notes - 9
Relationships between the characteristic times τ1 and τ2 and the alternative
parameters τ and ξ are
τ1 =
τ
ξ − ξ −1 2
(
=τ ξ + ξ 2 −1 ) (9.1.4)
τ2 =
τ
ξ + ξ −1
2
(
= τ ξ − ξ 2 −1 ) (9.1.5)
τ = τ 1τ 2 (9.1.6)
τ1 + τ 2
ξ= (9.1.7)
2 τ 1τ 2
τ 1 + τ 1 = 2τξ (9.1.8)
τ 1 − τ 1 = 2τ ξ 2 − 1 (9.1.9)
As in the first-order system, τ has the dimension of time and represents a
characteristic time of the system. We will take it to be positive;
dimensionless ξ in the second-order system, however, we will allow to be
positive or negative. Parameter ξ is called the damping coefficient, and
the response depends markedly on its value.
9.2 Over damped systems: ξ > 1
The poles are real, negative, and unequal. We have seen such a system in
(8.6.3), a system of unequal tanks in series. The responses are
Step: x (t ) = Au (t − t d )
 
( )
 ξ + ξ − 1 ξ −1 (t −td )τ 
2 2
M
 1 − ξ − ξ −1 
2
e 
 −ξ (t −td )  Kτ  2 ξ 2 −1 
y (t ) = AK u (t − td ) − e τ
 
( )
 ξ − ξ − 1 − ξ −1 (t −td )τ
2
   
2
M
− 1 − Kτ ξ + ξ − 1 
2
 e 
    2 ξ 2 −1 
4
Lecture Notes - 9
e
−ξ (t −td )
τ
 
 1 −

M
Kτ
( 
ξ − ξ − 1 e
2

)
ξ 2 −1 (t −td )
τ



y (t ) = AK 
 
2τ ξ − 1   
( )
− ξ 2 −1 (t −td )
2
 − 1 −
M  

ξ + ξ − 1 e
2 τ

   Kτ  
Sine:
Figure 9.1 shows step responses for various ξ, with M = 0. In contrast to

the first-order response, the trace has a sigmoid shape: a slow start and an
inflection point. Larger ξ make the response more sluggish.
1.2
1
1.1
1.5
2
0.8
ξ=3
y(t)
0.6
0.4
0.2
0
0 2 4 6 8 10 12
time
Figure 9.1. Response for step input to an overdamped second-order

system with M = 0.
5
Lecture Notes - 9
9.3 Critically damped systems: ξ = 1
For this special condition, the poles are real, negative, and equal. This is
the case for (8.5.2), two identical tanks in series. The responses are
Step: x (t ) = Au (t − t d )
 − ( t − td )   M  (t − td ) 
y (t ) = AK u (t − td ) − e τ
1 + 1 −  
   Kτ  τ 
Sine:
9.4 Underdamped systems: ξ < 1
The poles are complex conjugates, and the system will oscillate even for a
non-periodic disturbance. The overdamped responses may be modified by
substituting
ξ 2 − 1 = − (1 − ξ 2 ) = j 1 − ξ 2 (9.4.1)
and using Euler’s relation (3.2.3) to find
second-order
Step: x (t ) = Au (t − t d )
  M 
 −ξ (t −td )  (t − t ) ξ− (t − t ) 
( ) Kτ sin 1 − ξ 2
− − cos 1 − ξ +
τ 2 d d
 u t t e 
y (t ) = AK  d
 τ 1−ξ 2 τ 
  
 
or
  2

1
2 
  1 − ξ 2 +  ξ − M    
 2 (t − td ) −1 1 − ξ 
−ξ (t −td ) 2
   Kτ  
u (t − td ) − e τ
sin  1 − ξ + tan
y (t ) = AK   1−ξ 2  τ M 
  ξ − 
    K τ 
 
 
 
6
Lecture Notes - 9
Sine:
9.5 Unstable systems: ξ < 0
The exponential terms in Table 9-x, when the damping factor is less than
zero, increase with time, and the system is unstable to disturbances.
9.6 General map for second order
Solution Characteristic Map
5
4 overshoot
3
2
reciprocal zero*tau
pseudo 1st order

1
sigmoid
0
-1
inverse inverse
-2
-3
-4 oscillatory oscillatory
excursive overdamped
-5 unstable stable
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
dam ping coefficient
Figure 9-2. Map relating character of step response to model parameters
9.7 Measures that characterize the underdamped step response

Defined from Figure 9-3. (Definitions from Coughanowr and Koppel, Process Systems Analysis
and Control, McGraw-Hill, 1965.) Equations are given for the particular case of the 2nd order
system.
7
Lecture Notes - 9
1.8
1.6 p
1.4
normalized response
A
1.2
C
1
0.8
0.6
B
0.4
0.2
0 rise time response time
0 2 4 6 8 10 12 14 16
normalized time
Figure 9-3. Oscillatory response to step input.
overshoot – response exceeds the ultimate value; equal to A/B
−πξ
1−ξ 2
overshoot = e
decay ratio – ratio of successive peaks, equal to C/A
−2πξ
1−ξ 2
decay ratio = overshoot = e 2
rise time – time to first reach the ultimate value
response time – time until response remains within ±5% of ultimate value
period – time between peaks, or between alternate crossings of the ultimate value.
 1 2π  2πτ
p = =  =
 f ω  1−ξ 2
natural period – period if there is no damping. A step disturbance will cause an undamped
system to oscillate perpetually about its ultimate value. Notice that the time constant τ is directly
proportional to the natural period. Notice also that damping lengthens the period.
9.8 On to stability?
With the second-order system we enlarge our appreciation of system
dynamics to include inverse response, oscillation, and unstable response.
In the next lesson, we will consider how systems become unstable.
8
Lecture Notes - 10
Lesson 10. Stability defined and related to the poles of the transfer function.
10.0 Context
In lesson 9 we found that a second-order process could be brought to
instability by reducing the damping coefficient. Here we study criteria for
stability in general. In a chemical process, instability may mean unsafe
conditions, bad quality product, and loss of money.
10.1 An example, using a badly conceived process

Consider a tank with pumped output in which the input flow is directly
proportional to the liquid level in the tank. Thus a rising level leads to
increased inlet flow, and falling level cuts back on the inlet flow: each
change makes matters worse.
Fin
The float operates a

lever that opens the
valve when the
level rises, and
closes it when it Fout
falls.
Analyzing by material balance:
dh
A = kh − Fout (t )
dt
(10.1.1)
dh 1
−τ + h = Fout (t ) h(0) = hs
dt k
Here, the outlet flow is an arbitrary function of time, and the initial
condition is some tank level. Dividing by the inlet flow proportionality
defines a characteristic time τ. Proceeding by integrating factor, we obtain
the solution
t
−1 tτ − tτ t
h(t ) = e ∫ e Fout (t )dt + hs e τ (10.1.2)
kτ 0
Suppose Fout is constant; then
1
Lecture Notes - 10
Fout 
1 − e τ  + hs e τ
t t
h(t ) =
k  
(10.1.3)
F  F  t
= out +  hs − out e τ
k  k 
If Fout and hs are in perfect balance, the level will be steady at hs.
However, the slightest change in Fout will lead to exponential variation in
the liquid level. This is an unstable system, headed inevitably for
overflow or draining.
10.2 Define stability

If we disturb a system, will it return to good operation, or will it get out of
hand? This is asking whether the system is stable. We define stability as
"bounded output for a bounded input". That means that
• a ramp disturbance is not fair – even stable systems can get into
trouble if the input keeps rising.
• an impulse disturbance is fair – although it is briefly infinite, it
soon passes.
• a stable system should also handle a step change in input,
ultimately coming to some new steady state. (We must be
realistic, however. If a system is so sensitive that a small input
step leads to an unacceptably high, though steady, output, we
might declare it unstable for practical purposes.)
• it should also handle a sine input; here the result is in general not
steady state, because the output may oscillate. (Thus we
distinguish between 'steady state' and 'long-term stability'.)
Stability depends on
• the system – certainly; we will discuss the characteristics of the
system that determine stability
• the type of disturbance – yes, as discussed above. In addition, it is
possible that a system is stable to some bounded inputs, but not
others. For example, the first-order integrator (pumped outlet
tank) is stable to sine disturbances but not to steps. No partial
credit – we declare such a system unstable.
• the magnitude of disturbance – maybe. That is, we continue to
represent real systems by LINEAR models. While the magnitude
of the disturbance does not influence stability of the linear system,
a sufficiently large disturbance to the real system may move it to a
regime of operation that the linear approximation does not
describe.
10.3 Develop a criterion for stability of linear systems

The key is the poles of the LINEAR characteristic equation. If any pole
has a zero or positive real part, the system will be unstable to bounded
2
Lecture Notes - 10
disturbances. This happens because of the structure of the linear,
constant-coefficient, ordinary differential equations that we are using – the
solutions are exponential terms. Here's the general form
dny d n−1 y
a n n + a n−1 n−1 + ⋅ ⋅ ⋅ + a 0 y = K 1 x1 (t ) + K 2 x 2 (t ) + ⋅ ⋅ ⋅ (10.3.1)
dt dt
To obtain the homogeneous solution, we substitute a candidate solution y

= ert and set the right-hand side to zero. We find that the parameter r is
constrained to be a root of the characteristic equation
a n r n + a n−1 r n−1 + ⋅ ⋅ ⋅ + a0 = 0 (10.3.2)
These n roots will in general be complex numbers.
ri = α i + jβ i (10.3.3)
The homogeneous solution of (10.3.1) will be a sum of terms, each

containing a factor eαit. Any term in which αi is positive will grow
without bound, and thus render the entire solution unstable. Such
instability results from the very structure of the system itself (i.e., the
values of the coefficients ai in (10.3.1)), and not from the particular
disturbances on the right-hand side.
Notice that (10.1.1) resembles our well-known first-order lag system, but
differs in a crucial respect:
dy
τ + y = Kx(t) first-order lag
dt
dy
−τ + y = Kx(t) unstable system from (10.1.1)
dt
In the first-order lag, the dependent variable opposes change in itself: a

higher value of y forces a smaller rate of change. The opposite effect is
seen in (10.1.1), and this mathematical structure leads to the positive root
of the characteristic equation.
Considering the homogeneous solution has led us to regard positive real

parts of roots as unstable. However, we must become even more
restrictive, as shown by considering the disturbances. Taking the Laplace
transform of (10.3.1),
K1 K2
y(s) = x1 (s) + x2 (s) + ⋅⋅⋅ (10.3.4)
an s + ⋅ ⋅⋅ + a0
n
an s + ⋅ ⋅⋅ + a0
n
3
Lecture Notes - 10
Equation (10.3.2), the characteristic equation of (10.3.1), is the same as

the denominator of the transfer functions in (10.3.4). Of course, they both
represent the left-hand side of the differential equation, and the poles of
the transfer function are the same as the roots of the characteristic
equation. They have the same significance for the solution (as they must)
in that partial fraction expansion of the transfer function in terms of the
poles leads to exponential terms in the solution when the transform is
inverted.
Consider now a transfer function that has a zero pole.
K1
y(s) = x1 (s) (10.3.5)
s(s + b)
The root -b is a typical stable root, and more denominator terms of the
form (s+bi) would not change the outcome. Now suppose that x1(t) is a
step of magnitude A. Then
K1 A  1 1 1 
y(s) = = K1 A 2 − 2 + 2  (10.3.6)
s (s + b)
2
 bs b s b (s + b) 
Upon taking inverse transforms,
t 1 1 
y(t) = K 1 A − 2 + 2 e −bt  (10.3.7)
b b b 
Thus the step disturbance, by repeating the existing zero pole, elicits a
term that grows steadily with time. The zero pole would give a bounded
output for a sine disturbance, but its behavior for a persistent disturbance
of one sign is grounds for declaring it categorically unstable.
A system is unstable if any root of its characteristic equation (i.e., pole of its transfer function)
has a real part of zero or greater.
By the way, the system of (10.3.5) is an example of an integrator process,

in which the dependent variable has no influence on its time derivatives.
Its model equation, reasoning backward from (10.3.5), is
d2y dy
2
+b = K 1 x1 (t)
dt dt
We saw a first-order integrator in the pumped outlet tank of Lesson 1.
4
Lecture Notes - 10
10.4 Stability of a recycle structure
Consider a non-oscillating third-order system in a recycle structure. For
example, this might be three mixed tanks in series under feedback control,
in which the controlled variable is compared to a set point and used to
drive a manipulated input variable, as we discussed in Lesson 1. The
block diagram for the system is
r(s) u(s) Kp y(s)
-
Kc
(τs + 1)
3
The value of signal y is subtracted from the set point r. The difference,
which constitutes error, is fed to the controller. The controller operates
upon the error by a proportional gain to produce a signal u. Input u acts
upon output y according to the transfer function shown in the block. The
controller gain Kc is variable; process gain Kp and characteristic time τ are
fixed.
We derive the transfer function for the system by block diagram rules.
Kp
y(s) = u(s)
(τs +1)3
K p Kc
= (r(s) − y(s) )
(τs +1)3
K p Kc
(10.4.1)
=
(τs +1)3 r(s)
K p Kc
1+
(τs +1)3
K p Kc
= r(s)
(τs +1)3 + K p K c
We emphasize that transfer function (10.4.1) does not describe the third-
order process, but rather that process under feedback control by a
proportional controller. The transfer function shows how the controlled
variable y responds to changes in set point r. The poles of (10.4.1) are
most easily found by numerical methods. In Figure 10-1 we have plotted
the poles as a function of the value of KpKc, assuming τ = 1. The real part
of the complex conjugate roots becomes zero at KpKc = 8.
5
Lecture Notes - 10
2
1.5
1
imaginary part
0.5
0
Kc = 0
-0.5
-1
-1.5
-2
-3.5 -3 -2.5 -2 -1.5 -1 -0.5 0 0.5
real part
Figure 10-1. Poles of transfer function (10.4.1). With gain of zero, the
repeated root is -1. As gain increases (KpKc = 3, 5, 7, 8, 9) the
complex conjugate poles approach the imaginary axis with increasing
imaginary parts.
We can calculate the step response as we have before by substituting a step for r(s) and inverting
the transform y(s) in (10.4.1).
10.5 Stability of systems with dead time

In Lesson 8 we introduced the dead time. Consider feedback control of
the FODT process.
r(s) u(s) Kp y(s)

Kc e − sθ
- τs +1
By the same means as in Section 10.4, we derive the transfer function for
the system.
K p K c e − sθ
y(s) = r(s) (10.5.1)
τs +1+ K p K c e − sθ
This departs from the polynomial equations we used as a basis for the
stability criterion! We might salvage that method by approximating the
exponential term by an infinite polynomial series. We might also develop
another method – that’s for another lesson.
6
Lecture Notes - 10
10.6 Stability and control
Many chemical processes, left alone, are stable. Of course we want to
maintain the operation at specific conditions, so we add control. As
indicated in Section 10.4, however, we will find that it is possible, via
poorly applied control, to produce instability in a process that, without
control, would be stable.
7
Lecture Notes - 11
Lesson 11. Frequency response of dynamic systems.
11.0 Context
We have worked with step, pulse, and sine disturbances. Of course, there
are many sine disturbances, because the applied frequency may vary.
Surely the system response should differ for low- and high-frequency
inputs. In this lesson we will explore the frequency response of dynamic
systems.
11.1 Response to sine input

Recall a first order lag, such as a stirred tank reactor, mixer, or heater.
The transfer function between inlet and outlet is
Co* (s) K
G(s) = * = (11.1.1)
Ci (s) τs +1
Let's disturb the inlet in a way that remains bounded, as with a step
disturbance, but prevents approach to a steady state:
Ci* (t) = Asin(ωt) (11.1.2)
If we take the Laplace transform of this inlet disturbance, the transfer

function gives us
Co* (s) = G(s)Ci* (s)

K Aω (11.1.3)
=
τs +1 s + ω 2
2
From a transform-pair table, we find
KAωτ −t τ K
Co* (t) = e + Asin(ωt + φ )
1+ ω τ2 2
1 + ω 2τ 2 (11.1.4)
φ = tan −1 (−ωτ )
The first term dies out – it represents the system adjusting to the onset of
the disturbance. The second term is an oscillation – the system's
continuing response to the prevailing input. It looks like the disturbance
(11.1.2), except that the inlet amplitude is modified and the response lags
the input by a phase angle φ (φ will be negative). Both amplitude and
phase angle depend on properties of the system, parameters in the transfer
function (11.1.1).
1
Lecture Notes - 11
11.2 Frequency response
We began by talking about a sine wave disturbance. We have derived
(11.1.4), which shows that the long-term response is to multiply the
amplitude by a factor and delay the signal. Hence, we define the long-
term response in terms of an amplitude ratio (AR) and a phase angle φ.
We call it a frequency response, because both AR and φ depend on the
disturbance frequency ω: the system will react differently according to
how it is shaken.
Co* (t ) K
AR = =
Ci* (t ) 1 + ω 2τ 2 (11.2.1)
φ = tan −1 (−ωτ )
Consider the mixing tank. Make the inlet concentration vary slowly, so
slowly that the tank has plenty of time to mix each incremental change in
concentration and deliver it to the outlet. Then the outlet will follow the
inlet concentration closely. For small ω, (11.2.1) shows that the amplitude
ratio is the same as the tank gain K, and the phase angle is zero. The trace
of the outlet signal lies over that of the inlet.
Now make the variations faster, so fast that the tank concentration scarcely
begins to rise before the inlet concentration falls again. Such rapid
fluctuations at the inlet will never propagate to the outlet, but be lost in the
'capacitance' of the tank. For sufficiently large ω, (11.2.1) shows that the
amplitude ratio goes to zero, and the phase angle to -90°.
If you are responsible for a process that is disturbed by fluctuations in the

input stream, you might specify a surge tank to damp out those
fluctuations. Understanding frequency response will assist you in sizing
the tank (thereby specifying its time constant τ) to do an effective job on
the disturbance frequencies of interest.
11.3 The Bode plot

Frequency response information is commonly presented on a Bode (bo'-
dee) plot.
2
Lecture Notes - 11
Bode plot - first order system
1.2
amplitude ratio/gain
1.0
0.8
0.6
0.4
0.2
0.0
0 20 40 60 80 100
ωτ (radians )
Because frequency and amplitude may meaningfully vary over several

orders of magnitude, Bode plots use logarithmic axes.
1.00
amplitude ratio/gain
0.10
0.01
0.01 0.1 1 10 100
ωτ (radians )
AR begins at 1.0 at low frequency and drops off at high frequency with a
slope of -1. The high frequency region can be extrapolated to AR=1 at the
'corner frequency' ωc.
1
ωc = (11.3.1)
τ
The phase angle varies asymptotically between 0 and -90°.
3
Lecture Notes - 11
-20
phase angle (deg)
-40
-60
-80
-100
0.01 0.1 1 10 100
ωτ (radians )
The Bode plot was once used for calculations in control system design.
Now computers have replaced that function, but it is still useful as an
illustrative tool.
11.4 Calculate frequency response for a system

In 11.1 we obtained the frequency response by calculating the solution to
the system equation when disturbed by a sine input. That gave us, in
addition, the short-term transient response.
We can get frequency response directly from the Laplace domain transfer
function if we're willing to replace the transform-pair table with a bit of
complex variable algebra. First an example, and then the rule…
The first order transfer function is
Co* ( s ) K
G ( s) = = (11.4.1)
Ci ( s ) τs +1
*
Substitute jω for s in (11.4.1).
K
G ( jω ) = (11.4.2)
jτω +1
Multiply by the complex conjugate to remove j from the denominator.
4
Lecture Notes - 11
K 1− jτω
G ( jω ) =
1+ jτω 1− jτω
(11.4.3)
1− jτω
=K
1+ τ 2ω 2
This is a complex number whose real and imaginary parts are
K
Re[G( jω )] =
1+ τ 2ω 2
(11.4.4)
− Kτω
Im[G( jω )] =
1+ τ 2ω 2
The complex number can be written in the alternate polar form. The
amplitude is
2 2
 K   − Kτω 
G( jω ) =  2 2 
+ 2 2 
1+τ ω  1+τ ω 
K 1 + τ 2ω 2
= (11.4.5)
1 + τ 2ω 2
K
=
1 + τ 2ω 2
and the angle is
 − Kτω 
 
φ = ∠G( jω ) = tan −1  1+ τ ω
2 2

 K  (11.4.6)
 
 1+ τ 2ω 2 
= tan −1 (−τω )
You probably saw this coming. When jω is substituted for s in the transfer
function Eqn (1), the amplitude and angle of the resulting complex number
are the amplitude ratio and phase angle of the frequency response of the
transfer function. We demonstrated for first order, but it's true in general.
To summarize,
5
Lecture Notes - 11
y(s)
for x(t) = Asin ωt and G(s) =
x(s)
y f .r. (t) = A G( jω ) sin (ωt + φ )
y f .r. (t) (11.4.7)
= G( jω ) = Re[G ( jω )] + Im[G ( jω )]
2 2
AR =
A
 Im[G( jω )] 
φ = ∠G( jω ) = tan −1  
 Re[G( jω )] 
Notice that the mean value of both x and y is zero. This means that we
have implicitly assumed deviation variables in defining the frequency
response.
11.5 Frequency response by combining components

We can further simplify our task of determining frequency response if the
transfer function is assembled as a product of known components.
G(s) = G1 (s)G2 (s)... (11.5.1)
After substituting jω for s, the complex numbers can be expressed in polar

form
G( jω ) = G1 ( jω )G2 ( jω )...
= G1 ( jω ) e jφ1 G1 ( jω ) e jφ2 ... (11.5.2)
ARe jΦ = G1 ( jω ) G1 ( jω ) ...exp( j ∑ φi )
so that
AR = ∏ Gi ( jω ) Φ = ∑ φi (11.5.3)
i i
For example, the important case of first-order-plus-dead-time (FOPDT):
Ke −θs 1
G(s) = = K ⋅ e −θs ⋅
τs +1 τs +1
1
AR = G ( jω ) = K ⋅1 ⋅
τ 2ω 2 + 1 (11.5.4)
K
=
τ 2ω 2 + 1
φ = ∠G( jω ) = 0 + (−θω ) + tan −1 (−τω )
6
Lecture Notes - 11
The amplitude ratio is unchanged from that of the first order system.
However the phase delay is greatly increased at high frequencies. The
plot is for the case in which θ = τ.
Bode plot - first order plus dead time system
0
phase angle (deg)
-100
FOPDT
first order
-200
-300
0.01 0.1 1 10
ωτ (radians)
Graphical combination of transfer functions is convenient on Bode plots.

However, this technique is not so commonly used when computers are
available.
7
Lecture Notes -12
Lesson 12. Dynamics of a stirred tank reactor.
12.0 Context
Our mission has been to examine how systems change with time. We
have identified typical features of dynamic systems (such as input and
output variables), used the conservation equations to derive mathematical
models of the systems, developed methods to solve the equations, studied
the responses of simple systems and their combinations, and explored the
boundaries of stable response. Let’s do a grand example.
12.1 The problem

A second-order decomposition reaction occurs in an overflow stirred tank
reactor. The reactor is equipped with a heat transfer surface that contains
a condensing fluid, so that it remains at uniform temperature. The outlet
composition and temperature may vary as a result of disturbances in inlet
temperature, inlet composition, and coolant temperature.
Material and energy balances, plus supplementary equations, are the basis
for a model of the reactor system. The equations are nonlinear, so that
numerical solution is necessitated to predict how the reactor will behave
from arbitrary initial conditions. However, because we desire to operate
the reactor at steady state, it may be sufficient to examine the dynamics in
the vicinity of a steady state; hence linear approximations are useful.
12.2 The method

• Use Taylor series approximations for nonlinear terms in the equations
• write the model at steady state
• solve for any required operating parameters
• subtract from the model equations to leave deviation variables
• express the equations in standard form, identifying dynamic parameters of known
significance
• take Laplace transforms
• combine equations to isolate independent variables
• again, express these equations in standard form if the order of the system has changed.
• apply disturbances as Laplace transforms
• invert to obtain time response of the dependent variables
12.3 Material balance

Mole balance on the reactant A
dC A −E
V = FC Ai − FC A − Vk o e RT C A2 (12.3.1)
dt
Approximate the nonlinear term about a reference condition Ts and CAs
1
Lecture Notes -12
−E −E E − E RTs 2 −E
e RT
C A2 ≈ e RTs 2
C As + 2
e C As (T − Ts ) + 2e RTs C As (C A − C As )
RTs
(12.3.2)
Substitute into mole balance (12.3.1).
dC A E
V = FC Ai − FC A − Vk s C As
2
−V k C 2 (T − Ts ) − 2Vk s C As (C A − C As )
2 s As
dt RTs
(12.3.3)
where ks is the reference condition rate constant
−E
ks = koe RTs
(12.3.4)
Writing the mole balance at the reference condition
0 = FC Ais − FC As − Vk s C As
2
(12.3.5)
Solve (12.3.5) for the reference outlet concentration
− 1 + 1 + 4τ r k s C Ais
C As = (12.3.6)
2τ r k s
where τr is the residence time of the tank, important to reactor conversion.
V
τr = (12.3.7)
F
By L'Hospital's Rule, as ks goes to zero (no reaction), CAs approaches CAis.

Also, as CAis goes to zero, so does the outlet concentration.
Subtract the steady state (12.3.6) from the mole balance (12.3.3), and
express as deviation variables.
dC A* E
V + FC A* + 2Vk s C As C A* = FC Ai
*
−V 2
k s C As T* (12.3.8)
dt RTs2
This is a first-order lag equation, which is more apparent if it is placed into

standard form
2
Lecture Notes -12
dCa* τ *
τC + C A* = c C Ai − τ C K CT T * C A* (0) = 0
dt τr
V
τC = (12.3.9)
F + 2Vk s C As
2
E
K CT = k C2
2 s As
RTs
The time constant τC characterizes the dynamics of concentration change.

The group KCT multiplied by τC is the gain for the effect of temperature on
composition in the reactor.
12.4 Energy balance

The energy balance must account for the reaction and heat transfer.
dT −E
VρC p = FρC p (Ti − Tref ) − FρC p (T − Tref ) − ∆H rVk o e RT C A2 − UA(T − Tc )
dt
(12.4.1)
Enthalpies are defined with respect to a reference temperature. Heat

exchange occurs with a condensing fluid at Tc. The nonlinear term is
approximated as before. At steady state, the energy balance becomes
∆H r UA
0 = Tis − Ts − τ r k s C As
2
− (Ts − Tc ) (12.4.2)
ρC p FρC p
which may be solved for the reference outlet temperature.
∆H r UA
Tis − τ r k s C As
2
+ Tc
ρC p FρC p
Ts = (12.4.3)
UA
1+
FρC p
Because the reaction rate constant ks depends on Ts, (12.4.3) must be

solved iteratively.
Expressing the energy balance in deviation variables leads to
3
Lecture Notes -12
dT * τ
τT + T * = T Ti * − τ T K TC C A* T * (0) = 0
dt τr
τr
τT = (12.4.4)
∆H r E UA
1+ τr k C2 +
ρC p RTs 2 s As
FρC p
∆H r
K TC = 2 k s C As
ρC p
Once again, standard-form parameters have been defined. The thermal

time constant τT characterizes the rate of temperature change, and KTC
combines with τT to form the gain for concentration disturbances.
12.5 Laplace transforms and transfer functions

Now Laplace transforms may be performed on the mole balance (12.3.9)
and the energy balance (12.4.4).
τc *
(τ C s +1)C A* (s) = C Ai (s) − τ C K CT T * (s) (12.5.1)
τr
τ
(τ T s +1)T * (s) = T Ti * (s) − τ T K TC C A* (s) (12.5.2)
τr
These equations must be combined to isolate the dependent variables.

First, express them in a block diagram.
τC
CAi*(s) τr
τ C s +1
τT τ C K CT +
Ti*(s) τr + T*(s) - CA*(s)
τT s +1 - τC s +1
τ T K TC
τT s +1
This diagram illustrates how the disturbances in inlet temperature and

concentration affect the outlet concentration. The interdependence of
concentration and temperature is shown as a recycle loop. By reshaping
the diagram, we could form an analogous diagram that produces the outlet
temperature. Being based on the same equations, it is of course equivalent
to the one above.
From the block diagram, we derive the expression for CA*(s).
4
Lecture Notes -12
τC
τr τ C K CT *
C A* (s) = *
C Ai (s) − T (s)
τ C s +1 τ C s +1
τC  τ T 
τr τ K  τ τ K 
= *
C Ai (s) − C CT  r
Ti* (s) − T TC C A* (s)
τ C s +1 τ C s +1  τ T s +1 τ T s +1 
  (12.5.3)
τC
(τ s +1)
τr T
C A (s) =
* *
C Ai (s)
(τ C s +1)(τ T s +1) − τ Cτ T K CT K TC
τ τ
− C T K CT
τr
+ Ti * (s)
(τ C s +1)(τ T s +1) − τ Cτ T K CT KTC
We recognize this from the denominator as a second order system. It will

be helpful to place it into standard form.
τ Cτ 2 τ 2 K CT
(τ s +1) −
τ rτ T T τr
C A (s) = 2 2
*
C Ai (s) + 2 2
*
Ti * (s)
(τ s + 2ξτs +1) (τ s + 2ξτs +1)
τ Cτ T
τ= (12.5.4)
1 − τ Cτ T K CT K TC
τC +τT
ξ=
2 τ Cτ T (1 − τ Cτ T K CT K TC )
Substituting (12.5.4) into either of the Laplace transforms gives, after

some effort,
τ Tτ 2 τ 2 K TC
(τ C s +1) −
ττ τr
T * (s) = 2r 2C Ti * (s) + 2 2 *
C Ai (s) (12.5.5)
(τ s + 2ξτs +1) (τ s + 2ξτs +1)
12.6 Predicting transient response

Equations (12.5.4) and (12.5.5) are process models for outlet temperature
and concentration, as disturbed by inlet temperature and concentration.
For step changes
5
Lecture Notes -12
∆Ce − tC s
C (t) = ∆Cu(t − tC )
*
Ai C (s) =
*
Ai
s
− tT s
(12.6.1)
∆Te
Ti (t) = ∆Tu(t − tT )
*
Ti* (s) =
s
the outlet concentration transform becomes
τ 2 ∆Ce −t s C
τ 2 ∆Ce −t s C
C (s) =
*
+
τ r (τ 2 s 2 + 2ξτs +1) τ rτ T (s)(τ 2 s 2 + 2ξτs +1)
A
(12.6.2)
τ 2 K CT ∆Te −t s T
−
τ r (s)(τ 2 s 2 + 2ξτs +1)
The inverse transform is
  1−ξ 2  
 ∆C e
−ξ ( t −tC )
τ
sin  (t − tC )  
τ 1 − ξ 2  τ  
   
 
τ 2  ∆C  1 −ξ ( t −t C )  1−ξ 2 1 − ξ 2  
C A* (t) = + u (t − tC ) − e τ
sin  (t − tC ) + tan −1  
τ r  τT  1 − ξ 2  τ ξ  
  
 
  1 −ξ (t −tT )  1−ξ 2 1−ξ 2  
− K CT ∆T u(t − tT ) − e τ
sin  (t − tT ) + tan −1  
 1−ξ 2  τ ξ  
   
(12.6.3)
Proceeding similarly, the temperature is
  1−ξ 2  
 ∆T e
−ξ ( t −tT )
τ
sin  (t − tT )  
τ 1 − ξ 2  τ  
   
 
τ 2  ∆T  1 −ξ ( t −tT )  1−ξ 2 1 − ξ 2  
T * (t) = + u (t − tT ) − e τ
sin  (t − tT ) + tan −1  
τr  τC  1 − ξ 2  τ ξ  
  
 
   1−ξ 2  
−1 1 − ξ
−ξ (t −t C ) 2
1   

− KTC ∆C u(t − tC ) − e τ
sin (t − tC ) + tan
  1−ξ 2  τ ξ  
  
(12.6.4)
The solutions show a striking symmetry; this is a result of linearizing the

original model (so that material and energy balances had identical
6
Lecture Notes -12
structure) and expressing the approximate equations in standard form (so
that analogous dynamic parameters would be defined).
Notice as well that the solutions are written with the presumption that the
damping factor ξ is less than one. If for any particular operating
conditions this is not true, the complex terms would be modified through
Euler's relations to give corresponding overdamped forms of the solutions.
Presuming ξ is positive, the oscillatory terms decay to a long-term

response to the step disturbances ∆Ti and ∆CAi. These are
τ2  ∆C 
C A* (∞) =  − K CT ∆T  (12.6.5)
τr  τT 
τ2  ∆T 
T * (t) =  − KTC ∆C  (12.6.6)
τr  τC 
12.7 Parameter dependence

The virtue of a model, even a simplified one, is that it can offer insight
about the influence of the various parameters. Our second order system
depends on the standard time constant τ and damping factor ξ, as well as
parameters from the individual first order systems, such as τC and τT. Let's
look at these in more detail by defining three parameters that characterize
the reaction, physical properties, and heat transfer.
KC represents the extent of reaction at the initial steady state. It is non-

negative.
K C = 2k s C Asτ r (12.7.1)
KH represents the ratio of capabilities for heat transfer and thermal storage.
It is also non-negative.
UA
KH = (12.7.2)
FρC p
KR represents the energetics of the reaction. Its sign can vary according to
the heat of reaction.
∆H r E C As
KR = (12.7.3)
ρC p RTs2 2
The dynamic parameters are now expressed in these terms
7
Lecture Notes -12
τr
τ=
K R K C + ( K H + 1)( K C + 1)
(12.7.4)
K K + ( K H + 1) + ( K C + 1)
ξ= R C
2 K R K C + ( K H + 1)( K C + 1)
12.8 Limiting cases
Suppose there is no heat transfer. Then KH = 0, and
τr
τ nht =
K R K C + ( K C + 1)
(12.8.1)
K R KC + KC + 2
ξ nht =
2 K R K C + ( K C + 1)
The damping coefficient has a minimum value of 1 when either KC = 0 (no

reaction) or KR = -1. Thus the dynamic response is never underdamped
without heat transfer.
Suppose there is no reaction. Then KC = 0, and
τr
τ nr =
( K H + 1)
(12.8.2)
KH + 2
ξ nr =
2 ( K H + 1)
The damping coefficient has a minimum value of 1 when KH = 0 (no heat

transfer). Thus the dynamic response is never underdamped without
reaction.
Suppose there is reaction, but that the heat of reaction is negligible. Then
KR = 0, and
τr
τ nhr =
( K H + 1)( K C + 1)
(12.8.3)
K H + KC + 2
ξ nhr =
2 ( K H + 1)( K C + 1)
The damping coefficient has a minimum value of 1 when KH = KC. Thus

the dynamic response is never underdamped without heat of reaction.
8
Lecture Notes -12
12.9 Stability
The more interesting cases result from the interaction of energetic reaction
and heat transfer.
9
C H A P T E R 9
Feedforward Control
Objectives of the Chapter
• Describe how to use feedforward control to compensate for measured disturbances.

• Show how to use the material in chapters 3, 4, 6, and 7 to design and tune combined
feedforward/feedback control systems.
Prerequisite Reading
Chapter 3, “One-Degree of Freedom Internal Model Control”

Chapter 4, “Two-Degree of Freedom Internal Model Control”
Chapter 6, “PI and PID Controller Parameters from IMC Designs”
Chapter 7, “Tuning and Synthesis of 1DF IMC Controllers for Uncertain Processes”
221
222 Feedforward Control Chapter 9
9.1 INTRODUCTION
Combined feedforward plus feedback control can significantly improve performance over
simple feedback control whenever there is a major disturbance that can be measured before
it affects the process output. In the most ideal situation, feedforward control can entirely
eliminate the effect of the measured disturbance on the process output. Even when there are
modeling errors, feedforward control can often reduce the effect of the measured
disturbance on the output better than that achievable by feedback control alone. However,
the decision as to whether or not to use feedforward control depends on whether the degree
of improvement in the response to the measured disturbance justifies the added costs of
implementation and maintenance. The economic benefits of feedforward control can come
from lower operating costs and/or increased salability of the product due to its more
consistent quality.
Feedforward control is always used along with feedback control because a feedback
control system is required to track setpoint changes and to suppress unmeasured
disturbances that are always present in any real process.
Figure 9.1a gives the traditional block diagram of a feedforward control system
(Seborg et al., 1989). Figure 9.1b shows the same block diagram, but redrawn so as to show
clearly that the feedforward part of the control system does not affect the stability of the
feedback system and that each system can be designed independently. Figure 9.2 shows a
typical application of feedforward control. The continuously stirred tank reactor is under
feedback temperature control. Feedforward control is used to rapidly suppress feed flow
rate disturbances.
Feedforward
Controller Measured
Feedforward
qff Disturbance
Control d
Effort
uff Process pd
Feedback
Controller Total de
Setpoint ufb− y
PID
Control p Output
r − Effort
u
Feedback Control
Effort
Figure 9.1a Traditional feedforward/feedback control structure.

9.1 Introduction 223
Feedforward
Controller Measured
qff Disturbance
d
Control Effort uff

Feedforward
Process pd
de
−
p
Feedback dc
Controller
Feedback
Setpoint Control y
r PID Effort ufb
p Output
−
Figure 9.1b Block diagram equivalent to that of Figure 9.1a.
Feedforward
FFC
Controller
Feed
Temp. Setpoint Flow
FT
Transmitter
Temperature
TC TT TC Controller
Temperature
+ TT
Transmitter
Σ _ FFC FT
Cooling
Water
Product
Figure 9.2 Feedforward control on the feed to a continuously stirred tank reactor operating
under reactor temperature control.
9.2 CONTROLLER DESIGN WHEN PERFECT COMPENSATION IS

POSSIBLE
The transfer function between the process output y and the measured disturbance d
from Figure 9.1b is
dc ( p d − pq ff )d
y( s) = = . (9.1)
1 + PID ∗ p 1 + PID ∗ p
To eliminate the effect of the measured disturbance, we need only choose qff so that
pd − pq ff = 0. (9.2)
If the deadtime and relative order of pd are both greater than those of p, and p has no
right half plane zeros, then qff can be chosen as
q ff = ~
p −1 ~
pd . (9.3)
As in previous chapters, a ~ over a process transfer function indicates that it is a model of

the process.
Even in the above case, where the feedforward controller can perfectly compensate
the measured disturbance, it may not pay to implement a feedforward controller when the
process p is well approximated as a minimum phase system (i.e., one that contains no
deadtime or right half plane zeros). In this case, it is usually possible to design a 1DF or
2DF PID controller to suppress disturbances well enough so that there is little to be gained
by the addition of feedforward control (Morari and Zafiriou, 1989).
Whenever the relative order of ~ pd ( s ) is less than or equal to that of ~
p ( s ) , then the
noise amplification can be reduced by adding a filter, as in the design of an IMC controller,
so that Eq. (9.3) becomes:
q ff = ~
p −1 ~
pd f ; f ≡ 1 /(εs + 1) r (9.4)
The order r of the filter f is either the relative order of ~ pd−1 ~

p ( s ) or 0 if the relative
order of ~
p ~
−1
d p ( s ) is equal or less than zero. The filter time constant ε is chosen to limit noise
amplification, just as was done for IMC controllers for perfect models (see Chapter 3).
The PID controller in Figure 9.1 can be designed and tuned as either a 1DF or 2DF
controller using the techniques of chapters 6, 7, and 8 without regard to the addition of a
feedforward controller. Most often, however, the unmeasured disturbances are not severe
enough to justify the added complication of a 2DF PID controller.
9.3 Controller Design when Perfect Compensation is Not Possible 225
9.3 CONTROLLER DESIGN WHEN PERFECT

COMPENSATION IS NOT POSSIBLE
When the deadtime in the disturbance lag pd is less than that of the control effort lag p, then
the feedforward controller qff, given by Eq. (9.4), is not realizable. In this case, perfect
compensation is no longer possible. Much of the literature suggests designing the
feedforward controller by simply dropping the unrealizable part of the controller, as is done
for a single-degree of freedom IMC controller. However, as we shall soon see by example,
the foregoing is by no means the best design. A better feedforward controller can be
obtained using a 2DF design. To see why this is, we rewrite the expression for the net effect
of the measured disturbance on the output from Eq. (9.1) as
d c = (1 − pq ff pd−1 ) p d d = (1 − pq ff pd−1 )d e . (9.5)
Since, by assumption, the term ppd−1 contains a deadtime, there is no realizable choice
of the feedforward controller qff that makes the term in brackets in Eq. (9.5) zero. Therefore,
there will be a long tail to the response of dc to a step disturbance d if the lag of pd is on the
order of, or larger than, the lag of p, unless qff is chosen so that the zeros of
(1 − ~ pd−1 ) cancel the poles of ~
pq ff ~ p d . The procedure for such a choice of qff is the same as
that described in Chapter 4, except that here, the model is ~ p −1 rather than just ~
p~ d p, as in the
design of 2DF feedback controllers. Indeed, the 2DF controller design capability of
IMCTUNE can be used to determine qff. To illustrate, we return to Example 8.1, except now
the disturbance is measured and there is a deadtime of 0.5 time units before the disturbance
enters the disturbance lag.
Example 9.1 Comparison of 1DF and 2DF Feedforward

Controller Designs
The purpose of this example is to compare the performance of 1DF and 2DF feedforward
controllers when the process models are perfect. The process is
~
p ( s ) = p( s ) = e − s /( s + 1); ~
p d ( s ) = pd ( s ) = e −0.5 s /( 4 s + 1). (9.6)
For the above system, the 1DF and 2DF feedforward controllers from IMCTUNE,
and from equations (9.3) and (9.5) are
q ff ( s ) = ( s + 1) /( 4s + 1), (9.7a)
q ff ( s ) = ( s + 1)(.473s + 1) /( 4s + 1)(0.006 s + 1). (9.7b)
Notice that the 1DF controller, Eq. (9.7a), does not need a filter, as it is already a lag
that does not amplify noise. The 2DF controller, Eq. (9.7b), satisfies the requirement that
(1 − ~ pd−1 ) cancel the poles of ~

pq ff ~ pd . Only a small filter time constant of 0.006 is needed to
reduce the controller noise amplification to 20 because of the relatively large disturbance
lag. To obtain Eq. (9.7b) from IMCTUNE, we entered the process model as
~
p~pd−1 = (4 s + 1)e −.5 s /( s + 1) and the model “disturbance” lag as (4s + 1) . The part of the
model to invert is taken as (4s + 1) /( s + 1) , and the order of the controller filter is set to
zero. Entering a filter time constant of 0.006 yields Eq. (9.7b).
Figures 9.3 and 9.4 compare the responses of the process output with and without
feedback, using the IDF and 2DF controllers given by equations (9.7a) and (9.7b). The
output y, when the feedback loop is open, is the same as the compensated disturbance signal
dc, shown in Figure 9.1b and defined by Eq. (9.5). Notice that the addition of the feedback
system improves the response of the output using a 1DF feedforward controller, but
degrades the output response using a 2DF feedforward controller. The reason for the
degradation in response is that the PID controller attempts to suppress the pulse dc, created
by the feedforward control system. Of course it cannot, and the attempt creates another
pulse in the opposite direction that is diminished in amplitude but broader in time. The same
effect occurs with the 1DF feedforward controller. However, the response of this controller
is long enough with respect to the settling time of the feedback loop that the feedback loop
is capable of diminishing the effect of the disturbance on the output.
Compensated disturbance response, dc(t) (Eq. (9.5))
.12
.10
( s + 1)
0.08 q ff =
(4s + 1)
0.06
( s + 1)(.475s + 1)
q ff =
0.04 (4 s + 1)(.006s + 1)
0.02
−0.02
0 2 4 6 8 10 12 14 16 18 20
Time
Figure 9.3 Comparison of compensated disturbance responses for Example 9.1.

9.3 Controller Design when Perfect Compensation is Not Possible 227
0.12
0.10
( s + 1)
0.08 q ff =
(4s + 1)
0.06
0.04 ( s + 1)(.475s + 1)
Output
q ff =
(4s + 1)(.006s + 1)
0.02
0
−0.02
−0.04
−0.06
0 2 4 6 8 10 12 14 16 18 20
Time
Figure 9.4 Comparison of feedforward plus feedback control system responses to a unit step
disturbance for Example 9.1.
The PID controller used to obtain the responses shown in Figure 9.4 was obtained as
described in Chapter 6. For a perfect model, an IMC filter time constant of 0.3 yields the
following PID controller, which overshoots a step setpoint change by about 10%.
PID = 0.610(1 + 1 /(1.24 s ) + .179 s /(.090 s + 1)). (9.7c)
♦
It is possible to prevent the degradation in output response by making the feedback
loop blind to the pulse generated by the feedforward controller. This is accomplished by
subtracting the pulse from the feedback system, as shown in Figure 9.5.
Based on the responses given in Figures 9.3 and 9.4, the pulse-compensated diagram
should be used only when the feedforward control system settling time, with the feedback
loop open, is significantly faster than the feedback control system settling time. Thus, for
Example 9.1 the pulse compensation of Figure 9.5 improves the response of the 2DF control
system, which has a fast response, but degrades the response of the 1DF feedforward
control system, which has a slow response relative to that of the feedback system.
~
pd
Feedforward Measured
Model Disturbance
Controller
− ~
p qff
Process pd
Controller
Setpoint − −
PID p y
r −
Output
Figure 9.5 Pulse-compensated feedforward control system.
9.4 CONTROLLER DESIGN FOR UNCERTAIN PROCESSES
9.4.1 Gain Variations
Gain variations in either or both p(s) and pd(s) can result in a nonzero value for the steady
state effect of the compensated disturbance dc (see Figure 9.1b), on the process output. The
gain of the feedforward controller Kf should be chosen to minimize either the maximum
magnitude of the steady-state compensated disturbance dc(∞) or the ratio of the
compensated to the uncompensated disturbance dc(∞)/de(∞). Mathematically, the problem
can be expressed as
d c ( ∞ )opt . = min max K d − K p K f (9.8)

K f K p ,Kd
or
( d c ( ∞ ) / d e ( ∞ ))opt . = min max 1 − K p K f / K d . (9.9)
K f K p ,Kd
The maxima in equations (9.8) and (9.9) occur at the simultaneous upper bound of Kp and
lower bound of Kd, and at the lower bound of Kp and the upper bound of Kd. The values of
Kf that minimize these maxima are those values that equalize the values of the two extremes.
That is, for Eq. (9.8)
9.4 Controller Design for Uncertain Processes 229
 max( K d ) + min( K d )   max( K p ) + min( K p ) 

( K f ) opt =    
 2   2 , (9.10)
 
≡ ( K d ) ave /( K p ) ave
and for Eq. (9.9)
−1
 max( K p / K d ) + min( K p / K d ) 
( K f ) opt =   .

 2  (9.11)
≡ ( K p / K d ) −ave
1
The feedforward controller gain given by Eq. (9.11) assures that the magnitude of the
compensated disturbance effect dc(∞) is always less than the magnitude of the
uncompensated disturbance effect de(∞). That is, the action of the feedforward controller
always improves the output response over that which would have been achieved without
feedforward control. However, this feedforward controller gain also yields the same relative
improvement if the ratio of ( K d / K p ) is at its maximum or its minimum, and this weighting
tends to amplify the effect of large values of ( K d / K p ) on the output. On the other hand, the
feedforward controller gain given by Eq. (9.10) can cause the compensated disturbance to
be worse than the uncompensated disturbance in some situations. As we shall see from
Example 9.2, the choice between equations (9.10) and (9.11) depends to a large degree on
engineering judgment, and to some degree on the amount of uncertainty in the process and
disturbance lag gains.
Example 9.2 Variation in Process and Disturbance Lag Gains
Case 1, Modest Uncertainty:

0.8 ≤ K p , K d ≤ 1.2 ; d = 1
From Eq. (9.10), we have ( K f ) opt = 1, and from Eq. (9.8), the worst-case compensated
effect of the disturbance is
d c (∞) max . = max K d − K p ( K f ) opt d (∞) = K d − K p = Kd − K p = 0.4.

K p ,Kd K p =.8, K d =1.2 K p =1.2, K d =.8
The worst case uncompensated effect of the disturbance is de(∞)max = (Kd)maxd(∞) = 1.2.
3 / 2 + 2 / 3 −1
From Eq. (9.11) we have ( K f ) opt = ( ) = 12 / 13, and from Eq. (9.9.)
2
(1.2)(12 / 13) (.8)(12 / 13)
d c (∞) / d e (∞) max = max 1 − K p K f / K d = 1 − = 1− =.385.
K p ,Kd .8 1.2
Since de(∞)max = (Kd)maxd(∞) = 1.2, then

dc(∞)max = (.385)( de(∞)max) = .46.
For most engineering situations the feedforward controller gain from Eq. (9.10)
(i.e., K f = 1 ) would be preferred because this minimizes the maximum effect of the
compensated disturbance d c (∞) max on the output (0.4 versus .46).
Case 2, Large Uncertainty:
1 ≤ K p , Kd ≤ 3 ; d = 1
From Eq. (9.10) we have: ( K f ) opt = 1 and from Eq. (9.8), if Kd = 1, and Kp = 3, then the
compensated effect of the disturbance is
d c (∞) = 1 – 3 = –2.
However, the uncompensated effect of the disturbance is only 1.0 since

de(∞) = Kdd(∞) = 1.0
3 + 1 / 3 −1
From Eq. (9.11) we have ( K f ) opt = ( ) = 3 / 5, and from Eq. (9.9), if Kd = 1, and
2
Kp = 3, then the compensated effect of the disturbance is
d c (∞) / d e (∞) = 1 − K p ( K f ) opt / K d = (1– (3)(3/5)/1) = – 0.8
Thus for the situation where Kd = 1, and Kp = 3, the feedforward gain given by
Eq. (9.11) is clearly preferred. However, if the gains of Kd and Kp are reversed so that
Kd = 1, and Kp = 3, then the uncompensated effect of the disturbance is 3.0, and the
compensated effect of the disturbance for the feedforward gains given by Eq. (9.10) and
Eq. (9.11) are 2 and 2.4 respectively. So, in the first instance the gain given by Eq. (9.11) is
preferred, while in the second instance, the gain given by Eq. (9.10) is preferred. “You pays
your money and takes your choice.” The authors prefer the gain given by Eq. (9.11).
9.4.2 Deadtime Variations
This section considers the effect of deadtime variations on the choice of feedforward
controller deadtime. We assume that the process deadtime is on the order of, or greater than,
the disturbance deadtime. Further, the relative order of the process p(s) is the same as that of
the disturbance transfer function pd (s), and all time constants are known exactly. The
process description is therefore given by
−T p s
y ( s ) = Kg ( s )e u ( s ) + K d g d ( s )e −T s d ( s ),
d
(9.12)
where
g (0) = g d (0) = 1 , T p ≤ T p ≤ T p , T d ≤ Td ≤ Td , K ≤ K ≤ K , K d ≤ K d ≤ K d
The feedforward controller is given by
q f ( s ) = K f g d ( s )e − ∆s / g ( s ), (9.13)
where ∆ is a parameter that we need to find.

Substituting Eq. (9.13) into Eq. (9.5) gives
− (T p + ∆ ) s
d c ( s ) = ( K d e −Td s − KK f e ) g d ( s )d ( s ). (9.14)
When (Td > T p + ∆) , the maximum magnitude of the integral of dc(t) occurs when
(Td − (T p + ∆)) is a maximum. When (Td < (T p + ∆)) , the maximum magnitude of the
integral of dc(t) occurs when ((T p + ∆) − Td ) is a maximum. Choosing ∆ so that these two
maxima are equal gives
(Td + T d ) − (T p + T p )
∆= ≡ (Td ) ave − (T p ) ave . (9.15)
2
In Example 9.3 we present the results of simulations on two processes, each of which
are similar to Example 9.1. In the first example, the deadtime between the measured
disturbance and the output is on the order of the process deadtime, and hence perfect
feedforward compensation is possible for most choices of models in the uncertain set. For
this process we use a simple 1DF feedforward controller. In the second example, the
deadtime between the measured disturbance and the output is smaller than the process
deadtime, and hence perfect feedforward compensation is not possible. Therefore, for this
process we use a 2DF feedforward controller.
Example 9.3 Disturbance and Process Deadtimes of

Roughly the Same Size
−T p s
p( s ) = Ke /( s + 1); 0.8 ≤ K , T p ≤ 1.2 (9.16a)
pd ( s ) = K d e −Td s /( 4 s + 1); 0.8 ≤ K d , Td ≤ 1.2 (9.16b)
q f ( s ) = K f (4s + 1) /( s + 1) (9.16c)
The feedforward controller given by Eq. (9.16c) has a deadtime of zero, since the difference
between (Td)ave and (Tp)ave is zero (see Eq. 9.15).
Figures 9.6a and 9.6b show the worst-case feedforward-only responses for the above
process, using feedforward controller gains given by equations (9.10) and (9.11). Notice
that for a disturbance lag gain of 1.2 and a process gain of .8 (i.e., Kd = 1.2, K = .8), the
controller gain given by Eq. (9.10) gives the better response. However, when the gains are
reversed, (i.e., Kd = .8, K = 1.2), the controller gain given by Eq. (9.11) gives the better
response.
Compensated (dc) and Uncompensated (de) 1.2
1.0
Feedforward only Responses
de
0.8
0.6 dc, Kf = 0.923 from Eq. (9.11)
0.4
0.2 dc, Kf = 1 from Eq. (9.10)
−0.2
0 5 10 15 20 25 30
Time
Figure 9.6a Example 9.3 responses of the feedforward-only control system to a unit step
disturbance for Kd = 1.2, Kp = .8, Td = 1.2, Tp = .8.
1.0
Compensated (dc) and Uncompensated (de)
0.8
Feedforward Only Responses
de
0.6
0.4
0.2
0 dc, Kf = 1 from Eq. (9.10)
−0.2 dc, Kf = 0.923 from Eq. (9.11)
−0.4
0 5 10 15 20 25 30
Time
Figure 9.6b Example 9.3 responses of the feedforward-only control system to a unit step
disturbance for Kd = .8, Kp = 1.2, Td = 1.2, Tp = .8.
Figures 9.7a and 9.7b show worst-case feedforward plus feedback responses
corresponding to the responses in figures 9.6a and 9.6b. The 1DF PID feedback controller
was obtained using Mp tuning (see Chapter 7) and the IMC to PID controller approximation
method of Chapter 6. The IMC filter time constant required to achieve an Mp of 1.05 is
1.04, and the resulting PID controller is
PID = 0.610(1 + 1 /(1.24 s ) + .179 s /(.090 s + 1)). (9.17)
The responses in figures 9.7a and 9.7b should be compared to those of a well-tuned
2DF feedback-only control system, as shown in Figure 9.8. The controller used to generate
the responses in Figure 9.8 was obtained using the design and tuning methods described in
Chapter 8. Clearly, the performance of the combined feedforward/feedback control system
is superior to that of a standalone 2DF feedback control system.
0.15
Feedforward plus Feedback
0.10 Kf = 0.923 from Eq. (9.11)

Outputs, y(t)
0.05
0 Kf = 1 from Eq. (9.10)
−0.05
−0.10
0 5 10 15 20 25 30
Time
Figure 9.7a Example 9.3 output responses of the feedforward plus feedback control system
to a unit step disturbance Kd = 1.2, K = .8, Td = 1.2, Tp = .8.
Feedforward plus Feedback −0.02
−0.04 Kf = 0.923 from Eq. (9.11)

Kf = 1 from Eq. (9.10)
−0.06
Outputs, y(t)
−0.08
−0.10
−0.12
−0.14
−0.16
0 5 10 15 20 25 30
Time
Figure 9.7b Example 9.3 output responses of the feedforward plus feedback control system
to a unit step disturbance Kd = .8, Kp = 1.2, Td = 1.2, Tp = .8.
0.4
0.35
0.30
0.25
K = 0.8, Tp = 1.2
Output
0.20
K = 1, Tp = 1
0.15
.010
K = 1.2, Tp= 1.2
0.05
0
−.05
0 2 4 6 8 10 12 14 16 18 20
Time
Figure 9.8 Example 9.3 responses of a 2DF feedback control system to a unit step
disturbance with feedback controller = 3.52(1 + 1/(4.04s) + .965s) /(4.17s + 1).
♦
The next example is the same as Example 9.3, except that the disturbance transfer
function deadtime is always smaller than the process deadtime.
Example 9.4 Disturbance deadtime < process deadtime
−T p s
p( s ) = Ke /( s + 1); 0.8 ≤ K , T p ≤ 1.2 , (9.18a)
p d ( s ) = K d e −Td s /(4 s + 1); 0.8 ≤ K d ≤ 1.2, .4 ≤ Td ≤ .6 , (9.18b)
~
p ( s ) = e − s /( s + 1); ~
p d ( s ) = e −.5 s /( 4 s + 1) . (9.18c)
The 1DF feedforward controller is the same as that of Eq. (9.7a):

q ff 1 ( s ) = ( s + 1) /(4s + 1) . (9.18d)
The 2DF feedforward controller is the same as that of Eq. (9.7b):

q ff 2 ( s ) = ( s + 1)(.473s + 1) /( 4s + 1)(0.006s + 1) . (9.18e)
The PID feedback controller is the same as that of Eq. (9.17):

PID = 0.610(1 + 1 /(1.24 s ) + .179 s /(.090 s + 1)). (9.18f)
As before, the following feedforward/feedback responses to a unit step disturbance

should be compared with the 2DF feedback-only responses given by Figure 9.8. Figure 9.9
repeats the perfect model responses for the feedback controller of Eq. (9.18f). Notice that
the undershoots in Figure 9.9 of the uncompensated controllers are less severe than those in
Figure 9.4. This occurs because the feedback controller, Eq. (9.18f) , is less aggressive than
that of Eq. (9.7c) because Eq. (9.9f) has been tuned to accommodate the specified process
uncertainty.
Figures 9.10 and 9.11 show typical worst-case responses of the 1DF and 2DF
feedforward/feedback control systems. Unlike Figure 9.8, there is little difference in the
responses of the compensated (via Figure 9.5) and uncompensated 2DF controllers, and
both controllers perform better than the 1DF controller.
Figure 9.12 shows typical best-case responses that result when the difference between
the disturbance deadtime and process deadtime is 0.2, which is the smallest in the set of
uncertain parameters. The poorer behavior of the 2DF controllers relative to the 1DF
controller arises from the fact that the 2DF feedforward controller, Eq. (9.18e), was
obtained for a difference between the disturbance deadtime and process deadtime of 0.5.
0.12
One-Degree of Freedom
0.10
Feedforward Controller, qff 1
0.08 Two-Degree of Freedom
0.06
uncompensated
Output

0.02 compensated as in Figure 9.5
−0.02
−0.04
0 2 4 6 8 10 12 14 16 18 20
Time
Figure 9.9 Responses of feedforward/feedback control systems to a unit step disturbance at

time 1.0 for Example 9.4, K = Kd = Tp = 1, and Td = 0.5.
0.25
0.20 Feedforward Controller, qff 1

0.10 uncompensated
Output
Two-Degree of Freedom
compensated as in Figure 9.5
0
−0.05
−0.10
0 2 4 6 8 10 12 14 16 18 20
Time

time 1.0 for Example 9.4, and K = Kd = Tp = 1.2, and Td = 0.4
0.25
0.2
Feedforward Controller, q ff 2
uncompensated
0.15
Output
0.1 compensated as in Figure 9.5
0.05
0
0 2 4 6 8 10 12 14 16 18 20
Time
Figure 9.11 Responses of feedforward/feedback control systems to a unit step disturbance
at time 1.0 for Example 9.4, K = .8, Kd = Tp = 1.2, and Td = 0.4.
0.04
0.03
0.01
0
−0.01
Output
−0.02 Feedforward Controller, q ff 2
−0.03 uncompensated
−0.04
Feedforward Controller, q ff 2
−0.05 compensated as in Figure 9.5
−0.06
0 2 4 6 8 10 12 14 16 18 20
Time

time 1.0 for Example 9.4, K = Kd = Tp = .8, and Td = 0.6.
Since the deviations from setpoint in Figure 9.12 are quite small relative to those in
figures 9.10 and 9.11, our conclusion from these figures is that the 2DF feedforward
controller design is preferred.
♦
9.5 SUMMARY
A properly designed feedforward/feedback control system will always improve performance

over a simple feedback control system, independent of the process uncertainty, provided
only that the measured disturbance does not enter directly (i.e., with a unity transfer
function) into the process output. However, as one might expect, the greater the process
uncertainty, the less the potential improvement in response for processes at the extremes of
the uncertainty ranges.
For an uncertain process, the feedforward controller gain should be chosen either
K = ( K d ) ave /( K p ) ave K = ( K p / K d ) −ave
1
as f (see Eq. (9.10)) or as f (see Eq. (9.11), depending
on the uncertainties in the gains and on control objectives. However, only the latter choice
(i.e. Eq. (9.11)) guarantees that the feedforward/feedback controller will perform better than
a simple feedback controller for all processes in the set of uncertain processes.
When the deadtime between the measured disturbance and the output is greater than
the deadtime between the control and the output, then a feedforward controller should be
designed like a single-degree of freedom IMC controller, except that the difference between
the deadtimes is included in the controller.
When the deadtime between the measured disturbance and the output is less than the
deadtime between the control and the output, then the feedforward controller should be
designed as a 2DF IMC controller. If there is also relatively little process uncertainty, better
performance can be achieved by modifying the feedforward control system structure to
prevent the pulse generated by the feedforward action from being propagated around the
feedback control loop.
Problems
Select and tune control systems for each of the following processes. The control objective is to control
the measured output y(t), and all disturbances di(t) are measured.
2( 1 − 2s)e −6 s 50e −4 s d(s)

9.1 y(s) = 2
u(s) +
( 2s + 1 ) ( 50 s + 1 )
e −10 s e −5 s
9.2 y(s) = y 2(s) + d1(s) y 2(s) = u(s) + d 2(s)
30 s + 1 4s + 1
9.5 Summary 239
Ke −4 s K e −5 s
9.3 y ( s) = u ( s) + d d ( s ) 1 ≤ K ≤ 3 ; 2 ≤ τ ≤ 5; 3 ≤ K d ≤ 4
s (τs + 1) s (τs + 1)
K1e −Ts
9.4 y ( s) = y 2 ( s ) + d1 ( s ) 1 ≤ K1 ≤ 3, 4 ≤ T ≤ 6
5s + 1
K 2e− s
y2 ( s ) = u(s) + d 2 (s) 1 ≤ K 2 ≤ 5, 0 ≤ u (t ) ≤ 10
3s + 1
K1e−Ts
9.5 y(s) = y2(s) + d1(s) 1 ≤ K1 ≤ 5; 2 ≤ T ≤ 4
3s +1
K 2e − s
y2 ( s) = u (s) + d 2 1 ≤ K 2 ≤ 5
2s + 1
K1e −Ts e −Ts d1 ( s )
9.6 y ( s) = y1 ( s ) + 1≤ K ≤ 5 2 ≤T ≤ 4
3s + 1 3s + 1
K 2 (−2s + 1) d
y1 ( s ) = 2
u ( s) + 2
(2 s + 1) s +1
References
Seborg, D. E., T. F. Edgar, and D. A. Mellichamp. 1989. Process Dynamics and Control, John Wiley
& Sons, NY.
C H A P T E R 1 0
Cascade Control
Objectives of the Chapter
• To review classical cascade control.

• To present an alternate way of thinking about cascade control that leads to improved
performance.
• To introduce controller design methods that accommodate process uncertainty.
Prerequisite Reading
Chapter 3, “One-Degree of Freedom Internal Model Control”
Chapter 4, “Two-Degree of Freedom Internal Model Control”
Chapter 5, “MSF Implementations of IMC Systems”
Chapter 6, “PI and PID Controller Parameters from IMC Design”
Chapter 7, “Tuning and Synthesis of 1DF IMC Controllers for Uncertain Processes”
Chapter 8, “Tuning and Synthesis of 2DF Control Systems”
241
242 Cascade Control Chapter 10
10.1 INTRODUCTION
Cascade control can improve control system performance over single-loop control whenever
either: (1) Disturbances affect a measurable intermediate or secondary process output that
directly affects the primary process output that we wish to control; or (2) the gain of the
secondary process, including the actuator, is nonlinear. In the first case, a cascade control
system can limit the effect of the disturbances entering the secondary variable on the
primary output. In the second case, a cascade control system can limit the effect of actuator
or secondary process gain variations on the control system performance. Such gain
variations usually arise from changes in operating point due to setpoint changes or sustained
disturbances.
A typical candidate for cascade control is the shell and tube heat exchanger of
Figure 10.1.
Steam
PR
FR
TR
Feed
Effluent
Condensate
Figure 10.1 A shell and tube heat exchanger.
The primary process output is the temperature of the tube side effluent stream. There
are two possible secondary variables, the flow rate of steam into the exchanger and the
steam pressure in the exchanger. The steam flow rate affects the effluent temperature
through its effect on the steam pressure in the exchanger. The steam pressure in the
exchanger affects the effluent temperature by its effect on the condensation temperature of
the steam. Therefore, either the steam flow rate or the steam pressure in the exchanger can
be used as the secondary output in a cascade control system. The choice of which to use
depends on the disturbances that affect the effluent temperature.
If the main disturbance is variations in the steam supply pressure, due possibly to
variable steam demands of other process units, then controlling the steam flow with the
control valve is most likely to be the best choice. Such a controller can greatly diminish the
effect of steam supply pressure variations on the effluent temperature. However, it is still
10.1 Introduction 243
necessary to have positive control of the effluent temperature to be able to track effluent
temperature setpoint changes and to reject changes in effluent temperature due to feed
temperature and flow variation. Since there is only one control effort, the steam valve stem
position, traditional cascade control uses the effluent temperature controller to adjust the
setpoint of the steam flow controller, as shown in Figure 10.2.
Steam
Flow Temperature
FRC Setpoint TRC Setpoint
Feed
Effluent
Condensate
Figure 10.2 Cascade control of effluent temperature via steam flow control.
If feed flow and temperature variations are significant, then these disturbances can be
at least partially compensated by using the exchanger pressure rather than the steam flow as
the secondary variable in a cascade loop, as shown in Figure 10.3.
Steam
Pressure Temperature
PRC TRC
Setpoint Setpoint
Feed
Effluent
Condensate
Figure 10.3 Cascade control of effluent temperature via shell side pressure control.
The trade-off in using the configuration of Figure 10.3 rather than that of Figure 10.2
is that the inner control loop from the steam pressure to the valve stem position may not
suppress variations in valve gain as well as with an inner loop that uses the valve to control
the steam flow rate. This consideration relates to using a cascade control system to suppress
the effect of process uncertainty, in this case the valve gain, on the control of the primary
process variable, the effluent temperature. We will have a lot more to say about using
cascade control systems to suppress process uncertainty in the following sections.
To repeat, cascade control has two objectives. The first is to suppress the effect of
disturbances on the primary process output via the action of a secondary, or inner control
loop around a secondary process measurement. The second is to reduce the sensitivity of the
primary process variable to gain variations of the part of the process in the inner control
loop.
As we shall demonstrate, cascade control can be usefully applied to any process where
a measurable secondary variable directly influences the primary controlled variable through
some dynamics. We will also demonstrate that despite frequent literature statements to the
contrary, inner loop dynamics do not have to be faster than the outer loop dynamics.
However, the traditional cascade structure and tuning methods must be modified in order for
cascade control to achieve its objectives when the inner loop process has dynamics that are
on the order of, or slower than, the primary process dynamics.
10.2 CASCADE STRUCTURES AND CONTROLLER DESIGNS
Figure 10.4 shows the traditional PID cascade control system block diagram (Seborg et al.,
1989). This is the cascade structure associated with figures 10.2 and 10.3. For Figure 10.2,
the secondary process variable y2 is the steam flow rate, while for Figure 10.3, it is the shell-
side steam pressure. In both cases, the primary variable y1 is the effluent temperature.
Disturbances
d2 d1
Saturation
Setpoint y2 y1
PID1 PID2 p2 p1
Process
Inner loop Outer loop
Figure 10.4 Traditional cascade block diagram.

10.2 Cascade Structures and Controller Designs 245
One of the objectives of this section is to present methods for obtaining the parameters
of the PID controllers of Figure 10.4 from a well-designed and well-tuned IMC cascade
control system, just as we did for single-loop control systems in Chapter 6.
Figure 10.5 shows an IMC cascade block diagram that accomplishes the same
objectives as Figure 10.4. There are other equivalent IMC cascade structures to that given
by Figure 10.5 (Morari and Zafiriou, 1989). However, the configuration of Figure 10.5 is
convenient because it suggests that controller q2 should be designed and tuned solely to
suppress the effect of the disturbance d2 on the primary output y1, and also convenient
because both controller outputs u1 and u2 enter directly into the actuator. As we shall see
later, this last point facilitates dealing with control effort saturation. However, for the
remainder of this section we shall ignore the saturation block in order to study the design
and tuning of IMC controllers. These IMC controllers will then be used to obtain the PID
controller parameters in Figure 10.4, as was done in Chapter 6 for single-loop control
systems.
d2 d1
r u1 u us y2 y1
q1 ( s ) p2 ( s ) p1 ( s )
− −
u2 ~
p2 ( s ) − ~
p1 ( s ) −
~
q2 ( s )
d2
~
d1
Figure 10.5 IMC cascade structure.
From Figure 10.5, the transfer functions between the inputs to the inner loop, u1 and
d2, and the secondary process output y2 are
p ( s )u1 ( s ) + (1 − ~
p 2 ( s )q 2 ( s ))d 2 ( s )
y2 (s) = 2 . (10.1)
(1 + ( p ( s ) − ~
2 p ( s ))q ( s )
2 2
The transfer functions between the setpoint and disturbances and the primary process output
y1 are
p p q r ( s ) + (1 − ~
p 2 q 2 ) p1d 2 ( s ) + (1 − ~
p1 p 2 q1 + ( p2 − ~
p 2 ) q 2 ) d1 ( s )
y1 ( s ) = 1 2 1 ~ ~ . (10.2)
(1 + ( p − p ) p q + ( p − p )q )
1 1 2 1 2 2 2
In Eq. (10.2) we have suppressed the dependency of all transfer functions on the Laplace
variable s to keep the equation on one line. Based on equations (10.1) and (10.2) we observe
the following:
(1) If the lag time constants of the primary process p1(s) are large relative to those of
the secondary process p2(s) then the inner loop controller q2(s) should be chosen so that the
zeros of (1 − ~ p2 ( s )q2 ( s )) cancel the small poles (i.e., large time constants) of ~
p1 ( s ) as
~
outlined in Chapter 4. Otherwise, q2(s) should simply invert a portion of p2 ( s ) as described
in chapters 3 and 7.
(2) The outer loop controller q1(s) should approximately invert the entire process
model ~p1 ~
p2 ( s ) , as described in chapters 3 and 7.
(3) The IMCTUNE software can be used to design and tune both q1(s) and q2(s).
We recommend tuning q2(s) with the outer loop open, and then tuning q1(s) with the
inner loop closed. That is, first find the filter time constant ε2 for q2(s), and then find ε1 for
q1(s). According to the denominator of Eq. (10.2), the tunings for q1(s) and q2(s) interact.
Therefore, some adjustment of ε2 may be necessary after obtaining ε1.
Having obtained the IMC controllers for Figure 10.5, we would like to use these
controllers to obtain the PID controllers in Figure 10.4 in a manner similar to that for single-
loop controllers described in Chapter 6. Unfortunately, however, we can do so only very
approximately. Figure 10.5 can be rearranged, ignoring the saturation block, as given by
Figure 10.6.
d2 d1
r q2 ( s )
y2 y1
q1q 2−1 ( s ) p2 ( s ) p1 ( s )
(1 − ~
p2 q2 ( s ))
~
p1 ( s )
~
d1
Figure 10.6 IMC cascade control with a simple feedback inner loop.
The controller given by q2 ( s ) /(1 − ~

p2 ( s )q2 ( s )) can often be well approximated by a
PID controller, as described in Chapter 6. Again, IMCTUNE can be used to obtain this
controller. However, obtaining PID1 in Figure 10.4 is not so straightforward. Collapsing the
feedback loop through ~p1 ( s ) , while leaving the inner loop alone, yields Figure 10.7.
d2 d1
(1 + ( p2 ( s) − ~p2 ( s)) q2 ( s))
(1 − ~
p1 ( s ) p2 ( s)q1 ( s) + ( p2 ( s ) − ~
p2 ( s )) q2 ( s ))
r q2 ( s ) y2 y1
C1 ( s )q 2−1 ( s ) p2 ( s ) p1 ( s )
(1 − ~p2 q2 ( s ))
C1 ( s ) ≡ q1 ( s)(1 + ( p2 ( s ) − ~
p2 ( s)) q2 ( s )) /(1 + ( p2 ( s ) − ~
p2 ( s)) q2 ( s ) − ~
p1 ( s ) p2 ( s )q1 ( s ))
Figure 10.7 Standard feedback form of Figure 10.6.
The controller C1(s) in Figure 10.7 cannot be realized because it contains the process
transfer function p2(s), which is uncertain and cannot be made part of the controller. We can
however approximate p2(s) with its model ~ p2 ( s ) . In this case C1(s) becomes
C1 ( s ) ≅ q1 ( s ) /(1 − ~
p1 ( s ) ~
p2 ( s )q1 ( s )). (10.3)
Another difference between figures 10.6 and 10.7 is that even if the model ~ p1 ( s ) is a
perfect representation of the process, the pulse created by the inner loop response to the
disturbance d2(s) (i.e., d 2 ( s ) /(1 − ~
p1 ( s ) ~
p 2 ( s )q 2 ( s )) for a perfect model ~
p2 ( s ) ) feeds back
around the outer loop of Figure 10.7. Since the primary controller cannot suppress this
pulse, it continues around the loop until it dies out.
Even using the approximation given by Eq. (10.3) to obtain a PID controller does not
reduce Figure 10.7 to the standard PID cascade diagram of Figure 10.4 because C1(s) in
Figure 10.7 is multiplied by q2−1 ( s ) . If q2−1 ( s ) is a lead (which will generally occur only if
the process description is quite uncertain), then q2−1 ( s ) can be approximated by a
polynomial via a Taylor’s series. This polynomial can be multiplied into the PID controller
obtained from C1(s) to obtain a new PID controller after dropping higher order terms. Even
if q2−1 ( s ) is a lag, it may still be possible to approximate the term C1(s) q2−1 ( s ) by a PID
controller. However, the necessary approximations will have to be carried out by hand,
following procedures in Chapter 6, as the current version of IMCTUNE does not carry out
the necessary manipulations.
Two rather long examples of cascade control of uncertain processes follow. The
individual processes in both examples are first-order lags plus dead time and have
significant process uncertainty. In the first example, the secondary process output dynamics
are significantly faster than the primary process dynamics. In the second example, the
primary and secondary process dynamics have similar dynamic behavior.
Example 10.1 Secondary Process has Faster Dynamics

than the Primary Process
K 1e − T s
1
p1 ( s ) = ; 0.8 ≤ K1 ≤ 1.2, 17.5 ≤ T1 ≤ 22.5, 14 ≤ τ 1 ≤ 16 (10.4a)

τ1 s + 1
K 2 e −T s2
p2 ( s) = ; 0.6 ≤ K 2 ≤ 1.8, 2 ≤ T2 ≤ 4, 1 ≤ τ 2 ≤ 3 (10.4b)

τ2s + 1
(a) IMC System Design

Following the suggestions in chapters 7 and 8, we use the upper-bound gains and dead times
and the lower-bound time constants for the process models.
−22 .5s
~ 1.2e
p1 ( s ) = (10.4c)
14 s + 1
−4s
~ 1.8e
p2 ( s ) = (10.4d)
s+1
Computing the 2DF feedback controller for the inner loop (see Figure 10.5), using
IMCTUNE with the outer loop, open gives
( s + 1)(9.05s + 1)
q2 ( s) = . (10.5a)
1.8(4.4s + 1) 2
Figure 10.8 shows the tuning curves, while Figure 10.9 shows typical time responses to a
step disturbance in the inner loop. Data for both figures was obtained from IMCTUNE.
101
| Partial Sensitivity Function |
100
Upper-bound
10−1
Lower-bound
10−2
10−3
10−2 10−1 100 101
Frequency (rad/unit time)
Figure 10.8 Cascade inner loop tuning using controller given by Eq. (10.5a).
0.5
0.4
K1 τ1 T1 K2 τ2 T2 = 1.2 14 17.5 .6 1 2
0.3
K1 τ1 T1 K2 τ2 T2 = .8 16 17.5 .6 1 2
Output
0.2
0.1
0
K1 τ1 T1 K2 τ2 T2 = 1.2 14 17.5 1.8 3 4
−0.1
0 50 100 150 200 250 300
Time
Figure 10.9 Responses to a step inner loop disturbance (d2) with the outer loop open.
Having obtained the inner loop controller, the outer loop controller can be obtained from the
cascade facility of IMCTUNE, and is
(15s + 1)
q1 ( s ) = . (10.5b)
2.16(16.87 s + 1)
The tuning curves for the outer loop of the cascade, using Eq. (10.5b), are shown in
Figure 10.10. Also in this figure are the closed-loop upper-bound and lower-bound curves
for a single-loop controller for a model and controller of
−26 .5s
~ 2.16 e (15s + 1)
p(s) = q( s) = . (10.6)
15s + 1 2.16 (14.3s + 1)
Recall from equations (10.4a) and (10.4b) that the overall process is
Ke −T s
p( s) = 0.48 ≤ K ≤ 2.16, 19.5 ≤ T ≤ 26.5, 14 ≤ τ 1 ≤ 16 1 ≤τ 2 ≤ 3.
( τ1 s + 1)( τ 2 s + 1)
Upper and Lower Bounds of the Magnitude
of the Complementary Sensitivity Function
101
100
10−1
Cascade Control
10−2 Single-Loop Control
10−3
10−2 10−1 100 101
Figure 10.10 Comparison of closed-loop setpoint to output responses.
Based on the closed-loop frequency responses we can conclude that the fastest
responses of the single-loop system are slightly faster than those of the cascade system, but
more importantly, the slowest responses are significantly slower. Figures 10.11 and 10.12
support these conclusions. Note the different time axes in figures 10.11 and 10.12.
1.4
1.2
0.8
Output
K1 τ1 T1 K2 τ2 T2
0.6
1.2 16 22.5 1.8 3 4
0.4
1.2 14 17.5 1.8 3 4
0.2
0.8 14 17.5 0.6 1 2
0
0 50 100 150 200 250
Time
Figure 10.11 Step setpoint responses for the cascade control system of Figure 10.5.
1.4
1.2
0.8
Output
K τ1 τ2 T
0.6
2.16 16 3 26.5
0.4
2.16 14 1 19.5
0.2
0.48 14 1 19.5
0
0 100 200 300 400 500 600 700
Time
Figure 10.12 Step setpoint responses for the single-loop control system, using Eq. (10.6).
The reason for the improved setpoint response of the cascade system is that the inner
loop of the cascade reduces the effect of gain uncertainty in the inner loop process. To show
that this is so, Figure 10.13 compares the closed-loop frequency responses of the cascade
system with that of a single-loop controller. The process is the same as that given by
equations (10.4a) and 10.4b), except that instead of a lower-bound of 0.48, the lower
bounds (lb) are 1.1 and 1.44. That is, the single-loop process is
Ke −T s
p( s) = lb ≤ K ≤ 2.16, 19.5 ≤ T ≤ 26.5, 14 ≤ τ 1 ≤ 16 1 ≤τ 2 ≤ 3. (10.7)
( τ1 s + 1)( τ 2 s + 1)
The model and controller for the process of Eq. (10.7) are the same as given in Eq. (10.6)
and are repeated for convenience:
−26 .5s
~ 2.16 e (15s + 1)
p(s) = q( s) =
15s + 1 2.16 (14.3s + 1)
A lower-bound gain of 1.44 corresponds to a secondary process (i.e., ~ p2 ( s ) ) with a

gain of 1.8 and no gain uncertainty. A lower-bound gain of 1.1 corresponds to a secondary
process whose gain varies between 1.375 and 1.8. In other words, the effect on the outer
loop of the ratio of the maximum to minimum gain variation of the secondary process has
been reduced from a ratio of 3 to a ratio of 1.3. The slowest time responses are compared in
Figure 10.14.
101
Upper and Lower Bounds of the Magnitude
of the Complementary Sensitivity Function
Upper Bounds
10 0
10 −1
Lower Bounds
Cascade Control
10 −2
Single-Loop Control Lower-Bound Gain = 1.44
−3
10
10 −3 10 −2 10 −1 100
Figure 10.13 Comparison of cascade and single-loop control systems.

1.4
1.2
1.0
Cascade Control
0.8
Output
0.6
Process parameters
0.4 K1 τ1 T1 K2 τ2 T2 = .8 14 17.5 .6 1 2
0.2
0
0 50 100 150 200 250
Time
Figure 10.14 Comparison of slowest responses to a step setpoint change.
We now return to the cascade control system responses to a step disturbance to the
inner loop, but this time with the outer loop closed. The time responses for the same
processes as in Figure 10.9 are shown in Figure 10.15. From this figure we conclude that
there is no need to retune the inner loop.
Figure 10.16 shows the effect of using the single-degree of freedom IMC controller
given by Eq. (10.8) on the response to a step disturbance in the inner loop. These responses
should be compared with those of Figure 10.15.
( s + 1)
q2 ( s) = . (10.8)
1.8(4.18s + 1)
The filter time constant of 4.18 in Eq. (10.8) yields an Mp of 1.05. That is, the
controller is tuned so that the worst-case overshoot of the inner loop output y2 to a step
setpoint change to the inner loop is about 10% with the controller q2 in the forward path.
This controller is then used in the feedback path of the inner loop in Figure 10.5.
0.5
0.4
K1 τ1 T1 K2 τ2 T2 = 1.2 14 17.5 .6 1 2
0.3
Output
0.2
K1 τ1 T1 K2 τ2 T2 = .8 16 17.5 .6 1 2
0.1
0
K1 τ1 T1 K2 τ2 T2 = 1.2 14 17.5 1.8 3 4
−0.1
−0.2
0 50 100 150 200 250 300
Time
Figure 10.15 Responses to a step inner loop disturbance (d2) with the outer loop closed.
0.6
0.5
K1 τ1 T1 K2 τ2 T2 = 1.2 14 17.5 .6 1 2
0.4
0.3
Output
0.2
K1 τ1 T1 K2 τ2 T2 = .8 16 17.5 .6 1 2
0.1
0
K1 τ1 T1 K2 τ2 T2 = 1.2 14 17.5 1.8 3 4
−0.1
0 50 100 150 200 250 300
Time
Figure 10.16 Responses to a step inner loop disturbance using the controller given by
Eq. (10.8).
While the inner loop disturbance responses using the single-degree of freedom
controller Eq. (10.8) are significantly slower than the 2DF controller given by Eq. (10.5a),
the responses of the output y1(t) to setpoint changes to the outer loop are only slightly
slower than those given in Figure 10.11.
(b) PID Cascade Controller Designs
Section 10.2 discusses methods for approximating the IMC cascade control system with the
traditional cascade system of Figure 10.4. Figure 10.7 shows the IMC equivalent
configuration. For convenience, this figure is repeated in Figure 10.17.
d2 d1
(1 + ( p2 ( s ) − ~ p2 ( s ))q2 ( s ))
(1 − ~
p1 ( s ) p2 ( s )q1 ( s ) + ( p2 ( s ) − ~
p2 ( s ))q2 ( s ))
r q2 ( s ) y2 y1
C1 ( s )q 2−1 ( s )
(1 − ~p2 q2 ( s ))
p2 ( s ) p1 ( s )
C1 ( s ) ≡ q1 ( s )(1 + ( p2 ( s ) − ~
p2 ( s )) q2 ( s )) /(1 + ( p2 ( s ) − ~
p2 ( s )) q2 ( s ) − ~
p1 ( s ) p2 ( s )q1 ( s ))
Figure 10.17 Standard feedback form of an IMC cascade control system.
Recall that the controller C1(s) in Figure 10.7 is not realizable because it contains
terms involving the inner loop process p2(s), which varies within the uncertainty set and
cannot be part of the controller. Therefore we suggested replacing p2(s) with its
model, ~
p 2 ( s ) . This gives
C1 ( s ) ≅ q1 ( s ) /(1 − ~
p1 ( s ) ~
p2 ( s )q1 ( s )). (10.9a)
IMCTUNE provides the following PID controllers from the IMC controllers obtained
previously:
q2 ( s)
Inner loop: ≅ PID2 = 1.79(1 + 1 /(12.05s ) + 1.68s ) /(14.7 s + 1). (10.9b)
(1 − ~
p 2 q 2 ( s ))
Outer loop: C1 ( s ) ≅ PID1 = .234(1 + 1 /( 23.77 s ) + 5.35s /(.29 s + 1)), (10.9c)
q2−1 = 1.8(4.4 s + 1) 2 /(( s + 1)(9.05s + 1)). (10.9d)
Figures 10.18 and 10.19 show the disturbance responses for the configuration of
Figure 10.17 using the controllers given in equation sets (10.9) and (10.10). Notice that
q2(s) in Eq. (10.9d) is from a 2DF design, and for this reason the responses are labeled
Cascade 2.
0.4
IMC Cascade (2DF Inner Loop)
0.3 PID Cascade 2 (2DF Inner Loop)
PID Cascade 1 (1DF Inner Loop)
0.2
Process parameters
K1 τ1 T1 K2 τ2 T2 = .8 16 17.5 .6 1 2
Output
0.1
−0.1
−0.2
0 50 100 150 200 250 300
Time
Figure 10.18 Comparison of responses to a step disturbance in the inner loop.
0.6
0.5 IMC Cascade (2DF Inner Loop)
0.2
Output
Process parameters
K1 τ1 T1 K2 τ2 T2 = 1.2 14 17.5 .6 1 2
0.1
0
−0.1
−0.2
−0.3
0 50 100 150 200 250 300
Time

Using the 1DF IMC controller for q2(s), given by Eq. (10.8) and repeated below,
yields the inner loop PID controller given by Eq. (10.10a).
( s + 1)
q2 ( s) =
1.8(4.18s + 1)
Inner loop: PID2 = .134(1 + 1 /(1.98s ) + .319 s /(.016 s + 1)). (10.10a)
The outer loop controller remains the same as in Eq. (10.9b) because q1(s) has not
changed. In figures 10.18 and 10.19 the responses using Eq. (10.10a) are labeled Cascade 1.
These responses show the benefits of an IMC outer loop over a PID outer loop. The outer
loop PID controller in the responses in figures 10.18 and 10.19 is cascaded with the term
q2−1. Since q2−1 is a lead, it can be approximated by the Taylor series as the polynomial
1.8(−3.18s 2 + 3.18s + 1) . Multiplying this polynomial into Eq. (10.10a) and dropping terms
higher than second order gives, after some rearrangement,
Outer loop: PID1 = .4956 (1 + 1 /(26.96 s ) + 7.404 s /(.37 s + 1)). (10.10b)
The PID controller given by Eq. (10.10b) can be used in place of equations (10.8) and
(10.10a) and gives virtually identical results. The advantage of Eq. (10.10b) is that it can be
used in the traditional cascade configuration of Figure 10.4. Of course, the disturbance
response will be that of the 1DF cascade of figures 10.18 and 10.19.
♦
The purpose of the next example is to demonstrate the advantages of cascade control
even when the dynamics of the secondary process are on the same order as the dynamics of
the primary process. A common literature fallacy is that the dynamics of the secondary
process have to be fast relative to those of the primary process in order to get improved
performance from a cascade control system. This fallacy probably arose from the methods
used for designing and/or tuning PID cascade control systems. Traditionally, the outer loop
controller was designed and tuned assuming that the inner loop is so fast that it can be
approximated as a unity gain. When this assumption is not true, the inner and outer loop
designs and/or tunings interact, and there existed no good methods of designing and tuning
the controller parameters that significantly improved performance over that of a single-loop
controller. In an IMC cascade configuration (see Figure 10.5) the inner and outer loops
interact mainly by the fact that the inner loop process gain variations are reduced by the
action of the inner loop controller. Such interaction is desirable and, as we shall show, does
not preclude arriving at controller designs so that cascade performance is significantly better
than single-loop performance.
Example 10.2 Primary and Secondary Processes have

Similar Dynamics
The following process was obtained by reducing the time constant and dead time of the
primary process of Example 10.1 by a factor of five. This gives the following system:
K1e −T1 s
p1 ( s ) = ; .8 ≤ K1 ≤ 1.2, 3.5 ≤ T1 ≤ 4.5, 2.8 ≤ τ1 ≤ 3.2 . (10.11a)
τ 1s + 1
K 2 e −T2 s
p2 ( s) = ; .6 ≤ K 2 ≤ 1.8, 2.0 ≤ T2 ≤ 4.0, 1.0 ≤ τ 2 ≤ 3.0. (10.11b)
τ2s + 1
(a) IMC System Design

Again following the suggestions in chapters 7 and 8, we use the upper-bound gains and
dead times and the lower-bound time constants for the process models.
−4.5s −4s
~ 1.2e 1.8e
p1 ( s ) = and ~
p2 ( s) = . (10.11c)
2.8s + 1 s +1
The controllers associated with the IMC cascade structure of Figure 10.5 are
( s + 1) (3.8s + 1)
q2 ( s) = and q1 ( s ) = . (10.11d)
1.8(2.8s + 1) 2.16(5.24s + 1)
An initial attempt at designing a 2DF controller for the inner loop resulted in the filter
time constant reaching the primary process model time constant of 2.8 before achieving an
Mp of 1.05 for the partial sensitivity function. In such a situation the inner loop feedback
controller is chosen as a 1DF controller with the filter time constant tuned using the partial
sensitivity function just as for a 2DF design. This controller is given by Eq. (10.11d).
Equation (10.11d) also shows the outer loop controller that achieves an Mp of 1.05 for the
complementary sensitivity function. Figures 10.20 and 10.21 show the disturbance and
setpoint responses of the IMC cascade control system with models and controllers given by
equations (10.11c) and (10.11d).
1.2
1.0
K1 τ1 T1 K 2 τ 2 T2 = 1.2 3.2 4.5 1.8 3 4

0.8
Output
0.6
K1 τ1 T1 K 2 τ 2 T2 = .8 3.2 4.5 .6 3 4
0.4
K1 τ1 T1 K 2 τ 2 T2 = .8 2.8 3.5 .6 1 2
0.2
−0.2
0 50 100 150
Time
Figure 10.20 Unit step disturbance (d2) responses for the IMC cascade control system.
1.4
1.2
K1 τ1 T1 K2 τ2 T2 = 1.2 3.2 4.5 1.8 3 4
1
0.8
Output
0.6
K1 τ1 T1 K2 τ2 T2 = 0.8 3.2 3.5 0.6 1 2
0.4
K1 τ1 T1 K2 τ2 T2 = 1.2 2.8 4.5 1.8 1 2
0.2
0
0 50 100 150
Time
Figure 10.21 Unit step setpoint responses for the IMC cascade control system.
The responses in figures 10.20 and 10.21 should be compared with those of a well-
tuned single-loop control system for the process given by equations (10.11a) and 10.11b)
and rewritten as equations (10.12a) and (10.12b). Equations (10.12c) and (10.12d) give the
associated model and controller:
K1 K 2 e −Ts
p( s) = , (10.12a)
(τ 1 s + 1)(τ 2 s + 1)
where .8 ≤ K1 ≤ 1.2, .6 ≤ K 2 ≤ 1.8, 2.8 ≤ τ 1 ≤ 3.2, 1 ≤ τ 2 ≤ 3, 5.5 ≤ T ≤ 8.5 ,
K1
pd ( s) = , (10.12b)
(τ 1 s + 1)
~ 2.16e −8.5 s
p ( s) = , (10.12c)
(3.8s + 1)
(3.8s + 1)
q( s) = . (10.12d)
2.16(6.31s + 1)
Notice that Eq. (10.12b) ignores the disturbance deadtime since this term changes
only the effective arrival time of the disturbance and so cannot be distinguished from the
disturbance itself. Also, the model given by Eq. (10.12c) approximates the process lags as a
first-order system whose time constant is the sum of the time constants of the two first order
process lags. Finally, the controller given by Eq. (10.12d) is a 1DF controller because we
are using a single loop controller in spite of the fact that the disturbance, d2, enters into the
primary output through the lag given by Eq. (10.12b).
The single-loop responses given in figures 10.22 and 10.23 are roughly twice as slow
as those of the cascade control system shown in Figures 10.20 and 10.21. Notice that the
time scales in figures 10.22 and 10.23 are from 0 to 300 whereas the time scales in
figures 10.20 and 10.21 are from 0 to 150. Also, the disturbance peak heights in
Figure 10.22 are higher than those of the cascade control system in Figure 10.20.
1.2
0.8
K1 K2 τ1 τ2 T = 0.8 .6 3.2 3 8.5

Output
0.6
0.4
0.2 K1 K2 τ1 τ2 T = 0.8 .6 2.8 1 5.5
0
K1 K2 τ1 τ2 T = 1.2 1.8 3.2 3 8.5
−0.2
0 50 100 150 200 250 300
Time
Figure 10.22 Single-loop control system of Eq. (10.12) responses to a step disturbance in d2.
1.4
1.2
K1 τ1 T1 K2 τ2 T2 = 1.2 3.2 4.5 1.8 3 4
1
0.8
Output
K1 τ1 T1 K2 τ2 T2 = 1.2 2.8 4.5 1.8 1 2

0.6
0.4
K1 τ1 T1 K2 τ2 T2 = 0.8 3.2 3.5 0.6 1 2
0.2
0
0 50 100 150 200 250 300
Time
Figure 10.23 Step setpoint responses for the single-loop control system is given by
Eq. (10.12).
(b) PID Cascade Controller Designs

Replacing the IMC inner loop with a feedback controller q2 ( s ) /(1 − ~
p 2 q 2 ( s )), as in
Figure 10.6, and then approximating the feedback controller with the PID controller given
by Eq. (10.13a) does not change the setpoint and disturbance responses of
figures 10.20 and 10.21. That is, there is no degradation of the performance of the mixed
IMC-PID cascade control system.
PID2 ( s ) = .178(1 + 1 / 2.18s + .456 s /(0.0228s + 1)). (10.13a)
Approximating the controller C1 by Eq. (10.3), multiplying it by the Maclaurin series

approximation to q2−1 ( s ), and finally approximating the term C1 q2−1 ( s ) as a PID controller,
as in Example 10.1, gives
PID1 ( s ) = .485(1 + 1 / 8.0s + 2.45s /(.122 s + 1)). (10.13b)
Figure 10.24 shows the inner loop disturbance d2 response for the traditional cascade
configuration of Figure 10.4, using the PID controllers given by equations (10.13a and
(10.13b). The response for a process with upper-bound parameters is too oscillatory. The
step setpoint response for the same process shows a 21% overshoot. The reason for this
poorer behavior is probably the interaction between the inner and outer loops.
1.0
0.8
IMC Cascade
0.6 K1 τ1 T1 K2 τ2 T2 = 0.8 3.2 3.5 0.6 3 2
0.4
K1 τ1 T1 K2 τ2 T2 = 1.2 3.2 4.5 1.8 3 4
0.2
Output
0
−0.2
−0.4 PID Cascade
−0.6 K1 τ1 T1 K2 τ2 T2 = 0.8 3.2 3.5 0.6 3 2
−0.8 K1 τ1 T1 K2 τ2 T2 = 1.2 3.2 4.5 1.8 3 4
−1.0
0 20 40 60 80 100 120 140 160 180 200
Time

10.3 Saturation Compensation 263
Conceptually, it is possible to extend the cascade feature of IMCTUNE to

accommodate the PID cascade configuration of Figure 10.4 and to automatically increase
the filter time constant of q1(s) to tune the outer loop to give a specified Mp. After tuning,
IMCTUNE, or any program like it, should be able to provide the PID approximation to the
term C1 q2−1 ( s ) . Unfortunately, such an extension does not yet exist, and the only method
that we can suggest to improve the responses in Figure 10.24 is a rather tedious trial-and-
error method wherein one increases the filter time constant of q1(s), re-computes PID1, and
then checks the responses of the processes with the upper-bound parameters. This assumes
that the worst-case responses will always be those for the upper-bound parameters, which of
course may not always be true.
♦
10.3 SATURATION COMPENSATION
10.3.1 IMC Cascade
Figure 10.5 provides the simplest starting point for a discussion of control effort saturation
in cascade control systems. For convenience, this figure is reproduced in Figure 10.25.
d2 d1
Saturation
r u1 u us y2 y1
q1 ( s ) p2 ( s ) p1 ( s )
− −
u2 ~
p2 ( s ) − ~
p1 ( s ) −
~
q2 ( s )
d2
~
d1
Figure 10.25 IMC cascade structure.
The effect of control effort saturation on the inner loop of Figure 10.25 can be minimized by
implementing the inner loop as a model state feedback (MSF) system, as shown in
Figure 10.26a.
d2
Saturation
u1 u us y2
p2 ( s )
−
u2 ~
p2 ( s )
−
D2−1 ( s ) N 2 ( s )e −Ts
q2 ( s )
~
d2
−
K(s)
Ksp
Figure 10.26a An MSF implementation of the inner loop of Figure 10.25.
The only difference between Figure 10.26a and Figure 5.5 of Chapter 5 is that there is no
setpoint in Figure 10.26a.
Saturation compensation for the outer loop of Figure 10.25 is not quite so
straightforward. One problem is that the outer loop controller is designed to invert portions
of both inner loop and outer loop models (i.e., ~ p1 ( s ) ~
p2 ( s ) ). However, there is no such
transfer function, since the output of the inner loop model is not the input to the outer loop
model. One solution is to create a new transfer function, 1/ D1 (s) , where D1(s) contains the
denominator of the transfer function that the controller q1(s) inverts. Figure 10.26b shows an
MSF implementation of the outer loop controller, q1(s), using this approach. This figure
includes inner loop control system of Figure 10.26a, as it is necessary to remove the inner
loop control effort u2 from the signal used to compute the feedback portion of the outer loop
control effort u1.
10.3 Saturation Compensation 265
d2
q1 ( s )
Saturation
r u1 u us y2
K1SP p2 ( s )
u2 ~
p2 ( s )
K1(s) D1−1 ( s ) D2−1 ( s ) N 2 ( s )e −Ts
~
~ d2
d1
K2(s)
K2SP
Disturbance
estimate from the
outer loop
Figure 10.26b MSF implementation of both inner loop and outer loop controllers of
Figure 10.25.
10.3.2 IMC/PID Cascade
In the absence of saturation there is usually little performance loss if the inner loop of the
IMC cascade structure is replaced by a simple feedback loop, as shown in Figure 10.6.
However, the method used in Figure 10.26b to compensate the outer loop for control effort
saturation does not readily carry over to the outer loop of the cascade structure in
Figure 10.6. The problem is that in the structure of Figure 10.6, there is no explicit
calculation of an inner loop control effort, u2, as there is in figures 10.5 and 10.26b. For this
reason, we recommend implementing the outer loop as shown in Figure 10.27. The limits of
saturation block in this figure would ideally be set to the limits of the actual control effort
less the contribution of the inner loop control effort u2 to the total control effort. However,
since u2 is not available without additional calculations, we recommend simply setting the
limits to those of the actual control effort. This is, of course, equivalent to assuming that u2
is zero. Notice that the saturation block in Figure 10.27 is not on the outer loop control
effort u1 but rather only on the input to the inverse of the numerator of q1(s), which is D1(s).
The reason is that the role of the structure in Figure 10.27 is only to attempt to compensate
for saturation in the inner loop, and not to limit the setpoint sent to the inner loop. Finally,
we recommend replacing the IMC controller, q2 ( s ) /(1 − ~ p 2 q 2 ( s )), in Figure 10.6 with a
standard anti-reset windup PID controller, as described in Chapter 6.
q1 ( s )
r u1 Setpoint to
Ksp q2−1 ( s ) inner loop
− −
Saturation
D1−1 ( s )
~ K(s) Disturbance
estimate from
d1 the outer loop
Figure 10.27 Compensating the outer loop of Figure 10.6 for control effort saturation.
10.3.3 PID Cascade
Saturation compensation for the standard PID cascade structure of Figure 10.4 is usually
accomplished by either of two methods. The preferred method is to use logic statements that
stop the integration in both the inner loop and outer loop PID controllers whenever the
control effort reaches a limit, and start it again whenever the error signals change sign or the
control effort comes off saturation. The second, and possibly more common, method is to
use a standard anti-reset windup controller in the inner loop and implement the integral
portion in the outer loop PID controller, as shown in Figure 10.28.
Disturbances
d2 d1
y2 y1
Setpoint
PD1 PID2 p2 p1
− −
Saturation
Process
−
1
( τ I1s + 1)
Figure 10.28 Feedback method of outer loop anti-reset windup for PID cascade.
10.4 Summary 267
The difficulty with the implementation of Figure 10.28 is that is the outer loop integral
time constant is not τI1, but rather a complicated function of τI1 and the parameters of the
inner loop transfer function. This complication can make it inadvisable to use the IMC-
generated outer loop PID parameters developed in this section. The first method, which
simply stops the integration on control effort saturation, does not have the foregoing
drawback.
10.4 SUMMARY
To achieve the best disturbance rejection and setpoint tracking the inner loop of the cascade
should be designed and tuned as a 2DF controller. The outer loop process lag plays the role
of the disturbance lag in the controller design. The outer loop should be implemented as an
MSF IMC system. The inner loop can be implemented as either a PID control system or, in
the case of very little process uncertainty, in IMC MSF form.
There is no need for the inner loop process to be faster than the outer loop process in
order for a well-designed cascade control system to provide significant performance
advantages over a single-loop control system.
The techniques of this chapter can often be used to obtain the PID parameters for the
traditional cascade structure. However, anti-reset windup for the outer loop should be
implemented by stopping integration when the control effort saturates in order to use the
calculated integral time constant. If the anti-reset windup for the outer loop is implemented
via a lag around the inner loop, then the lag time constant is not necessarily the same as the
computed integral time constant. Outer loop setpoint tracking and disturbance rejection is
generally better than that achievable with a single-loop control system because the inner
loop serves to reduce the apparent gain uncertainty of the inner loop process.
Problems
Design and tune cascade control systems for each of the following processes. The primary output is y,
and the measured secondary output is y2. Also, compare the performance your cascade control system
with that of the feedforward control systems found for the problems of Chapter 9. The problems in
Chapter 9 are the same as those below except that all the disturbances were considered to be measured
whereas now only the process outputs y(s) and y2(s) are measured.
e −10 s e −5 s
10.1 y (s) = y 2 ( s ) + d1 ( s ) y2 ( s) = u (s) + d 2 ( s)
30 s + 1 4s + 1
K1e −Ts
10.2 y ( s) = y 2 ( s ) + d1 ( s ) 1 ≤ K1 ≤ 3, 4 ≤ T ≤ 6
5s + 1
K 2e−s
y 2 ( s) = u ( s) + d 2 ( s) 1 ≤ K 2 ≤ 5, 0 ≤ u (t ) ≤ 10
3s + 1
K1e −Ts
10.3 y ( s) = y 2 ( s ) + d ( s ) 1 ≤ K1 ≤ 5, 2 ≤ T ≤ 4
3s + 1
K 2e − s
y2 ( s) = u (s) + d 2 1 ≤ K 2 ≤ 5
2s + 1
K1e −Ts e −Ts d1 ( s )

10.4 y (s) = y1 ( s ) + 1 ≤ K ≤ 5, 2 ≤ T ≤ 4
3s + 1 3s + 1
K 2 (−2s + 1) d
y1 ( s ) = u ( s) + 2
(2 s + 1) 2 s +1
References
Morari, M., and E. Zafiriou, E. 1989. Robust Process Control. Prentice Hall, NJ.
Seborg, D. E., T. F. Edgar, and D. A. Mellichamp. 1989. Process Dynamics and Control. John Wiley
& Sons, NY.

PC Lessons On Process Control

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PC Lessons On Process Control

Uploaded by

Copyright:

Available Formats

10.

450 Process Dynamics, Operations, and Control

1.1 Surge tank

from other processes

Stream wi has some steady design flowrate, but in practice it varies,

1.2 Designing the tank

(1) characterizing the inlet stream

(2) specifying limits on the outlet stream

(3) making a suitable process model to connect inlet with outlet

Process model: by instinct a chemical engineer writes a material balance.

where height is in m, area in m2, time in minutes, and density in kg m-3.

Substituting (1.2.4) into (1.2.2), we find

Therefore, we should protect our tank with some automatic process

1.3 Definitions to get us started

final control element

1.4 That’s process control?

'Process' is a good notion; another useful notion is that of 'system'. A

inputs system outputs

2.2 Systems within systems

Considering the relationship of these component systems, we recognize

When we add a controller to a process, we create a single time-varying

manipulated variable (intermediate

final control element

To do our job well, we must understand more about system dynamics --

2.3 Dynamics of a tank, without any control

2.4 Response to rectangular pulse at time td

Inserting the disturbance (2.4.1) into process model (2.3.1), we compute

2.5 Impulse at time td

Again, the level never recovers from the brief disturbance.

2.6 Step at time td

normalized input and output 3

2.8 Typical disturbances

3.1 Quadratic equation

3.2 complex numbers

z = a + jb, where j = − 1 , Re(z) = a, Im(z) = b (3.2.1)

The same number may be written in polar form

e ± jϕ = cos ϕ ± j sin ϕ (3.2.3)

Each complex number z has a conjugate z*

The conjugates have the same magnitude.

zz* = |z|2 =|z*|2 = a2 + b2 (3.2.6)

Use the conjugate to eliminate j in the denominator of a ratio.

3.3 Oscillatory behavior

Inserting 2π into the argument is necessary because we are attempting to

x(t) = B sin (ωt ) and y(t) = C sin (ωt + ϕ ) (3.3.2)

Representing an oscillation with a phase angle is quite useful, but on

However, most calculators and spreadsheets will process (3.2.2) to give

Calculators and spreadsheets tend to work between -90° and +90°.

ϕ = ϕ calculator − 180o , 0 ≤ ϕ calculator ≤ 90o (3.4.2)

By the way, the conversion factor between degrees and radians is

3.5 First-order, linear, variable-coefficient ODE

In writing (3.5.1) we have arranged the coefficient functions to isolate the

and integrating to find

3.6 First-order ODE, special case for process control applications

or for initial time at zero,

3.7 First-order ODE, special case for no disturbance (forcing function)

3.8 First-order ODE, special case for missing dependent variable

Separable equations are convenient to solve by straightforward function

integration. The surge tank of Chapters 1 and 2 was described by a

separable equation of form (3.8.1).

3.9 First-order ODE, delayed disturbance

0. Then from (3.5.3)

K  d p(t)(0) p(t)(1)  p(0) y(0)

3.10 Second-order, linear, constant-coefficient ODE