You are on page 1of 179

Lesson 1: Processes and Systems

1.0 context and direction

Process control is an application area of chemical engineering - an
identifiable specialty for the ChE. It combines chemical process
knowledge (how physics, chemistry, and biology work in operating
equipment) and an understanding of dynamic systems, a topic important to
many fields of engineering. Thus study of process control allows
chemical engineers to span their own field, as well as form a useful
acquaintance with allied fields. Practitioners of process control find their
skills useful in design, operation, and troubleshooting - major categories of
chemical engineering practice.

Process control, like any coherent topic, is an integrated body of

knowledge - it hangs together on a multidimensional framework, and
practitioners draw from many parts of the framework in doing their work.
Yet in learning, we must receive information in sequence - following a
path through multidimensional space. It is like entering a large building
with unlighted rooms, holding a dim flashlight and clutching a vague map
that omits some of the stairways and passages. How best to learn one’s
way around?

In these lessons we will attempt to move through a significant portion of

the structure - say, half a textbook - in about two weeks. Then we will
repeat the journey several times, each time inspecting the rooms more
thoroughly. By this means we hope to gain, from the start, a sense of
doing an entire process control job, as well as approach each new topic in
the context of a familiar path.

1.1 the job we will do, over and over

We encounter a process, learn how it behaves, specify how we wish to
control it, choose appropriate equipment, and then explore the behavior
under control to see if we have improved things.

1.2 introducing a simple process

A large tank must be filled with liquid from a supply line. One operator
stands at ground level to operate the feed valve. Another stands on the
tank, gauging its level with a dipstick. When the tank is near full, the stick
operator will instruct the other to start closing the valve. Overfilling can
cause spills, but underfilling will cause later process problems.
Lesson 1: Processes and Systems

To learn how the process works, we write an overall material balance on

the tank.

UV UFi (1.2-1)

The tank volume V can be expressed in terms of the liquid level h. The
inlet volumetric flow rate Fi may vary with time due to supply pressure
fluctuations and valve manipulations by the operator. The liquid density
depends on the temperature, but will usually not vary significantly with
time during the course of filling. Thus (1.2-1) becomes

A Fi ( t ) (1.2-2)

We integrate (1.2-2) to find the liquid level as a function of time.

1 t
A ³0
h h ( 0)  Fi ( t )dt (1.2-3)

1.3 planning a control scheme

Clearly the liquid level h is important, and we will call it the controlled
variable. Our control objective is to bring h quickly to its target value hr
and not exceed it. (To be realistic, we would specify allowable limits ± įh
on hr.) We will call the volumetric flow Fi the manipulated variable,
because we adjust it to achieve our objective for the controlled variable.

The existing control scheme is to measure the controlled variable via

dipstick, decide when the controlled variable is near target, and instruct
Lesson 1: Processes and Systems
the valve operator to change the manipulated variable. The scheme suffers
x delay in measurement. Overfilling can occur if the stick operator
cannot complete the measurement in time.
x performance variations. Both stick and valve operators may vary in
attentiveness and speed of execution.
x resources required. There are better uses for operating personnel.
x unsafe conditions. There is too much potential for chemical exposure.

A new scheme is proposed: put a timer on the valve. Calculate the time
required for filling from (1.2-3). Close the valve when time has expired.

The timing scheme would no longer require an operator to be on the tank

top, and with a motor-driven valve actuator the entire operation could be
directed from a control room. These are indeed improvements. However,
the timing scheme abandons a crucial virtue of the existing scheme: by
measuring the controlled variable, the operators can react to unexpected
disturbances, such as changes in the filling rate. Using knowledge of the
controlled variable to motivate changes to the manipulated variable is a
fundamental control structure, known as feedback control. The proposed
timing scheme has no feedback mechanism, and thus cannot accommodate
changes to h(0) and Fi(t) in (1.2-3).

An alternative is to build on the feedback already inherent in the two-

operator scheme, but to improve its operation. We propose an automatic
controller that behaves according to the following controller algorithm:

h  h near Fi Fmax
hr  h (1.3-1)
h ! h near Fi Fmax
h r  h near

Algorithm (1.3-1) is an idealization of what the operators are already

doing: filling occurs at maximum flow until the level reaches a value hnear.
Beyond this point, the flow decreases linearly, reaching zero when h
reaches the target hr. The setting of hnear may be adjusted to tune the
control performance.

1.4 choosing equipment

We need a sensor to replace the dipstick, a valve actuator to replace the
valve operator, and a controller mechanism to replace the stick operator.
We imagine a buoyant object floating on the liquid surface. The float is
linked to a lever that drives the valve stem. When the liquid level is low,
the float rests above it on a structure so that the valve is fully open.
Lesson 1: Processes and Systems

Content removed due to copyright restrictions.

(To see a cut-away diagram of a toilet, go to

1.5 process behavior under automatic control

Typically these things work quite well. We predict its performance by
combining our process model (1.2-2) with the controller algorithm (1.3-1),
which eliminates the manipulated variable between the equations. We
take the simple case in which Fmax does not vary during filling due to
pressure fluctuations, etc. For h less than hnear,

A Fmax h ( 0) known
dt (1.5-1)
h h (0)  max t

Equation (1.5-1) can be used to calculate tnear, the time at which h reaches
hnear. For h greater than hnear,

dh hr  h
A Fmax h(t near ) h near
dt h r  h near
ª  h r (t  t near ) º
h h r  h r  h near exp « »
«¬ t fill h r  h near »¼
Lesson 1: Processes and Systems

where the parameter tfill is the time required for the level to reach hr at
flow Fmax, starting from an empty tank.

Ah r
t fill (1.5-3)

The plot shows the filling profile from h(0) = 0.10hr with several values of
hnear/hr. Certainly the filling goes faster if the flow can go instantaneously
from Fmax to zero at hr; however this will not be practical, so that hnear will
be less than hr.

hnear/hr = 0.95
1 0.75




0 0.2 0.4 0.6 0.8 1 1.2 1.4

1.6 defining ‘system’

In Section 1.2, we introduced a process - a tank with feed piping - whose
inventory varied 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.

'Process' is a good notion, important to chemical engineers. 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 input and output signals. 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
Lesson 1: Processes and Systems
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" or “stimuli”; outputs are "things that are affected" or “responses”.

inputs system outputs

(causes) (responses)

To approach the problem of controlling our filling process in Section 1.3,

we thought of it in system terms: the primary output was the liquid level h
-- not a stream, certainly, but an important response variable of the system
-- and inlet stream Fi was an input. And peculiar as it first seems, if the
tank had an outlet flow Fo, it would also be an input signal, because it
influences the liquid level, just as does Fi.

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 (Fo as an outlet
stream) to write the material balance and make fluid mechanics
calculations. View it as a system (Fo as an input) to analyze the dynamic
behavior implied by that material balance and make control calculations.

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. 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.

1.7 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. In
some cases it may help to 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.

inputs outputs
Lesson 1: Processes and Systems

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.

1.8 the system of single-loop feedback control

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 a repertoire of relatively
few control schemes (relationships between process and controller)
suffices for myriad process applications. Using the terms we defined in
Section 1.3, we represent a control scheme called single-loop feedback
control in this fashion:


other inputs other


final control sensor


set point

Figure 1.8-1 The single-loop feedback control system and its


We will see this structure repeatedly. 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 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 (often a pressure, temperature, flow rate, liquid level, or
composition) and provide a suitable sensor to measure it. From the inputs,
we choose a manipulated variable (often a flow rate) and install an
appropriate final control element (often a valve). The measurement is fed
to the controller, which decides how to adjust the manipulated variable to
keep the controlled variable at the desired condition: the set point. The
Lesson 1: Processes and Systems
other inputs are potential disturbances that affect the controlled variable,
and so require action by the controller.

1.9 conclusion
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 a single-loop feedback structure and
arrive at a new dynamic system that responds in different ways to the
inputs. If we do our job well, it responds in better ways, so to justify all
the trouble.
Lesson 2: Mathematics Review
2.0 context and direction
Imagine a system that varies in time; we might plot its output vs. time. A
plot might imply an equation, and the equation is usually an ODE
(ordinary differential equation). Therefore, we will review the math of the
first-order ODE while emphasizing how it can represent a dynamic
system. We examine how the system is affected by its initial condition
and by disturbances, where the disturbances may be non-smooth, multiple,
or delayed.

2.1 first-order, linear, variable-coefficient ODE

The dependent variable y(t) depends on its first derivative and forcing
function x(t). When the independent variable t is t0, y is y0.

a (t)  y( t ) Kx ( t ) y( t 0 ) y0 (2.1-1)

In writing (2.1-1) we have arranged a coefficient of +1 for y. Therefore

a(t) must have dimensions of independent variable t, and K has
dimensions of y/x. We solve (2.1-1) by defining the integrating factor p(t)

p(t ) exp ³ (2.1-2)
a(t )

Notice that p(t) is dimensionless, as is the quotient under the integral. The

p( t 0 ) y( t 0 ) K p( t ) x ( t )
p( t ) t³0 a ( t )
y( t )  dt (2.1-3)
p( t )

comprises contributions from the initial condition y(t0) and the forcing
function Kx(t). These are known as the homogeneous (as if the right-hand
side were zero) and particular (depends on the right-hand side) 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).

2.2 first-order ODE, special case for process control applications

The independent variable t will represent time. For many process control
applications, a(t) in (2.1-1) will be a positive constant; we call it the time
constant W.

W  y( t ) Kx ( t ) y( t 0 ) y0 (2.2-1)

The integrating factor (2.1-2) is

Lesson 2: Mathematics Review
dt t
p( t ) exp ³ e W

and the solution (2.1-3) becomes

 t  t 0 K t W t W
y( t ) y 0e W
 e ³ e x ( t )dt (2.2-3)
W t0

The initial condition affects the system response from the beginning, but
its effect decays to zero according to the magnitude of the time constant -
larger time constants represent slower decay. If not further disturbed by
some x(t), the first order system reaches equilibrium at zero.

However, most practical systems are disturbed. K is a property of the

system, called the gain. By its magnitude and sign, the gain influences
how strongly y responds to x. The form of the response depends on the
nature of the disturbance.

Example: suppose x is a unit step function at time t1. Before we proceed

formally, let us think intuitively. From (2.2-3) we expect the response y to
decay toward zero from IC y0. At time t1, the system will respond to being
hit with a step disturbance. After a long time, there will be no memory of
the initial condition, and the system will respond only to the disturbance
input. Because this is constant after the step, we guess that the response
will also become constant.

Now the math: from (2.2-3)

 t  t 0 K t W t W
y( t ) y0e W
 e ³ e U( t  t1 )dt
W t0
 t  t 0  t  t1
§ ·
y 0e W
 KU( t  t1 )¨1  e W
© ¹

Figure 2.2-1 shows the solution. Notice that the particular solution makes
no contribution before time t1. The initial condition decays, and with no
disturbance would continue to zero. At t1, however, the system responds
to the step disturbance, approaching constant value K as time becomes
large. This immediate response, followed by asymptotic approach to the
new steady state, is characteristic of first-order systems. Because the
response does not track the step input faithfully, the response is said to lag
behind the input; the first-order system is sometimes called a first-order
Lesson 2: Mathematics Review



0 1 2 3 4 5 6

2y 0




0t 0 1 2t 1 3 4 5 6

Figure 2.2-1 first-order response to initial condition and step


2.3 piecewise integration of non-smooth disturbances

The solution (2.2-3) is applied over succeeding time intervals, each
featuring an initial condition (from the preceding interval) and disturbance

­  t  t 0 t
K t W t W
° 0y ( t ) e W

e ³t e x (t )dt t 0  t  t1
° 0

°  t  t1
K t t
y( t ) ® y ( t 1 )e W
 e W ³ e W x ( t )dt t1  t  t 2 (2.3-1)
° W t1

Example: suppose
Lesson 2: Mathematics Review
y x y ( 0) 0
­0 0  t 1 (2.3-2)
x ®2 t  1 1  t  2
°0 2t

In this problem, variables t, x, and y should be presumed to have

appropriate, if unstated, units; in these units, both gain and time constant
are of magnitude 1. From (2.3-1),

­0 0  t 1
y( t )

®2 t  2  e
 t 1
1 t  2 (2.3-3)
°2e 1e  t  2 2t

With a zero initial condition and no disturbance, the system remains at

equilibrium until the ramp disturbance begins at t = 1. Then the output
immediately rises in response, lagging behind the linear ramp. At t = 2,
the disturbance ceases, and the output decays back toward equilibrium.
Lesson 2: Mathematics Review




0 1 2 3 4 5 6







0 1 2 3 4 5 6

2.4 multiple disturbances and superimposition

Systems can have more than one input. Consider a first-order system with
two disturbance functions.

W  y( t ) K1x1 ( t )  K 2 x 2 ( t ) y( t 0 ) y0 (2.4-1)

Applying (2.2-3) and distributing the integral across the disturbances, we

find that the effects of the disturbances on y are additive.

 t  t 0 K t
t K t t
y( t ) y0e W
 1 e W ³ e W x1 ( t )dt  2 e W ³ e W x 2 ( t )dt (2.4-2)
W t0
W t0

This additive behavior is a happy characteristic of linear systems. Thus

another way to view problem (2.4-1) is to decompose it into component
problems. That is, define

y y H  y1  y 2 (2.4-3)
Lesson 2: Mathematics Review
and write (2.4-1) in three equations. We put the initial condition with no
disturbances, and each disturbance with a zero initial condition.

dy H
W  yH (t) 0 yH (t 0 ) y0
W 1  y1 ( t ) K1x1 ( t ) y1 ( t 0 ) 0 (2.4-4)
dy 2
W  y 2 (t) K 2 x 2 (t) y2 (t 0 ) 0

Equations and initial conditions (2.4-4) can be summed to recover the

original problem specification (2.4-1). The solutions are

 t  t 0
y H (t ) y 0e
K1  t W t W
y1 ( t ) e ³ e x1 ( t )dt (2.4-5)
W t0
K 2 t W t W
y 2 (t ) e ³ e x 2 ( t )dt
W t0

and of course these solutions can be added to recover original solution

(2.4-2). Thus we can view the problem of multiple disturbances as a
system responding to the sum of the disturbances, or as the sum of
responses from several identical systems, each responding to a single

Example: consider

1 dy 1 3
 y U ( t  1)  U ( t  3) y ( 0) 2 (2.4-6)
4 dt 4 4

We first place the equation in standard form, in which the coefficient of y

is +1.

y 3U ( t  1)  4 U ( t  3) y ( 0) 2 (2.4-7)

Equation (2.4-7) shows us that the time constant is 1, and that the system
responds to the first disturbance with a gain of 3, and to the second with a
gain of -4. The solution is

y 2e  t  3U( t  1) 1  e  ( t 1)  4U( t  3) 1  e  ( t 3) (2.4-8)

Lesson 2: Mathematics Review
In Figure 2.4-1, the individual solution components are plotted as solid
traces; their sum, which is the system response, is a dashed trace. Notice
how the first-order lag responds to each new disturbance as it occurs.



0 2 4 6 8

0 2 4 6 8

Figure 2.4-1 first-order response to multiple disturbances

Writing the step functions explicitly in solution (2.4-8) emphasizes that

particular disturbances do not influence the solution until the time of their
occurrence. For example, if they were omitted, some deceptively correct
but inappropriate rearrangement would lead to errors.

2e  t  3 1  e  ( t 1)  4 1  e  ( t 3)
2e  t  3  3e ( t 1)  4  4e ( t 3) (do not do this!) (2.4-9)
t  ( t 1)  ( t 3)
1  2e  3e  4e

This notation at least implies that two of the exponential functions have
delayed onsets. However, further correct-but-inappropriate rearrangement
makes things even worse.
Lesson 2: Mathematics Review
y 1  2e  t  3e  ( t 1)  4e  ( t 3)
1  2e  t  3e1e  t  4e 3e  t (do not do this!) (2.4-10)

1  2  3e1  4e 3 e  t
The incorrect solutions are plotted with (2.4-8) in Figure 2.4-2. Equation
(2.4-9) has become discontinuous - the response takes non-physical leaps
at the onset of each new disturbance. Equation (2.4-10) has lost all
dependence on the disturbances and decays from a non-physical initial
condition. Even with the mistakes, both incorrect solutions lead to the
correct long-term condition.




0 2 4 6 8

solution eq 2.4.9 eq 2.4.10

Figure 2.4-2 comparison of correct and incorrect solutions

2.5 delayed response to disturbances

Consider a system that reacts to a disturbance, but only after some
intervening time interval ș has passed. That is

W  y( t ) Kx ( t  T) y( t 0 ) y0 (2.5-1)

Equation (2.5-1) shows the dependence of y, at any time t, on the value of

x at earlier time t - ș. The solution is written directly from (2.2-3).

 t  t 0 t
K t t
y( t ) y0e W
 e W ³ e W x ( t  T)dt (2.5-2)
W t0

We must integrate the disturbance considering the time delay. Take as an

example a disturbance x(t) occurring at time t1. The plot shows the
Lesson 2: Mathematics Review
disturbance, as well as the disturbance as the system experiences it, which
begins at time t1 + ș. We could express this disturbance-as-experienced as
some new function x1(t), occurring at time t1 + ș.

disturbance as
experienced by
as it occurs

x(t) x(t - ș) = x1(t)

t0 t1 t1 + ș


t0 - ș t1 - ș t1

Alternatively, we could define a new time variable

[ tT (2.5-3)

and write the input as x(ȟ). The integral in (2.5-2) becomes, then,

t t [
t t [T

³ e x (t  T)dt
³ e x1 (t )dt
t 0 T
x ([)d[ (2.5-4)

Therefore, solution (2.5-2) becomes

 t  t 0 [
K t W T W [
y( t ) y0e W
 e e ³ e W x ([)d[ (2.5-5)
W t 0 T

Example: consider a step disturbance at time t = 2 that affects the system

3 time units later.

 y x ( t  3) y(0) 0
dt (2.5-6)
x ( t ) U( t  2)

Using (2.5-5)
Lesson 2: Mathematics Review

y e  t e 3 ³ e [ U([  2)d[
0 3

ª2 [
e  t e 3 « ³ e [ U([  2)d[  ³ e [ U([  2)d[»
«¬03 2 »¼
e  t e 3 ª U ([  2)e [ º

«¬ 2»¼
 t 3 t 3
U( t  3  2)e e e  e 2 > @
U( t  5) e t  3 t  3
e 2 t 3
U( t  5)>1  e  ( t 5 )
Figure 2.5-1 shows that a typical first-order lag step response occurs 3
time units after being disturbed at t = 2.



0 2 4 6 8 10





0 2 4 6 8 10

Figure 2.5-1 step response of first order system with dead time

The time delay in responding to a disturbance is often called dead time.

Dead time is different from lag. Lag occurs because of the combination of
y and its derivative on the left-hand side of the equation. Dead time
Lesson 2: Mathematics Review
occurs because of a time delay in processing a disturbance on the right-
hand side.

2.6 conclusion
Please become comfortable with handling ODEs. View them as systems;
identify their inputs and outputs, their gains and time parameters.
Lesson 3: The Blending Tank
3.0 context and direction
A particularly simple process is a tank used for blending. Just as promised
in Section 1.1, we will first represent the process as a dynamic system and
explore its response to disturbances. Then we will pose a feedback control
scheme. We will briefly consider the equipment required to realize this
control. Finally we will explore its behavior under control.


3.1 math model of a simple continuous holding tank

Imagine a process stream comprising an important chemical species A in
dilute liquid solution. It might be the effluent of some process, and we
might wish to use it to feed another process. Suppose that the solution
composition varies unacceptably with time. We might moderate these
swings by holding up a volume in a stirred tank: intuitively we expect the
changes in the outlet composition to be more moderate than those of the
feed stream.

F, CAi

F, CAo

volume V

Our concern is the time-varying behavior of the process, so we should

treat our process as a dynamic system. To describe the system, we begin
by writing a component material balance over the solute.

VC Ao FC Ai  FC Ao (3.1-1)

In writing (3.1-1) we have recognized that the tank operates in overflow:

the volume is constant, so that changes in the inlet flow are quickly
duplicated in the outlet flow. Hence both streams are written in terms of a
single volumetric flow F. Furthermore, for now we will regard the flow as
constant in time.

Balance (3.1-1) also represents the concentration of the outlet stream, CAo,
as the same as the average concentration in the tank. That is, the tank is a
perfect mixer: the inlet stream is quickly dispersed throughout the tank
volume. Putting (3.1-1) into standard form,
Lesson 3: The Blending Tank

V dC Ao
 C Ao C Ai (3.1-2)
F dt

we identify a first-order dynamic system describing the response of the

outlet concentration CAo to disturbances in the inlet concentration CAi.
The speed of response depends on the time constant, which is equal to the
ratio of tank volume and volumetric flow. Although both of these
quantities influence the dynamic behavior of the system, they do so as a
ratio. Hence a small tank and large tank may respond at the same rate, if
their flow rates are suitably scaled.

System (3.1-2) has a gain equal to 1. This means that a sustained

disturbance in the inlet concentration is ultimately communicated fully to
the outlet.

Before solving (3.1-2) we specify a reference condition: we prefer that CAo

be at a particular value CAo,r. For steady operation in the desired state,
there is no accumulation of solute in the tank.

V dC Ao
0 C Ai,r  C Ao,r (3.1-3)
F dt r

Thus, as expected, steady outlet conditions require a steady inlet at the

same concentration; call it CA,r. Let us take this reference condition as an
initial condition in solving (3.1-2). The solution is

W t
t e t
C Ao ( t ) C A ,r e W

W ³e
C Ai ( t )dt (3.1-4)

where the time constant is

W (3.1-5)

Equation (3.1-4) describes how outlet concentration CAo varies as CAi

changes in time. In the next few sections we explore the transient
behavior predicted by (3.1-4).

3.2 response of system to steady input

Suppose inlet concentration remains steady at CA,r. Then from (3.1-4)
Lesson 3: The Blending Tank
W t
t e t
C Ao C A ,r e W
 C A ,r We W
W 0 (3.2-1)
§¨ e  1·¸
t t t
C A ,r e W
 C A ,r e W W
C A ,r
© ¹

Equation (3.2-1) merely confirms that the system remains steady if not

3.3 leaning on the system - response to step disturbance

Step functions typify disturbances in which an input variable moves
relatively rapidly to some new value and remains there. Suppose that
input CAi is initially at the reference value CA,r and changes at time t1 to
value CA1. Until t1 the outlet concentration is given by (3.2-1). From the
step at t1, the outlet concentration begins to respond.

 ( t  t1 ) W t
e t
C Ao C A ,r e W
 C A1 We W
t ! t1
W t1

 ( t  t1 )
§ tW t t1
C A ,r e W
 C A1e ¨e  e W ¸ W
© ¹
 ( t  t1 )
§  ( t  t1 )
W ·
C A ,r e W
 C A1 ¨1  e ¸
© ¹

In Figure 3.3-1, CA,r = 1 and CA1 = 0.8 in arbitrary units; t1 has been set
equal to IJ. At sufficiently long time, the initial condition has no influence
and the outlet concentration becomes equal to the new inlet concentration.
After time equal to three time constants has elapsed, the response is about
95% complete – this is typical of first-order systems.

In Section 3.1, we suggested that the tank would mitigate the effect of
changes in the inlet composition. Here we see that the tank will not
eliminate a step disturbance, but it does soften its arrival.
Lesson 3: The Blending Tank



0 1 2 3 4 5 6





0 1 2 3 4 5 6

Figure 3.3-1 first-order response to step disturbance

3.4 kicking the system - response to pulse disturbance

Pulse functions typify disturbances in which an input variable moves
relatively rapidly to some new value and subsequently returns to normal.
Suppose that CAi changes to CA1 at time t1 and returns to CA,r at t2. Then,
drawing on (3.2-1) and (3.3-1),

°C 0  t  t1
° A ,r
°  ( t  t1 )
§  ( t  t1 )
W ·
C Ao ®C A ,r e W
 C A1 ¨1  e ¸ t1  t  t 2 (3.4-1)
° © ¹
°ª § ·º ( t  t 2 ) W § W ·
 ( t 2  t1 )  ( t 2  t1 ) ( t  t 2 )
° «C A , r e W
 C A1 ¨1  e W
¸» e  C A , r ¨1  e ¸ t2  t
¯¬ © ¹¼ © ¹

In Figure 3.4-1, CA,r = 0.6 and CA1 = 1 in arbitrary units; t1 has been set
equal to IJ and t2 to 2.5IJ. We see that the tank has softened the pulse and
reduced its peak value. A pulse is a sequence of two counteracting step
changes. If the pulse duration is long (compared to the time constant IJ),
Lesson 3: The Blending Tank
the system can complete the first step response before being disturbed by
the second.



0 1 2 3 4 5 6



0 1 2 3 4 5 6

Figure 3.4-1 first-order response to pulse disturbance

3.5 shaking the system - response to sine disturbance

Sine functions typify disturbances that oscillate. Suppose the inlet
concentration varies around the reference value with amplitude A and
frequency Ȧ, which has dimensions of radians per time.

C Ai C A , r  A sin Zt (3.5-1)

From (3.1-4),

AZW  t W
C Ao C A ,r  e 

sin Zt  tan 1  ZW (3.5-2)
1 Z W
2 2
1 Z W
2 2
Lesson 3: The Blending Tank
Solution (3.5-2) comprises the mean value CA,r, a term that decays with
time, and a continuing oscillation term. Thus, the long-term system
response to the sine input is to oscillate at the same frequency Ȧ. Notice,
however, that the amplitude of the output oscillation is diminished by the
square-root term in the denominator. Notice further that the outlet
oscillation lags the input by a phase angle tan-1(-ȦIJ).

In Figure 3.5-1, CA,r = 0.8 and A = 0.5 in arbitrary units; ȦIJ has been set
equal to 2.5 radians, and IJ to 1 in arbitrary units. The decaying portion of
the solution makes a negligible contribution after the first cycle. The
phase lag and reduced amplitude of the solution are evident; our tank has
mitigated the inlet disturbance.


input and response





0 2 4 6 8

input decaying part continuing part solution

Figure 3.5-1 first-order response to sine disturbance

3.6 frequency response and the Bode plot

The long-term response to a sine input is the most important part of the
solution; we call it the frequency response of the system. We will
examine the frequency response for an abstract first order system.
(Because we wish to focus on the oscillatory response, we will write (3.6-
1) so that x and y vary about zero. The effect of a non-zero bias term can
be seen in (3.5-1) and (3.5-2).)
Lesson 3: The Blending Tank
system : IJ  y Kx
input : x A sin(Zt )
freq resp : y fr y fr sin(Zt  I)
KA (3.6-1)
amplitude : y fr
1 Z W2 2

phase angle : I tan 1 (ZW)

y fr K
amplitude ratio : R A
x 1  Z2 W 2

The frequency response is a sine function, characterized by an amplitude,

frequency, and phase angle. The amplitude and phase angle depend on
system properties (IJ and K) and characteristics of the disturbance input (Ȧ
and A). It is convenient to show the frequency dependence on a Bode
plot, Figure 3.6-1.

The Bode plot abscissa is Ȧ in radians per time unit; the scale is
logarithmic. The frequency may be normalized by multiplying by the
system time constant. Thus plotting Ȧ is good for a particular system;
plotting ȦIJ is good for systems in general.

The upper ordinate is the amplitude ratio, also on logarithmic scale. RA is

often normalized by dividing by the system gain K. The lower ordinate is
the phase angle, in degrees on a linear scale.

In Figure 3.6-1, the coordinates have been normalized to depict first-order

systems in general; the particular point represents conditions in the
example of Section 3.5.

For a first order system, the normalized amplitude ratio decreases from 1
to 0 as frequency increases. Similarly, the phase lag decreases from 0 to
-90º. Both these measures indicate that the system can follow slow inputs
faithfully, but cannot keep up at high frequencies.

Another way to think about it is to view the system as a low-pass filter:

variations in the input signal are softened in the output, particularly for
high frequencies.

The slope of the amplitude ratio plot approaches zero at low frequency;
the high frequency slope approaches -1. These two asymptotes intersect at
the corner frequency, the reciprocal of the system time constant. At the
corner frequency, the phase lag is -45º.
Lesson 3: The Blending Tank
Zcorner (3.6-2)

amplitude ratio/gain


0.01 0.1 1 10 100
phase angle (deg)

0.01 0.1 1 10 100
ȦIJ (radians)

Figure 3.6-1: Bode plot for first-order system

3.7 stability of a system

If we disturb our 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
x a ramp disturbance is not fair – even stable systems can get into
trouble if the input keeps rising.
x a stable system should handle a step change in input, ultimately
coming to some new steady state. (We must be realistic, however.
If the 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.)
Lesson 3: The Blending Tank
x 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'.)

The solutions for the typical bounded step, pulse, and sine disturbances,
given in Sections 3.3 through 3.5, show no terms that grow with time, so
long as the time constant IJ is a positive value. For these categories of
bounded input, at least, a first-order system appears to be stable. We will
need to examine stability again when we introduce automatic control to
our process.

3.8 concentration control in a blending tank

In Section 3.1 we described how variations in stream composition could
be moderated by passing the stream through a larger volume - a holding
tank. Let us be more ambitious and seek to control the outlet composition:
we add a small inlet stream Fc of concentrated solution to the tank. This
will allow us to adjust the composition in response to disturbances.

Fc, CAc

F, CAi

F, CAo

volume V

Our analysis begins as in Section 3.1 with a component material balance.

FC Ai  Fc C Ac  F  Fc C Ao
VC Ao (3.8-1)

As before, we place (3.8-1) in standard form (response variable on the left

with a coefficient of +1).

V C Ac
F dC Ao  C 1
C  F Fc (3.8-2)
F Fc Ai F
1  c dt 1 1 c

Notice that our equation coefficients each contain the input variable Fc.
Notice, as well, that for dilute CAo and concentrated CAc stream Fc
Lesson 3: The Blending Tank
(however it may vary) will not be very large in comparison to the main
flow F. If this is the case, we may be justified in making an engineering
approximation: neglecting the ratio Fc/F in comparison to 1. Thus

V dC Ao C Ac
 C Ao C Ai  Fc (3.8-3)
F dt F

Now we have a linear first-order system. Comparison with (3.1-2) shows

the same time constant V/F and the same unity gain for inlet concentration
disturbances. There is a new input Fc, whose influence on CAo (i.e., gain)
increases with high concentration CAc and decreases with large
throughflow F.

3.9 use of deviation variables in solving equations

In process control applications, we usually have some desired operating
condition. We now write system model (3.8-3) at the target steady state.
All variables are at reference values, denoted by subscript r.

C Ac
C Ao,r C Ai, r  Fc ,r (3.9-1)

We recognize that deviations from these reference conditions represent

errors to be corrected. Hence we recast our system description (3.8-3) in
terms of deviation variables; we do this by subtracting (3.9-1) from (3.8-

V d C Ao  C Ao,r
 C Ao  C Ao,r C  C Ai,r  Fc  Fc,r
C Ac
F dt F
V dC 'Ao C
 C 'Ao C 'Ai  Ac Fc'
F dt F

where we indicate a deviation variable by a prime superscript. The target

condition of a deviation variable is zero, indicating that the process is
operating at desired conditions. Using deviation variables

x makes conceptual sense for process control because they indicate

deviations from desired states
x makes the mathematical descriptions simpler

Thus we shall use deviation variables for derivations and modeling. For
doing process control (computing valve positions, e.g.) we will return to
the physical variables. We can recover the physical variable by adding its
deviation variable to its reference value. For example,

C Ao ( t ) C Ao,r  C'Ao ( t ) (3.9-3)

Lesson 3: The Blending Tank

where we emphasize the variables that are time-varying.

3.10 integration from zero initial conditions

As a rule, we will presume that our systems are initially at the reference
condition. That is, the initial conditions for our differential equations are
zero. Integrating (3.9-2) we find

t t
W t W t
e t C Ac t W ' e
³ e C ( t )dt 
W F ³0
' '
C e Fc ( t )dt (3.10-1)
W 0
Ao Ai

Equation (3.10-1) shows how the outlet composition deviates from its
desired value CAo,r under disturbances to inlet composition CAi and the
flow rate of the concentrated makeup stream Fc, where both of these are
also expressed as deviations from reference values. Equation (3.10-1) is
analogous to (3.1-4) for the simpler holding tank.

3.11 response to step changes

Proceeding as in Section 3.3, we presume a step in inlet composition of
ǻCAi at time t1 and of ǻFc in makeup flow at time t2.

t t t t
e t e C Ac t
U( t  t1 )'C Ai ³ e W dt 
F t³2
C '
U( t  t 2 )'Fc e W
§  ( t  t1 )
W ·
C Ac § ( t  t 2 )
W ·
U( t  t1 )'C Ai ¨1  e ¸  U( t  t 2 )'Fc ¨1  e ¸
© ¹ F © ¹

CAoƍ exhibits a first-order response to each of these step inputs.

Example: try these numbers:

V = 6 m3 t1 = 0 s
F = 0.02 m3 s-1 ǻCAi = 1 kg m-3
Fcs = 10-4 m3 s-1 t2 = 120 s
CAis = 8 kg m-3 ǻFc = -5×10-5 m3 s-1
CAos = 10 kg m-3
CAc = 400 kg m-3

First, verify the steady-state material balance (3.9-1) for the desired

kg kg kg (0.0001)
(10) 3 (8) 3  (400) 3 s
m m m (0.02) m3
Lesson 3: The Blending Tank
(Notice that the exact steady-state balance, derived from (3.8-2), is
satisfied to within 1%, so that our approximation in deriving (3.8-3)
appears to be reasonable.) The time constant for our process is

V ( 6) m 3
W 300 s (3.11-3)
F (0.02) m3

Substituting values into (3.11-1) we obtain

kg § ( t 0 )
300 ·
kg (5 u 10 5 ) §  ( t 120 )
300 ·
C 'Ao U ( t  0)(1) 3 ¨
1  e ¸  U ( t  120 )( 400 ) ¨1  e ¸
m © ¹ (0.02) © ¹
300 ·
(1) 3 §¨1  e 300 ·¸  U ( t  120)(1) 3 §¨1  e
kg t kg  ( t 120 )
m © ¹ m © ¹

where t must be computed with units of seconds. In Figure 3.11-1, we can

see that the reduction in make-up flow at 120 s compensates for the earlier
increase in inlet composition. Now we are ready to consider control.


3.12 developing a control scheme for the blending tank

A control scheme is the plan by which we intend to control a process. A
control scheme requires:

1) specifying control objectives, consistent with the overall objectives

of safety for people and equipment, environmental protection,
product quality, and economy
2) specifying the control architecture, in which various of the system
variables are assigned to roles of controlled, disturbance, and
manipulated variables, and their relationships specified
3) choosing a controller algorithm
4) specifying set points and limits

3.13 step 1 - specify a control objective for the process

Our control objective is to maintain the outlet composition at a constant
value. Insofar as the process has been described, this seems consistent
with the overall objectives.

3.14 step 2 - assign variables in the dynamic system

The controlled variable is clearly the outlet composition. The inlet
composition is a disturbance variable: we have no influence over it, but
must react to its effects on the controlled variable. We do have available a
variable that we can manipulate: the make-up flow rate.
Lesson 3: The Blending Tank
We specify feedback control as our control architecture: departure of the
controlled variable from the set point will trigger corrective action in the
manipulated variable. Said another way, we manipulate make-up flow to
control outlet composition.

make-up flow deviation (m3 s -1)

inlet comp deviation (kg m-3)

1.5 0.000075
1 0.000050
0.5 0.000025
0 0.000000
-0.5 -0.000025
-1 -0.000050
-1.5 -0.000075
-100 0 100 200 300 400 500 600 700 800

inlet composition make-up flow

outlet composition deviation (kg m )







-100 0 100 200 300 400 500 600 700 800

Figure 3.11-1 outlet composition response to opposing step inputs

3.15 step 3 - introduce proportional control for our process

The controller algorithm dictates how the manipulated variable is to be
adjusted in response to deviations between the controlled variable and the
set point. We will introduce a simple and plausible algorithm, called
Lesson 3: The Blending Tank
proportional control. This algorithm specifies that the magnitude of the
manipulation is directly proportional to the magnitude of the deviation.

Fc  Fbias K gain C Ao ,setpt  C Ao (3.15-1)

In algorithm (3.15-1) the controlled variable CAo is subtracted from the set
point. (Subtracting from the set point, rather than the reverse, is a
convention.) Any non-zero result is an error. The error is multiplied by
the controller gain Kgain. Their product determines the degree to which
manipulated variable Fc differs from Fbias, its value when there is no error.
The gain may be adjusted in magnitude to vary the aggressiveness of the
controller. Large errors and high gain lead to large changes in Fc.

We must consider the direction of the controller, as well as its strength:

should the outlet composition exceed the set point, the make-up flow must
be reduced. Algorithm (3.15-1) satisfies this requirement if controller gain
Kc is positive.

3.16 step 4 - choose set points and limits

The set point is the target operating value. For many continuous processes
this target rarely varies. In our blending tank example, we may always
desire a particular outlet concentration. In other cases, such as a process
that makes several grades of product, the set point might be varied from
time to time. In batch processes, moreover, the set point can show
frequent variation because it provides the desired trajectory for the time-
varying process conditions.

Several sorts of limits must be considered in control engineering:

safety limits: if a variable exceeds these limits, a hazard exists. Examples

are explosive composition limits on mixtures, bursting pressure in a
vessel, temperatures that trigger runaway reactions.

These limits are determined by the process, and the control scheme must
be designed to abide by them.

expected variation: it is necessary to estimate how much variation might

be expected in a disturbance variable. This estimate is the basis for
specifying the strength of the manipulated variable response. In Section
3.11, our system model (based on the material balance) showed us how
much variation in make-up flow, at specified make-up composition, was
required to compensate for a particular change in the inlet composition.

These limits are determined by the process and its environment. No

amount of controller design can compensate for a manipulated variable
that is unequal to the disturbance task.
Lesson 3: The Blending Tank

tolerable variation: ideally the controlled variable would never deviate

from the set point. This, of course, is unrealistic; in practice some
variation must be tolerated, because
x obtaining enough information on the process and disturbance is
usually impossible, and in any case too expensive.
x exerting sufficient manipulative strength to suppress variation in the
control variable might be expected to require large variations in the
manipulated variable, which can cause problems elsewhere in the

Tolerable limits are determined by the safety limits, above, and then an
economic analysis that considers the cost of variation and the cost of
control. We do not expect to achieve perfect control, but good control is
usually worth spending some money.

For the blending tank example, then, we select:

x set point: CAo,setpt = 10 kg m-3. This would be determined by the user
of the stream.
x safety limits: none apparent from problem statement
x expected variation: ±1 kg m-3; such a specification might come from
historical data or engineering calculations. The steady-state material
balance (e.g., (3.11-1) applied at long times) shows that the make-up
flow must vary at least ±5×10-5 m3 s-1 to compensate such
disturbances. However, might we need more capability during the
course of a transient??
x tolerable variation: ±0.1 kg m-3. This specification depends on the
user of the stream.


3.17 type of equipment needed for process control

Figure 3.17-1 shows our process and control scheme as two
communicating systems. The system representing the process has two
inputs and one output. Of these only one is a material stream; however,
we recall that systems communicate with their environment (and other
systems) through signals, and in the blending process the outlet
composition responds to the inlet composition and make-up flow rate.

The system representing feedback control describes the needed operations,

but we have not described the nature of the equipment – could there be a
single device that takes in a composition measurement and puts out a
flow? Can we find a vendor to make such a device to execute controller
algorithm (3.15-1)? Can we have the gain knob calibrated in units
consistent with those we want to use for flow and composition?
Lesson 3: The Blending Tank

system representing process

inlet composition outlet composition
-tank to hold liquid
make-up flow -agitator to mix contents
-inlet and outlet piping

system representing controller and other equipment

adjust multiply subtract

make-up gain and from set
flow add bias point

set point

Figure 3.17-1: Closed loop feedback control of process

We will address these questions in later lessons. For now, we assume that
there will be several distinct pieces of equipment involved, and that they
work together so that

Fc  Fbias K c C Ao ,setpt  C Ao (3.17-1)

where we use the conventional symbol Kc for controller gain. In the case
of (3.17-1), we notice that the dimensions of Kc are volume2 mass-1 time-1.

In good time we will improve our description of both equipment and

controller algorithms. When we do, however, we will find that the overall
concept of feedback control is the same as presented in Figure 3.17-1: the
controlled variable is measured, decisions are made, and the manipulated
variable is adjusted to improve the controlled variable.


3.18 closing the loop - feedback control of the blending process

Our next task will be to combine our controller algorithm with our system
model to describe how the process behaves under control. We begin by
expressing algorithm (3.17-1) in deviation variables. At the reference
condition, all variables are at steady values, indicated by subscript r.

Fc, r  Fbias K c C Ao,setpt ,r  C Ao, r (3.18-1)

Lesson 3: The Blending Tank

Presumably the reference condition has no error, so that the set point is
simply the target outlet composition CAo,r. Thus we learn that Fbias, the
zero-error manipulated variable value, is simply Fc,r. Subtracting (3.18-1)
from (3.17-1), we find

Fc  Fc ,r  Fbias  Fbias K c C Ao,setpt  C Ao,setpt ,r  C Ao  C Ao,r

Fc' K c C 'Ao,setpt  C 'Ao

If the set point remains at CAo,r, the deviation variable CƍAo,setpt will be
identically zero.

We replace the manipulated variable in system model (3.9-2) with

controller algorithm (3.18-2) to find

K c C 'Ao,setpt  C 'Ao
dC 'Ao C Ac
W  C 'Ao C 'Ai  (3.18-3)
dt F

On expressing (3.18-3) in standard form, we arrive at a first-order

dynamic system model representing the process under proportional-mode
feedback control, as shown in Figure 3.17-1.

C Ac K c
W dC 'Ao 1 F
 C 'Ao C'  C' (3.18-4)
C Ac K c dt C Ac K c Ai C Ac K c Ao,setpt
1 1 1

Equation (3.18-4) describes a dynamic system (process and controller in

closed loop) in which the outlet composition varies with two inputs: the
inlet composition and the set point. Figure 3.18-1 compares (3.18-4) with
the process model (3.9-2) alone; we see that
x the closed loop responds more quickly because the closed loop time
constant is less than process time constant IJ.
x the closed loop has a smaller dependence on disturbance CƍAi because
the gain is less than unity. Both time constant and gain are reduced by
increasing the controller gain Kc.

3.19 integration from zero initial conditions

In Section 3.10, we integrated our open-loop system model to find how
CƍAo responded to inputs CƍAi and Fƍc. Now we integrate closed-loop
system model (3.18-4) in a similar manner.

 t  t
W CL t W CL t
e t e t
C '
K CL ³ e W CL
C dt 
K SP ³ e W CL
C 'Ao,setpt dt (3.19-1)
Ao Ai
0 0
Lesson 3: The Blending Tank


C Ac K c
W 1 F
WCL K CL K SP (3.19-2)
C Ac K c C Ac K c C Ac K c
1 1 1

dC 'Ao C Ac '
open-loop behavior W  C 'Ao C 'Ai  Fc
dt F
(the process without control)
C Ac K c
W dC 'Ao 1 F
closed-loop behavior  C 'Ao C'  C'
C Ac K c dt C Ac K c Ai C Ac K c Ao ,setpt
1 1 1
(the process under control) F F F

time constant disturbance other input

variable gain

Figure 3.18-1: Comparing open- and closed-loop system descriptions

3.20 closed-loop response to pulse disturbance

We test our controlled process by a pulse ǻC in the inlet composition that
begins at time t1 and ends at t2. We find

°0 0  t  t1
° §  ( t  t1 )
W CL ·
C*Ao '
® Ai CL ¨
C K ¨ 1  e ¸¸ t1  t  t 2 (3.20-1)
° © ¹
°«'C Ai K CL §¨1  e W CL · º
 ( t 2  t1 ) ( t  t 2 )
¨ ¸
¸» e t2  t
°¯¬ © ¹¼

Figure 3.20-1 shows both uncontrolled (open-loop) and controlled (closed-

loop) process responses for the same operating conditions used in Section
3.11. We see the faster response and smaller error that we expected when
we examined (3.18-4) in Section 3.18. These characteristics improve as
gain increases. Increasing gain also elicits stronger manipulated variable
action. Thus automatic control appears to have improved matters.
inlet comp deviation (kg m-3) Lesson 3: The Blending Tank



0 100 200 300 400 500
0 100 200 300 400 500
make-up flow deviation (10-5 m3 s -1)

Kc = 0 m6 kg-1 s-1




outlet composition deviation (kg m-3)


Kc = 0 m6 kg-1 s-1

0.15 0.0001



0 100 200 300 400 500

Figure 3.20-1: Response to pulse input under proportional control.

3.21 closed-loop response to step disturbance - the offset phenomenon

Integrating (3.19-1) for a step of ǻC, we obtain

§ t
C 'Ao 'C K CL ¨1  e WCL ¸ (3.21-1)
© ¹
Lesson 3: The Blending Tank

Figure 3.21-1 shows open- and closed-loop step responses. Notice that for
no case does the controlled variable return to the set point! This is the
phenomenon of offset, which is a characteristic of the proportional control
algorithm responding to step inputs.

offset longterm response - set point

C Ao (f)  C Ao,setpt
C 'Ao (f)

Recalling (3.19-2), increasing the controller gain decreases the closed-loop

disturbance gain KCL, and thus decreases the offset.

We find that offset is implicit in the proportional control definition

(3.15-1). An off-normal disturbance variable requires the manipulated
variable to change to compensate. For the manipulated variable to differ
from its bias value, (3.15-1) shows that the error must be non-zero. Hence
some error must persist so that the manipulated variable can persist in
compensating for a persistent disturbance.

3.22 response to set point changes

We apply (3.19-1) to a change in set point.

§ t
C 'Ao 'C Ao,setpt K SP ¨1  e WCL ¸ (3.22-1)
© ¹

We recall from (3.19-2) that KSP is less than 1. Thus, the outlet
composition follows the change, but cannot reach the new set point. This
is again offset due to proportional-mode control. Increasing controller
gain increases KSP and reduces the offset.

3.23 tuning the controller

Choosing values of the adjustable controller parameters, such as gain, for
good control is called tuning the controller. So far, our experience has
been that increasing the gain decreases offset - then should we not set the
gain as high as possible?

We should not jump to that conclusion. In general, tuning positions the

closed-loop response between two extremes. At one extreme is no control
at all, gain set at zero (open-loop). At the other is too much attempted
control, driving the system to instability. In the former case, the
controlled variable wanders where it will; in the latter case, over-
aggressive manipulation produces severe variations in the controlled
variable, worse than no control at all. Tuning seeks a middle ground in
Lesson 3: The Blending Tank
which control reduces variability in the controlled variable. This means
both rejection of disturbances and fidelity to set point changes.
make-up flow deviation (10-5 m3 s -1) inlet comp deviation (kg m )



0 100 200 300 400 500
0 100 200 300 400 500
Kc = 0 m6 kg-1 s-1



outlet composition deviation (kg m-3)



Kc = 0 m6 kg-1 s-1


0.20 0.0003
0.10 0.0009
0 100 200 300 400 500

Figure 3.21-1: proportional control step response, showing offset

Lesson 3: The Blending Tank
Recall Figures 3.20-1 and 3.21-1. In these, we achieved our ±0.1 kg m-3
specification on outlet concentration at a gain between 0.0003 and 0.0009
m6 kg-1 s-1. Hence we will use our model, which predicts the response to
disturbances, to guide us in tuning. In operation, we would use the
predicted value as a starting point, and make further adjustments, if
required, in the field.

3.24 stability of the closed-loop system

In Section 3.23, we said that tuning positions the controlled process
between non-control and instability. We must therefore inquire into the
stability limit. Because (3.19-1) describes the closed-loop system, we
should be able to seek the conditions under which it becomes unstable.
We invoke the notion of stability to bounded inputs that we introduced in
Section 3.7, and we come to the same conclusion we reached there: a first-
order system is stable to all bounded inputs, and we have not changed the
order of the system by adding feedback control in the proportional mode.

So “theoretically” we can increase gain as much as we like with no

possibility of reaching instability. Equation (3.19-2) shows that in the
limit of infinite controller gain, the response will be instantaneous (IJCL =
0), disturbances will be completely rejected (KCL = 0) and set points will
be faithfully tracked (KSP = 1).

Practically, we will not be surprised to find that this is NOT true. No

automatically-controlled chemical process will really be first order.
Increasing the gain in real processes will ultimately lead to instability. We
will explore this point further in future lessons.

3.25 conclusion
We have done quite a lot:
x used conservation equations to derive a dynamic system model of a
x identified three characteristic disturbances to test system responses
x introduced the frequency response and Bode plots
x discussed how to formulate a control scheme
x introduced proportional-mode control and explored its behavior
x compared open- and closed-loop response
x learned how tuning fits between limits of no control and instability

We will elaborate each of these topics in later lessons.

Lesson 4: Two Tanks in Series
4.0 context and direction
In Lesson 3 we performed a material balance on a mixing tank and derived
a first-order system model. We used that model to predict the open-loop
process behavior and its closed-loop behavior, under feedback control. In
this lesson, we complicate the process, and find that some additional
analysis tools will be useful.


4.1 math model of continuous blending tanks

We consider two tanks in series with single inlet and outlet streams.

F, CAi F, CA1

volume V1 F, CA2

volume V2

Our component A mass balance is written over each tank.

V1C A1 FC Ai  FC A1
V2 C A 2 FC A1  FC A 2

As in Lesson 3, we have recognized that each tank operates in overflow:

the volume is constant, so that changes in the inlet flow are quickly
duplicated in the outlet flow. Hence all streams are written in terms of a
single volumetric flow F. Again, we will regard the flow as constant in
time. Also, each tank is well mixed.

Putting (4.1-1) into standard form

dC A1
W1  C A1 C Ai
dC A 2
W2  C A 2 C A1
Lesson 4: Two Tanks in Series
we identify two first-order dynamic systems coupled through the
composition of the intermediate stream, CA1. If we view the tanks as
separate systems, we see that CA1 is the response variable of the first tank
and the input to the second. If instead we view the pair of tanks as a single
system, CA1 becomes an intermediate variable. The speed of response
depends on two time constants, which (as before) are equal to the ratio of
volume for each tank and the common volumetric flow.

We write (4.1-2) at a steady reference condition to find

C A ,r C Ai ,r C A1,r C A 2,r (4.1-3)

We subtract the reference condition from (4.1-2) and thus express the
variables in deviation form.

dC 'A1
W1  C 'A1 C 'Ai
dC 'A 2
W2  C 'A 2 C 'A1

4.2 solving the coupled equations - a second-order system

As usual, we will take the initial condition to be zero (response variables
at their reference conditions). We may solve (4.1-4) in two ways:

Because the first equation contains only CƍA1, we may integrate it directly
to find CƍA1 as a function of the input CƍAi. This solution becomes the
forcing function in the second equation, which may be integrated directly
to find CƍA2. That is

1  t W1 t W1 '
C'A1 e ³ e C Ai dt (4.2-1)
W1 0

1  t W2 t W 2 ª 1  t W1 t W1 ' º
t t
C '
e ³ e « e ³ e C Ai dt » dt (4.2-2)
W2 ¬ W1
0 0 ¼

On defining a specific disturbance CƍAi we can integrate (4.2-2) to a


Alternatively, we may eliminate the intermediate variable CƍA1 between

the equations (4.1-4) and obtain a second-order equation for CƍA2 as a
function of CƍAi. The steps are

(1) differentiate the second equation

(2) solve the first equation for the derivative of CƍA1
Lesson 4: Two Tanks in Series
(3) solve the original second equation for CƍA1
(4) substitute in the equation of the first step.

The result is

d 2 C 'A 2 dC 'A 2
W1W 2 2
 W1  W 2  C 'A 2 C 'Ai (4.2-3)
dt dt

Two mass storage elements led to two first-order equations, which have
combined to produce a single second-order equation. A homogeneous
solution to (4.2-3) can be found directly, but the particular solution
depends on the nature of the disturbance:

t t
W1 W2
C'A 2 A1e  A 2e  C'A 2,part C 'Ai (4.2-4)

where the constants A1 and A2 are found by invoking initial conditions

after the particular solution is determined.

4.3 response of system to step disturbance

Suppose a step change ǻC occurs in the inlet concentration at time td.
Either (4.2-2) or (4.2-4) yields

ª W1  (t  t d )
W2  (t  t d )
W2 º
C'A2 U t  t d 'C «1  e W1
 e » (4.3-1)
¬ W1  W2 W1  W2 ¼

Each tank contributes a first-order response based on its own time

constant. However, these responses are weighted by factors that depend
on both time constants.

The result in Figure 4.3-1 looks somewhat different from the first-order
responses we have seen. We have plotted the step response of a second-
order system with IJ1 = 1 and IJ1 = 1.5 in arbitrary units. At sufficiently
long time, the initial condition has no influence and the outlet
concentration will become equal to the new inlet concentration; in this
respect it looks like the first-order system response. However, the initial
behavior differs: the outlet concentration rises gradually instead of
abruptly. This S-shaped curve, often called “sigmoid”, is a feature of
systems of order greater than one. Physically, we can understand this by
realizing that the change in inlet concentration must spread through two
tanks, and it reaches the second tank only after being diluted in the first.
Lesson 4: Two Tanks in Series

C* Ai/ C


0 1 2 3 4 5 6


C*A2/ C



0 1 2 3 4 5 6
Figure 4.3-1: Response to step change in inlet composition

4.4 introducing the Laplace transform

We bother with the Laplace transform for two reasons:
x after the initial learning pains, it actually makes the math easier, so we
will use it in derivations
x some of the terminology in linear systems and process control is based
on formulating the equations with Laplace transforms.

Definition: the Laplace transform turns a function of time y(t) into a

function of the complex variable s. Variable s has dimensions of
reciprocal time. All the information contained in the time-domain
function is preserved in the Laplace domain.

y(s) L^y( t )` ³ y ( t )e
dt (4.4-1)

(In these notes, we use the notation y(s) merely to indicate that y(t) has
been transformed; we do not mean that y(s) has the same functional
dependence on s that it does on t.)

Functional transforms: textbooks (for example, Marlin, Sec. 4.2) usually

include tables of transform pairs, so these derivations from definition
(4.4-1) are primarily to demonstrate how the tables came to be.
Lesson 4: Two Tanks in Series
L^CU ( t  t d )` ³ CU(t  t d )e st dt
C ³ e st dt
td (4.4-2)
 C st f
s td

Ce st d

L Ce  at
` ³ Ce  at st
e dt

 C (s  a ) t f
e (4.4-3)
sa 0


Operational transforms: this allows us to transform entire equations, not

just particular functions.

­ df ( t ) ½ df ( t ) st

¯ dt ¿
¾ ³
e dt

³ «¬sf (t )e
 fe »dt
d st º
dt ¼ (4.4-4)
s ³ f ( t )e st dt  f ( t )e st

sL^f ( t )` f (0)

­t ½ f
§t · st
L ®³ f [ d[¾ ³0 ¨¨© ³0 f [ d[ ¸¸¹e dt
¯0 ¿
1 §¨ §
f t
· st ·
³ f t e st
dt  ³ d ¨ ³ f [ d [ ¸e ¸
s0 s 0 ¨© ¨© 0 ¸ ¸
¹ ¹ (4.4-5)
1§ ·t
L^f t ` ¨¨ ³ f [ d[ ¸¸e st
s s©0 ¹ 0

L^f t `
Lesson 4: Two Tanks in Series

Inverting transforms: use the tables to invert simple Laplace-domain

functions to their time-domain equivalents. To simplify the polynomial
functions often found in control engineering we may use partial fraction
expansion. The complicated ratio in (4.4-6) can be inverted if is expanded
into a series of simpler fractions.

N (s) N (s) C1 C2 C3
 n 1
 ...   ... (4.4-6)
D(s) (s  D1 ) n (s  D 2 ) m ... (s  D1 ) n
(s  D1 ) (s  D 2 ) m

In (4.4-6), Į1 and Į2 are repeated roots of the denominator. The inverse

transform of each term will involve an exponential function of the root Di.

C Ct n 1e Dt
f (s) Ÿ f (t) (4.4-7)
(s  D) 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
(s  D1 )(s  D 2 )(s  D 3 ) (s  D1 ) (s  D 2 ) (s  D 3 )

If a root is repeated, it requires a term for each repetition.

N(s) C1 C2 C3
(s  D1 )(s  D 2 ) 2 (s  D1 ) (s  D 2 ) 2
(s  D 2 )

Some roots may appear as complex conjugate pairs, so that, for example

D1 a  jb
D2 a  jb

where j is the square root of -1.

how to solve for the coefficients - it’s only algebra

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 coefficients determined, it may be easiest to substitute
arbitrary values for s to get equations in the unknown coefficients.
Lesson 4: Two Tanks in Series
3) For repeated roots, either
(3a) multiply the expansion by s and take the limit as s o f.
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, solving for one coefficient is enough. The other
coefficient will be the complex conjugate.

4.5 solving linear ODEs with Laplace transforms

We return to (4.1-4), the two equations that describe concentration in the

dC 'A1
W1  C 'A1 C 'Ai
dC 'A 2
W2  C 'A 2 C 'A1

We perform the Laplace transform on the entire first equation. It

distributes across addition, and constant IJ1 may be factored out.

­ dC ' ½
L ®W1 A1  C 'A1 ¾ L C 'Ai ^ `
¯ dt ¿
­ dC ' ½
^ ` ^ `
W1L ® A1 ¾  L C 'A1 L C 'Ai
¯ dt ¿

We next perform an operational transform on the derivative. Because the

functional forms of the variables CƍA1 and CƍAi are not yet known, we
simply indicate a variable in the Laplace domain.

^ `
W1 sC'A1 (s)  C'A1 (0)  C'A1 (s) C'Ai (s)
W1sC (s)  C (s) C'Ai (s)

We can easily solve (4.5-2) for CƍA1. If we similarly treat the second
equation in (4.1-4), we arrive at the equivalent formulation in the Laplace
Lesson 4: Two Tanks in Series
C 'A1 (s) C 'Ai (s)
W1s  1
' 1
C (s) C 'A1 (s)
W 2s  1

Equations (4.5-3) are not solutions - we have not solved anything! We

merely have a new formulation of problem (4.1-4), a formulation that is
more abstract (what on earth is Laplace domain?) and yet simpler, by
virtue of being algebraic. It is important to remember that all the
information held in the differential equations (4.1-4) is preserved in the
Laplace domain formulation (4.5-3).

We proceed toward solution by eliminating the intermediate variable in

(4.5-3). We find

1 1
C 'A 2 (s) C 'Ai (s) (4.5-4)
W 2s  1 W1s  1

With (4.5-4) we have gone as far as we can without knowing more about
the disturbance. That is, we cannot invert the right-hand side of (4.5-4)
until we can actually substitute a functional transform for the variable
CƍAi(s). In this sense, (4.5-4) resembles (4.2-2) and (4.2-4): a solution
needing more specification.

As in Section 4.3, suppose a step change ǻC occurs in the inlet

concentration at time td.

C 'Ai ( t ) U t  t d 'C (4.5-5)

We must take the Laplace transform,

'C  t d s
C 'Ai (s) e (4.5-6)

which we may substitute into (4.5-4).

1 1 'C  t ds
C 'A 2 (s) e (4.5-7)
W 2s  1 W1s  1 s

This IS the solution, the step response of the two tanks in series. Of
course, it really must be inverted to the time domain. We treat the
polynomial denominator from either the tables or partial fraction
Lesson 4: Two Tanks in Series
­ 1 1 1½ W1 t
W1 W2 t
L1 ® ¾ 1  e  e (4.5-8)
¯ W 2s  1 W1s  1 s ¿ W1  W 2 W1  W 2

Next we apply the time delay

­ 1 1 1  t ds ½ ª W1 ( t  t d )
W2 ( t  t d )
W2 º
L1 ® e ¾ U t  t d «1  e W1
 e » (4.5-9)
¯ W 2s  1 W1s  1 s ¿ ¬ W1  W 2 W1  W 2 ¼

Remembering the constant factor, we complete the inverse transform of


ª W1 ( t  t d )
W1 W2 ( t  t d )
W2 º
C'A 2 U( t  t d )'C«1  e  e » (4.5-10)
¬ W1  W 2 W1  W 2 ¼

which, of course, is identical to (4.3-1), derived in the time domain.

4.6 describing systems with transfer functions

In Section 4.1 we derived a system model to describe transient behavior in a tanks-in-series
process. Then in Section 4.5 we used Laplace transforms to solve it. Let us now do the same
procedure in the abstract. Begin with the first-order lag, written in deviation variables:

dy '
W  y' Kx ' ( t ) y ' (0) 0 (4.6-1)

After taking Laplace transforms, we relate input and output by an algebraic equation:

K '
y ' (s) x (s) (4.6-2)
Ws  1

The ratio in (4.6-2) multiplies xƍ(s) (the transform of disturbance xƍ(t)) and in the process
converts that signal into yƍ(s) (the transform of the response yƍ(t)). We call this ratio the transfer
function G(s).

y ' (s) K
G (s) (4.6-3)
x ' (s) Ws  1

G(s) contains all the information about the ODE (4.6-1). We should from now recognize it,
when we see it, as a first-order lag. Should we want to know how the first-order lag behaves in
response to some disturbance, we transform the disturbance, multiply it by the first-order lag
transfer function, and then take the inverse transform of the result.

Let us generalize (4.5-4), which described two first-order lags in series:

Lesson 4: Two Tanks in Series
K1 K2
y ' (s) x ' (s) (4.6-4)
W1s  1 W 2s  1

The transfer function for this second-order system is a product of two first-
order lags

y ' (s) K1 K2
G1 (s)G 2 (s) G (s) (4.6-5)
x ' (s) W1s  1 W 2s  1

We shall consider a transfer function to be a completely satisfactory

description of a dynamic system. We shall learn to notice its gain (long-
term steady-state relationship between yƍ and xƍ), time constants, and poles
(roots of the denominator).

Table 4.6-1: Characteristics of systems we have studied

type equation transfer poles steady
function state
1st order lag dy' K -W-1 K
W  y ' ( t ) Kx ' ( t )
dt Ws  1
2 order d 2 y' dy' K -W1-1
overdamped* W W
1 2 2
 W1  W 2  y ' Kx ' ( t )
( W1s  1)(W 2s  1) -W2-1
dt dt
*why “overdamped”? There are other 2 order forms to be encountered.

4.7 describing systems with block diagrams

The block diagram is a graphical display of the system in the Laplace domain.

xƍ(s) yƍ(s)

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 a description of more complicated
dynamic behavior. Remember that the transfer function incorporates all the dynamic
information in the system equations. This diagram implies the Laplace domain relationship

y ' (s) G (s) x ' (s) (4.7-1)

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

The Block Diagram Rules (see Marlin, Sec.4.4):

Lesson 4: Two Tanks in Series
x exclusion: only one input and output to a block.

inputs system outputs

x summing: two signals may be summed at an explicit summing

junction. The algebraic sign is indicated at the junction (if omitted,
is presumed to be positive).
x'‘ 1(s) y‘' 1(s)
+ x' 3(s) y' 3(s)
x' 2(s)
y' 2(s)

x multiple assignment: a single signal may feed its value to multiple

blocks. This does NOT indicate that the signal is divided up
among the blocks.
y 2(s)
x 1(s)
x' y 1(s)
G1(s) G2(s)

y' 3(s)
y' 1(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

y 3' (s) G 3 (s)x 3' (s)

G 3 (s) y1' (s)  y '2 (s)
G 3 (s) G1 (s)x1' (s)  G 2 (s)x '2 (s)

G 3 (s)G1 (s)x1' (s)  G 3 (s)G 2 (s)x '2 (s)

In the multiple assignment diagram

y '2 (s) G 2 (s) y1' (s)

G 2 (s)G1 (s) x1' (s)
y 3' (s) G 3 (s) y1' (s)
G 3 (s)G1 (s) x1' (s)
Lesson 4: Two Tanks in Series

Similarly, equations may be turned into block diagrams. System (4.7-2) has two inputs and thus
requires at least 2 blocks.

x' 1(s)
G1(s) G3(s)
+ y' 3(s)
x' 2(s)
G2(s) G3(s)

System (4.7-3) has two outputs for one input. Input x*1 is not split – its full value is sent to each
of two blocks.

y' 2(s)
x 1(s)
x' G1(s) G2(s)

y 3(s)
G1(s) G3(s)

This pair of block diagrams is equivalent to the pair from which they were derived.

As a further illustration, we apply the block diagram rules to the two-tank

system in (4.5-3):

C' Ai(s) 1 C' A1(s) 1 C' A2(s) C' Ai(s) 1 1 C' A2(s)
W1s  1 W 2s  1 W1s  1 W 2 s  1

4.8 frequency response from the transfer function

In Section 3.5 we derived the system response to a sine input by
integrating the differential equation. We learned that the frequency
response - that is, the long-term oscillation - could be characterized by its
amplitude ratio and phase angle; these quantities were expressed on a
Bode plot.

Alternatively, we may derive the frequency response directly from the

transfer function by substituting jȦ for s, where j is the square root of -1
and Ȧ is the radian frequency of the sine input. For the second-order
transfer function (4.6-5),
Lesson 4: Two Tanks in Series
K 1K 2
G ( jZ)
1  W1Zj 1  W2Zj
1  W1Zj 1  W 2Zj
K 1K 2
1  W Z 1  W
2 2
Z2 (4.8-1)

1  W1W 2 Z  Z W1  W 2 j
K 1K 2

1  W1 Z2 1  W 2 Z2
2 2

The function G(jȦ), although perhaps daunting to behold, is simply a
complex number. As such, it has real and imaginary parts, and a
magnitude and a phase angle. It turns out that the magnitude of G(jȦ) is
the amplitude ratio of the frequency response.

G ( jZ)
1  W W Z  Z W 2 2
 W 2

1  W Z 1  W Z
1 2 1
K1K 2 2 2 2 2
1 2

1  Z W  W  Z W W
2 2 2 4 2 2

1  W Z 1  W Z
1 2 1 2
K 1K 2 2 2 2 2
1 2

K 1K 2
1  W1 Z2 1  W 2 Z2
2 2

Furthermore, the phase angle of G(jȦ) is the frequency response phase


Im G ( jZ)
‘G ( jZ) tan 1
Re G ( jZ)
 Z W1  W 2
tan 1
1  W1W 2 Z2

In Figure 4.8-1, the Bode plot abscissa has been normalized by the square
root of the product of the time constants. We see that the amplitude ratio
of a second-order system declines more swiftly than that of a first-order
system: the slope of the high-frequency asymptote is -2. Unbalancing the
time constants further decreases the amplitude ratio. The phase angle can
reach -180º. It is symmetric about -90º.
Lesson 4: Two Tanks in Series
amplitude ratio/gain


0.01 0.1 1 10 100
-30 5
phase angle (deg)

-60 W1



0.01 0.1 1 10 100
Ȧ(IJ1IJ2)0.5 (radians)

Figure 4.8-1: Bode plot for overdamped second order system

4.9 stability of the two-tank system, and linear systems in general

The step response (4.5-10) shows no terms that grow with time, so long as
the time constants are positive. Furthermore, the amplitude ratio in Figure
4.8-1 is bounded. Thus, a second-order overdamped system appears to be
stable to bounded inputs. In practical terms, a concentration disturbance at
the inlet should not provoke a runaway response at the outlet.

We have seen first- and second-order linear systems. Let us generalize to

arbitrary order:

d n y' d n 1 y ' dy '

a n n  a n 1 n 1  "  a1  y' x' (4.9-1)
dt dt dt

The function xƍ represents all manner of bounded disturbances, expressed

in deviation form. The system properties, however, reside on the left-hand
side of (4.9-1), and its stability behavior should be independent of the
particular nature of the disturbances xƍ. Hence, we may examine the
homogeneous equation
Lesson 4: Two Tanks in Series

d n y' d n 1 y ' dy '

an n
 a n 1 n 1
 "  a 1  y' 0 (4.9-2)
dt dt dt

The solution to (4.9-2) is the sum of n terms, each containing a factor eĮit,
where Įi is a real root, or the real part of a complex root, of the
characteristic equation.

a n r n  a n 1r n 1  "  a 1r  1 0 (4.9-3)

Hence if all the Įi are negative, the solution cannot grow with time and
will thus be stable. If we take the Laplace transform of (4.9-1),

y ' (s) 1
G (s) n 1
x ' (s) a n s  a n 1s
 "  a 1s  1

we see that the denominator of the transfer function is identical to the

characteristic equation. Hence, the stability of the system is determined
by the poles of the transfer function: poles with zero or positive real parts
indicate a system unstable to bounded disturbances. Table 4.6-1 shows
that first-order lag and overdamped second-order systems are stable if
their time constants are positive.


4.10 step 1 - specify a control objective for the process

Our control objective is to maintain the outlet composition CA2 at a
constant value.

4.11 step 2 - assign variables in the dynamic system

The controlled variable is clearly CA2. The inlet composition CAi is a
disturbance variable. The composition CA1 is an intermediate variable.
We have no candidate manipulated variable; hence we decide to add a
concentrated make-up stream as we did with the single tank in Lesson 3.
We could add the stream to the first or the second tank - we seek advice:

One advisor says, “Add it to the first tank, where the disturbance enters.
That way, the manipulation can interact thoroughly with the disturbance;
it’s a matter of success through cooperation.” A second advisor says,
“Add it to the second tank. That way the manipulated variable can affect
the controlled variable more directly, through one tank instead of two.”
This advice brings to mind the difference we have seen between first- and
second-order responses. Yet the first advisor is better-dressed, friendlier,
and has a comforting manner - we decide to add the make-up stream to the
first tank.
Lesson 4: Two Tanks in Series

Material balances on the solute give

V1 dC A1 F C
 C A1 C Ai  Ac Fc
F  Fc dt F  Fc F  Fc
V2 dC A 2
 CA 2 C A1
F  Fc dt

We expect to use a relatively small make-up flow Fc of concentration CAc;

hence we make the approximation that F + Fc ~ F. Hence

V1 dC A1 C
 C A1 C Ai  Ac Fc
F dt F
V2 dC A 2
 C A 2 C A1
F dt

As is our custom, we write (4.11-2) at a steady reference condition and

subtract this reference to leave the variables in deviation form. After
taking Laplace transforms, we eliminate intermediate variable CƍA1(s) to

1 C Ac F
C 'A 2 (s) C 'Ai (s)  Fc' (s) (4.11-3)
W1s  1 W 2s  1 W1s  1 W 2s  1
The time constants are the usual volume-to-flow ratios. The poles of the
two transfer functions are negative, so adding a make-up stream has not
made the system unstable. Compare (4.11-3) to (4.5-4), which describes
the two tanks without makeup flow. Figure 4.11-1, the block diagram of
(4.11-3), emphasizes that controlled variable CA2 is influenced by both
disturbance CAi and manipulated variable Fc.

C'Ai (s) 1
W1s  1 W 2s  1

Fc' (s ) C Ac F C'A 2 (s)

W1s  1 W2s  1

Figure 4.11-1: Block diagram of two-tank mixing process

4.12 step 3 - proportional control

We will use proportional control. Although we recognize the
disadvantage of offset, as demonstrated in Lesson 3, we feel confident in
Lesson 4: Two Tanks in Series
our ability to manage it at an acceptable level by increasing the controller
gain. The proportional algorithm is

Fc  Fbias K gain C A 2,setpt  C A 2 (4.12-1)

Should the outlet composition fall below the set point, the error will
become positive. A positive controller gain Kgain will increase make-up
flow Fc above the bias value, which will act to increase the outlet

4.13 step 4 - choose set points and limits

As in Section 3.16, we identify any applicable safety limits, choose the
desired operating point, specify limits of tolerable variation about it, and
supply make-up flow in sufficient quantity to counteract the anticipated


4.14 components of the feedback control loop

We should consider the equipment more realistically. Figure 4.14-1 is a
block diagram showing the four components found in most process control
applications: process, sensor, controller, and final control element. Each
block contains a transfer function that relates output to input. The signals
between blocks are Laplace transforms of deviation variables. We
recognize that the process will typically comprise multiple blocks for
disturbance and manipulated variable inputs. An example is the mixing
tank process that shown in Figure 4.11-1.



x’m(s) y’c (s)


final sensor
control Gv(s) Gs(s)
y’s (s)
y’ İ’ (s) -

Gsp(s) y’sp(s)

Figure 4.14-1: General block diagram of feedback control loop

Lesson 4: Two Tanks in Series
The output signal that we have designated as the controlled variable yƍc is
measured by a sensor. In our mixing tank example, we need some
measurement that reliably indicates composition; depending on the nature
of the solution, it might be a chromatograph, conductivity meter,
spectrometer, densitometer, etc.

The controller subtracts the sensor signal yƍs from the set point yƍsp and
executes the controller algorithm Gc(s) on the error. The subtraction is
sometimes said to take place in the comparator.

The controller output yƍco drives a final control element to produce

manipulated variable xƍm. In chemical process industries this is most often
a valve in a pipe carrying some fluid stream, but it could also be a motor, a
heater control, etc.

4.15 transfer functions for loop components

Although we are not yet ready to address hardware, we can improve our
description of equipment behavior by posing transfer functions for each of
the closed loop components indicated in Figure 4.14-1. Notice that a
transfer function has two main parts: steady and dynamic.
x The steady part is the gain; gain indicates the magnitude of the effect
of the transfer function on the input signal, and it performs the units
conversion between input and output. In (4.11-3), the transfer function
that converts make-up flow Fc into outlet concentration CA2 has a gain
of CAc/F. The gain depends on these two process parameters, and the
units are chosen to be consistent with those of the input and output
signals. The other transfer function in (4.11-3) has a dimensionless
unity gain, independent of process parameters.
x The dynamic part is everything else - all the system time constants and
the functions of the Laplace variable s. The dynamic part
characterizes the way that an input signal is processed in time. In
(4.11-3), both transfer functions feature second-order dynamics.

Let us presume that the sensor is fast - really fast - so that negligible time
elapses between a change in the controlled variable yc and its
measurement ys. Then the transfer function is

y s' K s y 'c (4.15-1)

Being really fast means that the transfer function has NO dynamic part.
Such a transfer function indicates a “pure gain” process, one in which
changes in the input are “instantaneously” seen in the output. The
dimensions of the gain Ks are
Lesson 4: Two Tanks in Series
controller _ in
Ks ( ) (4.15-2)
controlled var iable

where the sensor is presumed to deliver the measurement to the controller

in suitable “controller_in” units.

We see that a proportional controller is also a pure gain process between
error signal and controller response.

y 'co K cH' (4.15-3)

where the error is conventionally defined with the set point as positive:

H' '
y sp  y s' (4.15-4)

For now we will leave the controller signal dimensions unspecified.

However, we can be sure that the gain has dimensions of

controller _ out
Kc ( ) (4.15-5)
controller _ in

set point
The error signal has dimensions suitable for the controller, which implies
that ysp and ys have the same units. However, the operator might prefer to
have the set point expressed in the units of the controlled variable (so-
called engineering units), which implies that ysp,e and yc have the same
units. The set point transfer function performs this unit conversion; it is a
pure gain process with gain identical to that of the sensor. That is

G sp Ks (4.15-6)

final control element

The valve is a mechanical device that takes some time to move. We might
imagine that a valve can change more quickly than a large chemical
process vessel, and thus that for many control applications the valve
dynamics can be neglected. In our case, however, we will assume that the
valve operates with first-order dynamics, such that

Kv '
x 'm y co (4.15-7)
W vs  1

The valve gain has dimensions of

Lesson 4: Two Tanks in Series
Kv ( ) (4.15-8)
controller _ out


4.16 assembling the components into a closed loop

The closed loop block diagram in Figure 4.14-1 shows how the loop
components are arranged. The transfer functions for the various
components are defined in Section 4.15. We will now use the block
diagram rules to derive the equations for the closed loop, in three steps:
x in general, good for any application of Figure 4.14-1
x applying the choices we made in Section 4.15
x adapting the general nomenclature of Section 4.15 to the two-tank
mixing process

We begin with the controlled variable, which is the output of the closed
loop system, and work backward through the diagram until all paths are
traced and the inputs appear.

y 'c (s) G d x 'd (s)  G m x 'm (s)

G d x 'd (s)  G m G v G c H ' (s) (4.16-1)
G d x 'd (s)  G m G v G c G sp y sp
, e (s)  G s y c (s)

At this point, we collect the controlled variable on the left-hand side.

y 'c (s) 1  G m G v G c G s G d x 'd (s)  G m G v G c G sp y sp

,e (s)

Gd G m G v G c G sp (4.16-2)
y 'c (s) x 'd (s)  '
y sp (s)
1  G m G v G c G s 1  G m G v G c G s ,e
Equation (4.16-2) shows how a controlled variable responds to a
disturbance and set point inputs. It is derived from Figure 4.14-1 and
applies to any system that can be represented by the figure.

We now specialize (4.16-2) with transfer functions we defined in Section

4.15. These use the general nomenclature of Figure 4.14-1, but depend on
assumptions we made about fast sensors and first-order valves.

Gm K cKs
Gd W vs  1
y 'c (s) x d (s) 
' '
y sp ,e (s) (4.16-3)
§ Kv · § Kv ·
¨¨1  G m K c K s ¸¸ ¨¨1  G m K c K s ¸¸
© W v s  1 ¹ © W v s  1 ¹
Lesson 4: Two Tanks in Series
Equation (4.16-3) begins to show how the general transfer functions
become specific functions of the Laplace variable s. Now we further
specialize (4.16-3) to the two-tank problem by substituting the process
transfer functions and specific nomenclature from (4.11-3) or,
equivalently, Figure 4.11-1.

1 C Ac F Kv
C 'A 2 (s)
W1s  1 W2s  1 C Ai (s) 
' W1s  1 W 2s  1 W vs  1
C 'A 2,sp (s)
§ C Ac F Kv · § C Ac F Kv ·
¨¨1  K c K s ¸¸ ¨¨1  K c K s ¸¸
© W1s  1 W 2s  1 W v s  1 ¹ © W1s  1 W 2s  1 W vs  1 ¹

Beneath its apparent complexity, (4.16-4) simply tells how the outlet
concentration reacts to disturbances and to a set point input. This closed-
loop response is compared in Figure 4.16-1 to the open-loop response of
the process.

transfer function for disturbance other inputs

1 C Ac F
C 'A 2 (s) C 'Ai (s)  Fc' (s)
W1s  1 W 2s  1 W1s  1 W2s  1
1 C Ac F Kv
K cKs
C 'A 2 (s)
W1s  1 W2s  1 C Ai (s) 
' W1s  1 W 2s  1 W v s  1
C 'A 2,sp (s)
§ C Ac F Kv · § C Ac F Kv ·
¨¨1  K c K s ¸¸ ¨¨1  K c K s ¸¸
© W1s  1 W 2s  1 W v s  1 ¹ © W1s  1 W2s  1 W vs  1 ¹

Figure 4.16-1: Comparing open- and closed-loop responses

It is clear that the disturbance response has a different character, because

the transfer function has changed. IF we made a good decision on control
algorithm, and IF we tune the controller properly, the closed-loop response
should be better.

4.17 some perspective on how we derived the closed-loop response

Remember what we did: we proposed a block diagram of feedback control
and derived the associated transfer functions between inputs and output.
Then we substituted the component transfer functions appropriate to our
particular problem.

Instead of the block diagram algebra, we could have combined the Laplace
domain equations of Section 4.15 directly, eliminating intermediate
variables until we arrived at (4.16-4).
Lesson 4: Two Tanks in Series
Furthermore, we could have proceeded entirely in the time domain, as we
did in Lesson 3. That is, the second-order process ODE could have been
combined with the first-order valve ODE and algebraic equations for
sensor and controller to arrive at an ODE for the controlled variable with
disturbance and set point forcing functions.

We have used new tools - the Laplace transform and the block diagram -
but the underlying objective, and the relationships between inputs and
outputs, were the same as working in the time domain. This is not

4.18 calculating closed-loop responses

But how does the closed-loop perform? We approach this by simplifying

W vs  1
W1W 2 W v
C 'A 2 (s) C 'Ai (s)
§ 1 1 1 · § W W W · 1  K v K c K s C Ac F
s 3  ¨¨   ¸¸s 2  ¨¨ 1 2 v ¸¸s 
© W1 W 2 W v ¹ © W1W 2 W v ¹ W1W 2 W v
1  K v K c K s C Ac F
W1W 2 W v
 C 'A 2,sp (s)
§1 1 1· § W  W  W · 1  K v K c K s C Ac F
s 3  ¨¨   ¸¸s 2  ¨¨ 1 2 v ¸¸s 
© W1 W 2 W v ¹ © W1W 2 W v ¹ W1W 2 W v

Clearly we did not succeed at that... As with (4.5-4) and (4.11-3), we

would like to substitute a particular disturbance for CƍAi(s) in (4.18-1) and
invert the result to obtain the time response of CA2. Here we encounter an
obstacle: our transform tables do not feature anything as complicated as
(4.18-1). Furthermore, to use partial fraction expansion we must find the
roots of the cubic equation; however, we are unlikely to find an analytical
expression for these. That is unfortunate, because it would be helpful to
know how the transfer function poles depend on the controller gain Kc.

We resort to numerical methods. Here is our plan:

x do a partial-fraction expansion of each transfer function in (4.18-1) in
terms of the poles
x multiply each term in the expansion by the disturbance of interest and
invert to find the response
x for a particular value of controller gain Kc, find the poles numerically
x repeat for different values of Kc to map out the behavior

This expedient of using numerical calculations does not show us the

functional dependence of the response on the parameters, but it does get
the job done.
Lesson 4: Two Tanks in Series

We will begin with the disturbance transfer function in (4.18-1). The

polynomial ratio part of the transfer function is written as the sum of

W vs  1
§1 1 1· § W  W  W · 1  K v K c K s C Ac F
s 3  ¨¨   ¸¸s 2  ¨¨ 1 2 v ¸¸s 
© W1 W 2 W v ¹ © W1W 2 W v ¹ W1W 2 W v
W vs  1
s  D1 s  D 2 s  D 3
C1 C2 C3
s  D1 s  D 2 s  D 3

The poles of the transfer function are Įi, and the coefficients Ci depend on
these poles, as well as the numerator. We will keep in mind that both Įi,
and Ci depend on the system time constants and gains, as well as the
controller gain Kc. Solving for the coefficients is an algebra problem. The
results are

W v D1  1
D1  D 2 D1  D 3
WvD 2  1
C2 (4.18-3)
D 2  D1 D 2  D 3
WvD3  1
D 3  D1 D 3  D 2
Thus numerical values of coefficients Ci can be computed for each set of
poles Įi.

Now we use expansion (4.18-2) to rewrite (4.18-1) for a disturbance

response. With no change in set point, CƍA2,sp(s) is identically zero.

1 ª C1 C2 C3 º '
C 'A 2 (s) «   » C Ai (s)
W1W 2 W v ¬ s  D1 s  D 2 s  D 3 ¼
ª 1 1 1 º (4.18-4)
« C C C3 »
1 2
« 1  2  3 » C 'Ai (s)
W1W 2 W v «  1 s  1  1 s  1  1 s  1»
«¬ D1 D2 D3 »¼

Now, as an example, we pose a step disturbance in the inlet composition.

Lesson 4: Two Tanks in Series
C 'Ai ( t ) 'CU ( t  t d )
'C  t ds (4.18-5)
C 'Ai (s) e

We substitute the step disturbance (4.18-5) into the system model (4.18-4)
to obtain

ª 1 1 1 º
« C1 C2 C3 »
'C « D1 D2 D3 » e  t ds
C 'A 2 (s)   (4.18-6)
W1W 2 W v « §  1 · § 1 · § 1 · »
« ¨¨ s  1¸¸s ¨¨ s  1¸¸s ¨¨ s  1¸¸s »
«¬ © D1 ¹ © D2 ¹ © D3 ¹ »¼

Each term inverts to a step response. Remembering to apply the time

delay, we find the result

'CU(t  t d ) ª C1 C 2  C3 º
CcA 2 (t)
W1W2 W v
1  eD1 (t  t d ) 
1  eD2 (t  t d ) 
1  eD3 (t  t d ) » (4.18-7)
¬ D1 D2 D3 ¼

Examining (4.18-7) we learn that the exponential terms contribute

according to the magnitude of the pole Įi: small poles (larger time
constants) cause the term to persist. We see that there will be offset,
because CƍA2 does not go to zero at long times. The amount of offset will
depend on the magnitude of the coefficients Ci; our experience in Lesson 3
would suggest that these will become smaller as the controller gain Kc

4.19 calculating the response for a particular example

We begin with similar parameter values to our example in Lesson 3:
F = 1.2 m3 min-1
Fc,r = 6×10-3 m3 min-1
V1 = 6 m3 (thus IJ1 = 5 min)
V2 = 4 m3 (thus IJ2 = 3.33 min)
CAi,r = 8 kg m-3
CAo,r = 10 kg m-3
CAc = 400 kg m-3
IJv = 0.1 min
Kv = 0.01 m3 min-1 controller_out-1
Ks = 0.5 controller_in m3 kg-3

We may calculate roots in (4.18-2) with calculators, spreadsheets, or

computer code. For example, using matlab we obtain roots of polynomial
s2 + 3s +4 by

>> roots ([1 3 4])

Lesson 4: Two Tanks in Series

ans =

-1.5000 + 1.3229i
-1.5000 - 1.3229i

Table 4.19-1: poles of closed loop disturbance transfer function

Kc (out in-1) Į1 (min-1) Į2 (min-1) Į3 (min-1)
0 -10 -0.2 -0.3
0.012 -10.0001 -0.2143 -0.2856
0.024 -10.0003 -0.2437 -0.2561

Notice that zero controller gain leads to poles equal to the negative inverse
of the three system time constants. Thus our closed-loop transfer function
reduces to describe the behavior of the process alone, under open-loop
conditions. After using the poles in Table 4.19-1 to compute the solution
(4.18-7) we obtain a plot of the response behavior. Indeed controller gain
can be increased to reduce the effects of the input disturbance.
Lesson 4: Two Tanks in Series

inlet comp deviation (kg m


00 10
10 20
20 30
30 40
make-up deviation (10 m s )

Kc = 0
3 -1




outlet composition deviation (kg m)






0 10 20 30 40
time (min)

Figure 4.19-1: Step response of two tanks under proportional control

4.20 surprise - increasing gain introduces oscillations!

We have been very tentative with the gain setting, so we act more
aggressively to suppress offset. The poles become complex!
Lesson 4: Two Tanks in Series

Table 4.20-1: complex poles of transfer function

Kc (out in-1) Į1 (min-1) Į2 (min-1) Į3 (min-1)
0.1 -10.001 -0.249 + j0.088 -0.249 - j 0.088
10 -10.10 -0.198 + j1.00 -0.198 - j1.00
50 -10.48 -0.011 + j2.20 -0.011 - j2.20

We must modify (4.18-7) to accommodate complex numbers. The roots

Į2 and Į3 are complex conjugates, and (recalling our discussion of partial
fraction expansion) so are coefficients C2 and C3. We define four new real
quantities to replace the complex ones:

D2 a  jb D3 a  jb
C2 C3 (4.20-1)
A  jB A  jB
D2 D3

We substitute these definitions into (4.18-7), recalling Euler’s relation,

e a  jb t e at cos bt  j sin bt (4.20-2)

to obtain

'CU ( t  t d ) ª  C1 º
C 'A 2 ( t )
W1W 2 W v
1  e D1 ( t  t d )  2A  2e a ( t  t d ) A cos b( t  t d )  B sin b( t  t d ) » (4.20-3)
¬ D1 ¼

Parameters A, B, a, and b are not variables with time, in the sense of CA2,
but they do depend on the value of the controller gain Kc. Parameters a
and b are found (via the root-finding procedure) in Table 4.20-1. A and B
come via complex algebra from (4.18-3).

 bW v a 2  b 2  D1b  2ab
> @
2b a 2  b 2 D1  a  b 2

W a D
 b 2 1  a  a D1  a  b 2
> @

2b a 2  b 2 D1  a  b 2

Taking data from Table 4.20-1 and using (4.20-4), we can plot response
Lesson 4: Two Tanks in Series
inlet comp deviation (kg m)



00 10
10 20
20 30
30 4400
make-up flow deviation (10 m s )
3 -1




outlet composition deviation (kg m)

Kc = 0.1






0.10 10
0.00 50
0 10 20 30 40
time (min)
Figure 4.20-1: Oscillatory step response

For the single tank of Lesson 3, the closed loop behavior was qualitatively
the same as that of the process itself. Here, however, closing the loop has
introduced behavior we would NOT see in the process alone: the response
variable oscillates in response to a steady input. The key is the third-order
characteristic equation, which can admit complex roots. The transfer
function is third order because the second-order process was placed in a
feedback loop with a first-order valve. If the system mathematics provide
Lesson 4: Two Tanks in Series
a true representation of the process equipment, we will see oscillations in

4.21 closed-loop stability by the Bode criterion

Figure 4.20-1 shows us that, despite the onset of oscillations, increasing
gain suppresses the offset. However, we must still be mindful of the
possibility of instability. In Figure 4.21-1 we plot the poles on a complex
plane. We see that increasing gain brings the closed loop to the point of
oscillation and then increases the imaginary (oscillatory) component. We
observe also that the real component approaches zero. Recalling that
stability depends on the real parts of the poles being negative, we see that
the closed loop will become unstable at a controller gain between 50 and


arrows show increasing Kc
from 0 to 100
imaginary (min )






-12 -10 -8 -6 -4 -2 0 2
real (min )

Figure 4.20-1: Root locus plot

This root locus plot provides a map of the stability limit; this is
particularly helpful, because we were unable to derive a single expression
that showed the effect of controller gain on the real parts of the poles. We
might imagine that finding the poles will only become more difficult as we
consider more complicated processes and controllers. Hence we introduce
an alternative means of predicting stability: the Bode criterion.

We develop the criterion intuitively; recalling Figure 4.14-1, we begin by

realizing that instability in a previously stable system happens because of
feedback in a closed loop. Suppose that the output signal yƍc contains
some fluctuating component at a particular frequency Ȧc. This component
is inverted in the comparator by being subtracted from the set point, and is
Lesson 4: Two Tanks in Series
fed to the controller. If the loop (process, controller, sensor, valve, etc.)
contributes a phase delay of -180q at frequency Ȧc, the inverted signal
returns to the output in phase to reinforce the fluctuation. If in addition
the amplitude ratio at Ȧc is greater than one, the fluctuation will grow.

We are not contriving a circumstance here. In any realistic process there

will be small disturbances, fluctuations over a wide domain of frequency.
The loop processes these fluctuations according to its frequency response
characteristics. Depending on the amplitude ratio, signals at some
frequencies may be amplified. Assuredly, signals at high frequencies will
be delayed by at least 180q. If these conditions overlap, there will be an
oscillating signal that will grow. We need not supply it; the system will
select it from the spectrum of background noise. This intuitive
development is described in more detail by Marlin (Sec.10.6).

To turn our intuition into a method, we return to the general description of

the closed loop transfer function in (4.16-2). Recall that the poles are the
roots of the characteristic equation, which is the denominator of the
transfer function. This characteristic equation is always of the form 1 plus
the product of the transfer functions around the loop. For convenience, we
will call this product the loop transfer function GL.

characteristic equation 1  G s G c G v G m
1 GL

It is the amplitude ratio and phase angle, that is, the frequency response, of
GL that determines whether signals will grow in the loop. First we find Ȧc,
the frequency at which the phase delay is -180q. At this crossover
frequency, we inspect the amplitude ratio; if it is less than one, the system
will attenuate reinforced disturbances, and thus be stable.

Thus the Bode criterion evaluates the stability of (1 + GL)-1 from the
frequency response of GL.

For our process, (4.16-4) gives

C Ac F Kv
G L (s) KK (4.21-2)
W1s  1 W2s  1 W vs  1 c s
The phase angle of GL is the sum of the phase angles of the various
elements in (4.21-2).
Lesson 4: Two Tanks in Series
C Ac 1 1 1
‘G L ( jZ) ‘ KcKsK v  ‘ ‘ ‘
F W1s  1 W 2s  1 W vs  1 (4.21-3)
0  tan 1  ZW1  tan 1  ZW 2  tan 1  ZW v

Equation (4.21-3) may be solved for the crossover frequency Ȧc; that is,
the frequency at which the loop delays the signal by -180º.

 180D tan 1  Zc W1  tan 1  Zc W 2  tan 1  Zc W v (4.21-4)

The amplitude ratio of GL is the product of the amplitude ratios of the

various elements in (4.21-2).

C Ac 1 1 1
G L ( jZ) KcKsK v u u u
F W1s  1 W 2s  1 W vs  1
C Ac 1 1 1
KcKsK v
F 1  Z2 W1 1  Z2 W 2 1  Z2 W v
2 2 2

We are particularly interested in the amplitude ratio at the crossover

frequency, RAc.

C Ac 1 1 1
R Ac KcKsK v (4.21-6)
F 1  Zc W1 1  Zc W 2 1  Zc W v
2 2 2 2 2 2

Using the data in Section 4.19, we find the crossover frequency from
(4.21-4) to be 2.25 radians minute-1. We notice that phase lag in the loop
depends only on the tanks and valve; the proportional controller, being a
“pure gain” system, contributes no lag to the dynamic response of the
loop. Hence the crossover frequency does not vary with the controller
gain setting.

Using the crossover frequency and further data from Section 4.19, we find
from (4.21-6) that the crossover amplitude ratio will be 1 when the
controller gain Kc is 52.55. The effect of controller gain is to amplify the
signals in the loop. Around Kc = 52.55, therefore, the system output will
oscillate unabated at frequency Ȧc. At higher gain settings, the amplitude
of the oscillation will grow in time. (The frequency of these oscillations
depends on the poles of the transfer function.)

Figure 4.21-1 is a Bode plot for the loop transfer function GL, showing
gains below, at, and above the instability threshold. The stability
threshold (amplitude ratio = 1, phase angle = -180°) is shown by a single
point at the crossover frequency
Lesson 4: Two Tanks in Series


amplitude ratio


Kc = 100
0.01 0.1 1 10 100

phase angle (deg)


-150 crossover point



0.01 0.1 1 10 100
Ȧ (radians min )

Figure 4.21-1: Bode plot for loop transfer function

Figure 4.21-2 shows unstable step responses at gains of 60 and 100. The
latter response quickly gets out of hand. Notice how the make-up flow
varies in response to the increasing error in the outlet composition.
make-up flow deviation (10 m s ) inlet comp deviation (kg m) Lesson 4: Two Tanks in Series



00 55 10
10 15
15 20
20 25
25 30
-5 3 -1


outlet composition deviation (kg m)

Kc = 100




0 5 10 15 20 25 30
time (min)

Figure 4.21-2: Step response for two-tank process at high gains

4.22 tuning based on stability limit - gain and phase margin

We tune a controller seeking good performance, somewhere between the
extremes of “no control” and “instability”. One method of tuning is
simply to maintain a reasonable distance from the instability limit and
presume that the result is an improvement over having no control. Thus,
Lesson 4: Two Tanks in Series
we find the instability threshold and tune the controller to leave margins
between these conditions and normal operation. The margins are
indicated in Figure 4.22-1, which shows a Bode plot for the loop transfer
function GL at some arbitrary controller setting.


GM > 1

PM > 0

Ȧ1 Ȧc
Figure 4.22-1: Illustration of gain margin and phase margin at a
single controller setting

The gain margin and phase margin are the distances shown on the
ordinates. Their definitions are

! 1 for stability
R Ac (4.22-1)
PM 180  I1 D
! 0 for stability
The controller setting determines the amplitude ratio and phase angle
curves. From those curves we then calculate the margins to see if they are

(1) use a phase angle of -180° to find the crossover frequency Ȧc

(2) use an amplitude ratio of 1 to find the frequency Ȧ1
(3) use Ȧc to find the amplitude ratio RAc, and thus GM
(4) use Ȧ1 to find the phase angle I1, and thus PM

In Figure 4.22-1, the system is stable. However, as the controller gain is

increased, the Bode plot will shift so that Ȧ1 and Ȧc approach each other.
At the instability threshold, Ȧ1 equals Ȧc, the gain margin is 1, and phase
margin is zero.
Lesson 4: Two Tanks in Series
A procedure for tuning proportional controllers by stability margin is:

(1) use a phase angle of -180° to find the crossover frequency Ȧc

(2) at Ȧc, find the gain that makes RAc = 1 (stability limit)
(3) at Ȧc, reduce the gain to make RAc = 1/GM (gain margin)
(4) use a phase angle of PM -180° to find the frequency Ȧ1
(5) at Ȧ1, find the gain that makes RA = 1 (phase margin)

Marlin (Sec.10.8) recommends tuning to maintain GM ~ 2 and PM ~ 30°.

Typically one or the other will be limiting.

Figure 4.22-2 shows the results of calculations for our tank example.
Earlier, we found the crossover frequency from (4.21-4). Then (4.21-6)
was solved for the controller gain that gave RAc = 0.5 (thus GM = 2). The
result was Kc = 26.3. Then (4.21-3) was solved for the frequency Ȧ1 to
give a phase angle of -150° (thus PM = 30°). Then (4.21-5) was solved
for the controller gain that gave an amplitude ratio of 1 at frequency Ȧ1.
The result was a much lower gain of 7. Therefore for our system, PM is
limiting, and the lower gain would be chosen. For reference, Figure 4.22-
2 also shows the stability limit determined earlier.

The gain and phase margins have given us a tuning criterion for selecting
a controller gain. Using the chosen gain, we can now predict the
performance in response to disturbances and set point changes. The
calculations would be similar to those illustrated in Sections 4.19 and
4.20: a partial fraction expansion leading to an expression for the
response, with parameter values based on numerical root-finding.
Lesson 4: Two Tanks in Series


amplitude ratio

Kc = 52.55

0.01 0.1 1 10 100

phase angle (deg)





0.01 0.1 1 10 100
Ȧ (radians min )
Figure 4.22-2: Bode plot illustrating GM and PM limits on gain

4.23 conclusion
We have completed dynamic analysis and control of a more complicated
process than in Lesson 3. In doing so we have introduced new tools for
analysis - the Laplace transforms and block diagrams - and developed
stability and tuning criteria.

Was it a good idea to listen to the appealing advisor and put the make-up
flow into the first tank? A good way to examine the question would be to
repeat the full analysis for the other case. Even without doing that,
however, we might reflect how removing one lag from the system might
affect the Bode stability criterion for the closed loop…

4.24 reference
Marlin, Thomas E. Process Control. 2nd ed. Boston, MA: McGraw-Hill, 2000.
ISBN: 0070393621.

4.25 nomenclature
a constant
Lesson 4: Two Tanks in Series
A constant
A1,A2 constants of integration
b constant
B constant
C constant
C1… constants in partial fraction expansion
CA1 intermediate stream concentration of solute A
CA2 exit stream concentration of solute A
CAc make-up stream concentration of solute A
CAi inlet stream concentration of solute A
CAs reference concentration of solute A at steady state
ǻC change in solute concentration
F volumetric flowrate
Fc volumetric flowrate of make-up stream
f function
G transfer function
Im operator that takes imaginary part of complex number
j square root of -1
K gain (time-independent part of transfer function)
L Laplace operator
N(s) polynomial in s
r dummy variable in polynomial characteristic equation
Re operator that takes real part of complex number
RA the amplitude ratio of the loop transfer function
RAc the amplitude ratio of the loop transfer function at the crossover frequency
s complex Laplace domain variable
t time
td time at which disturbance occurs
U unit step function
V1 volume of tank 1
V2 volume of tank 2
x(t) input signal to system
y(t) output signal from system
Į1 … roots of polynomial in s
İ error; set point minus controlled variable
IJ1 time constant of tank 1
IJ2 time constant of tank 2
IJv time constant of valve
ȟ dummy variable of integration
Ȧ radian frequency (has dimensions of radians time-1)
Ȧc crossover frequency, at which loop transfer function lag is -180°
Lesson 5: Heated Tank
5.0 context and direction
From Lesson 3 to Lesson 4, we increased the dynamic order of the
process, introduced the Laplace transform and block diagram tools, took
more account of equipment, and discovered how control can produce
instability. Now we change the process: our system models have
previously depended on material balances, but now we will write the
energy balance. We will also introduce the integral mode of control in the


5.1 a heated tank

We consider a tank that blends and heats two inlet streams. The heating
medium is a condensing vapor at temperature Tc in a heat exchanger of
surface area A.

F1 T1

F2 T2

F To

For the present, we continue to assume constant mass in an overflow tank.

Writing the material balance,

UF1  UF2 UF (5.1-1)

This is not yet the time for complications: we will approximate the
physical properties of the liquid (density, heat capacity, etc.) as constants.
We will also simplify the problem by assuming that the flow rates remain
constant in time. The energy balance is

UVC p (To  Tref ) UF1C p (T1  Tref )  UF2C p (T2  Tref )  UA(Tc  To )  UFCp (To  Tref ) (5.1-2)
Lesson 5: Heated Tank
where the overall heat transfer coefficient is U and the thermodynamic
reference is Tref. We identify a steady-state operating reference condition
with all variables at their desired values.

UVC p (Tor  Tref ) 0 UF1C p (T1r  Tref )  UF2C p (T2r  Tref )  UA(Tcr  Tor )  UFCp (Tor  Tref )

We subtract (5.1-3) from (5.1-2), define deviation variables, and rearrange

to standard form.

UVC pdTo' UF1C p UF2 C p UA

 To' T1'  T2'  Tc' (5.1-4)
UFC p  UA dt UFC p  UA UFC p  UA UFC p  UA

To make some sense of the equation coefficients, define the tank residence

WR (5.1-5)

and a ratio of the capability for heat transfer to the capability for enthalpy
removal by flow.

E (5.1-6)

ȕ thus indicates the importance of heat transfer in the mixing of the fluids.
We now use (5.1-5) and (5.1-6) to define the dynamic parameters: time
constant and gains.

W (5.1-7)
1 E

Thus the dynamic response of the tank temperature to disturbances is

faster as heat transfer capability (ȕ) becomes more significant. For no heat
transfer (ȕ = 0) the time constant is equal to the residence time.

K1 F (5.1-8)
1 E
K2 F (5.1-9)
1 E
Lesson 5: Heated Tank
Gains K1 and K2 show the effects of inlet temperatures T1 and T2 on the
outlet temperature. For example, a change in T1 will have a small effect
on To if the inlet flow rate F1 is small compared to overall flow F.

K3 (5.1-10)
1 E

Gain K3 shows the effect of changes in the temperature Tc of the

condensing vapor. For high heat transfer capability, ȕ is large, and gain
K3 approaches unity.

Our system model of the heated tank is finally written

W  To' K1T1'  K 2 T2'  K 3Tc' (5.1-11)

which shows a first-order system with three inputs. As is our custom, we

will take the initial condition on Tƍo as zero.

5.2 solving by laplace transform

Solution by Laplace transform is straightforward:

­ dTo' ½
L ®W ^
 To' ¾ L K1T1'  K 2 T2'  K 3Tc' `
¯ dt ¿
­ dT ' ½
^ ` ^ ` ^ `
WL® o ¾  L To' K1L T1'  K 2 L T2'  K 3 L Tc' ^ ` (5.2-1)
¯ dt ¿

W sTo' (s)  To' (0)  To' (s) K1T1' (s)  K 2 T2' (s)  K 3Tc' (s)
K1 ' K K
To' (s) T1 (s)  2 T2' (s)  3 Tc' (s)
Ws  1 Ws  1 Ws  1

Output Tƍo is related to three inputs, each through a first-order transfer


5.3 response of system to step disturbance

Suppose the tank is disturbed by a step change ǻT1 in temperature T1. We
have studied first-order systems, so we already know what the first-order
step response looks like: at the time of disturbance td, the output Tƍo will
depart from its initial zero value toward an ultimate value equal to the
product of gain K1 and the step ǻT1. The time required depends on the
magnitude of time constant IJ: when time equal to one time constant has
passed (i.e., t = td + IJ) the outlet temperature will be about 63% of its way
toward the long-term value.
Lesson 5: Heated Tank
However, doing the formalities for two step disturbances and one steady

T1' 'T1U t  t d1 T2' 'T2 U t  t d 2 Tc' 0 (5.3-1)

We take Laplace transforms of these input variables:

'T1  t d1s 'T2  t d 2s

T1' (s) e T2' (s) e Tc' (s) 0 (5.3-2)
s s

Then we substitute (5.3-2) into the system model (5.2-1).

K1 'T1  t d1s K 2 'T2  t d 2s K3

To' (s) e  e  0 (5.3-3)
Ws  1 s Ws  1 s Ws  1

We must invert each term; this is most easily done by processing the
polynomial first and then applying the time delay. Thus

­ 1 1½ t
L1 ® ¾ 1 e W (5.3-4)
¯ Ws  1 s ¿
­§ 1 1 ·  t d1s ½ § ( t  t d1 )
W ·
L1 ®¨ ¸e ¾ U t  t d1 ¨1  e ¸ (5.3-5)
¯© Ws  1 s ¹ ¿ © ¹

and finally

§  ( t  t d1 )
· § ( ttd 2 )
W ·
To* K1'T1U t  t d1 ¨1  e W
¸  K 2 'T2 U t  t d 2 ¨1  e ¸ (5.3-6)
© ¹ © ¹

Figure 5.3-1 shows a sample calculation for opposing input disturbances.

Notice that the gains K1 and K2 are less than unity. This is because we
dilute each disturbance with other streams: streams of matter (two input
flows) and energy (through the heat transfer surface). From the plot, can
you tell which gain is greater?
Lesson 5: Heated Tank

T'1 and T'2 (K)

-5 0 50 100 150 200 250 300

T'o (K)




0 50 100 150 200 250 300
t (s)

Figure 5.3-1: Response of outlet temperature to steps in two inlets

5.4 stability of the heated tank

System model (5.2-1) has one negative pole, s = -IJ-1, because the time
constant is a positive quantity. Hence, the system is stable to all bounded
inputs. Therefore, a blip in the inlet temperature will not set off wild
temperature excursions at the outlet.

5.5 numerical solution of ODEs

Analytical solutions to equations are desirable, but many useful equations
simply cannot be solved. When we “solve an ODE numerically” we
execute a set of calculations that results in a “numerical solution”: in
effect, a table of numbers. One column in that table contains values of the
independent variable, and the other column holds associated values of the
dependent variable. A table of numbers lacks the economy of presentation
and conceptual insight offered by an analytical expression. However, a
table of numbers is much better than no solution, and it can certainly be

Matlab offers a suite of functions for solving differential equations. For

example, the following file contains code to produce a plot
indistinguishable from that of the analytical solution in Figure 5.3-1.
Lesson 5: Heated Tank

function heated_tank (tauR, beta, F1frac)

% program to solve Eqn (5.1-11)
% the temperature of the heated tank is disturbed by the temperatures of
% the inlet streams and the condensing temperature of the vapor in the
% heat exchanger bundle

% INPUT variables
% tauR residence time in seconds
% beta heat transfer significance parameter
% F1frac fraction of flow in stream 1

% OUTPUT variables
% To the deviation in outlet temperature is plotted

% NOTE: all system variables are in deviation form

% define equation parameters

tau = tauR/(1+beta) ; % time constant in seconds
K1 = F1frac/(1+beta) ; % stream 1 gain
K2 = (1 - F1frac)/(1+beta) ; % stream 2 gain
K3 = beta/(1+beta) ; % heat exchange gain

% define the disturbances - first interval

tspan = [0, 10] ; % set the time interval
Toinit = 0 ; % start out at reference condition
T1 = 0; T2 = 0; Tc = 0; % no disturbances

% integrate the equation

[t,To] = ode45(@hot_tank,tspan,Toinit,[],tau,K1,K2,K3,T1,T2,Tc) ;

% plot the solution

plot (t,To)
hold on % allow plot to be updated with further plotting

% define the disturbances - second interval

tspan = [10, 50] ; % set the time interval
Toinit = To(size(To,1)); % start out at most recent value
T1 = 10; T2 = 0; Tc = 0; % introduce step in T1

% integrate the equation

[t,To] = ode45(@hot_tank,tspan,Toinit,[],tau,K1,K2,K3,T1,T2,Tc) ;

% plot the solution

plot (t,To)

% define the disturbances - third interval

tspan = [50, 250] ; % set the time interval
Toinit = To(size(To,1)); % start out at most recent value
T1 = 10; T2 = -5; Tc = 0; % introduce step in T2

% integrate the equation

[t,To] = ode45(@hot_tank,tspan,Toinit,[],tau,K1,K2,K3,T1,T2,Tc) ;

% plot the solution

plot (t,To)
Lesson 5: Heated Tank
hold off % turn off the plot hold

end % heated_tank

% define the differential equation in a subfunction

function dTodt = hot_tank(t,To,tau,K1,K2,K3,T1,T2,Tc)
dTodt = (K1*T1 + K2*T2 + K3*Tc - To)/tau ;

The Matlab function ode45 is one of several routines available for

integrating ordinary differential equations. The output arguments are a
column vector for time t and another for temperature deviation To. The
input argument @hot_tank directs ode45 to the subfunction in which
the equation is defined. Notice the way in which hot_tank is written:
the equation is arranged to calculate the derivative of output variable To
as a function of To and the input variables. ode45 calls hot_tank
repeatedly as it marches through the time interval denoted by the row
vector tspan.

5.6 scaled variables versus deviation variables

We introduced deviation variables so that any non-zero variable - positive
or negative - would be seen as a departure from the ideal reference
condition. Deviation variables also allowed us to define and use transfer
functions in expressing our system models through Laplace transforms.

We must, however, finally read real instruments and control real

processes. For these purposes, it is common to represent sensor readings,
controller outputs, and valve-stem positions on a scale of 0 to 100. Such
scaled variables obscure actual values, but immediately reveal context.
That is, a visitor to a control room would not know the significance of a
particular tank level reading of 2.6 m, but could interpret 96% easily. An
everyday example of a sensor that presents a scaled variable is the fuel
gauge in an automobile.

To scale physical variable y, for example, we identify the range in which

we expect it to vary: from ymin to ymax. We then subtract some bias value
from y and divide the difference by the range:

y  yb
y* { 100% (5.6-1)
y max  y min

If the bias value yb is set to the minimum ymin, then y* varies between 0
and 100%. This is the typical control room presentation. If instead the
bias value is set to the reference value yr, then y* varies from
Lesson 5: Heated Tank
y min  y r y  yr
100% d y* d max 100% (5.6-2)
y max  y min y max  y min

The scaled range is still 100% wide, but includes both positive and
negative regions, depending on where yr lies between ymin and ymax. Of
course, yb may be set to any arbitrary value between the limits, but ymin
and yr are generally the most useful.

We will use primes (ƍ) to denote deviation variables and asterisks (*) for
scaled variables. Unadorned variables will be presumed to be physical.
To convert a deviation variable yƍ (from an analytical solution, e.g.) for
presentation as a scaled variable y*, the definitions are combined:

y* {
y  y  y
r b
100% (5.6-3)
y max  y min

For example, Figure 5.6-1 shows a temperature trace expressed in

physical, deviation, and scaled form. Because these are linear
transformations, the basic character of the variable is unchanged.

5.7 just when I was getting accustomed to deviation variables!

We will tend to use deviation variables for analytic solutions and
derivations because of the convenience of zero initial conditions. We will
use scaled variables for our numerical work, in which each time step
moves the solution to a new value from the “initial condition” of the
previous step.
Lesson 5: Heated Tank









Figure 5.6-1 Expressing the variable in physical, deviation, and scaled forms


5.8 step 1 - specify a control objective for the process

Our control objective is to maintain the outlet temperature To at a constant
value. In particular, we prefer not to have offset in response to step-like
inputs. This means we must do something besides proportional mode.

5.9 step 2 - assign variables in the dynamic system

The controlled variable is clearly To. The inlet temperatures T1 and T2 are
disturbance variables.

By the model, we are left with the condensing vapor temperature as the
manipulated variable. But what sort of valve adjusts temperature? We
will discuss this below when we select equipment.

5.10 step 3 - PI (proportional-integral) control

We introduced proportional control as an intuitively appealing mechanism
- the response increases with the severity of the error. However,
Lesson 5: Heated Tank
proportional control suffers from offset. To counter this defect, we
introduce a second mode of control, called integral, to complement the
proportional behavior.

§ 1
x *co  x *co , b K *c ¨¨ H*  ³ H*dt ¸¸ (5.10-1)
© TI 0 ¹

where x*co is the controller output and the controlled variable error is

H* y*sp  y* (5.10-2)

Before we discuss the integral mode, let us understand why we have

written the controller algorithm in scaled variables. The controller is not
specific to a process; it merely produces a response according to the error
it receives. Hence it is most reasonable to build it so that error and
response are expressed as percentages of a range; whatever y may be as a
physical variable, in Equation (5.10-2) the scaled controlled variable y* is
subtracted from y*sp, which is the preferred position on the scale. The
resulting error İ* is processed by the controller in Equation (5.10-1) to
generate a scaled output x*co. The bias output x*co,b is the controller’s
resting state, achieved when there has been no error.

Integral mode integrates the error, so that the controller output x*co, which
drives the manipulated variable in the loop, increases with the persistence
of error İ*, in addition to its severity. The influence of the integral mode is
set by the magnitude of the integral time TI. In the special case of a
constant error input to the controller, TI is the time in which the controller
output doubles. Thus decreasing TI strengthens the controller response.
Very large TI disables the integral mode, leaving a proportional controller.
The dimensionless controller gain K*c acts on both the proportional and
integral modes.

As an example, let us test a controller in isolation, so that we supply a

controlled variable independently, and the controller output has no effect.
Suppose we set K*c to 1 and TI to 1 minute. Suppose the set point is 40%
and the bias output 50%. Initially, there is no error, and so the controller
output is 50%. Figure 5.10-1 shows the controller response to a step
increase in the controlled variable from 40 to 60%. Because the error
suddenly becomes -20%, the proportional mode calls for the output to
change by -20% (gain of +1). Thus x*co becomes 30%. Over the next
minute, there is no change in the controlled variable. Hence the integral
mode calls for more controller output. In 1 minute, x*co decreases by
another 20% and so reaches 10%.
Lesson 5: Heated Tank
At one minute, we contrive to return the controlled variable to set point, so
that the error is again zero. The proportional mode then ceases to call for
output. However, the integral mode still remembers the earlier error, and
so the output returns to 30%, not the original 50%. We will see later that
when the controller is placed in a feedback loop, not isolated as we have
used it here, the integral mode acts to eliminate offset.








Figure 5.10-1: response of isolated PI controller to an input pulse

For our analytical work, we will want to express the controller algorithm
in deviation variables. We proceed by substituting from definition (5.6-1)
into the algorithm (5.10-1) and (5.10-2).
Lesson 5: Heated Tank
y sp  y min y  y min
H* 100%  100%
y max  y min y max  y min

y  y
y max  y min


y y ' 100% '

y max  y min

y max  y min

We can see that İ* represents the physical variable error İ multiplied by a

ratio of scaling ranges. Continuing with the controller output

x co  0% x  0%
x *co  x *co ,b 100%  co ,b 100%
100  0% 100  0%
x co  x co ,b (5.10-4)
x co  x co ,r
x 'co

In (5.10-4) we recognize that the controller bias is most reasonably

thought of as its value at the reference condition. Also, because controller
output ranges from 0 to 100%, its “physical” value is identical to its scaled
value. Rewriting algorithm (5.10-1)

100% § ·
x '
K *
¨ H  ³ Hdt ¸
co c
'y ©¨ ¸
TI 0 ¹
§ 1
K c ¨¨ H  ³ Hdt ¸¸
© TI 0 ¹

The dimensionless controller gain K*c (the setting that would actually be
found “on” the controller itself) is multiplied by the ratio 100% ǻy-1 to
produce a dimensional quantity Kc. Kc converts the dimensions of the
error İ to the % units of controller output xco. If the error İ were expressed
in the units of a physical variable (a liquid level, for example), ǻy would
perhaps be some number of centimeters. If error were instead expressed
in terms of the output of a signal transducer on the measuring instrument,
ǻy might be in volts or milliamps.

The Laplace transform of controller algorithm (5.10-5) is

Lesson 5: Heated Tank
§ 1 ·
x 'co (s) K c ¨¨ H(s)  H(s) ¸¸ (5.10-6)
© TIs ¹

5.11 step 4 - choose set points and limits

The heated blending tank might require an occasional change of set point,
depending on product grade, time of year, other process conditions, etc.
At present we are presuming that the tank is part of a continuous process,
so that its normal operating mode is to maintain temperature at the set
point. Such a tank could also be used in a batch process, however. In this
case, the set point might be an active function of time, according to the
recipe of the batch.

Limits placed on temperature can be both high and low, depending on the
process. Reasons for imposing high limits are often undesirable chemical
changes: polymerization, product degradation, fouling, side reactions.
Both high and low limits may be imposed to avoid phase changes: boiling
and freezing for liquids.

We institute regulatory control, using, e.g., proportional-integral

controllers, to keep controlled variables within acceptable operating limits.
However, when there are safety limits to be enforced, regulatory control
may be superseded by a safety control system. Thus a reactor temperature
may normally be regulated within operating limits, but some higher value
will trigger an audible alarm in the control room, and some yet higher
value will initiate emergency response or shutdown procedures that
override normal regulatory control.


5.12 the sensor in the feedback control loop

Temperature may be measured with a variety of instruments that respond
to temperature with an electrical signal, including thermocouples,
thermistors, RTDs (resistance thermometry devices), etc. In this section,
we address both the static (calibration) and dynamic (time response)
characteristics of temperature sensors, with reference to our heated tank

Calibration of the sensor is determining the relationship between the actual

quantity of interest (the temperature at some location in the fluid) and the
output given by the sensor (which can be a voltage, a current in a circuit, a
digital representation, etc., depending on the instrument). When we speak
of a sensor, we usually refer to both the sensing element (such as the
bimetallic junction of a thermocouple) and signal conditioning electronics.
It is this latter component that produces a linear relationship between
Lesson 5: Heated Tank
temperature and sensor output, even though the behavior of the sensing
element itself may be nonlinear.

Thus we relate the physical temperature To and its sensor reading Ts by

Ts K s To  b s (5.12-1)

In a handheld digital thermometer, the electronics are adjusted so that gain

Ks is unity and bias bs is zero: 26ºC produces a reading of 26ºC. In a
control loop, however, we are more likely to have To produce an electric
current that ranges over 4 to 20 mA, where these limits correspond to the
expected range of temperature variation. Current loops are a good way to
transmit signals over the sorts of distances that separate operating
processes from their control rooms.

The sensor range is adjusted by varying Ks and bs. For example, suppose
that we wish to follow To over the range 50 to 100ºC. Then

4 mA K s 50 D C  b s
20 mA K s 100 D C  b s (5.12-2)
Ÿ Ks 0.32 mA K 1
bs  12 mA
We express the sensor calibration in deviation variables by subtracting the
reference state from (5.12-1). Suppose we wish to use 75ºC as a reference
operating condition. At the reference, the sensor output will be 12 mA.

Ts' K s To' (5.12-3)

We obtain a 0 to 100% range by taking the scaled variable reference as the

minimum temperature.

* Ts  4 mA To  50 D C
T 100% 100% (5.12-4)
20  4 mA 100  50 D C
From (5.12-4), we see that the scaled sensor output may be defined in
terms of the sensor reading or the controlled variable itself. A temperature
of 75ºC causes a sensor output of 12 mA, or 50% of range.

We now consider the dynamic response of our sensor. Suppose we place

the sensor in fluid at 50ºC and allow it to equilibrate, so that its output is 4
mA. We now move the sensor suddenly to another fluid at 100ºC; how
quickly does the sensor respond to this step input? How long until the
output becomes 20 mA? Classic textbook treatments of thermocouples
Lesson 5: Heated Tank
and thermometers indicate that the sensor response is first-order. Hence,
we may write the sensor transfer function as

Ts' (s) To' (s) (5.12-5)
Ws s  1

In (5.12-5) Ks is the sensor gain that was determined by calibration in

(5.12-2) and (5.12-3). The sensor has a time constant IJs that depends on,
e.g., the mass of the sensor element and the rate of heat transfer to the
sensor. It will follow a step input in Tƍo with the familiar first-order
exponential trace.

However, we plan to immerse the sensor in a large stirred tank. Therefore,

we are unlikely to encounter step changes in To. In fact, it is often the case
that the sensor time constant IJs is small in comparison to the tank time
constant IJ, given in (5.1-7). This implies that Ts will keep up with changes
in To, so that we can simplify (5.12-5) to a pure gain process.

Ts' (s) K s To' (s) (5.12-6)

5.13 the valve in the feedback control loop

Physically, the controller output dictates the opening of a valve that admits
the heating fluid to the heat exchanger. The figure shows a steam supply
header, a line to the heat exchanger, and a steam trap at the exit, which
delivers condensate to the condensate return header.

steam supply header

heat exchanger
bundle in tank

steam trap

condensate return

Viewing the control valve as a dynamic system, we think of the valve

converting the controller output (measured as %out) to the temperature of
the condensing steam (measured in degrees). This is not quite as
farfetched as it may sound:
Lesson 5: Heated Tank
x the controller output (between 0 and 100%) specifies the position of the valve stem
(between closed and open).
x The valve opening varies the resistance presented to the flow.
x Resistance relates any flow through the valve to the pressure drop required by that
x The pressure drop across the valve relates the steam supply pressure to the pressure at
which the steam condenses in the heat exchanger.
x The prevailing pressure in the heat exchanger determines the condensing temperature.
x The condensing temperature determines the rate of heat transfer from vapor to tank.
x The heat transfer rate determines the rate at which steam condenses to supply the
Thus the flow admitted by the valve is the flow that is able to condense at
a temperature high enough to transfer that heat of condensation to the
liquid in the tank. In short, we open the valve to supply more heat.

Let us not become lost among momentum equation, vapor pressure

relationship, empirical flow resistance relationships, etc., which we
invoked above. Our purpose is to describe the action of the controller
upon the condensing temperature, and so we need a gain (from steady-
state relationships) and dynamic description (from the rate processes).
Will a simple first-order description be sufficient?

Tc' (s) x 'co (s) (5.13-1)
W vs  1

We will address this question more thoroughly in Lesson 6. For the

present, while not dismissing our skepticism, we will accept (5.13-1).
That is, we presume that over some range of operating conditions, at least,
the change in condensing temperature is directly proportional to a change
in controller output. The underlined phrase is a key one: our linear (i.e.,
constant gain) model is an approximation, and part of our engineering job
is to determine how far our model can be trusted.

Regarding dynamic response, the first-order response of Tc to xco is also

an approximation. In Lesson 4 we used this description of our valve, and
we saw that our second-order process combined with the valve to produce
a third-order closed loop. In this lesson, however, we have other topics to
explore, so we will presume that the characteristic time IJv for changing the
condensing temperature is much smaller than the mixing tank time
constant IJ, so that

Tc' (s) K v x 'co (s) (5.13-2)

Lesson 5: Heated Tank
We have said, then, that we expect the dynamic response of the closed
loop to depend primarily on the dynamic characteristics of the process,
and very little on the characteristics of sensor (Section 5.12) and valve
(Section 5.13).

5.14 numerical controller calculations

In Lessons 3 and 4, we have expressed controller behavior in
mathematical terms and predicted the closed loop behavior by solving
process and controller equations simultaneously. We posit “step
disturbance” and then make a plot of what the equations say. This
dynamic system representation is useful if it can help us manage real

We understand that our process is actually a tank, but what does the
controller look like? For many years, the controller was a physical box
that manipulated air flow with bellows and dampers; its output was an air
pressure that positioned a valve stem. Coughanowr and Koppel (chap.22)
describe such mechanisms.

Now, the controller is most often a program that runs on a microprocessor.

The program input is numbers that represent physical signals from the
sensor. The algorithm (5.10-1) is computed numerically. The output
numbers are fed to a transducer that makes physical adjustments to a

Whereas our analytical solutions represent continuous connection between

controller and process, the computer program controller samples the
process values at intervals. Marlin (chap.11) discusses the effect of
sampling frequency on controller performance.

A very simple controller program outline is given below:

set up arrays to hold variables
set controller parameters (Kc*, TI)

loop for automatic control

interrogate sensor to determine controlled variable at t_now
use comparator to determine set point and calculate error
use controller algorithm to calculate controller output at t_now
administer controller output to process
update monitor plot with values at t_now
wait for time interval ǻt to pass
set new value of t_now and return to beginning of loop
Lesson 5: Heated Tank
In this course, we will embellish this scheme with coding to simulate the
process, as well, so that we can make numerical simulations of processes
under control.


5.15 closed loop transfer functions

We recall the general feedback block diagram introduced in Lesson 4.
Here we modify it to show our two disturbance inputs, but the control
architecture is unchanged.




x'm(s) y' (s)


control Gv(s) Gs(s)
y's (s)
x'co(s) İ' (s) -

Gsp(s) y'sp,s(s)

From the diagram, our general closed loop model is

G d1 Gd2 G m G v G c G sp
y 'c (s) x 'd1 (s)  x 'd 2 (s)  y ' (s) (5.15-1)
1  G m G v G c G s 1  G m G v G c G s 1  G m G v G c G s sp
We apply (5.15-1) to our heated tank by inserting the particular transfer
functions from Sections 5.2, 5.12, 5.13, and 5.14. Thus the disturbance
transfer function is

T (s)
o KsKcK vK3
T (s) WTI 1  KsKcK vK3
1 s 2  TI s 1
KsKcK vK3 KsKcK vK3

(and similarly for disturbance T2) and the set point transfer function is
Lesson 5: Heated Tank
To' (s) TIs  1
Tsp' (s) WTI 1  KsKcK vK3
s 2  TI s 1
KsKcK vK3 KsKcK vK3

Adding integral control to a first order process has resulted in a closed

loop with second-order dynamics.

5.16 closed-loop behavior - set point step response

Responses to disturbance inputs and set point changes will depend on the
poles of (5.15-3). These are

§ 4W ·
 1  K s K c K v K 3 r 1  ¨¨ 2  ¸¸K s K c K v K 3  K s K c K v K 3

© TI ¹
s1 , s 2 (5.16-1)

We observe that the poles could be complex, so that the closed loop
response could be oscillatory. The tendency toward a negative square
root, and thus oscillation, is exacerbated by reducing the integral time TI.
We also observe that the real part of the poles is negative, indicating a
stable system.

We illustrate set point step response for real poles, where the set point is
changed by magnitude ǻT:

ª 1 1 º
«  TI  TI »
s1 s2
To' 'T «1  es1t  es 2 t » (5.16-2)
« 1 1 1 1 »
«¬ s 2 s1 s 2 s1 »¼

The response is written in terms of poles s1 and s2. Because they are
negative, the two exponential terms decay in time, leaving the long-term
change in set point as ǻT. Thus we requested that the tank temperature
change by ǻT, and the tank temperature changed by ǻT. There is no
offset - this is the contribution of the integral mode of control.

5.17 second order systems

In Lesson 4, we described the two tanks in series as an overdamped
second-order system; now our heated tank with integral-mode control is
also second order, but perhaps not overdamped. We defer further
exploration of our closed loop behavior until we learn more about the
properties of second-order systems. Thus we consider
Lesson 5: Heated Tank
d2y dy dx
D 2
 E  y( t ) Kx ( t )  M
dt dt dt
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 Į and ȕ.

We solve (5.17-1) by Laplace transform.

y(s) Ms  K
x (s) Ds 2  E s  1

The behavior of the solution depends fundamentally on the poles of (5.17-

2). From the properties of the quadratic equation we recall that the poles
are real for

D (5.17-3)

A map relating system stability to coefficients Į and ȕ is given in Figure

Lesson 5: Heated Tank

unequal, negative
poles: stable, non-
2 equal, negative
poles: stable, non-
complex negative
poles: stable,
unequal poles of
opposite sign:

0 excursively
unstable complex positive imaginary poles:
poles: unstable, persistent
-1 oscillating oscillation

equal, positive
poles: excursively
-2 unstable
unequal, positive
poles: excursively
-1.6 -1.2 -0.8 -0.4 0 0.4 0.8 1.2 1.6 2

Figure 5.17-1. Second-order system stability related to equation coefficients

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 IJ1 and IJ2 may be positive or negative.

E r E 2  4D
W1 , W 2 (5.17-4)

We compare side-by-side second-order and first-order systems:

Table 5.17-1. Second-order system (IJ1, IJ2 form) – real poles

second-order first-order
equation dy
W1W 2 2  W1  W 2  y( t ) Kx ( t )  M
d y dy dx
W  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) W1s  1 W 2s  1 x (s) Ws  1
Poles 1 1 1
s  , s 
W1 W 2 W
Lesson 5: Heated Tank
Zeroes K none

The system will be stable if both IJ1 and IJ2 are positive, as they were for the
mixing tanks of Lesson 4. A second-order step response is

ª §  t  t d  t  t d º
M · W1 § M · W2
y( t ) AKU t  t d «1  ¨¨1  ¸¸ e W1
 ¨¨1  ¸¸ e W2
» (5.17-5)
¬ © KW1 ¹ W1  W 2 © KW 2 ¹ W1  W 2 ¼

When Į, the coefficient of the second derivative, is positive (right-hand-

side of Figure 5.17-1), it is customary to express the second-order system
in an alternative way:

Table 5.17-2. Second-order system (IJ, ȟ form) – complex poles

second-order first-order
equation d y dy dx dy
W 2 2  2[W  y( t ) M  Kx ( t ) W  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 x (s) W s  2[Ws  1
2 2
x (s) Ws  1
poles 1
s  [ r [2 1
zeroes K none

Relationships between the characteristic times IJ1 and IJ2 and the alternative

parameters IJ and ȟ are

[  [ 1 2

W [  [2  1 (5.17-6)

[  [ 1 2

W [  [2  1 (5.17-7)

W W1W2 (5.17-8)

W1  W 2
[ (5.17-9)
2 W1W 2

W1  W2 2W[ (5.17-10)
Lesson 5: Heated Tank

W1  W 2 2W [ 2  1 (5.17-11)

As in the first-order system, IJ has the dimension of time and represents a

characteristic time of the system. We will take it to be positive;
dimensionless ȟ in the second-order system, however, we will allow to be
positive or negative. Parameter [ is called the damping coefficient, and
the character of the response depends markedly on its value, as we will
explore in the next few sections.

5.18 overdamped systems: [ > 1

The poles are real, negative, and unequal. We have seen such a system in
Lesson 4, a system of unequal tanks in series. The overdamped response
to an input step of magnitude A is

ª ­§

· [  [  1 [ 1 t  t d W ½º
2 2
« °¨1  [  [ 1 ¸
e °»
«  [ t  t d °© KW ¹ 2 [2 1 °»
y( t ) AKU t  t d «1  e W
® ¾» (5.18-1)
« ° §

· [  [  1  [ 1 t  t d W °»
2 2
°  ¨ 1  KW [  [  1 ¸
« e °»
«¬ ¯ © ¹ 2 [2 1 ¿»¼

Figure 5.18-1 shows step responses for various ȟ, with M = 0. In contrast

to the first-order response, the trace has a sigmoid shape: a slow start and
an inflection point. Larger ȟ make the response more sluggish.

5.19 critically damped systems: [ = 1

For this special condition, the poles are real, negative, and equal. This is
the case for two identical tanks in series. The step response is

ª  t  t d ­
§ M · t  t d ½º
y( t ) AKU t  t d «1  e W
®1  ¨1  ¸ ¾» (5.19-1)
¬ ¯ © KW ¹ W ¿¼
Lesson 5: Heated Tank







0 2 4 6 8 10 12

Figure 5.18-1. Step response for overdamped second-order system with M = 0.

5.20 underdamped systems: [ < 1

The poles are complex conjugates, and the system will oscillate even for a
non-periodic disturbance. The overdamped response (5.18-1) may be
modified by substituting

[2  1
 1  [2 j 1  [2 (5.20-1)

and using Euler’s relation to find

ª ­ M ½º
2 t  t d KW sin 1  [ 2 t  t d °» (5.20-2)
« [ t  t d °
y( t ) AKU t  t d «1  e W
®cos 1  [  ¾»
« ° W 1  [2 W °»
¬« ¯ ¿¼»
Lesson 5: Heated Tank
or equivalently

ª § 2
2 º
« §
¨ 1  [2  ¨ [  M · ¸ ­ ½ »
AKU t  t d «1  e
[ t  t d
W ¨ © KW ¹ ¸ °
sin ® 1  [ 2
t  t  tan 1
1  [ 2
° »
M °»
y( t ) ¨ ¸
« 1 [ 2
° W [ »
¨ ¸
« ¨ ¸ ¯ K W ¿ »
¬ © ¹ ¼

5.21 unstable systems: [ < 0

When the damping factor is less than zero, the exponential terms in (5.20-
2) and (5.20-3) increase with time, and the system is unstable to

5.22 inverse responses

When the disturbance rate-of-change factor M is non-zero, and of opposite
sign to the gain K, the system can show an inverse response; that it, the
initial response of the system is opposite to its ultimate direction. Figure
5.22-1 shows a second-order response that initially is negative, but
ultimately oscillates around a positive change. The usual textbook
example is that of level control in a boiler – adding water will initially
lower the liquid level, as measured by the sensor, because bubbling is
suppressed. Such inverse responses present challenges to controllers,
because they start off by making the problem worse.
Lesson 5: Heated Tank

Step Response



x(t) and y(t)




0 2 4 6 8 10 12 14 16


input pole 1 first order pole 2 first order response

Figure 5.22-1. Second-order system; inverse response.

5.23 general map for second order

Figure 5.23-1 relates the qualitative behavior of a second-order system to
its damping coefficient and MK-1IJ-1, a group that is the reciprocal of the
product of the time constant and the transfer function zero. The major
divisions of behavior correspond to the value of ȟ, but various other
features depend on this other group.
Lesson 5: Heated Tank

Solution Characteristic Map

4 overshoot
reciprocal zero*tau

pseudo 1st order

inverse inverse
-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
damping coefficient

Figure 5.23-1. Map relating character of step response to model parameters

5.24 Measures that characterize the underdamped step response

Oscillatory responses are common. We have focused on second-order as a
type for an oscillatory system: we have derived equations and inferred
behavior. However, much process control work is conducted in the
opposite direction: we observe oscillatory sensor readings, and we must
try to infer a workable system model, as well as diagnose faults.

Practitioners use several measures to characterize an observed response

(Coughanowr and Koppel, 1965). For the particular case of the 2nd order
system, we can supplement these definitions with equations, but for an
unknown system we can only compute the measures from recorded data.
Lesson 5: Heated Tank

1.6 p
normalized response

0 rise time response time
0 2 4 6 8 10 12 14 16
normalized time

Figure 5.24-1. Oscillatory response to step input.

overshoot – response exceeds the ultimate value; equal to A/B

1 [ 2
overshoot e

decay ratio – ratio of successive peaks, equal to C/A

2 S[
2 1 [ 2
decay ratio overshoot e

rise time – time to first reach the ultimate value

response time – time until response remains within r5% of ultimate value

period – time between peaks, or between alternate crossings of the

ultimate value.

§ 1 2S · 2SW
p¨ ¸
© f Z¹ 1  [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 W is directly proportional to the
natural period. Notice also that damping lengthens the period.

5.25 disturbance response for the heated tank

Armed with more knowledge of second order systems, we see that
allowing the response to be underdamped will return the controlled
Lesson 5: Heated Tank
variable to the vicinity of the set point more quickly than in an
overdamped response. However, if the damping coefficient becomes too
small, the oscillation amplitude and persistence may be unacceptable. We
return to the disturbance transfer function in (5.15-2) and recognize the
characteristic time and damping coefficient.

Wcl (5.25-1)
KsKcK vK3

1  KsKcK vK3 TI
[ (5.25-2)
2 WK s K c K v K 3

From (5.25-1) we see that increasing controller gain Kc and decreasing the
integral time TI tend to speed the loop response. Both these adjustments
move the controller in the direction of aggressive tuning. The results of
aggressive tuning are mixed on the damping coefficient - decreasing TI
increases oscillation, but increasing Kc suppresses it. The lower limit of ȟ,
however, is zero, so our second-order closed loop can be unstable
(theoretically) only in the limit of zero integral time.

For a step disturbance in T1 of magnitude ǻT, we find from (5.15-2),

(5.25-1), and (5.25-2)

K s K c K v K 3 'T
To' (s) (5.25-3)
W cl2 s2  2W cl[s  1 s

which may be inverted to give

K1TI 'T 1 t
To' ( t ) e W cl sin 1  [ 2 (5.25-4)
K s K c K v K 3 W cl 1  [ 2 W cl

Equation (5.25-4) shows that the response oscillates about, and decays to,
zero; as with the set point response we calculated in (5.16-2), there is no
offset in the controlled variable, in spite of the permanent change in input.
Thus integral-mode control has improved our ability to control the outlet

Figure 5.25-1 shows responses for several controller tunings – that is,
several choices of parameters Kc (represented within the loop gain K) and
TI (scaled to the process time constant). Upon reducing the integral time,
we reduce the amplitude of the error but undergo more oscillation. By
increasing the gain, we speed the decay and thus reduce both the
Lesson 5: Heated Tank
amplitude of the error and time spent away from set point: a 10 K inlet
disturbance affects the outlet temperature by less than 1 K.


T'o (K)


0 100 200 300 400
t (s)

K=1 TI/tau=0.5 K=1 TI/tau=0.1 K=3 TI/tau=0.1

Figure 5.25-1. Step response of heated tank under PI control.

5.26 tuning the controller - measures of performance

In Lesson 4 we tuned by keeping our distance from the stability boundary,
but did not consider whether performance was satisfactory. In Section
5.25 we saw that we could adjust two controller parameters independently
to affect the response, and that there was a trade-off between amplitude
and oscillation. It is time to introduce standard measures for the
controlled variable that allow us to compare different tunings

integral error (IE) -- can be deceptively small if errors are balanced

IE ³ H ( t)dt

integral absolute error (IAE) -- accounts for any deviation

IAE ³ H ( t ) dt

integral square error (ISE) -- emphasizes large errors

Lesson 5: Heated Tank

³ >H ( t )@ dt
ISE (5.26-3)

integral time absolute error (ITAE) -- emphasizes persistent errors

ITAE ³ t H ( t) dt

Of course, these integral error measures can be defined for scaled error İ*,
as well. The latter three will always increase as the controlled variable
spends time away from the set point, so in general, smaller means better
control. The measures can be calculated from plant data to compare the
results of different tunings. They can also be used to compare the results
of simulations. For example, we could calculate IAE for the three traces
in Figure 5.25-1.

Furthermore, we could optimize by seeking a combination of tuning

parameters that minimized one of the error measures. Thus, the controller
would be tuned by an objective performance criterion. For response
(5.25-4), of course, such an optimization would be expected to lead to
infinite gain and zero integral time, because (as we determined by
examining the poles of the transfer function in Section 5.16) the closed
loop is stable for all tunings. In a less idealized system, however, an
optimum is more likely to exist.

5.27 stability by Bode criterion

We were assured of closed loop stability in Section 5.16. Even so, we
examine the frequency response of the loop transfer function.

KsK cK vK3 § 1 ·
GL ¨¨1  ¸ (5.26-1)
Ws  1 © TIs ¸¹

From the table of frequency response components in Marlin (Sec.10.6) we

1 2 2
R A K sK cK vK 3 (5.26-2)
1  W 2 Z2


§ 1 ·
I tan 1  ZW  tan 1 ¨¨ ¸¸ (5.26-3)
© ZTI ¹
Lesson 5: Heated Tank

The Bode plot is made for three ratios of tank time constant IJ to controller
setting TI. For the least aggressive tuning with large TI, the response
resembles that of a first order system, although we notice different
behavior at low frequencies, due to the integral controller. Decreasing the
integral time depresses the phase angle toward a uniform -90º and
increases the low-frequency amplitude ratio. The amplitude ratio on the
plot is normalized by the loop gain; high controller gain settings would
directly increase the amplitude of low-frequency disturbances.

amplitude ratio/gain


0.01 0.1 1 10 100

-30 0.01
phase angle (deg)


IJ/TI = 1


0.01 0.1 1 10 100
ȦIJ (radians)

Because the phase lag never reaches -180º, the closed loop will remain
stable. However, we see that the integral mode contributes phase lag at
low frequencies, as well as boosts the amplitude ratio. Combined with
other lags in a closed loop, integral control would tend to destabilize the

5.28 numerical simulation of closed loop response

Numerical simulation of a first-order process under PI control was
performed for a long pulse disturbance. The traces appear in scaled
variables. The response is slightly underdamped.
Lesson 5: Heated Tank
Lesson 5: Heated Tank
5.29 conclusion
Integral control has been a big help by removing offset. However, it
contributes phase lag to the closed loop, particularly at low frequencies,
where it also boosts the amplitude ratio. Having two control parameters
has allowed us more influence over the shape of the response.

5.30 references
Coughanowr, Donald R., and Lowell B. Koppel. Process Systems Analysis and Control.
New York, NY: McGraw-Hill, 1965. ISBN: 0070132100.
Marlin, Thomas E. Process Control. 2nd ed. Boston, MA: McGraw-Hill, 2000. ISBN: 0070393621.
Lesson 5: Operability of Processes
5.0 context and direction
In Lesson 4, we encountered instability. We think of stability as a
mathematical property of our linear system models. Now we will embed
this mathematical notion within the practical context of process
operability. That is, we must not forget that our system models help us
operate processes. Along the way, we will encounter a special category of
instability/inoperability: the non-self-regulating process.


5.1 remember the stability criterion for linear systems

In Section 4.9, we introduced a stability criterion for a linear system: non-
negative poles in the transfer function (5.1-1) indicate that the system
output y(t) will not remain stable in response to a system input x(t).

y ' (s) 1
G (s) n 1
x ' (s) a n s  a n 1s
   a 1s  1

As a simple example, consider a first-order system:

W  yc Kxc yc 0 0 (5.1-2)

We know that the Laplace transform representation is completely


yc s xc s (5.1-3)
Ws  1

The transfer function in (5.1-3) has a single pole at -IJ-1. If the time
constant IJ is a positive quantity (as in our tank), the pole is negative and
the response is stable (as we have seen in Lesson 3).

If the time constant were a negative quantity, however, the pole would be
positive. As we saw in Section 4.9, the response would be unstable
because of the exponential term in the solution of (5.1-2)

yc(t) e W

This unbounded response could be in a positive or negative direction,

depending on the sign of the gain K. We will address on another occasion
what sort of system might have a negative time constant; for now we
recognize that encountering an unstable linear system should cause us to
look carefully at the process whose behavior it represents.
Lesson 5: Operability of Processes
5.2 remember that feedback control can make stable systems go unstable
Recall from Section 4.21 that we took a perfectly stable second-order
process, placed it in a feedback loop with a first-order valve, and applied
proportional-mode control. By increasing the controller gain too far, we
could drive the system to instability. We could use our linear stability
criterion to predict the onset of instability as we did in Section 5.1. That
is, we compute the poles of … not the process transfer function, but the
transfer function that represents the process in feedback loop with other

5.3 the special case of zero poles

Now consider a system with a single pole whose value is zero.

yc s xc s (5.3-1)

This is a non-negative pole; we claim this indicates an unstable system. If

we apply a step disturbance to (5.3-1), we obtain upon inversion:

yc K (5.3-2)

Certainly y’ increases without bound, so that it qualifies as unstable. You

should try different bounded disturbances in (5.3-1), though, to explore
whether this is always the case.

5.4 the integrator and the non-self-regulating process

The system of Section 5.3, far from being an oddball case, is actually quite
important. It is known as an integrator. An example is shown in Figure



Figure 5.4-1: Tank with independent inlet and outlet flows

The inlet flow is simply given, out of our control, and the outlet is
pumped. The material balance gives
Lesson 5: Operability of Processes
A Fi  Fo (5.4-1)

Both inlet and outlet flow disturbances affect the tank level h. We
envision a steady reference condition in which the flows are balanced at Fr
and the level is at hr. Expressing (5.4-1) in deviation variables, we obtain

A Fic  Foc (5.4-2)

Integrating (5.4-2) from an initial deviation of zero, we obtain

1 t
Fic  Foc dt
A ³0
hc (5.4-3)

Thus the name integrator: the response variable is simply the integral of
whatever inputs are fed to it. This can be a big problem. Suppose that this
is your tank. You observe that Fi is running quite steadily, so you adjust
Fo to match it and go home for the night. Just after you leave, Fi increases
to a new steady value. We derive the results from (5.4-3):

Fic  Foc
hc t (5.4-4)

In this simple case, the numerator is a positive constant. You can

substitute numbers into (5.4-4) to help you estimate what time of night
you will receive a telephone call.

The integrator has a zero pole; if we generalize from (5.4-2) we see a

corollary of this property:

z f y (5.4-5)

Recall that y is a quantity that can accumulate in a system; it could be total

mass, amount of a chemical species, or energy, for example. Equation
(5.4-5) tells us that the rate of accumulation of y is not affected by the
inventory: your tank level may be rising fast, but the flow keeps on
coming. An equivalent statement of (5.4-5) is that the net inflow of y
across the system boundary does not depend on how much y is within the

Processes that contain an integrator are known as non-self-regulating.

This term comes directly from the property expressed by (5.4-5). By
contrast, the first order system in (5.1-2) will come to a new steady value
Lesson 5: Operability of Processes
after a disturbance. It can do this because its rate of accumulation does
depend on the inventory y, and we call it self-regulating. Clearly a non-
self-regulating process will require frequent attention.

5.5 getting practical - when self-regulation is not

We have identified formal stability with negative poles in the transfer
function. Of the non-negative poles, we have given a special name -
integrator - to systems with zero poles, and we know to watch out for
them. These results of linear stability theory are quite useful. However,
we must never forget that our models are only approximations of the real

As an illustration, we cite Shinskey (2002), who asserts that non-self-

regulating behavior is a matter of degree, not strictly confined to systems
with zero poles. To see this, we rearrange the first-order system of (5.1-

W dyc 1
 yc xc yc 0 0 (5.5-1)
K dt K

We suppose that a particular system features a time constant and gain that
are both large. Under these circumstances, the term involving the
inventory y’ might be neglected in comparison to the others, leaving

W dyc
xc yc 0 0 (5.5-2)
K dt

In the limit of (5.5-2), our system is an integrator. The fact that it is not
strictly an integrator is irrelevant if its behavior approaches that of (5.5-2):
in the matter of midnight phone calls, the process might well be
considered non-self-regulating.

Let us illustrate by a tank with gravity-driven outlet flow.



Figure 5.5-1: Tank with gravity-driven outlet flow

Lesson 5: Operability of Processes
We take the simplest course and assume that the outlet flow is directly
proportional to the liquid level h. The proportionality constant k is small
when the outlet valve is constricted, and increases as the valve is opened.
A material balance gives

A dhc 1
 hc Fic (5.5-3)
k dt k

where we base deviation variables on a steady reference condition.

Formally, (5.5-3) represents a self-regulating process. However, consider

for a tightly constricted valve, k is small. In system terms, this means that
gain 1/k and time constant A/k are both large. The large gain means that
small disturbances Fi’ have a large effect on response h’. The large time
constant means that the response will be slow.

In physical terms, the constricted outlet forces the liquid level to rise
significantly to respond to relatively small increases in the inlet flow. It
will take a while to reach a level sufficient to push the flow out, so that
there will have been a lot of accumulation in the tank. If the constriction
is sufficiently severe, you will still receive your midnight telephone call,
even though your linear system model (5.5-3) has a negative pole. The
process is effectively non-self-regulating.

5.6 we want our processes to be operable

When we speak of “instability”, we speak of unbounded change in the
response variable, as in (5.1-4), (5.3-2), or (5.4-4). Of course, nothing
really became unbounded: the tank level rose only until it spilled over, or
drained until it was empty. What really matters to us is process
operability. A process that wakes us at midnight is not operable; it will
always need attention, or else run into troublesome limits. We prefer that
our processes be operable, and process control should contribute to that
end; therefore, a non-self-regulating process should be placed under
automatic control.

Linear system stability calculations are useful indications of operability;

however, Section 5.5 taught us that operability depends on the context of
the operation. Thus “operability” is “stability” made practical. A control
engineer must ask, “Under what conditions can this process become

x Part of the answer may come from applying stability theory, either
through formal calculations, or by knowing what to look for while
tuning a controller. Most practical control loops will have a stability
limit. Stability calculations in Section 4.21 helped us select a
Lesson 5: Operability of Processes
controller gain for operability. In this way, we seek to determine a
degree of control.
x Another part of the answer may come from an intuitive examination of
the process, seeking to understand what affects the inventory of each
conserved quantity (mass, species, energy). This is especially useful
for processes that contain multiple operations. Such an examination
can reveal both where control is needed, and where control might
cause a problem. Here we are seeking a scheme of control.

Engineering often requires a blend of rigorous analysis and informed

intuition. For the remainder of Lesson 5 we will examine several process
features that affect operability. Much of this treatment is based on Downs

5.7 we must control the liquid level integrator

Suppose that we desire to maintain a constant level in the tank of Figure
5.4-1. To make this process self-regulating, we must apply automatic
control. We must choose a control scheme; perhaps conditioned by
steady-state thinking, we consider managing inventory by managing the
inlet and outlet flows: after all, doesn’t IN = OUT at steady state? We
arrange in Figure 5.7-1 to measure the instantaneous inlet flow and adjust
the outlet flow to match.



Figure 5.7-1: Feedforward control of inventory by manipulating

outlet flow

Such a scheme may allow you to stay at home all night, because it will
tend to reduce imbalance between flows. However, any discrepancy
between the flow rates - through calibration error, instrument drift,
shortcomings of the controller, insufficient adjustment of the pump motor
speed - will contribute to accumulation through (5.4-3). If that
discrepancy is sufficiently biased, over time the level will creep to an
undesired value.
Lesson 5: Operability of Processes
If our objective is to control level, then we should measure the level! The
scheme of Figure 5.7-1 did not correct the basic fault of the integrator; in
Figure 5.7-2 we apply a feedback loop to allow the inventory to affect the
rate of accumulation. Thus the process (under control) becomes self-
regulating: a rise in level will trigger an increase in the outlet flow rate,
and so forth.



Figure 5.7-2: Feedback control of inventory

(By the way, in a later lesson we will return to feedforward control

schemes and find them to be very useful when appropriately applied.)

5.8 inventory of individual chemical components

Downs (1992) recommends that a key to achieving good process
operability is ensuring that the inventory of each chemical species is self-
regulating. Thus we consider further the tank of Figure 5.7-2, letting the
inlet flow comprise chemical species A and B. We write the component
material balance for A, assuming that the tank is well-mixed:

VCAo i Ai  Fo C Ao
FC (5.8-1)

We saw this relationship in Lesson 3. Here, we notice particularly that the

inventory of component A (VCAo) does influence the outlet path of A
(FoCAo) through the composition, so that (5.4-5) does not apply: the
system is self-regulating. This is the behavior we recall from the
disturbance responses we computed in Lesson 3.

When we explore more extensive processes, we may not be able to write

the component balance in such concise form as (5.8-1). However, we can
still apply the principle of (5.4-5) and examine qualitatively whether the
inlet and outlet paths for a component depend on its inventory.

5.9 example of a solvent recovery process

Figure 5.9-1 shows a scrubber and distillation column used to recover
volatile solvent A from an inert gas stream. We have already installed
Lesson 5: Operability of Processes
plausible level control schemes (indicated by feedback loops with the
control valves omitted) to ensure that individual vessels are self-regulating
with respect to mass.







Figure 5.9-1 Basic process scheme for recovery of solvent A

We proceed methodically:
examine scrubber for component A
x inlet streams are independent of the inventory of A
x overhead gas stream is independent of A
x flow of A in the bottoms stream will increase as the bottoms
concentration of A rises, as in Section 5.8
By virtue of the bottoms stream, the scrubber is self-regulating for
component A. As a practical illustration, a rise in the inlet composition of
A will, over time, result in a higher flow of A from the bottom of the
scrubber. (We remark here that this is a qualitative assessment that
presumes good equipment operation, such as sufficient scrubber
performance and a well-tuned level-control loop. Our purpose here is not
to perform a detailed design or simulation; rather, it is to determine
whether this process can, in principle, be self-regulating.)

examine scrubber for water

x the conclusion for water is identical to that for A

examine scrubber for N2

Lesson 5: Operability of Processes
x the inventory of N2 is indicated by the pressure in the vessel. Both
inlet and outlet flows of gas will be affected by the vessel pressure.
Hence, the scrubber is self-regulating for N2.

examine the distillation unit (including accumulator)

x the flow of components in the bottoms and distillate streams
depend on the composition of the column sump and accumulator
As with the tank of Section 5.8, the level-control loops make the
distillation unit self-regulating for both components.

examine the entire process

x the two inlet streams are independent of the process
x the scrubber overhead gas stream regulates N2
x the distillate and bottoms streams regulate water and component A
The two unit operations are simply arranged in series, so that nothing
about their combination affects the conclusions we had drawn from the
individual units.

5.10 the process becomes non-self-regulating

A reasonable process objective would be to recover component A in a
desired, not arbitrary, concentration. Hence, we install a composition
controller on the distillate stream. In addition, we can reduce utility costs
by recycling the bottoms stream to the scrubber. The results of these
changes are shown in Figure 5.10-1.


water water



Lesson 5: Operability of Processes
Figure 5.10-1 Recovery process with composition control and recycle

We repeat the examination of Section 5.9. We see quickly that the

scrubber is unaffected: addition of a new independent inflow stream does
not change the ability of the level-control loop to regulate the inventories
of A and water.

examine the distillation unit (including accumulator) for A

x the inlet stream is independent of conditions in the unit
x the distillate composition is now fixed, but the removal of A can
still vary in response to conditions because the distillate flow can
Hence distillation is self-regulating for component A. For example,
should the inflow of A increase to the column, the distillate flow can
increase to remove it.

examine the distillation unit (including accumulator) for water

x the inlet stream is independent of conditions in the unit
x fixing the distillate composition means that water can be removed
overhead only at a rate in proportion to the removal of A. Hence,
this stream no longer depends on the inventory of water.
x water may still be removed in the bottoms, as required.
By virtue of the level-control loop on the bottoms, distillation is self-
regulating for water.

examine the entire process

x the two inlet streams are independent of the process
x the scrubber overhead gas stream regulates N2
x the distillate stream regulates component A
x the distillate stream does not regulate water.

Although each individual unit operation is satisfactory, the combination

can no longer regulate the inventory of water. Water enters
independently; water leaves only in proportion to the flow of A. To
illustrate, suppose there is a step increase in water feed. The scrubber
level controller responds by increasing the feed to the distillation column.
The column responds by increasing the bottoms flow, and thus the recycle
to the scrubber. The two sump controllers engage in a battle that one must
lose, as the water inventory continues to rise.

To correct this problem, we must make at least one of these streams

respond to the amount of water in the process. Downs (1992) gives the
solution: as shown in Figure 5.10-2, control the scrubber sump with the
water make-up flow. If we are designers, we still have further decisions to
make about equipment sizing and control structure, but at this point we
Lesson 5: Operability of Processes
have avoided one problem, at least, that no amount of post-design
controller tuning could have solved.


water water




Figure 5.10-2 Control scrubber inventory with make-up water stream

Sections 5.9 and 5.10 have presented a lengthy example. Let us

summarize the approach that was illustrated:
x ensure that overall mass inventory is regulated
x methodically examine each inlet and outlet stream for each process
unit, as well as the whole process, with respect to each chemical
species. Remember that good individual units can work poorly in
x determine the qualitative relationship between the flow of each
component and the inventory of that component. Certainly, there
may be no relationship.
x we call the relationship qualitative because we are not considering
the actual timing or magnitude of response to disturbances, nor
details of equipment performance (e.g., whether the trays might
flood, or if there is enough packing to remove all the solvent from
the gas). These are matters of detailed design, to be addressed
after we have established our basic process structure. Therefore,
presume that all equipment works well, including any control loops
that have already been specified.
x look out for composition controllers, because these restrict the flow
of the impurity components. Look out for recycle streams, because
Lesson 5: Operability of Processes
these eliminate an outlet path. These are not bad process features,
but the control scheme must accommodate them.
x if any component has no stream that responds to its inventory, the
process must be altered to supply one. Often this is accomplished
by a change in control scheme.

5.11 chemical reactions provide paths in and out of a process

A reactant may not leave a process by any outlet flow path at all, but
rather be entirely consumed within it. Therefore, in applying the method
of Section 5.10 to a process with chemical reactors, we must extend the
terms ‘stream’ and ‘flow’ to include consumption and generation. The
general principle of (5.4-5) still applies: if the rate of reaction of a species
does not depend on the inventory of that species, then the process may be

Kinetic expressions most often do depend on the concentration, so that

problem may actually arise outside the reactor, and because of the control
scheme. Downs (1992) gives an example of a reactor whose level,
temperature, and composition are all controlled. The level is controlled by
manipulating the outflow stream. The temperature is controlled by
manipulating flow of a service fluid. The composition is controlled by
manipulating the flow of a reactant recycle stream.

feed recycle

Figure 5.11-1 Reactor with three control loops

Let us examine each in/out path for a reactant A:

x the inflow of A in the feed is independent
x the inflow of A in the recycle stream does depend on the inventory
of A, because it is manipulated to keep the concentration at set
x the outflow of A varies to keep the inventory constant, because the
level controller manipulates the flow rate
Lesson 5: Operability of Processes
x the consumption of A depends on the volume of the reactor (set by
the level controller), the temperature (set by the temperature
controller), and the concentration of A (set by the composition
controller). For well-behaved control loops, it is constant.
Everything seems fine. In fact, because of the level and composition
controllers, we are actively working to keep the inventory of A constant.

The problem does not become apparent until we consider the reactor in the
context of other process units. Suppose the concentration of A in the feed
stream undergoes a step reduction. The level and composition controllers
respond: the composition controller increases the supply of recycle to
return the total inflow of A to normal, and the level controller adjusts the
outlet flow as needed to keep the volume the same. Thus A reacts at the
normal rate, and therefore excess A departs the reactor at the normal rate.

Somewhere downstream, a separation unit is being fed A at the normal

rate, but being asked to return A at a higher than normal rate. It can do
this only while its inventory lasts. The reactor has put the problem off
onto other units, but there remains a problem nonetheless. We discover it
by examining the component inventories, but doing so in the context of
the larger process.

How would you solve this problem?

5.12 conclusion
We enlarged our notions of stability to develop a concept of what makes a
process operable. In doing this, we have ranged qualitatively over a
variety of chemical processes. In some cases these have included multiple
operations and multiple control loops. In the next lessons, we will return
to analysis of a single operation; as we deepen our understanding, try not
to forget the broad perspective we have attempted here.

5.13 references
Downs, J. J. "Distillation Control in a Plant-wide Control Environment." In
Practical Distillation Control. Edited by W. L. Luyben. New York, NY: Van
Nostrand Reinhold, 1992. ISBN: 0442006012.

Shinskey, Greg F. Process Control Systems: Application, Design, and Tuning.

4th ed. New York: NY: McGraw-Hill, 1996. ISBN: 0070571015.
Lesson 6: Exothermic Tank Reactor
6.0 context and direction
A tank reactor with an exothermic reaction requires a more elaborate
system model, because its outputs are temperature and composition.
Furthermore, the model is nonlinear, which forces us to make a linear
approximation to solve it. We will add the derivative mode to our PI
controller to increase both stability and responsiveness. The closed loop
will show how automatic control can stabilize an inherently unstable


6.1 exothermic chemical reaction in a stirred tank reactor

A second-order dimerization reaction occurs in an overflow stirred tank
reactor. The reactor is equipped with a heat transfer surface (perhaps
jacket, coils, or bayonet) that contains a flow of cooling water. We wish
to know how the outlet composition and temperature may vary with time.

6.2 dynamic model of the reactor

With two output variables, we face two balances, as well as several
supporting relationships. The mole balance on the reactant A

FC Ai  FC A  V  rA
dC A
V (6.2-1)

requires a second-order kinetic rate expression for the rate of

disappearance of A, including Arrhenius temperature dependence.

§ 1 dN A · E
 rA ¨  ¸ kC 2A k oe RT
C 2A (6.2-2)
© V dt ¹

The energy balance must account for the reaction and heat transfer.

FUC p (Ti  Tref )  FUC p (T  Tref )  'H R V  rA  Q

V UC p (6.2-3)

Once again, we will regard physical properties as independent of

temperature. Enthalpies are defined with respect to an arbitrary
thermodynamic reference temperature. For an exothermic reaction, the
heat of reaction ǻHR will be a negative quantity, and will thus tend to raise
the reactor temperature T. The rate of heat transfer Q depends on the
logarithmic temperature difference

(T  Tci )  (T  Tco )
Q UoAo (6.2-4)
T  Tci
T  Tco
Lesson 6: Exothermic Tank Reactor

in which the well-mixed tank temperature T is uniform and the coolant

temperature Tc varies from inlet to outlet. We will presume that the
coolant supply temperature Tci is quite stable and thus not consider it as a
disturbance. The overall heat transfer coefficient depends on the film
coefficients on the inner and outer surfaces of the heat transfer barrier; we
will neglect any conduction resistance in the barrier itself.

§ 1 A ·
Uo ¨¨  o ¸¸ (6.2-5)
© h o Aih i ¹

The outer film coefficient ho depends on the rate of stirring in the tank, as
well as the variation of physical properties with temperature. With
constant physical properties, there is no reason for ho to vary. Inner
coefficient hi depends on the flow of coolant. Invoking typical internal-
flow behavior, we write

h i Di
Re n Pr m (6.2-6)

If we write (6.2-6) at a reference condition, we can express the flow

dependence of hi as
§F ·
hi h ir ¨¨ c ¸¸ (6.2-7)
© Fcr ¹

For flow in tubes, n is often about 0.8.

The main structure of the model is given by the balances (6.2-1) and (6.2-
3). These relate the outlet temperature and composition to their inlet
values. Supplementary equations are needed to describe the reaction
kinetics and heat transfer. We see that the two balances will be coupled
through the temperature dependence of the reaction rate parameter k in
(6.2-2). Heat transfer is described by equipment performance equation
(6.2-4) and the empirical relationship (6.2-7) that describes convective
heat transfer in conduits. These latter equations show how the coolant
flow Fc influences the reactor outlet temperature T.

Even so, we are not finished, because we have not yet accounted for the
outlet coolant temperature Tco in (6.2-4). Therefore, we must write an
energy balance on the coolant.

d Tc
VcUc C pc FcUc C pc (Tci  Tref )  FcUc C pc (Tco  Tref )  Q (6.2-8)
Lesson 6: Exothermic Tank Reactor

where ‹T c› is the average coolant temperature in the coolant volume.

Vc ³ T dV
c (6.2-9)

To proceed with (6.2-8), we must express the average temperature in

terms of the inlet and outlet temperatures. This would be an entertaining
diversion, but for the primary purposes of Lesson 6 we will assume that
heat exchanger outlet temperature adjusts much more quickly than does
the tank temperature T, so that we can neglect the accumulation term in
(6.2-8) and write it as

FcU c C pc (Tco  Tci ) Q (6.2-10)

Justifying this assumption would involve comparing the characteristic

times of (6.2-3) and (6.2-8). We have implicitly made a similar
assumption already in (6.2-4), in which we have said that the rate of heat
transfer depends on the instantaneous values of the inlet and outlet
temperatures according to a relationship that was derived for the steady

Taken together, the equations of this section describe how the outlet
temperature and composition vary in time due to disturbances in inlet
temperature, inlet composition, and coolant flow rate. The coolant outlet
temperature is an intermediate variable in the system.

6.3 we encounter nonlinear equations

The equations in Section 6.2 are nonlinear: the outlet composition is
squared, the temperature is an argument in exponential and logarithmic
functions, the coolant flow is raised to a power. Even (6.2-10) is
nonlinear, because it contains the product of variables Fc and Tco.
Analytical solution is highly improbable. We may attack the problem by
numerical simulation, or by simplifying it. We will first follow the latter
course for two reasons:
x A numerical solution can show how individual cases behave, but not
general behavior. An analytical solution, by virtue of its mathematical
expression, will usually instruct us in parameter dependence.
x Simplifying by linear approximation will often give us a reasonable
approximation near the reference set point. If the control objective is
to maintain the process at set point, then the approximate description
may be entirely satisfactory for achieving good control.
Lesson 6: Exothermic Tank Reactor
6.4 making linear approximations with Taylor series
Given a function f, we specify some reference value of the independent
variables, and represent the function in the neighborhood of that reference
point as a series of terms. For a function of one variable:

f (x) f (x r ) 
(x  x r )  O (x  x r )2 (6.4-1)
dx x r

For a function of more than one variable:

f ( x , y,...) f ( x r , y r ,...)  (x  x r )
wx x r , y r ,...


( y  y r )  ...  O ( x  x r ) 2 , ( y  y r ) 2 ,...
wy x
r , y r ,...

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.

6.5 linear approximation to the material balance

We apply (6.4-2) to material balance (6.2-1).

w  rA
dC A
FCAi  FCA  V  rA r  V CA  CAr  V w  rA T  Tr (6.5-1)
dt wC A r wT r

At the reference condition, (6.5-1) becomes

FC Air  FC Ar  V  rA r
dC Ar
V 0 (6.5-2)

We subtract (6.5-2) from (6.5-1) and define deviation variables.

dC'A w  rA ' w  rA '

V FC'Ai  FC'A  V CA  V T (6.5-3)
dt wC A r wT r

We obtain the partial derivatives in (6.5-3) from (6.2-2).

w  rA E
2k o e CA (6.5-4)
wC A
Lesson 6: Exothermic Tank Reactor
w  rA E
2k o e C Ar 2k r C Ar (6.5-5)
wC A r

w  rA E  E RT 2
ko e CA (6.5-6)
wT RT 2

w  rA E  E RTr 2 E 2
ko e C Ar kr C Ar (6.5-7)
wT r RTr2 RTr2

Here kr is the rate constant at the reference condition.

kr k oe (6.5-8)

We substitute (6.5-5) and (6.5-7) into linearized material balance (6.5-3).

dC 'A E 2 '
V FC'Ai  FC'A  2Vk r C Ar C 'A  Vk r C Ar T (6.5-9)
dt RTr2

This is a first-order lag equation, which is more apparent if it is placed into

standard form

dC 'A WC '
WC  C 'A C Ai  K CT T ' C 'A (0) 0 (6.5-10)
dt WR


WC (6.5-11)
1  2W R k r C Ar
K CT WC k r C 2Ar

WR is the residence time of the tank, important to reactor conversion. The

time constant WC (smaller than WR) characterizes the dynamics of
composition change. The group KCT is the gain for the effect of
temperature on composition in the reactor. The negative sign shows that
an increase in operating temperature will reduce the exit concentration of
reactant A (by way of increasing the reaction rate constant).
Lesson 6: Exothermic Tank Reactor
The steady state mole balance in (6.5-2) has significance beyond serving
as a reference condition for deviation variables. It also constrains the
relationship between the tank composition and temperature (within the
rate constant kr) at steady state.

6.6 similarly approximating the energy balance

We apply (6.4-2) to energy balance (6.2-3).

VUC p FUC p Ti  FUC p T
w  rA
 'H R V  rA r  'H R V C A  C Ar  'H R V w  rA T  Tr (6.6-1)
wC A r wT r
 Qr  T  Tr  wQ Fc  Fcr
wT r wFc r

At the reference condition, (6.6-1) becomes

FUC p Tir  FUC p Tr  'H R V  rA r  Q r

V UC p 0 (6.6-2)

We subtract (6.6-2) from (6.6-1) and define deviation variables.

dT ' w  rA ' w  rA '

VUC p FUC p Ti'  FUC p T '  'H R V C A  'H R V T
dt wC A r wT r
wQ ' wQ '
 T  Fc
wT r wFc r

The reaction rate partial derivatives are given in (6.5-5) and (6.5-7). To
obtain the heat transfer rate partial derivatives, we combine the heat
transfer expressions (6.2-4) through (6.2-7) with the coolant energy
balance (6.2-10) to eliminate intermediate variable Tco.

ª ­°  A § 1 A Fn ·
Q FcUc C pc T  Tci «1  exp ® o
¨¨  o cr n ¸¸ ¾» (6.6-4)
¬ °̄ FcUc C pc © h o A i h ir Fc ¹ °¿»¼

The partial derivatives are

wQ ª ­°  A § 1 A Fn ·
FcUc C pc «1  exp ® o
¨¨  o cr n ¸¸ ¾» (6.6-5)
wT « °̄ FcUc C pc © h o A i h ir Fc ¹ °¿»¼
Lesson 6: Exothermic Tank Reactor
Fcr Uc C pc >1  E r @ (6.6-6)
wT r

wQ ª ­°  A § 1 A Fn ·
Uc C pc T  Tci «1  exp® o
¨¨  o cr n ¸¸ ¾»
wFc « °̄ FcUc C pc © h o A i h ir Fc ¹ °¿»¼
A T  Tci § 1 A Fn ·
­°  A § 1 A Fn ·
½°ª § 1 A o Fcrn · A o nFcrn º

 o ¨¨  o cr n ¸¸ exp® o
¨¨  o cr n ¸¸ ¾«1  ¨¨  ¸
n ¸
°̄ FcUc C pc © h o A i h ir Fc « i ir c »
Fc © h o A i h ir Fc ¹ ¹ °¿¬ © o h A i ir c ¹
h F A h F

wQ E r U or A o Tr  Tci ª U or A o n º
Uc C pc Tr  Tci >1  E r @  «1  » (6.6-8)
wFc r
Fcr ¬ A i h ir ¼

in which
§ 1 A ·
U or ¨¨  o ¸¸ (6.6-9)
© h o A i h ir ¹


­°  U or A o ½°
Er exp® ¾ (6.6-8)
°̄ Fcr Uc C pc °¿

The argument of the exponential function in (6.6-8) is the “number of

transfer units”, used in models of heat exchangers (Incropera and DeWitt,
Sec. 11.4). We substitute (6.5-5), (6.5-7), (6.6-6), and (6.6-8) into
linearized energy balance (6.6-3) to obtain another first order lag equation.

dT ' WT '
WT  T' Ti  K TC C 'A  K ht Fc' T ' (0) 0 (6.6-10)
dt WR

Lesson 6: Exothermic Tank Reactor
Fcr Uc C pc
1 1  E r  WR 'H r E
k r C 2Ar
FUC p UC p RTr2

'H r
K TC 2 W T k r C Ar (6.6-11)
UC p
WT Uc C pc Tr  Tci ª E r U or A o § nU or A o · º
K ht « ¨¨1  ¸¸  1  E r »
WR FUC p «¬ Fcr Uc C pc © A i h ir ¹ »¼

Once again, standard-form parameters have been defined. The thermal

time constant WT characterizes the dynamics of temperature change, and
KTC and Kht are gains for composition and heat transfer disturbances.

6.7 deriving transfer functions by Laplace transform and block diagram

Laplace transforms may be performed on the mole balance (6.5-10) and
the energy balance (6.6-10).

WC '
( WCs  1)C 'A (s) C Ai (s)  K CT T ' (s) (6.7-1)
WT '
( WT s  1)T ' (s) Ti (s)  K TC C 'A (s)  K ht Fc' (s) (6.7-2)

These equations must be combined to isolate the dependent variables T

and CA. First, we express them in a block diagram to show the
dependencies more clearly. System inputs are on the left, and outputs on
the right. Notice that both outputs depend on all three inputs.

CAi '(s) Wr CA'(s)
W Cs  1
W Cs  1

WT WT s  1
Ti'(s) Wr T'(s)
WTs  1

K ht
WTs  1

Figure 6.7-1 Block diagram of stirred reactor

Lesson 6: Exothermic Tank Reactor
We isolate CAƍ(s) either by eliminating Tƍ(s) between (6.7-1) and (6.7-2),
or by tracing the dependency through the block diagram:

W R WCs  1 W W s  1 WT s  1 ' WCs  1 WTs  1 F' (s) (6.7-3)
C 'A (s) C 'Ai (s)  R C Ti (s)  c
1 1 1
WCs  1 WT s  1 WCs  1 WTs  1 WCs  1 WT s  1
After simplifying the individual transfer functions in (6.7-3), we recognize
a second-order system

W2 W 2 K CT W 2 K CT K ht
WTs  1
C 'A (s) C 'Ai (s)  2 2 R C Ti' (s)  2 2 T C Fc' (s) (6.7-4)
W s  2W[s  1
2 2
W s  2W[s  1 W s  2W[s  1

in which the coefficients in the characteristic equation are

W2 (6.7-5)

2W[ (6.7-6)
1  K CT K TC

We similarly isolate Tƍ(s) from the equations or the diagram to find

W2 W 2 K TC W 2 K ht
WCs  1 WCs  1
T ' (s) Ti (s)  2 2
C Ai (s)  2 2
Fc' (s) (6.7-7)
W s  2W[s  1
2 2
W s  2W[s  1 W s  2W[s  1

From (6.7-4) and (6.7-7) we see that the material and energy balances
combine to produce a second-order dependence for each output variable.
The characteristic equation and poles are the same for both T and CA.
Through (6.7-4) and (6.7-7) we can predict how T and CA will respond to
particular disturbances.

6.8 multiple steady-state operating conditions

Before we consider transient response any further, though, we should look
more carefully at the steady-state reference condition. As in previous
Lessons, we think of the reference condition as some desired set of steady
operating conditions. Now, however, our reference state is described by
nonlinear material and energy balances, (6.5-2) and (6.6-2), repeated here:

0 FCAir  FCAr  Vk r C 2Ar (6.8-1)

Lesson 6: Exothermic Tank Reactor

0 FUC p Tir  FUC p Tr  'H R Vk r C 2Ar  FcrUc C pc Tr  Tci 1  Er (6.8-2)

For some combinations of design (V), reaction system (kr, ǻHR, etc.), and
operating inputs (CAir, Fcr, etc.), it is possible for three distinct pairs of
reactor temperature Tr and composition CAr to satisfy (6.8-1) and (6.8-2).
Marlin (App. C) illustrates this behavior by plotting the term of (6.8-2)
that represents heat generation by reaction

 'H R Vk r C 2Ar (6.8-3)

along with the heat flow and transfer terms

FUC p Tir  FUC p Tr  Fcr U c C pc Tr  Tci 1  E r (6.8-4)

both versus reactor temperature Tr. Any intersection of the two curves
(point A, B, or C) satisfies (6.8-2).

energy C
time reaction

heat transfer


The relative slopes of the curves are significant: at points A and C, an

increase in Tr would exhibit less heat generation than heat transfer, tending
to cool the reactor toward the reference condition. At B, by contrast, an
increase in Tr will tend to generate more heat than can be removed, so that
reactor temperature will rise, moving toward condition C. The conclusion
is that reference states A and C are stable, and B is not: B is in balance,
but a slight disturbance will cause an excursive transient toward either A
or C.

Our linearized dynamic models should capture this behavior in the vicinity
of the reference state. That is, the poles of (6.7-4) or (6.7-7) should show
Lesson 6: Exothermic Tank Reactor
negative real parts for conditions such as A and C, and positive real parts
for B.

Let us examine a specific case (in which the parameters are chosen more
for illustration than realism):

physical properties design and operating reference conditions

ȡ= 850 kg m-3 Uor = 438 W m-2 K-1 Tir = 330 K
Cp = 3800 J kg-1 K-1 Ao = 18 m2 CAir = 2300 mol m-3
ǻHR = -450 kJ mol-1 V = 0.785 m3 Fcr = 0.008 m3 s-1
E/R = 8200 K F = 0.015 m3 s-1
ko = 1100 m3 mol-1 s-1 Tci = 300 K

From these parameters, three reference states are possible:

steady reference output dynamic characteristics character of response
Tr = 326.7 K, CAr = 2296 mol m IJ = 49.7 s, ȟ = 1.001 stable (little reactant conversion),
very slightly overdamped
Tr = 443.7 K, CAr = 1335 mol m-3 IJ2 = -978 s2 (better conversion) excursively
Tr = 578.5 K, CAr = 227 mol m-3 IJ = 13.5 s, ȟ = 1.72 stable (best conversion),

Here our linearized models agree with the stability assessment from the
nonlinear steady state balances. We could not expect a reactor to stay, on
its own, at the second condition.

6.9 response of system to step disturbance

Equations for second-order step response were given in Lesson 5. For
stable systems, it is instructive here to consider the long-term value of the
step change, which is given by the magnitude of the step multiplied by the
appropriate transfer function gain in (6.7-4) and (6.7-7). For a step change
ǻC in the inlet composition CAi, the long-term effects are

W2 W 2 K TC
C 'A (f) 'C T ' (f) 'C (6.9-1)

For a step change ǻT in the inlet temperature Ti,

W 2 K CT W2
C 'A (f) 'T T ' (f ) 'T (6.9-2)

For a step change ǻF in the cooling water flow rate ǻFc,

Lesson 6: Exothermic Tank Reactor
W 2 K CT K ht W 2 K ht
C 'A (f) 'F T ' (f ) 'F (6.9-3)

Equations (6.11-1) through (6.11-3) predict steady-state operating

conditions after a step input disturbance. However, because these
predictions are made with an approximate model -- a linear approximation
valid only at the reference conditions CAr and Tr -- they are unlikely to
satisfy the material and energy balances (6.2-1) and (6.2-3) applied at the
new steady-state. Therefore, if we use our approximate model as a basis
for tuning a controller at a reference condition, we should be skeptical
about our tuning if we must operate the process away from that reference.

6.10 do the linearized models describe instability?

We must not claim that our approximate model will describe the full path
of a reactor transient. However, we do feel that our prediction of the
stability threshold by the poles of (6.7-4) and (6.7-7) can alert us that a
contemplated operating condition might be troublesome. If we can avoid
the onset of instability, we need never describe an instability transient.


6.11 step 1 - specify a control objective for the process

Our control objective is to maintain the outlet temperature T and
composition CA at constant values. Especially we must be concerned with
the possibility of runaway reaction.

6.12 step 2 - assign variables in the dynamic system

We seem to have two desired controlled variables (T and CA), two
disturbances (Ti and CAi), and a single manipulated variable (Th). We
have not encountered a system with multiple outputs before. It is clear
from (6.7-4) and (6.7-7) that the manipulated variable affects both T and
CA. Furthermore, we know that we cannot set T and CA independently,
because they are related through the material and energy balances.
Therefore, we will attempt to control one of them - perhaps T, because it is
easier to measure - intending to obtain satisfactory behavior of CA as a

6.13 step 3 - PID (proportional-integral-derivative) control

Proportional control reacts to the magnitude of the present error, and
integral mode reacts to its persistence over time. Integral mode removes
offset in the controlled variable, but tends to make the closed loop less
stable, as well as make the response sluggish. We now address these
concerns with the derivative mode.
Lesson 6: Exothermic Tank Reactor
§ 1
dH * ·
x *co  x *co ,b K *c ¨¨ H*  ³ H*dt  TD ¸
¸ (6.13-1)
© TI 0 dt ¹

where x*co is the controller output and the controlled variable error,
expressed in scaled variables, is

H* y*sp  y* (6.13-2)

Equation (6.13-1) describes the ideal PID (proportional-integral-

derivative) controller algorithm. It adds the derivative mode to the
proportional and integral modes we have seen before.

Derivative mode is an early warning of error; by reacting to the change in

the error signal, it can dictate a significant response from the manipulated
variable before the error has grown sufficiently to evoke a similar
response via proportional mode. The influence of the derivative mode is
set by the magnitude of the derivative time TD. Increasing TD strengthens
the controller response.

We can express algorithm (6.13-1) in deviation variables.

100% § dH ·
x 'co K *c ¨ H  ³ Hdt  TD ¸
'y ¨© TI 0 dt ¸¹
§ 1
dH ·
K c ¨ H  ³ Hdt  TD ¸¸
© TI 0 dt ¹

The Laplace transform of (6.13-3) is

§ '
x 'co (s)
K c ¨¨ y sp
(s)  y ' (s) 
1 '

y sp (s)  y ' (s)  TDs y sp
(s)  y ' (s) ·¸¸
© TIs ¹
§ · '
K c ¨¨1 

 TD s ¸¸ y sp (s)  y ' (s)
© TIs ¹

6.14 step 4 - choose set points and limits

Both the set point and operating limits for temperature may depend on a
number of considerations, including reaction kinetics (desired reaction
rate), reaction equilibrium (possible conversion), the possibility of side-
products or degradation reactions, the vapor pressure of solvents, limits of
construction materials, etc. Because our process may be open-loop
unstable, we ask whether control can stabilize it.

Lesson 6: Exothermic Tank Reactor

6.15 adjusting heat transfer with a valve

Physically, the controller output dictates the opening of a valve that admits
the cooling water to the heat exchanger. The figure shows a cooling water
supply header, a line to the heat exchanger, and another line to the cooling
water return header.

cooling water supply header lines to other

users of CW

heat exchanger
bundle in tank

cooling water return

As in Lesson 5, we will represent our valve dynamics as “relatively fast”

and expect that adding dynamic lags to the valve transfer function would
have negligible influence on the closed loop behavior.

6.16 control valve mechanism

A valve is a variable flow resistance manipulated by some actuator
mechanism. In the sink, one turns the valve actuator by hand, which turns
a threaded stem to raise and lower a plug with respect to a seat. When the
plug is seated, there is no opening for flow, and the hydraulic resistance
coefficient is infinite (no flow for finite pressure difference).

In a control valve, the most common actuator mechanism is a chamber in

which air pressure on one side of a diaphragm opposes a spring on the
other. The position of the diaphragm determines the position of the valve
stem, and thus the opening between plug and seat. The schematic
illustrates this concept (of course, real valve actuators will be more
Lesson 6: Exothermic Tank Reactor

Thus the controller output must vary the air pressure at the diaphragm.
Conversion of controller output to air pressure requires another device
between controller and valve, called a transducer. The transducer supplies
air to the valve in sufficient quantity to achieve a pressure proportional to
the controller signal. By convention, a zero controller output is
represented to the valve as 3 psig; 100% output produces 15 psig.

6.17 control valve failure mode

Should the air supply fail, the spring will force the diaphragm to an
extreme position. In the valve schematic above, the valve would be fully
open. Thus this valve might be called fail-open, or air-to-close. By a
different arrangement of spring and air, an alternative fail-closed, or air-
to-open valve can be produced.

In selecting the failure mode of a valve, the engineer considers how best to
protect people and equipment under off-normal conditions. General
guidelines would include cooling water failing open, steam valves failing
closed, reactor feed failing closed, vessel effluent failing open. Of course,
exceptions to these cases could be found, too.

The failure mode of a valve determines the sign of its gain. For example,
suppose that we represent the combined valve and transducer by a transfer
function between the controller output and the flow rate through the valve:

F ' (s ) K v x 'co (s) (6.17-1)

For an air-to-open valve, the flow increases with controller output, so that
the gain Kv is positive. For an air-to-close valve, shown in the schematic,
the flow decreases with controller output, so that the gain is negative.

We write (6.17-1) in physical variables to show the bias term.

F  Fb
K v x co  0 (6.17-2)

For an air-to-open valve, the gain Kv is positive and the bias flow Fb is
Lesson 6: Exothermic Tank Reactor
F x co (6.17-3)

For an air-to-close valve, the gain is negative, and bias flow is the
maximum flow.


100%  x co (6.17-4)

Equations (6.17-3 and 4) are suitable for use in simulator calculations.

6.18 positive closed loop gain and the sense of the controller
In these lessons, we have occasionally checked the sign of the gain in our
dynamic systems. Finding in Section 6.17 that an air-to-close valve
necessarily has a negative gain motivates us to examine the gains in a
closed feedback loop.

As a basis, we want feedback to be negative. That is, if the controlled

variable becomes too large, we want the feedback loop to reduce it. The
alternative positive feedback will tend to increase the already-too-high
controlled variable. A common example of positive feedback occurs
when the output of a loudspeaker is fed back into a microphone, amplified,
and delivered to the speaker - cover your ears!

Because we have defined error to be the set point less the controlled
variable, a high controlled variable gives a negative error. If this error is
acted upon by a positive gain around the loop, the feedback to the
controlled variable is negative. Hence we want a positive loop gain.

The loop gain KL is the gain component of the loop transfer function
GL(s). Thus, the loop gain is the product of the sensor, controller, valve,
and manipulated variable (process) gains.

KL KsKcK vK m (6.18-1)

x Of these, the gain Ks for most sensors is positive - the mercury rises
with temperature.
x The process gain Km is determined by the process itself - in these
lessons, it has typically been positive, such that an increase in
manipulated variable causes an increase in the response variable.
However, we might in principle run across an opposite case.
x The sign of the valve gain Kv is a function of a safety analysis, as
discussed in Section 6.17.
x Because Km and Kv may be either positive or negative, and can be so
for independent reasons, we must therefore reserve the ability to
Lesson 6: Exothermic Tank Reactor
choose the sign of the controller gain Kc. This is done with the
controller sense switch, which might be a physical switch on an analog
controller, or an input value in software.

6.19 ambiguity!
The adjectives “direct-acting” and “reverse-acting” are used with a
controller to indicate the position of the sense switch. Alternative
adjectives are “increase/increase” and “increase/decrease”. However, the
adjectives are not consistently used! Hence, look at the controller
carefully, and ensure that you know the algebraic sign of the gain.


6.20 closed loop transfer function

From Lesson 5, we borrow the general closed loop block diagram.
However, we add an additional, uncontrolled, output yu to represent the
outlet composition. The portion of Figure 6.20-1 labeled “process” is
equivalent to Figure 6.7-1, but redrawn to reflect the algebraic
manipulations of Section 6.7.




yu' (s)



x'm(s) y' (s)


control Gv(s) Gs(s)
y's (s)
x'co(s) e' (s) -

Gsp(s) y'sp,s(s)

Figure 6.20-1 Block diagram for control of one of two outputs

Lesson 6: Exothermic Tank Reactor

Also from Lesson 5, the controlled variable is related to the inputs by

G d1 Gd2 G m G v G c G sp
y ' (s) x 'd1 (s)  x 'd 2 (s)  '
y sp (s) (6.20-1)
1  G m G v G c G s 1  G m G v G c G s 1  G m G v G c G s
We specialize the nomenclature and the transfer functions for our stirred
reactor case, using especially temperature model (6.7-7).

x 'd1 (s) Ti' (s)

WCs  1 (6.20-2)
G d1 (s)
W 2s 2  2W[s  1

x 'd 2 (s) C 'Ai (s)

W 2 K TC (6.20-3)
G d 2 (s)
W s  2W[s  1
2 2

x 'm (s) Fc' (s)

W 2 K ht
WCs  1 (6.20-4)
G m (s)
W 2s 2  2W[s  1

G v (s) Kv G s (s) G sp (s) Ks

§ 1 ·
G c (s) K c ¨¨1   TD s ¸¸ (6.20-5)
© TIs ¹
y ' (s) '
T (s)

Motivated by Section 6.18, let us examine the loop gain:

x The sensor gain Ks is positive for temperature measurement.
x The process gain Km is given in (6.20-4). For a stable process, IJ2, IJC,
and Kht are all positive, but IJT is negative. Hence, Km is negative,
implying that an increase in the heat exchanger coolant flow will lower
the reactor operating temperature.
x Because we wish to provide cooling water even in the event of an air
supply failure, we choose an air-to-close valve. In Section 6.19, we
saw that such a valve featured a negative gain.
x Because the product of the other three gain components is positive, our
controller sense must be set to positive gain. If the reactor is too cold,
Lesson 6: Exothermic Tank Reactor
T < Tsp and İ > 0. Positive Kc directs the controller output to increase,
closing the air-to-close valve, restricting the cooling water, and thus
allowing T to rise.

6.21 closed-loop behavior - laplace transform solution

From the equations in Section 6.20, we can derive the transfer function for
disturbances in the inlet temperature:

WCs  1 s
T (s) W R K v K c K s K Th
Ti' (s) ­ § WT · § 2[WT T · ½
°WC ¨¨  TD ¸¸s 3  WC ¨¨  D  1¸¸s 2 °
° © K v K c K s K ht ¹ © WK v K c K s K ht WC ¹ °
® ¾
° W §¨ WT 1 1·
1 °
° C ¨ W 2 K K K K  W  T ¸s  T °
¯ © v c s ht C I ¹ I ¿

The closed loop characteristic equation is third order, because the integral
mode has increased the process order by one. Partial fraction expansion
will show us that the step response will be the sum of three exponential
terms (for 3 real roots) or an exponential oscillation (for 1 real and 2
complex roots). We could proceed as in Lesson 4, in which we calculated
poles numerically and found the onset of oscillation, and then instability
with increasing gain. Here our tuning task would be more complicated,
because we have three controller parameters to vary, instead of just the

The derivative mode affects the coefficients of the two higher order terms.
Increasing D will increase the curvature of the characteristic function,
which (other parameters unchanged) can increase the likelihood of three
abscissa-crossings - thus three real roots, suppressing oscillation in the

The integral mode affects the lower order terms. Increasing TI will result
ultimately in reducing the order of the characteristic equation, which will
allow offset in the step response. Increasing the controller gain Kc will
reduce the transfer function gain (the coefficient in the numerator) and
reduce the magnitude of the dynamic term coefficients in the denominator.

6.22 Bode criterion for closed loop stability

We invoke the Bode stability criterion, as we did in Lesson 4, with one
important provision: the Bode criterion does not apply if the process is
open-loop unstable. Therefore, if we are attempting to operate at such a
condition, we must use more advanced methods to determine the closed-
loop stability limits.
Lesson 6: Exothermic Tank Reactor
For stable open-loop conditions, we may apply the Bode criterion to the
loop transfer function. The amplitude ratio is the magnitude of the loop
transfer function, which may be found as the product of the magnitudes of
the component transfer functions. Using a table of amplitude ratios, such
as that of Marlin (2000), we find

1  WC Z
§ 1 · W 2 K ht

RA KsKc 1  ¨¨ TD Z  ¸¸ K v (6.22-1)
© TI Z ¹ WC WT 1  W Z  2W[Z
2 2 2 2

Similarly, the phase angle is the sum of the component phase angles.

§ 1 · §  2W[Z ·
I tan 1 ¨¨ TD Z  ¸¸  tan 1 WC Z  tan 1 ¨ 2 2 ¸
© TI Z ¹ ©1 W Z ¹

The derivative mode opposes the phase lag due to the integral mode and
stabilizes the closed loop. In the figure, the controller parameters are set
to give P, PD, PI, and PID controllers at a stable open-loop condition, as
described in Section 6.8. Controllers using integral mode are shown with
dashed lines; solid lines refer to P and PD controllers.
Lesson 6: Exothermic Tank Reactor


amplitude ratio



0.00001 0.0001 0.001 0.01 0.1

phase angle (deg)





0.00001 0.0001 0.001 0.01 0.1
Ȧ (radians s -1)


The effect of integral mode is to increase the amplitude ratio and increase
the phase lag at low frequency. Both these actions are generally
destabilizing to a closed loop, although in this case, there is no crossover
frequency in evidence (the phase lag never goes beyond -180º). The effect
of derivative mode is to decrease the phase lag at high frequency. This
can often help push a crossover frequency to a higher value, at which the
amplitude ratio will tend to be smaller. In this way, derivative mode can
help to stabilize a closed loop.

In this process, perhaps the most significant contributor to stability is the

presence of a “lead”, the positive phase angle represented by the middle
term in (6.22-2). This term arises from the manipulated variable transfer
function (6.20-4), because coolant flow rate affects reactor temperature
not only by heat transfer, but by the rate of exothermic reaction. The
numerator term in (6.20-4) affects the coefficients of the characteristic
equation, and thus the poles that are possible. This third-order process, at
the reference condition we have chosen, seems to be intrinsically stable
under closed loop control.
Lesson 6: Exothermic Tank Reactor
6.23 tuning the controller
Given the process model, we may tune the controller by simulation, in
which we vary the three parameters and assess the effects by comparing IE
and IAE for the responses. The simulation can be done by supplying the
partial fraction expressions with numerically computed roots, or by
returning to the differential equations for numerical solution, either in
linearized or original nonlinear form.

6.24 conclusion
We encountered a more complicated process in this lesson, both because it
required two coupled equations, and because reaction kinetics and heat
transfer made them nonlinear. We introduced a formal approximation
process to make the model linear, and then were able to treat it with tools
we had previously developed. Even so, nonlinearity could not be escaped,
because we found that the behavior of the approximate description
depended on the reference conditions we chose. Furthermore, we found
that some conditions admitted multiple steady states. All of this should
inspire us to maintain a healthy skepticism toward our results.

The positive news was that it is possible to maintain the process at an

inherently unstable condition through feedback control. We have further
increased our feedback capabilities through the derivative mode, which
complements proportional and integral modes to comprise the PID
controller, widely used and justly respected in the chemical process

6.25 references
Incropera, Frank P., and David P. DeWitt. Fundamentals of Heat and Mass Transfer.
5th ed. New York, NY: J. Wiley, 2002. ISBN: 0471386502.

Marlin, Thomas E. Process Control. 2nd ed. Boston, MA: McGraw-Hill, 2000.
ISBN: 0070393621.
Lesson 7: High Order Overdamped Processes
7.0 context and direction
Chemical processing plants are characterized by large time constants and
time delays. For control engineering, we can often approximate these
high-order systems by the FODT (first-order-dead-time) model. Dead
time in a process increases the difficulty of controlling it.


7.1 big and slow - high-order overdamped systems

We began our study of process control by considering a mixed tank.
Applying a material or energy balance to a well-mixed tank produces a
first-order lag system. We subsequently combined two balances to
produce a second-order system. In one case, two material balances
described storage of material in two tanks. In another case, a single tank
stored both material and energy. Energy and material balances show that
the tank causes a dynamic lag between input and output, because it takes
time to adjust the amount of mass or energy distributed throughout the
tank. We might thus expect that more storage elements would lead to
higher-order behavior, and require higher-order equations to describe

The classic illustration of a high-order system is a set of n tanks in series:

each tank feeds the next, and a change in the inlet stream composition CA0
must propagate through multiple tanks to be felt at the output CAn. The
individual tank models are

Vi C 'Ai FC 'Ai 1  FC 'Ai (7.1-1)

They are combined by eliminating the interior stream variables to produce

a single transfer function between input and output.

C 'An (s) 1
C 'A 0 (s) W1s  1 W2s  1 " Wn s  1
Let us illustrate high-order behavior and (7.1-2) by first imagining a single
well-mixed overflow tank of time constant IJ. If we introduce a step
increase in the inlet concentration, we will (by the well-mixed assumption)
immediately detect a rise in the outlet stream – the familiar first-order lag
response. If we have instead two tanks in series, each half the volume of
the original, we will detect a second-order, sigmoid response at the outlet.
Lesson 7: High Order Overdamped Processes

Each tank has a smaller individual time constant, and their sum is the time
constant IJ of the original tank. 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 7.1-1.


step response




0 0.5 1 1.5 2 2.5 3 3.5 4 4.5

Figure 7.1-1. Step response for tanks in series; equal time constants
Lesson 7: High Order Overdamped Processes
The step response shows that high-order systems have a longer start-up
period before rising toward the final value.

7.2 the FODT approximation to high-order step response

There may be occasions when detailed dynamic analysis of a process is
warranted. Often, however, it is sufficient to obtain a simplified dynamic
model that gives a reasonable approximation to the process behavior.
Figure 7.1-1 indicates that high-order step responses might be represented
as a first-order rise following a period of delay. The dynamic model that
would behave this way is called the FODT (first-order-dead-time) model;
it turns out to be suitable for describing the dynamic response of many
chemical processes.

7.3 dead time is delay

Before we examine the FODT model, we will look at dead time by itself.
Chemical processes require that material be moved from one location to
another: in conduits, on conveyer belts, through vessels. The
transportation time is finite; it implies a delay between the onset of a
disturbance at one location and its observation at another. This delay is
often called dead time; it is familiar to anyone who has waited at the
faucet for the hot water to arrive.

Consider a pipe carrying a liquid. A pulse of solute added at the entrance

will be observed at the exit only after the solute is transported through the

x(t) seen by observer

at entrance of pipe y(t) seen by observer
at exit of pipe


td td + ș

The transit time depends on the liquid velocity and the length of the pipe.
The figure indicates faithful transmission of the input signal x(t) from inlet
to outlet, as if every particle in the pipe moved at the same velocity.
However, in real chemical processes the solute pulse y(t) would become
distorted through diffusive and dispersive effects. Nonetheless, a simple
description of transport using the average fluid velocity is often sufficient
to represent dead time in a process:
Lesson 7: High Order Overdamped Processes
T (7.3-1)
v F

Thus the dead time is the residence time in the pipe.

Consider again Figure 7.1-1 and the series of tanks. If taken to the limit of
an infinite number of tanks, each infinitesimally small, we finally obtain a
pure delay, in which at time IJ the full step disturbance appears at the

7.4 dead time and lag are different

In casual conversation, one might not distinguish between lag and delay;
however, in process control these two terms have distinct meanings. A lag
process is illustrated by a mixed tank, and a delay (or dead time) process
by a pipe.

Table 7.4-1. Comparison of lag and dead time processes

first-order lag pure dead time

defining dy ' y' x ' ( t  T)

equation W  y' x ' (t)
transfer y ' (s) 1 y ' (s)
function G (s) G (s) e Ts
x (s) Ws  1
x ' (s)
step response

IJ ș

In the tank step response, the output lags behind the input, but there is no
delay between input and a response at the output. In the pipe, by contrast,
the output is delayed.

7.5 frequency response of a dead time process

The sinusoidal input

x(t) A sin Zt (7.5-1)

is faithfully reproduced at the output of the dead time process, but will be
delayed. Thus the amplitude ratio is unity, and the phase lag depends on
the input frequency and the dead time. Inserting jȦ into the Laplace
transform in Table 7.4-1,
Lesson 7: High Order Overdamped Processes
RA e  Tj Z 1
I ‘e TjZ TZ

Thus dead time delays a signal in an unbounded manner, but does not
diminish its amplitude. From what we recall of the Bode criterion for
evaluating closed loop stability, we might speculate that dead time could
be particularly significant.

7.6 the FODT model

To represent the dynamic behavior of a complex process as first order plus
dead time is to say that the many storage elements and pipelines in the
process can be represented by a single tank and pipe in series (the order
does not matter). The FODT model is a first order ODE with a delayed

dy '
W  y' Kx ' t  T (7.6-1)

The FODT time constant IJ comes from the tank, and the dead time ș from
the pipe. Taking the Laplace transform of (7.6-1)

y ' (s) K Ts

G (s) e (7.6-2)
x ' (s) Ws  1

7.7 Bode plot of FODT

With dead time, we finally see how the phase angle can become
significant: although the first-order lag phase angle is limited to -90°, the
dead time contribution to phase lag is unbounded. In the plot, the dead
time has been set equal to the time constant.

If a FODT process is placed in a feedback loop with a controller, a

crossover frequency will inevitably be found, and the controller settings
will be limited by the onset of instability..
Lesson 7: High Order Overdamped Processes

Bode plot - first order plus dead time system

phase angle (deg)

first order

0.01 0.1 1 10

7.8 identification - obtaining an FODT to represent a process

Arriving at a suitable approximate model (both form and parameter
values) is known as identifying a model. Most often, the FODT
approximation would be derived from an experimental test of the dynamic
system. For example, a step input xdata(t) would be contrived, and the
response ydata(t) measured. Then the experimental data would be
compared with the FODT step response ymodel(t), and the parameters K, IJ,
and ș adjusted to achieve a satisfactory match. The match might be done
by eye; alternatively, a least-squares criterion could provide an objective
comparison between different sets of parameter values. Marlin (2000)
describes further methods for obtaining parameter values from the data.

If the step input can be maintained sufficiently long to see the response
become virtually constant, the gain K is relatively easy to determine.
However, the response variable remains subject to other disturbances,
which may distort the experimental data through noise and confounding
inputs. Such confounding inputs may force a given step test to be short.
Furthermore, in some systems it may be impossible to approximate a step
input. In others, it may be undesirable to force a system away from the
desired operating condition for a lengthy period. Thus other forms of
input disturbance suitable for identifying a system model are discussed by
Seborg et al (2004).

Figure 7.8-1 shows experimental data for a system with negative gain.
The noise in the response variable trace makes the choice of dead time
somewhat uncertain. Furthermore, the short duration of the run means
that different combinations of gain and time constant can be used to fit the
observed trace; a longer run would have distinguished these two
parameters more clearly.
Lesson 7: High Order Overdamped Processes

disturbance and response

time (s)
50 150 250 350 450
scaled traces

-0.4 Fc


Figure 7.8-1 Experimental step response and FODT fit

We make use of a process model in closed loop calculations. For

example, we have used the manipulated variable transfer function Gm(s) to
compute the loop transfer function. In an experimental test, however, we
are likely to obtain Gm in combination with the sensor and valve transfer
functions. This is because manipulating the valve with the controller
output in manual is the most common way of creating a step input in the
manipulated variable. Furthermore, we obtain the value of the response
variable only by measuring it, usually with the installed sensor. Hence the
FODT model actually includes the behavior of valve and sensor, as well.
Notice in Figure 7.8-1 that the input is not quite a step, but rather a rapid
first-order response. These dynamics become part of the ostensible
process model.

In Lesson 6 we encountered a process with underdamped behavior. In

such a case, it may be possible to identify a suitable second-order-dead-
time model, thus augmenting the more varied responses of second order
with dead time. We will not pursue this idea further in this lesson.
Lesson 7: High Order Overdamped Processes
7.9 digital calculation of dead time
Dead time is computed by defining arrays for the input and output
variables. The arrays represent the variation of the inputs and outputs over
time. The dead time is represented by a difference in the array index
between input and output variables. For example, suppose that the
variables are computed every five seconds, so that x_input(i+1)
represents the value of the input 5 s after that stored in x_input(i).
Then a dead time of 20 s would require that a change in the input variable
occurring at index value i not be introduced to the output variable array
until the index reaches i+4.

7.10 an example process with dead time

Consider the blending tank we studied in Lesson 3; however, recognize
that the composition sensor may be placed some distance from the tank, so
that there is a delay between achieving a composition in the tank and
presenting its value for measurement.

Fc, CAc

F, CAi

F, CAo

volume V

The system model, adapted from Lesson 3, is

dC 'Ao ( t ) C
W  C 'Ao ( t ) C 'Ai ( t  T)  Ac Fc' ( t  T) (7.10-1)
dt F

where the time constant IJ is given by the ratio of tank volume to

volumetric flow, and the dead time ș by the ratio of the length of pipe
between tank and sensor to the average velocity of the outlet stream in the
pipe. Notice that in this particularly simple case, we have derived the
model from our understanding of the process, not experimental data.

The responses to changes in either of the inputs CƍAi or Fƍc will resemble
those computed in Lesson 3, except that they will be observed only after a
time interval of ș has followed the occurrence of the input.

Lesson 7: High Order Overdamped Processes

7.11 the control scheme

step 1 - specify a control objective for the process
Our control objective is to maintain CA at set point CAs.

step 2 - assign variables in the dynamic system

The controlled variable is CA. From (7.10-1) we see that the manipulated
variable Fc affects the controlled variable through a transfer function that
includes dead time. The disturbance transfer function for input CAi
includes dead time, as well. For feedback control, that does not matter -
we observe that a disturbance has occurred only when the controlled
variable begins to deviate from set point. (In other control schemes,
however, we will be able to make use of measurements of the disturbance
variable. In this case, it will be useful to know the dead time for

step 3 - PID (proportional-integral-derivative) controller algorithm

Derivative mode can help to stabilize the loop. However, we expect to
find that dead time will force us to detune (that is, make less aggressive)
the controller.

step 4 - choose set points and limits

The parameter values, also from Lesson 3, are

V = 6 m3
F = 0.02 m3 s-1
IJ = 300 s
ș = 60 s
Fcs = 10-4 m3 s-1
CAis = 8 kg m-3
CAos = 10 kg m-3
CAc = 400 kg m-3

We will suppose that CAi may vary between 4 and 10 around its reference
value of 8 kg m-3. Therefore, from the steady state material balance, Fc
must be capable of varying between 0 and 2×10-4 m3 s-1 to maintain CA at
set point.


7.12 sensor transmitter range

The sensor must respond to the controlled variable (as mercury rises in the
glass), and its associated transducer or transmitter must convert this
response to an input signal for the controller. For old-generation
pneumatic controllers, this would be an air pressure; for electronic
controllers and computers, it is often transmitted as an electric current that
Lesson 7: High Order Overdamped Processes
varies between 4 and 20 mA. The controller will convert the current to a
percentage of the maximum input. Thus we assign to the sensor a gain
with dimensions of

mA %in
Ks ( ) (7.12-1)
unit of controlled var iable mA

The magnitude of Ks depends on the sensitivity, or the range, of the

sensor. Increasing the sensitivity of the sensor will cause a larger signal to
the controller for a given deviation in controlled variable. An increase in
sensitivity, and thus sensor gain, increases loop gain in the same way as
increasing the controller gain.

Some processes must employ multiple ranges. For example, a process

with a set point of 500°C would have a low sensitivity (wide range) for
monitoring the controlled variable during start-up from ambient conditions
and a higher sensitivity (narrow range) for normal operation near the set
point. Such an application might be done with multiple sensors, with
variable sensitivity, or with multiple control loops, according to the
particular case.

7.13 valve saturation

There is not unlimited power to manipulate a process: even though a
control algorithm might calculate an output greater than 100%, this can
direct the valve to be no more than fully open, or fully closed. The control
engineer must size equipment so that the manipulated variable is sufficient
to exert a countervailing influence to the anticipated disturbances.

For example, if the controlled variable is a reactor temperature and the

manipulated variable is the flow of cooling water to the reactor jacket, the
maximum flow rate must be sufficient to cool the reactor under the most
unfavorable anticipated disturbance conditions. Then the valve must be
selected, along with other piping components, to supply this flow. No
controller tuning can compensate for a valve that is too small.

7.14 proportional gain and proportional band

We have worked with the gain of the controller, Kc. An alternative
convention is also used: the proportional band. Proportional band is the
inverse of the non-dimensional controller gain, multiplied by 100.

PB (7.14-1)

Thus a gain of 1 %out %in-1 would be equivalent to a proportional band of

100. Large proportional band implies low gain.
Lesson 7: High Order Overdamped Processes
7.15 integral time and reset rate
Just as proportional mode has an inverse convention, the integral time is sometimes replaced
with the reset rate. The reset rate is simply the reciprocal of the integral time, and has
dimensions of repeats time-1. Using a large reset rate is equivalent to a small integral time,
implying aggressive integral mode control.

7.16 reset windup in integral mode

The integral mode integrates the error over time; a persistent error leads to
a growing integral-mode contribution to the controller output. In our
definition of the standard PID algorithm, there is no limit to the growth of
this contribution. Hence under some circumstances the computed
controller output could become significantly greater than 100%. The
valve would be no more than fully open (or closed) as described in Section
7.13, but the controller output would be increasing. If the error were
finally reversed, requiring the valve direction to reverse, the controller
would be unable to direct it to do so until the integral mode contribution
had been reduced by persistent error of the opposite sign.

Such a condition is called reset windup, and could occur under prolonged
severe input step disturbances, or a fault in the loop (such as a mistakenly
closed manual valve) that prevented the manipulated variable from
affecting the controlled variable. Controller devices and algorithms
generally include windup protection to prevent the unbounded growth of
the integral mode.

7.17 protecting set point changes from derivative spikes

The standard PID algorithm defined the derivative mode in terms of the
time derivative of the error. However, derivative mode need not be
employed in response to set point changes, so the definition is changed to
apply derivative only to changes in the controlled variable. Thus the
algorithm becomes

§ 1
dy ·
x co  x co,bias K c ¨¨ H  ³ Hdt  TD ¸¸ (7.17-1)
© TI 0 dt ¹

Here y represents the controlled variable, and the sign in front of TD has
become negative, because error is always defined as the difference
between set point and controlled variable.

Algorithm (7.17-1) is the basis of practical controller algorithms.

However, for derivations and some illustrations we will still use the
Laplace transform of original algorithm.

7.18 filtering the derivative mode

The derivative mode, by reacting to the rate of change of the controlled
variable, is subject to noise. Applying derivative mode to a noisy signal
Lesson 7: High Order Overdamped Processes
can introduce disturbances by needless variation of the manipulated
variable. The effect of noise can be reduced by filtering the derivative
mode; this is usually shown by altering the derivative-mode term in the
PID algorithm:

TD s
DTD s  1

Equation (7.17-1) may be thought of as the transfer function that describes

sending input İ(s) first through a differentiator TDs followed by a first-
order filter (ĮTDs+1)-1. In the time domain, this would be represented as

dH df dH
DTD  H df TD (7.18-2)
dt dt

where the variable İdf is the differentiated, filtered error signal that is
combined with the original error and the integrated error in the PID
algorithm. The filter parameter Į is often set between 0.1 and 0.2. If Į
were zero, İdf would be the unfiltered derivative of the error İ in the ideal
PID algorithm.


7.19 closed loop transfer function

The familiar closed loop diagram can be drawn, and the closed loop
transfer function derived. For the disturbance

C 'A (s) Ws  1 (7.19-1)
C 'Ai (s) § 1 · K
1  K s K v K c ¨¨1   TDs ¸¸ m e Ts
© TIs ¹ Ws  1

In (7.19-1) we have omitted the dead time from the disturbance transfer
function in the numerator because we (presume that we) have no
independent measure of CAi and thus only know that a disturbance has
occurred when we see the response in CA. We proceed as before to
resolve the fractions, and obtain

C 'A (s) K ds
C 'Ai (s) § 1·
Ws 2  s  K s K v K c K m ¨¨ TDs 2  s  ¸¸e Ts
© TI ¹

Here, we grind to a halt, because none of our Laplace transform

experience has prepared us to deal with the exponential term in the
Lesson 7: High Order Overdamped Processes
denominator. It is possible to substitute a polynomial approximation
(called a Padé approximation) for the exponential term, and thus obtain an
approximate transfer function that can be inverted, but we will not do this.

7.20 Bode criterion for closed loop stability

Instead, let us examine the stability of the closed loop by the Bode
criterion. Extracting the loop transfer function from the denominator in
(7.19-1) we obtain the frequency response.

§ 1 · 1
RA G L ( jZ) KsK vKcK m 1  ¨¨ TD Z  ¸ (7.20-1)
© TI Z ¸¹ 1  (WZ) 2

I ‘G L ( jZ) tan 1 (ZW)  tan 1 (TD Z  )  TZ (7.20-2)

From (7.20-2) we see that the crossover frequency depends on the system
model parameters IJ and ș, and it may be further influenced by controller
parameters TI and TD. These controller settings also influence the
amplitude ratio, along with the controller gain Kc.

Choosing sample process and controller parameters

KsKmKv = 0.5
IJ = 3 min
ș = 1 min
TI = 10 min
TD = 2 min

the crossover frequency is computed to be 3.088 radians min-1 and the

controller gain at instability to be 2.994. The Bode plot for the loop
transfer function is shown in Figure 7.20-1. Amplitude ratios are plotted
for gains below, at, and above the stability limit (which is indicated by a

Notice that the amplitude ratio becomes level at high frequencies, instead
of dropping off, as we have previously seen. This is because the
decreasing amplitude of the first order process is balanced by the
increasing amplitude of the derivative controller mode. The derivative
mode filter described in Section 7.18 would prevent this high derivative
gain at high frequencies.

The phase angle actually rises over some range of frequency, due to the
contributions of the controller, but ultimately the process dead time
dominates, and the phase angle crosses the -180º stability criterion.
Lesson 7: High Order Overdamped Processes


amplitude ratio



0.001 0.01 0.1 1 10

phase angle (deg)






0.001 0.01 0.1 1 10
Ȧ (radians min-1)

Figure 7.20-1. Bode plot for loop transfer function at gains 1, 3, and 5

A numerical calculation of closed loop response to a tiny step disturbance

is given in Figure 7.20-2. The system is initially stable, but the 1%
disturbance initiates a cycle of oscillation that increases in amplitude. In a
perfectly linear system, the amplitude would increase without bound. In a
practical closed loop, the valve output would oscillate between fully open
and fully closed, and the process would move into regions of operation not
well described by the model and parameters in use. While not strictly
unbounded, the practical system response is nonetheless undesirable, so
that the linear stability analysis has indicated a controller setting to be
Lesson 7: High Order Overdamped Processes



0 20 40 60 80 100 120

manipulated and controlled variables




0 20 40 60 80 100 120

manipulated controlled

Figure 7.20-1. Unstable response to small disturbance at gain of 3

7.21 tuning the controller via correlations

The Bode criterion has given us an estimate of the stability boundary; we
must still decide where to set the controller. We have previously tuned
controllers by several methods:
x stability preservation - adjusting gain to realize particular gain and
phase margins
x direct simulation with a process model to minimize integral
measures of error, such as IE and IAE, for various inputs
x direct simulation with a process model using less comprehensive
criteria, such as minimizing time to return to less than 2%

Now we introduce two correlations for setting controller parameters. In

each case, the method represents the process as an FODT and specifies
controller settings as a function of the model parameters. The correlations
give different results because their authors had different ideas of what
constituted “good control”.
Lesson 7: High Order Overdamped Processes
Probably the best known, it actually comes in two flavors. The first (open
loop) defines controller parameters in terms of the FODT process
parameters (p.347 in Marlin). The second (closed loop) depends on
knowing the controller gain that will push the closed loop to the point of
instability (constant amplitude oscillation), if only proportional mode is
used. Controller parameters are then defined in terms of the gain and the
frequency of the sustained oscillation (p.330 in Marlin). The independent
variables in these methods can be determined from an existing FODT
model, or from empirical plant testing (careful with that instability test,

This is Marlin’s recommendation. Read Chapter 9 for the story of its
development. Read particularly Example 9.5 for a good illustration of its
use. In short, the correlation provides robust tuning for an FODT process
(and thus for anything reasonably approximated by FODT). The closed
loop equations (FODT and PID controller) were solved numerically to
predict controlled variable response to disturbance and set point steps.
Controlled variable IAE was minimized by varying controller parameters
Kc, Ti, and Td. However, this optimization was broadened beyond a single
operating case. It accounted for model error and changes in operating
conditions by computing IAE for multiple cases, in which model
parameters were varied. In addition, constraints were placed on variation
in the manipulated variable. The correlation is presented as graphical
relations among nondimensional parameters.

§ T · Ti § T · Ti § T ·
Kc K p f1 ¨ ¸ f2 ¨ ¸ f3 ¨ ¸ (7.21-1)
©W T ¹ W T ©W T ¹ W T ©W T ¹

The controller gain is normalized by the product of the gains in the

remainder of the loop (recall that the product of the gains around the
feedback loop is dimensionless). The integral and derivative times are
normalized by the sum of the FODT time constant and dead time. The
independent variable is the fraction dead time in the process.

Other correlations could be used (many are available; Seborg et al (2004)

give a good overview), but the general idea should be clear: first get
knowledge of the process itself, such as a dynamic model and the values
of its parameters, and use this knowledge to choose controller settings.
The variety of results that can be obtained in practical systems is due
partly to the richness of the mathematical form of the PID algorithm in a
closed feedback loop, and partly to the approximate nature of our process
models - the real process simply has more potential for complexity than
our mathematical models can predict. Marlin (2000) gives further
discussion of the effects of error in process models.
Lesson 7: High Order Overdamped Processes
7.22 conclusion
We have asserted that an enormous variety of chemical processes may be
represented by the FODT approximation. With the FODT process we
have introduced the crucial issue of dead time. Dead time is one of the
distinguishing features of chemical process control, and exerts strong
influence on how well the process can be controlled.

Having such a standard system model available allows the development of

correlations for tuning controllers. The correlations point out that good
knowledge of the process is the basis for successful control. We will find
that as more detailed understanding of the process is obtained, more
sophisticated control strategies can be employed.

7.23 references
Marlin, Thomas E. Process Control. 2nd ed. Boston, MA: McGraw-Hill, 2000.
ISBN: 0070393621.

Seborg, Dale, Thomas Edgar, and Duncan Mellichamp. Process Dynamics and
Control. 2nd ed. Hoboken, NJ: Wiley, 2004. ISBN: 0471000779.
Lesson 8: Cascade and Feedforward Control Schemes
8.0 context and direction
Knowledge of the process is the basis for successful process control. In
Lesson 7, we used process knowledge to guide, through tuning
correlations, our choice of controller parameters. In this lesson, process
knowledge will guide our choice of control structure itself: by making
additional process measurements, we will augment the single-loop
feedback control scheme to give it greater capability.


8.1 a process with identifiable intermediate variable

We begin with a process that has three inputs, two of them disturbances,
and one output that we will wish to control. As usual, transfer functions
Gd1(s), Gd2(s), and Gm(s) may refer to the same assembly of equipment,
but specify how the output variable y depends on each particular input.



x'm(s) y' (s)


The Laplace domain process description is then

y ' (s) G m x 'm (s)  G d 2 x d' 2 (s)  G d1x 'd1 (s) (8.1-1)

We imagine a case in which process Gm(s) could be divided into two parts,
connected by a measurable intermediate variable xi: this could be as
simple as two tanks in series, as in Lesson 4. Having specified some of
the interior structure of Gm, we consider xd2 to be typical of disturbances
that affect the process further upstream and xd1 to affect the process
downstream, after the intermediate variable.



x'i (s) y' (s)

Gm2(s) Gm1(s)

The process description becomes

Lesson 8: Cascade and Feedforward Control Schemes
y ' (s) G m 2 G m1x 'm (s)  G d 2 a G m1x d' 2 (s)  G d1x 'd1 (s) (8.1-2)

Equations (8.1-1) and (8.1-2) describe the same process, so they must be
equivalent. Comparing them, we find

Gm G m 2G m1 (8.1-3)


G d2 G d 2 a G m1 (8.1-4)

Also, the intermediate variable is given by

x i' (s) G m 2 x 'm (s)  G d 2 a x d' 2 (s) (8.1-5)

8.2 response to disturbances

Suppose, for illustration, that we let each of these transfer functions be
first order. Then the responses of xi and y to a step in xd2 are shown in
Figure 8.2-1.


0.8 y




0 1 2 3 4 5 6 7

Figure 8.2-1. Step response of intermediate and output variables

We observe that the intermediate variable responds before the output.

Perhaps this can help us to improve the control of y.

Lesson 8: Cascade and Feedforward Control Schemes
8.3 step 1 - specify a control objective for the process
Our control objective is to maintain the outlet variable y at set point.

8.4 step 2 - assign variables in the dynamic system

The controlled variable is y. The manipulated variable affects the
controlled variable through the transfer function. These assignments are
familiar from previous Lessons. However, we have some new
assignments to make:
x By looking in more detail at the composition of this process, we
identify an intermediate variable that influences the controlled
variable, responds to disturbances before the controlled variable
does, and responds to the manipulated variable, as well. We can
use this extra information in a scheme called cascade control.
x Furthermore, suppose we can measure a disturbance variable that
frequently disturbs the process. By this means, we can forecast
when the controlled variable is about to be altered and forestall it
in a scheme called feedforward control.

8.5 step 3 - cascade control scheme

The idea is to insert a secondary feedback control loop between the
controlled variable y and manipulated variable xm. The secondary loop
controls intermediate variable xi. This variable must hold several
x it must respond to important disturbances (those that significantly
affect the controlled variable)
x it must also convey the effects of such disturbances to the
controlled variable
x it must respond to the manipulated variable

The intermediate variable xi is called the secondary variable, and the

control scheme now features a new secondary loop within the original
primary loop.
Lesson 8: Cascade and Feedforward Control Schemes


x'i (s) y' (s)

Gm2(s) Gm1(s)

Gv(s) Gs2(s) Gs1(s)

x'co(s) - -
Gc2(s) Gc1(s)


Figure 8.5-1. Block diagram for cascade control structure

The secondary loop controls the intermediate (secondary) variable xi by

adjusting the manipulated variable xm. The primary loop controls the
controlled variable y by manipulating the set point of the secondary
controller xo1. Thus we have the same controlled variable and set point as
before, but the valve has been augmented by an inner control loop.

Disturbances xd2ƍ are rejected by the secondary loop before they affect the
full process, and thus response is quicker and the impact on yƍ less. The
primary loop is necessary to handle the other disturbances, such as xd1ƍ,
that always exist. The extra layer of control does not degrade the response
to xd1ƍ, because the process is usually much slower than the controller.

Cascade can be carried to more nested levels. For example, in Figure 8.5-
2 the composition controller sets a temperature set point in the secondary
loop; the temperature controller in turn sets the flow set point for the
tertiary loop.

Figure 8.5-2. Three-level cascade control

Lesson 8: Cascade and Feedforward Control Schemes
Cascade control is still feedback control, performed with conventional PID
control algorithms. The improvement comes because we're looking inside
the process, discriminating among disturbances, and applying feedback
with increased deftness.

8.6 step 3 - feedforward control scheme

We get closer to the root of the problem if we react directly to the
disturbance, predicting what the manipulated variable should do, not
waiting for a process response. This is the topic of feedforward control.
We contrast simple feedback control in Figure 8.6-1 with feedforward
control in Figure 8.6-2:



x'm(s) y' (s)


Gv(s) Gs(s)



Figure 8.6-1. Feedback control diagram

In feedback control, disturbance xd2 proceeds through the process (Gd2) to

affect controlled variable y. The controller reacts to the resulting error and
adjusts the manipulated variable; the change in manipulated variable
proceeds through the process (Gm) after the fact to reduce the error.
Lesson 8: Cascade and Feedforward Control Schemes


x'm(s) y' (s)


Gsf(s) Gv(s)


Figure 8.6-2. Feedforward control diagram

In feedforward control, disturbance xd2 proceeds in parallel through the

process (Gd2) and through the feedforward controller (Gff). The controller
adjusts the manipulated variable to counteract the disturbance, so that
disturbance and manipulated variable affect output variable y together. In
the very best of cases, the manipulated variable would compensate the
disturbance step for step, so that the controlled variable would never be

We also contrast feedforward with the cascade structure of Section 8.5.

Feedforward control also adds another sensor and controller. However,
the concept differs from that of cascade in that the disturbance is
measured, but the manipulated variable does not affect it - there is no
feedback. Feedforward is thus more specific than cascade control: it is
designed to head off a particular disturbance. However, it cannot measure
how well it did, nor can it respond to other disturbances, such as xd1, that
might affect the controlled variable. Hence feedforward is to be applied in
conjunction with a conventional feedback loop. Both the feedback and
feedforward controllers adjust the manipulated variable in Figure 8.6-3.
Lesson 8: Cascade and Feedforward Control Schemes


x'm(s) y' (s)


Gsf(s) Gv(s) Gs(s)

Gff(s) Gc(s)


Figure 8.6-3. Feedforward/feedback control diagram

The feedforward algorithm is not conventional PID. Rather it is specific

to the process and the disturbance. We wish to specify the feedforward
controller transfer function Gff to minimize the effect of xd2 on y. Ideally,
we want yƍ(s) = 0, and from Figure 8.6-2 or 8.6-3 this requires that

G ff (8.6-1)
G sf G v G m

The parallel path through the feedforward controller makes use of advance
warning about the disturbance. Given perfect process models, plus the
ability to render those in transfer function Gff, the compensation can
completely negate the effect of xd2. Of course, perfection is unlikely, as
we will see later.

8.7 summary comparison between cascade and feedforward

On first encounter, one is apt to confuse cascade and feedforward with one
another. Table 8.7-1 shows a side-by-side comparison: assume an existing
process with a feedback controller, such as that in Figure 8.6-1. Call this
controller the primary controller, and compare adding either a secondary
feedback controller in a cascade scheme (Figure 8.5-1), or a feedforward
controller in a feedforward/feedback scheme (Figure 8.6-3). The table
highlights the similarities and differences of the two schemes.

We should conclude the comparison with a clarification of concept: think

of “cascade” and “feedforward” as ways to arrange controllers, not
confined to the specific arrangements given above. That is, it is possible
Lesson 8: Cascade and Feedforward Control Schemes
to have a feedforward controller that adjusts the set point of a feedback
controller. In this circumstance, we have both feedforward control and a
cascade structure; we might say that the feedforward controller “cascades”
to a secondary controller. We will show an example in the next section.

Table 8.7-1. Comparison between cascade and feedforward control

primary controller attempts to correct deviations from set point
primary controller set point unchanged
manipulated variable unchanged
add extra sensor and controller
differences cascade (Figure 8.5-1) feedforward (Figure 8.6-3)
extra measurement: intermediate variable disturbance variable
characteristics of the new intermediate variable measured disturbance
measured variable: x is affected by disturbances x is not affected by
and manipulated variable manipulated variable
x affects controlled variable x affects controlled variable
algorithm of added controller: PID specific to process model
manipulated variable: directed by secondary directed by both feedforward
controller and feedback controllers
primary controller action: varies set point of secondary varies manipulated variable, as
controller before
helps by: reducing the degree of anticipating effects of a
disturbance that reaches the particular disturbance and
controlled variable responding to compensate

8.8 example
A feed stream is pre-heated using a heating oil in a shell-and-tube
exchanger. The outlet temperature is controlled by manipulating the flow
rate of the oil.



Figure 8.8-1 Single-loop control of outlet temperature

The process is subject to several disturbances: the flow rate and inlet
temperature of both process (feed) and service (heating oil) streams may
vary. The latter disturbances are particularly troublesome; because the
heating oil is supplied from a header that feeds other, larger users, swings
Lesson 8: Cascade and Feedforward Control Schemes
in the supply pressure and temperature are frequent. We propose
employing additional measurements and control loops, arranged as shown
in Figure 8.8-2.



Figure 8.8-2 Multi-loop control scheme for outlet temperature

Flow control of the service stream is now effected by a secondary loop.

Two primary controllers cascade to this new secondary loop. The first is
the original feedback controller for the process temperature. The second is
a feedforward controller from the oil temperature.

Should the oil supply pressure decrease, the secondary flow controller will
respond by opening the supply valve. This will allow the oil flow rate to
return to its desired value. The cascade structure responds to the pressure
disturbance before the controlled variable is affected.

Should the oil supply temperature decrease, the feedforward controller

will respond by directing the secondary controller to increase the flow
rate. The higher flow rate of heating oil will tend to maintain the heat duty
in the heat exchanger even as the supply temperature falls.

The temperature controller on the oil requires a feedforward algorithm

because the manipulated variable (heating oil flow) does not affect the
measured variable (heating oil temperature). The temperature controller
on the process stream is a feedback controller that responds to any error in
that temperature, such as may arise from disturbances in the process
stream flow and temperature. Both controllers cascade to a secondary


8.9 practical feedforward controllers

Of course there is an overwhelming variety of process models, but if we
recall that a large class of processes may be successfully represented by
FODT, we can derive a feedforward controller with some claim to
generality (Marlin, 2000). Lumping the valve and sensor into the process
description, the transfer function (8.6-1) becomes
Lesson 8: Cascade and Feedforward Control Schemes
K d W m s  1 ( Td T m )
G ff  e (8.9-1)
K m Wds  1

We generalize this form by associating a controller gain with the ratio of

process gains, a lead time with the xm process path, a lag time with the xd2
process path, and a dead time with the difference between disturbance and
manipulated process dead times. These four parameters may then be
tuned, as with PID parameters, to improve controlled variable response to

Tlead s  1 Tff
G ff K ff e (8.9-2)
Tlag s  1

The effect of gain is to amplify the controller response to an input. The

gain satisfies the steady state relationship between disturbance and
manipulated variables. The effect of dead time is to delay the controller's
response, so that it will not affect the controlled variable prematurely.
Such a step is appropriate if the disturbance dead time exceeds that of the
process. However, if the disturbance dead time is less, perfect control
requires a negative Tff, which implies predicting the onset of future
disturbances! Such a time-machine would be very useful, for many
purposes, but we are not likely to find one - thus the parameter Tff would
be set to zero.

We can illustrate (8.9-2) by calculating the output of the controller upon

receiving a step input of magnitude A. Applying the step change and
inverting, we find

ª Tlead  Tlag  t Tff Tlag º

x coc AK ff «1  e » (8.9-3)
¬« Tlag ¼»

The lead/lag elements shape the development of the response, as shown in

Figure 8.9-1. For example, if Tlag is set to be greater than Tlead, the
controller output increases with time. This is appropriate if the
disturbance propagates more slowly through the process than does the
manipulated variable. If the reverse is true, then a more vigorous
manipulated variable response is in order from the start, and so Tlead is set
larger than Tlag.
Lesson 8: Cascade and Feedforward Control Schemes

change in controller output

lead = 2 x lag
2 lead = lag
lead = 0.5 x lag

0 5 10 15 20 25

Figure 8.9-1 Response of controller Gff to a step input

8.10 summing feedforward and feedback signals at one valve

Each controller specifies, at any time, how much the valve should be
moved from its previous position. These two outputs are summed and
then sent to the valve transducer.

It may be that the two controllers will oppose each other, so that the sum
of the outputs is a smaller movement than each individual controller
would have directed. This is not necessarily a problem: each controller
responds in its own area of expertise (feedback responding to present
error; feedforward forecasting the effects of present disturbance) and the
sum of the outputs addresses both concerns.

Of course, the sum of the controller outputs could fall outside the physical
0 - 100% range of the valve. In this case, the valve can move only to its
limit in response. If this is a frequent occurrence, the manipulated variable
may be insufficiently strong to overcome disturbances.

8.11 flow control

As we have already seen in examples, a very common application of a
cascade scheme is to use a flow controller as an inner cascade loop. This
allows the flow to be less affected by, for example, pressure differences,
so that it can be a more reliable manipulated variable in the outer loops.

We examine the flow control loop. The equipment comprises a valve, a

flowmeter, and connecting pipe.
Lesson 8: Cascade and Feedforward Control Schemes

Figure 8.11-1 Equipment to measure and manipulate flow

If we regard the valve as the final control element and the flowmeter as the
sensor, there is really no process left. Hence a block diagram will show
the process as a unity transfer function: the controlled variable and
manipulated variable are identical:



flow flow

Gv(s) valve flowmeter Gs(s)

controller output

set point
from outer loop
Figure 8.11-2 Block diagram of flow control loop

The valve affects the flow through the stem position, as described in
transfer function Gv. The diagram also identifies two disturbances that
affect flow: a change in the pressure difference across the valve (as might
result from variations in sources and sink conditions) and changes in
physical properties of the flowing fluid. All these transfer functions come,
as in other examples, from a linearized model of flow through a valve.

After our emphasis on distinguishing manipulated and controlled

variables, having the manipulated variable be identical to the controlled
variable may seem peculiar. An alternative point of view is to consider
the manipulated variable to be the valve stem position, as shown in Figure
Lesson 8: Cascade and Feedforward Control Schemes



stem position flow


1 flowmeter Gs(s)

controller output

set point
from outer loop
Figure 8.11-3 Alternative block diagram of flow control loop

In effect, the transfer function that describes how valve stem position
affects flow has moved from Gv in Figure 8.11-2 to Gm in Figure 8.11-3.
In this series of Lessons, we have preferred the point of view of Figure
8.11-2; that is, a controller acts through a final control element to produce
a flow, which may then be used to manipulate a process output.


8.12 cascade performance

From Figure 8.5-1, we derive the transfer function for the closed loop
cascade structure. With nested loops in block diagrams, it is best to begin
with the inner loop.

x i' (s) G d 2 x 'd 2 (s)  G m 2 x 'm (s)

G d 2 x d' 2 (s)  G m 2 G v G c 2 x co
1 (s )  G s 2 x i (s)

so that

Gd2 G G G
x i' (s) x d' 2 (s)  m 2 v c 2 x 'co1 (s) (8.12-2)
1  G L2 1  G L2

where the transfer function around the secondary loop is

G L2 G m 2G vG c 2G s 2 (8.12-3)
Lesson 8: Cascade and Feedforward Control Schemes
With xiƍ known, we continue in the outer loop.

y ' (s) G d1x d' 1 (s)  G m1x i' (s)

G m1G d 2 ' G G G G
G d1x 'd1 (s)  x d 2 (s)  m1 m 2 v c 2 x 'co1 (s) (8.12-4)
1  G L2 1  G L2

G d1x 'd1 (s) 

G m1G d 2 '
1  G L2
x d 2 (s)  m1 m 2 v c 2 G c1 G sp y sp
1  G L2

(s)  G s1 y ' (s)

Solving for yƍ

G d1 1  G L 2 ' G m1G d 2 G G G G G G '

y ' (s) x d1 (s)  x 'd 2 (s)  m1 m 2 v c 2 c1 sp y sp (s) (8.12-5)
1  G L 2  G L1 1  G L 2  G L1 1  G L 2  G L1

where the transfer function around the primary loop is

G L1 G m1G m 2 G v G c 2 G c1G s1 (8.12-6)

Equation (8.12-5) is general for a two-loop cascade structure with the

disturbances arranged as in Figure 8.5-1. It can be specialized by
substituting particular transfer functions for the components in the

8.13 cascade tuning

Marlin (2000) suggests tuning the inner loop first, with the outer loop in
manual setting. Then tune the outer loop with the inner loop in automatic

8.14 cascade closed loop stability

The characteristic equation for the cascade structure, obtained from (8.12-
5), contains the transfer functions for both controllers. Hence both
controllers affect the poles and thus stability of the cascade.

8.15 feedforward performance and stability

From Figure 8.6-3, we derive the transfer function for the closed loop
feedback/feedforward structure.

y ' (s) G d1x 'd1 (s)  G d 2 x 'd 2 (s)  G m x 'm (s)

G d1x 'd1 (s)  G d 2 x 'd 2 (s)  G m G v x 'cof (s)  x 'co (s) (8.15-1)
G d1x (s)  G d 2 x (s)  G m G v G ff G sf x (s)  G m G v G c G sp y (s)  G sp y (s)

Solving for controlled variable yƍ

Lesson 8: Cascade and Feedforward Control Schemes
G d1 ' G  G m G v G ff G sf ' G G GG '
y ' (s) x d1 (s)  d 2 x d 2 (s)  m v c sp y sp (s) (8.15-2)
1 GL 1 GL 1 GL

where the transfer function around the feedback loop is

GL G mG vG cG s (8.15-3)

Notice that the feedforward controller affects only the transfer function for
disturbance xd2; other disturbances and the set point have the usual
feedback loop transfer functions. By (8.6-1) we try to make the xd2
transfer function zero. However, if it is not, the feedback controller is also
available, through the transfer function denominator, to respond to xd2, as
it does to other disturbances.

Because the characteristic equation obtained from (8.15-2) does not

depend on the feedforward controller, adding a feedforward loop has no
effect on tuning the feedback controller for closed loop stability. Even so,
if the feedforward controller itself is unstable, the ensemble is likely to be

8.16 feedforward tuning

Ideally all the other disturbances would subside so that the feedback loop
could be put in manual and the feedforward loop tuned to respond to its
particular disturbance. One may not be so fortunate, however.

8.17 conclusion
By going to the trouble and expense of extra measurements, and obtaining
deeper knowledge of the process, we are enabled to improve on the
performance of the single-loop PID controller. This is not to say that
cascade or feedforward enhancements are always to be recommended -
they must be technically feasible (the measurements and character of the
new variables being appropriate) and economically justified. This of
course, is a familiar story to engineers.

Feedforward and cascade point the way to further control schemes that
make use of process models. Various forms of model-based control can
offer advantages in single control loops, but really come into their own
when we consider how individual control loops may interact - that is,
when realistic process models are MIMO - multiple input/multiple output.

8.18 reference
Marlin, Thomas E. Process Control. 2nd ed. Boston, MA: McGraw-Hill, 2000.
ISBN: 0070393621.