IBEHS 4A03

Section 10: Controller Design

Jake Nease
McMaster University
1
Topic Outline
▪ Designing controllers based on tuning correlations
– Cohen-Coon
– ITAE
– Ciancone

▪ Designing controllers based on desired closed-loop responses

– Direct Synthesis
– First and second-order systems

▪ Improving controller performance

– The mystical “tuning” procedure
– Optimization methods

Section10: Controller_Design 2
Measuring Controller Performance
Easy 12.

3
 1
The Tuning Procedure 𝐺𝐶 𝑠 = 𝐾𝐶 +
𝜏𝐼 𝑠
+ 𝐾𝐷 𝑠

▪ Tuning a PID controller involves the selection of the controller parameters 𝐾𝐶 , 𝜏𝐼 and
𝐾𝐷 in order to optimize the performance of the closed-loop system

▪ Optimal performance typically includes:

– Controlled variable performance (obviously, the ability for 𝑌 𝑠 to track 𝑅 𝑠 is very
important)
– Manipulated variable behaviour (the next most important variable in a closed-loop system
is 𝑈 𝑠 )
– Accounting for model error (most systems with approximated linear dynamics still need to
perform well when pushed away from the linearization point)

Section10: Controller_Design 4
The Tuning Procedure
1
𝐺𝐶 𝑠 = 𝐾𝐶 + + 𝐾𝐷 𝑠
𝜏𝐼 𝑠

▪ The process generally turns into an optimization problem wherein the objective is to
minimize some quantitative measure of the controller’s performance:

min 𝜙 = 𝑓 𝑡, 𝑦, 𝐾𝐶 , 𝜏𝐼 , 𝐾𝐷 …
𝐾𝐶 ,𝜏𝐼 ,𝐾𝐷

Minimize “over” the The “objective” is a This is the general

variables chosen via function of the representation of an
the optimization following variables, optimization problem
which must be
related

▪ Any ideas for 𝜙?

Section10: Controller_Design 5
Tuning Optimization Objectives 𝜙
▪ Integral Absolute Error (IAE) ∞ ∞
𝐼𝐴𝐸 = න 𝑟 𝑡 − 𝑦 𝑡 𝑑𝑡 = න 𝜖 𝑡 𝑑𝑡
– Most popular – penalizes all errors equally 0 0

▪ Integral Squared Error (ISE) ∞

2
∞
𝐼𝑆𝐸 = න 𝑟 𝑡 −𝑦 𝑡 𝑑𝑡 = න 𝜖 𝑡 2 𝑑𝑡
– Also popular – penalizes large errors more heavily 0 0

▪ Integral Time-Weighted Absolute Error (ITAE) ∞ ∞

𝐼𝑇𝐴𝐸 = න 𝑡 𝑟 𝑡 − 𝑦 𝑡 𝑑𝑡 = න 𝑡 𝜖 𝑡 𝑑𝑡
– More heavily penalizes errors at longer times 0 0

▪ Integral Time-Weighted Squared Error (ITSE) ∞

2
∞
𝐼𝑇𝑆𝐸 = න 𝑡 𝑟 𝑡 − 𝑦 𝑡 𝑑𝑡 = න 𝑡𝜖 𝑡 2 𝑑𝑡
– Very heavily penalizes large errors at long times 0 0

Section10: Controller_Design 6
Examples from MATLAB

∞
𝐼𝐴𝐸 = න 𝜖 𝑡 𝑑𝑡
0

∞
𝐼𝑆𝐸 = න 𝜖 𝑡 2 𝑑𝑡
0

∞
𝐼𝑇𝐴𝐸 = න 𝑡 𝜖 𝑡 𝑑𝑡
0

∞
𝐼𝑇𝑆𝐸 = න 𝑡𝜖 𝑡 2 𝑑𝑡
0

Section10: Controller_Design 7
Where to Start?
▪ This is all well and good… how do we come up with a starting point?

▪ We can use these measures of error (and the work of others) to choose a
combination of tuning parameters that should perform well for FO processes

▪ These methods aim to provide a good baseline error value… then, we can focus on
one of the optimization objectives we discussed to fine tune parameters

Section10: Controller_Design 8
Controller Design Correlations
Like a bladed sand wedge onto the green.

Ugly, but effective.

9
Some Important Notes
▪ The following methods are good for systems that can be approximated as first-order plus
dead time models

𝐾𝑒 −𝜃𝑠
𝐺𝑃 𝑠 ≈
𝜏𝑠 + 1

▪ It is also important that the “process” includes any measurement and actuation dynamics
(aka sensor dynamics 𝐺𝑀 and 𝐺𝑉 )
– If these are not known ahead of time, the following procedures will still work, but be suboptimal

▪ The Ziegler-Nichols tuning procedure is one that we have already covered!

– Applicable to any closed-loop stable system of any order
– Requires a frequency response analysis

Section10: Controller_Design 10
Some Important Notes
▪ These controller tuning correlations consider the controller transfer function to be in
ideal form

1
𝐺𝐶 𝑠 = 𝐾𝐶 1+ + 𝐾𝑑 𝑠
𝜏𝐼 𝑠

▪ If implementing in parallel form, make sure you “decouple” the effects of 𝐾𝐶 from
the other tuning parameters!

Section10: Controller_Design 11
 𝐾𝑒 −𝜃𝑠
Cohen-Coon Correlations 𝐺𝑃 𝑠 ≈
𝜏𝑠 + 1

▪ The Cohen-Coon correlations give an estimate of initial guesses for controller tuning
parameters based on process dynamics
– Attempt to achieve a decay ratio of ¼
– Typically, not “optimal” but offer a good starting point
– Based on the IAE metric
– Like ZN, parameters depend on controller structure

1𝜏 𝜃
P 𝐾𝐶 =
𝐾𝜃
1+
3𝜏

3𝜃
30+
PI
1𝜏 𝜃 𝜏
𝐾𝐶 = 0.9 + 𝜏𝐼 = 𝜃 20𝜃
𝐾𝜃 12𝜏 9+
𝜏

6𝜃
32+
PID
1𝜏 4 𝜃 𝜏 4
𝐾𝐶 = + 𝜏𝐼 = 𝜃 8𝜃 𝐾𝐷 = 𝜃 2𝜃
𝐾𝜃 3 4𝜏 13+ 11+
𝜏 𝜏

Section10: Controller_Design 12
 𝐾𝑒 −𝜃𝑠
Cohen-Coon Correlations 𝐺𝑃 𝑠 ≈
𝜏𝑠 + 1

▪ Discussion
– What happens to 𝐾𝐶 as 𝐾 increases?
– What happens to 𝜏𝐼 as 𝜏 decreases?
– What happens to 𝜏𝐼 as 𝜃 increases?
– NOTE that this correlation is inappropriate if 𝜃 = 0 (why?)

1𝜏 𝜃
P 𝐾𝐶 =
𝐾𝜃
1+
3𝜏

3𝜃
30+
PI
1𝜏 𝜃 𝜏
𝐾𝐶 = 0.9 + 𝜏𝐼 = 𝜃 20𝜃
𝐾𝜃 12𝜏 9+
𝜏

6𝜃
32+
PID
1𝜏 4 𝜃 𝜏 4
𝐾𝐶 = + 𝜏𝐼 = 𝜃 8𝜃 𝐾𝐷 = 𝜃 2𝜃
𝐾𝜃 3 4𝜏 13+ 11+
𝜏 𝜏

Section10: Controller_Design 13
 𝐾𝑒 −𝜃𝑠
ITAE Correlations 𝐺𝑃 𝑠 ≈
𝜏𝑠 + 1

▪ The ITAE correlations developed by Smith and Corripio (1997) are based on a general
minimization of ITAE
– Correlation depends on if the controller is tuned for set-point tracking or disturbance rejection

Objective Controller Mode A B

Disturbance Rejection P 0.859 -0.977

PI
I 0.674 -0.680
𝑌𝑃 = 𝐾𝐾𝐶
𝜏 P 1.357 -0.947
𝑌𝐼 =
𝐵 𝜏𝐼
PID I 0.842 -0.738
𝜃 𝑌𝐷 =
𝐾𝐷
𝑌=𝐴 𝜏 D 0.381 0.995
𝜏 Set-Point Tracking P 0.586 -0.916
PI
I 1.03* -0.165*
𝑌𝑃 = 𝐾𝐾𝐶
𝝉 𝜽 P 0.965 -0.85
𝒀𝑰 = = 𝑨 + 𝑩
𝝉𝑰 𝝉 PID I 0.796* -0.1465*
𝐾𝐷
𝑌𝐷 = D 0.308 0.929
𝜏

*Note the different relation for 𝑌𝐼 for the set-point tracking controller
Section10: Controller_Design 14
 𝐾𝑒 −𝜃𝑠
Ciancone Correlations 𝐺𝑃 𝑠 ≈
𝜏𝑠 + 1

disturbance setpoint
▪ Developed by Ciancone and Marlin (1997)
– Separate correlations for setpoint tracking and
disturbance rejection

▪ Based on dimensionless tuning coefficients

– These depend on the process gain, time constant,
and dead time

▪ These correlations are for PID control

– Different relations for PI (not included here, nor are
they necessary)

Section10: Controller_Design 15
Correlations - Conclusions
▪ Controller design using correlations has limitations
– These correlations are restricted to first-order plus dead time systems (except ZN from the
previous section)
· Good for process engineering, less likely for biomedical applications

– However, note that many higher-order systems can be approximated as first order,
especially if the system is overdamped
· Although not perfect by any means, something very important about these correlations must be
remembered…

▪ These correlations are for initial guesses of the tuning parameters only
– You will still need to “tinker”
– This is where your knowledge of tuning parameter effects on output are handy

Section10: Controller_Design 16
Workshop 10.1: Controller Design Using Tuning Correlations

▪ Consider the first-order plus dead time process, assuming this captures all sensor and
actuation dynamics:

3𝑒 −2𝑠
𝐺𝑃 =
5𝑠 + 1

▪ Design a PID controller for this process for efficient set-point tracking using the
correlations:
– Cohen-Coon ← we’ll do this now
– ITAE ← we’ll do this now
– Ciancone (in solution manual)
– Ziegler-Nichols (in solution manual)
▪ Compare the results of each set of tunings by simulating them in Simulink and
plotting/commenting on the results

Section10: Controller_Design 17
Workshop 10.1 Solution
▪ Cohen-Coon

Section10: Controller_Design 18
Workshop 10.1 Solution
▪ ITAE

Section10: Controller_Design 19
Workshop 10.1 Solution - Simulation

Section10: Controller_Design 20
Direct Synthesis
Taking the guess out of “Guess that’ll work”

21
The Idea
▪ Direct synthesis exploits the structure of the closed-loop feedback system to choose
𝐺𝐶
– Unlike a PID controller specifically, the structure of 𝐺𝐶 depends on the structure of the
process
– Can lead to controllers that “appear” as P/PI/PID or can be completely different

▪ We will “reverse-engineer” what 𝐺𝐶 must be in order for the closed-loop process to

exhibit a desired response

Section10: Controller_Design 22
The Idea
▪ Consider a typical feedback loop, ignoring the disturbance (we’ll just hope for the
best)

D(s)
Let’s take all this garbage to be 𝐺 𝑠 ,
which is the overall response of the
system that is fed back to 𝐺𝐶 𝑠
GD(s)

R(s) + E(s) U(s) + Y(s)

-
GC(s) GV(s) GP(s) +

YM(s)
GM(s)

Section10: Controller_Design 23
The Idea
R(s) + Y(s)
GC G
-

𝑌 𝑠 𝐺 𝑠 𝐺 𝑠
▪ 𝑅 𝑠
= 1+𝐺𝐶 → Can we isolate for 𝐺𝐶 𝑠 ?
𝐶 𝑠 𝐺 𝑠

▪ WORKSHOP 10.2: Direct Synthesis – Deriving an Expression for 𝐺𝑐

𝑌(𝑠)
1
– Show that 𝐺𝐶 𝑠 =
𝑅(𝑠)
𝐺(𝑠) 𝑌(𝑠)
1−𝑅(𝑠)

𝑌 𝑠
▪ Wait, I have a clever idea… If I want to choose 𝐺𝐶 (𝑠) so that has a certain shape, why
𝑅 𝑠
𝑌 𝑠
don’t I substitute a desired relationship for and solve for 𝐺𝐶 𝑠 ?
𝑅 𝑠
▪ Now that is a fascinating idea. This is called Direct Synthesis

Section10: Controller_Design 24
Workshop 10.3: Direct Synthesis – Perfect Control
▪ Determine the controller transfer function 𝐺𝐶 𝑠 if the desired response to a unit step
in 𝑅 𝑠 is a unit step, then tell me why this won’t work

y(t)
r(t)
y(t)

Section10: Controller_Design 25
Workshop 10.3: Direct Synthesis – Perfect Control
▪ Can we brainstorm a different “ideal” response?

𝐺𝐶 → ∞
Section10: Controller_Design 26
Direct Synthesis: FO CLTF Response
▪ OK, well that was a bit of a pipe dream.
▪ Instead, let’s try to produce a “good” response recognizing that matching 𝑟 𝑡 exactly
is impossible… There will need to at least be SOME sort of dynamics
▪ Here’s an idea: what if we target a first order response:

y(t)
r(t)
𝑌 𝑠 1 0.63
=
𝑅 𝑠 𝜏𝑐 𝑠 + 1

t
τc

Section10: Controller_Design 27
Direct Synthesis: FO CLTF Response
▪ You may ask: “what is 𝜏𝑐 ”?
– That’s a good question. It is the first-order time constant of the idealized closed-loop
response we are trying to create using our controller
– We can CHOOSE 𝜏𝑐 if we want!
▪ WORKSHOP 10.4: Direct Synthesis – FO CLTF for a FO Process
𝑌 𝑠 1 𝐾
– Derive 𝐺𝐶 𝑠 if =𝜏 and 𝐺 𝑠 = 𝜏𝑠+1
𝑅 𝑠 𝑐 𝑠+1

y(t)
r(t)
0.63

t
τc
Section10: Controller_Design 28
Workshop 10.4 Solution

Section10: Controller_Design 29
Workshop 10.4 Solution
▪ Is there anything familiar about 𝐺𝐶 𝑠 ?

𝜏 1
𝐺𝐶 𝑠 = 1+
𝐾𝜏𝑐 𝜏𝑠
Section10: Controller_Design 30
Direct Synthesis: Higher Order TF
▪ Second-order processes follow a similar path

▪ WORKSHOP 10.5: Direct Synthesis – FO CLTF for a SO Process

– Consider the following second-order process:

𝐾
𝐺 𝑠 =
𝜏1 𝑠 + 1 𝜏2 𝑠 + 1
𝑌 𝑠
– Determine a controller 𝐺𝐶 𝑠 using direct synthesis if (like the previous workshop) =
𝑅 𝑠
1
𝜏𝑐 𝑠+1

Section10: Controller_Design 31
Workshop 10.5 Solution
▪ Is there anything familiar about THIS 𝐺𝐶 𝑠 ?

𝜏1 + 𝜏2 1 𝜏1 𝜏2
𝐺𝐶 𝑠 = 1+ + 𝑠
𝐾𝜏𝑐 𝜏1 + 𝜏2 𝑠 𝜏1 + 𝜏2
Section10: Controller_Design 32
Direct Synthesis: Remarks
▪ We really only can choose 𝜏𝑐 , which impacts 𝐾𝑐 of the controller in ideal form
– Choosing a faster response is going to result in a higher gain
▪ Naturally, a high process gain 𝐾 leads to a lower controller gain 𝐾𝑐
– This makes sense intuitively – if the impact of the controller is high, we want the
aggressiveness of the controller to be low
▪ Typically, start with 𝜏𝑐 = 0.25𝜏
– If you are dealing with a second-order process, use the higher 𝜏 to be conservative
▪ Works for ANY system to provide initial guesses for controller parameters
– Will STILL need to tune!

Section10: Controller_Design 33
Optimizing Controller Performance
Minimum error, maximum effort

34
How do we Optimize the Response?
▪ Now that we have an initial guess, we can return to our optimization objectives to achieve
well-tuned performance
▪ All strategies involve:
1. Selecting a combination of tuning coefficients
2. Simulating the response (or performing tests on real system if required)
3. Evaluating the tuning objective 𝜙
4. Make some decision based on 𝜙 and return to (1)
▪ Step (4) can be performed in several ways:
– Graphical analysis (how does 𝜙 change for different decision vars?)
– Surface response analysis (use local changes in 𝜙 to predict what values of decision vars will
further lower 𝜙)
· AKA local approximated optimization
· AKA Design of Experiments (DOE)
– Formal optimization (certainly not required in this course, but there are some good ones you
could take)

Section10: Controller_Design 35
Graphical Analysis Examples
▪ When selecting one parameter (𝐾𝐶 here), we can plot the IAE for a variety of
simulations and choose the lowest

Section10: Controller_Design 36
Image from Marlin (2000)
Graphical Analysis Examples
▪ When selecting multiple parameters (𝐾𝐶 ant 𝜏𝐼 here), we can plot the IAE for a variety
of simulations and choose the lowest
– Can also be subjected to constraints

QUESTION – Where does this come

from? We have already done it before…

Section10: Controller_Design 37
Image from Marlin (2000)
Remarks
▪ Formal optimization is nice, but it requires an accurate process model
– We can sometimes treat the system as “black box”
– System identification is crucial in absence of rigorous dynamic models

▪ Can include constraints!

– Min/max value of 𝑦 𝑡
– Min/max value of 𝑢 𝑡
– Max allowed settling time
– Stability (duh!)

▪ Practically speaking, we often try to apply the known relationships between tuning
parameters
– Tweak or “adjust the knobs” of the controller to achieve “optimal” performance (loosely speaking)

Section10: Controller_Design 38
Conclusions: Controller Design
▪ Controller tuning is a detailed process
– Can use metrics (IAE, ISE, etc.) to help
▪ Initial guesses exist for FO processes
– Cohen-Coon
– Ciancone
– ITAE
▪ Zielger-Nichols is applicable to higher-order processes
– Requires a frequency analysis and a stable open-loop system
▪ Direct synthesis is another method that can be used on any process
– Unique for each type of system
– Make sure to account for sensor dynamics!
– Does not guarantee a typical PID control structure
▪ Regardless of your initial guess, optimization can be used to fine-tune

Section10: Controller_Design 39