Sie sind auf Seite 1von 32

EECE 460 : Control System Design

PID Controller Design and Tuning

Guy A. Dumont
UBC EECE

January 2011

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

1 / 32

Contents

Introduction

Control Specifications

Empirical Tuning Methods


Ziegler-Nichols
Cohen-Coon Method

Model-Based Methods
The Dahlin Controller
-Tuning
Haalmans Method
Internal Model Control

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

2 / 32

Introduction

PID Controller Design and Tuning


Vast literature on the topic. There is a plethora of techniques
As strm and Hgglund put it:
. . . there are many different types of control problems and consequently many
different design methods. To only use one method is as dangerous as to only
believe in empirical tuning rules
Empirical, historical methods
Ziegler-Nichols oscillation or frequency response method
The Cohen-Coon method

Beyond Ziegler-Nichols Tuning


Lambda-Tuning
Haalmans method
Skogestads IMC rules

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

3 / 32

Control Specifications

Control Specifications

For response to step setpoint changes, the criteria below have been often
used
Z

ITAE =

t|e(t)|dt;
0

ITE =

te(t)dt;
0

ITSE =

te2 (t)dt

Rise time
Settling time
Decay ratio
Overshoot ratio
Steady-state error

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

4 / 32

Control Specifications

Control Specifications
For attenuation of load disturbances
Z

IAE =

|e(t)|dt;

IE =

e(t)dt;

ISE =

e2 (t)dt

It can be shown that for a unit step disturbance


IE =

Ti
Kc

Sensitivity to measurement noise


Transmission of measurement noise to control signal is
Gnu = C/(1 + G0 C)
Normally at high frequencies G0 C 0 and Gnu C.
For a PID controller, the high frequency gain is Kc (1 + N) 10Kc
Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

5 / 32

Control Specifications

Control Specifications
Gain (or amplitude margin)
Am =

1
|G0 C(iu )|

where u is the ultimate frequency, i.e. arg G0 C(iu ) =


Phase margin
m = + arg G0 C(ic )
where c is the cross-over frequency, i.e. |G0 C(ic )| = 1
Robustness to model uncertainty
Ms = max |

1
| = max |S(i)|
1 + G0 (i)C(i)

where S is the sensitivity function


Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

6 / 32

Control Specifications

Control Specifications
1/Ms is the stability margin

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

7 / 32

Control Specifications

Control Specifications
A number of optimization-based techniques use Ms to perform the tuning

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

8 / 32

Empirical Tuning Methods

Ziegler-Nichols

Ziegler-Nichols Frequency Method


Only valid for open loop stable plants and it is carried out through the
following steps
1
2

Set the true plant under proportional control, with a very small gain.
Increase the gain until the loop starts oscillating. Note that linear
oscillation is required and that it should be detected at the controller
output.
Record the controller critical gain Kp = Kc and the oscillation period of
the controller output, Pc .
Adjust the controller parameters according to the Table 6.1 on next slide.

There is some controversy regarding the PID parameterization for which


the Z-N method was developed, but the version described here is, to the
best knowledge of the authors, applicable to the parameterization of
standard form PID:


Td s
1
+
C(s) = Kp 1 +
Tr s 1 + d s
Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

9 / 32

Empirical Tuning Methods

Ziegler-Nichols

Ziegler-Nichols Frequency Method

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

10 / 32

Empirical Tuning Methods

Ziegler-Nichols

Ziegler-Nichols Frequency Method


Consider a plant of the form
G0 (s) =

K0 es0
0 s + 1

where

0 > 0

Response is very sensitive to the ratio between the delay and the time
constant
Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

11 / 32

Empirical Tuning Methods

Ziegler-Nichols

Interpretation of the Frequency Response Method


The critical gain Kc is such that the Nyquist plot of the loop gain goes
through the 1 point at the oscillation frequency c , i.e.
Kc G0 (jc ) = 1
Method where one point of the Nyquist curve is moved

With PI control, ultimate point moved to (-0.4,+0.08i)


With PID control, ultimate point moved to (-0.6,-0.28i)
Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

12 / 32

Empirical Tuning Methods

Ziegler-Nichols

Interpretation of the Frequency Response Method

Generally better for PID than for PI


Quarter amplitude decay ratio gives poorly damped closed-loop system
No tuning parameter
Based on only one point on the Nyquist curve
Major shortcoming of the Z-N method is that it requires that the plant be
forced to oscillate with a non predictable amplitude.
This can be dangerous and expensive!

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

13 / 32

Empirical Tuning Methods

Ziegler-Nichols

Performance of Ziegler-Nichols Method

Define the normalized gain of a process as:




G0 (iu )


=
G0 (0)
and the normalized dead time as:
=

D
D+T

where D is the dead time and T is the dominant time constant

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

14 / 32

Empirical Tuning Methods

Ziegler-Nichols

Performance of Ziegler-Nichols Method

Small and : Easy to control. Other methods usually give better


performance than Ziegler-Nichols
Intermediate and : Primary range of use for Ziegler-Nichols tuning.
Setpoint weighting can reduce overshoot
and close to 1: Dynamics dominated by dead time. Ziegler-Nichols
should not be used. Actually, PID should probably not be used at all.

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

15 / 32

Empirical Tuning Methods

Cohen-Coon Method

Cohen-Coon Method
A simple model G0 (s) =
1

K0 es0
0 s+1

is built using the procedure below

With the plant in open loop, take the plant manually to a normal operating
point. The plant at y(t) = y0 for a constant u(t) = u0 .
At an initial time, t0 , apply a step change to the plant input, from u0 to u
(this should be in the range of 10 to 20% of full scale).
Record the plant output until it settles to the new operating point. Assume
you obtain the curve shown below (m.s.t. stands for maximum slope
tangent. This curve is known as the process reaction curve.

Compute the parameter model as follows:


K0 =

Guy A. Dumont (UBC EECE)

y y0
;
u u0

0 = t1 t0 ;

EECE 460 PID Tuning

0 = t2 t1
January 2011

16 / 32

Empirical Tuning Methods

Cohen-Coon Method

Cohen-Coon Method
Consider again a plant of the form
G0 (s) =

K0 es0
0 s + 1

where

0 > 0

Cohen-Coon propose the following tuning

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

17 / 32

Empirical Tuning Methods

Cohen-Coon Method

Cohen-Coon Method

Response is still quite sensitive to the ratio between the delay and the
time constant

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

18 / 32

Model-Based Methods

The Dahlin Controller

Dahlin Controller
Developed in 1968 by Dahlin and Higham
Used extensively in industry in its digital form, particularly in paper
machine control systems
Consider a feedback control loop with a process P(s) and controller
C(s). the closed-loop system can be written as:
Y(s) =

P(s)C(s)
1
Ysp (s) +
W(s)
1 + P(s)C(s)
1 + P(s)C(s)

with setpoint Ysp (s) and disturbance W(s).


If D(s) denotes the desired closed-loop transfer function then we want to
solve for C(s) such that
D(s) =

Guy A. Dumont (UBC EECE)

P(s)C(s)
1 + P(s)C(s)

EECE 460 PID Tuning

January 2011

19 / 32

Model-Based Methods

The Dahlin Controller

Dahlin Controller
The process transfer function P can be factored into two parts:
PN which contains the dead time and the right-half plane or poorly
damped zeros, i.e. elements which cannot be cancelled by the controller
PM which contains the minimum phase elements that can be cancelled by
the controller

Thus D(s) = PN (s)D (s) where D (s) is the arbitrary portion of the
desired closed-loop transfer function and is usually chosen as
D (s) =

1
1+s

Solving for the controller C(s) then gives


Dahlin Controller
C(s) =
Guy A. Dumont (UBC EECE)

1
D (s)
PM (s) 1 PN (s)D (s)
EECE 460 PID Tuning

January 2011

20 / 32

Model-Based Methods

The Dahlin Controller

Predictive PI Control
Consider the process
Kp sL
e
1 + Ts
The desired closed-loop transfer function is
P(s) =

D(s) =

esL
1+s

The Dahlin controller is then


Dahlin controller for FOPDT
C(s) =

1 + sT
Kp (1 + s esL )

This can be interpreted as a PI controller with dead-time compensation.


When L = 0, this becomes a simple PI controller.
Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

21 / 32

Model-Based Methods

-Tuning

Lambda Tuning for FOPDT


Recently popularized in pulp and paper industry by Bialkowski
Based on the Dahlin algorithm
Dead time approximated by rational transfer function
If dead time L is approximated by
esL 1 sL
the controller is
C(s) =

1 + sT
Kp ( + L)s

which is a PI controller with


PI -tuning for FOPDT
Ti = T

Guy A. Dumont (UBC EECE)

and

EECE 460 PID Tuning

Kc =

T
Kp ( + L)

January 2011

22 / 32

Model-Based Methods

-Tuning

Lambda Tuning for FOPDT


If dead time L is approximated by
esL
the controller is
C(s) =

1 sL/2
1 + sL/2

(1 + sL/2)(1 + sT)
Kp s(L + + s L/2)

which can be approximated by a PID controller


PID -tuning for FOPDT
CPID (s) =

Guy A. Dumont (UBC EECE)

(1 + sL/2)(1 + sT)
Kp s(L + )

EECE 460 PID Tuning

January 2011

23 / 32

Model-Based Methods

-Tuning

Typical Recommendations for Process Industries

Bialkowski1 makes the following recommendations when using


-tuning:
Flow control: = two to three times the process time constant
Temperature control: PID tuned with slightly smaller than larger
process time constant
Consistency control: > process time constant plus dead time
Tank level control: three dead times

This is actually a special case of pole placement, to be seen in more


details very soon!
A drawback of -tuning is that it cancels the process poles, which can
give poor load disturbance rejection characteristics

1 N.J.

Sell (Ed.), Process Control Fundamentals for the Pulp & Paper Industry, Tappi Press

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

24 / 32

Model-Based Methods

Haalmans Method

Haalmans Method

Haalman has suggested choosing an "ideal" loop transfer function


Gl = PC and then computing the controller C = Gl /P.
Haalman suggests choosing
Gl (s) =

2 sL
e
3Ls

This gives Ms = 1.9


Note that only the dead time influences the loop transfer function. All
process poles and zeros are cancelled which might lead to difficulties

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

25 / 32

Model-Based Methods

Haalmans Method

Haalmans Method

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

26 / 32

Model-Based Methods

Haalmans Method

Haalmans Method

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

27 / 32

Model-Based Methods

Haalmans Method

Haalmans Method
A problem with Haalman method is that it cancels all the poles and
zeros. Cancelling all poles and zeros may be bad.
Consider the plant G(s) = esL /(1 + sT) with the PI controller
C(s) = K(1 + sT)/sT, then we can write
dy(t)
1
= (u(t L) y(t))
dt
T


dy(t) y(t)
K
du(t)
= K
+
= u(t L)
dt
dt
T
T
With initial conditions y(0) = 1 and u(t) = 0 for L < t < 0 the
open-loop response is yo (t) = et/T
In closed-loop, equations above show that u(t) = 0, hence ycl = yol
The controller does nothing to reduce the error! We will discuss this
phenomenon in more details when presenting the Q-design
This is a problem for all design methods that cancel all process poles.
Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

28 / 32

Model-Based Methods

Haalmans Method

Haalman Method

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

29 / 32

Model-Based Methods

Internal Model Control

Internal Model Control Procedure


The IMC is a general control design procedure
It factorizes the plant model G0 (s) into an invertible minimum-phase
Gm (s) part and a non-invertible all-pass Ga (s) part
G0 (s) = Gm (s)Ga (s)
We then choose
T(s) = F(s)Ga (s)
where F(s) is a low-pass filter typically of the form F(s) = 1/(c s + 1)n
Knowing that T = G0 C/(1 + G0 C) we can solve for C
IMC Controller
C(s) = Gm (s)1

1
F(s) Ga (s)

By making assumptions about G0 (s), we can obtain PI and PID


controllers.
Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

30 / 32

Model-Based Methods

Internal Model Control

Skogestads IMC Rules


SIMC Rules
According to the author, "probably the best simple PID tuning rules in the
world"

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

31 / 32

Model-Based Methods

Internal Model Control

Summary

Limit PID to control of low-order systems and systems with small delay
Model-based tuning methods are preferrable to empirical ones
Dahlin control design, Haalmans method and IMC are simple to use but
cancel all process poles
These are special cases of more general design techniques
Pole placement
Q-design

We shall study those techniques in details

Guy A. Dumont (UBC EECE)

EECE 460 PID Tuning

January 2011

32 / 32

Das könnte Ihnen auch gefallen