Documente Academic
Documente Profesional
Documente Cultură
# 2006 EUCA
Performance and Robustness Improvement
in the IMC-PID Tuning Method
Alberto Leva
Dipartimento di Elettronica e Informazione, Politecnico di Milano Via Ponzio, 34/5, 20133 Milano, Italy
This manuscript presents an extension of the IMC-PID
tuning method. A higher order structure is adopted for
the process model than in the original method, where a
FOPDT (First-Order Plus Dead Time) model is used.
Then, the general IMC procedure is applied, and the so
obtained regulator is approximated with a (real) PID.
Simulations and experimental results indicate that
abandoning the FOPDT structure in the IMC approach
is beneficial both in terms of performance and
robustness.
Keywords: Internal Model Control; PID Control;
Process Control; Robust Autotuning
1. Introduction
In most process control applications, a good tuning of
the numerous PID loops at the low levels of the con-
trol system is very important for a satisfactory plant
operation. Indeed, no matter how complex a control
system is, its optimisation is a skyscraper whose
foundations are the low level loops, as recognised by
many audits and studies of which Ref. [10] is a
somehow historical example. A poorly tuned loop
invariantly means more hassle for the upper levels of
the control hierarchy, and having a loop tuned
incorrectly very often results in the necessity of taking
more complex solutions at the higher hierarchy levels,
or even of reducing the overall expectations [17].
Therefore, a huge research effort is being spent on
PID autotuning since many years, and as witnessed
for example, by [26], many different methods and
techniques are being experimented with.
Coming to the scope of this work, it is straightfor-
ward to apply the Internal Model Control (IMC)
principle to the synthesis of industrial PID regulators
[4,9,16,27,29,31], but this requires that a specific
structure be adopted for the process model [9].
Therefore, the model structure is dictated by the
type of regulator, not by the process dynamics.
Most frequently, FOPDT (First-Order Plus Dead
Time) models are used with satisfactory results.
Nonetheless, in some cases those models lead to poor
loop behaviour and are detrimental for the interpret-
ability of design parameter(s). In addition, the less the
model structure is close to reality, the more the
identification algorithm adopted is critical for the
tuning results.
To take profit of the IMC rationale for the synthesis
of industrial regulators, it is beneficial (and sometimes
necessary) to abandon the FOPDT model structure,
also while preserving the PID control law. Extending
the IMC-PID tuning method to non-FOPDT struc-
tures is not an easy task, however, and many works
can be found in which, to overcome the FOPDT
shortcomings, the choice is to abandon the PID law:
notable examples are Refs [11,12]. In any case,
extensions of the IMC-PID to non-FOPDT models
are practically absent in the applications [20,26],
especially if analytical tuning formul are required
E-mail: leva@elet.polimi.it
Received 19 November 2003; Accepted 25 November 2005
Recommended by F. Allgower and D. Clarke
[13]. This article proposes such an extension, based on
the preliminary results reported in Ref. [21].
2. Overview and Motivation of the
Presented Research
The general IMC scheme is reported in Fig. 1: P(s)
is the transfer function of the process, assumed here
asymptotically stable; M(s) is the process model; Q(s)
and F(s) are asymptotically stable transfer functions;
y
N
sM
N
s
M
D
s
, 2
where M
D
(s) is Hurwitz, M
D
0 M
N
0 M
N
1,
the roots of M
N
s in the LHP, and those of M
N
s
in the RHP. Then, make the empirical assumption
that M(s) describes the process satisfactorily, from
the standpoint of control synthesis, up to a
frequency .
M
. If M(s) is minimum-phase, .
M
is
selected as the maximum of the frequencies of the
roots of M
D
(s), M
N
s and M
N
s; in the opposite
case, .
M
is the minimum of the frequencies of the
roots of M
N
s. The model might have some poles
more or less cancelled by the zeros. If these (LHP)
cancellations are at high frequency with respect to the
relevant dynamics, then .
M
is excessively large.
To avoid this, the inverse of .
M
is constrained to
be larger than a fixed percentage (1/50) of the mea-
sured settling time. This also counteracts any other
meaningless high-frequency singularity owing to
overparametrisation. Finally, .
M
must be reduced if
the model has (meaningful) loosely damped poles or
LHP zeros, because those singularities will be can-
celled by the IMC regulator, making the PID
approximation critical. Therefore, if such a couple
of singularities exists with damping factor lower
than a fixed value (0.7 is used), then .
M
is further
constrained to be smaller than the characteristic
frequency of those singularities. Notice that the
degrees of M
D
(s), M
N
s and M
N
s allow to
compute their roots explicitly. Note also that
198 A. Leva
performing the identification in an interactive and
guided way would help a lot avoiding the mentioned
problems.
To obtain Q(s), set
Qs
M
D
s
jM
N
sQ
0
D
s
, 3
where
Q
0
D
s 1 s,.
Q
i
4
and .
Q
k
Q
.
M
, k
Q
being the first design parameter of
the proposed tuning method. This guarantees that
Q(s) be an acceptable approximation of the inverse of
the minimum-phase part of M(s) up to the reasonably
achievable control bandwidth. A larger k
Q
means the
request of a wider control bandwidth, thus of higher
performance. The standard value of 1 is adequate in
practice. The integer i is selected based on the degrees
of M
D
(s) and M
N
s so that the relative degree of Q(s)
be zero. Adopting the choice F(s) 1/(1 s/.
F
), the
IMC regulator turns out to be
Rs
M
D
s
jM
N
sQ
0
D
s1 s,.
F
M
N
s
, 5
where .
F
is set to k
.
.
Q
, k
.
2 (0, 1) being the second
design parameter of the proposed tuning method. The
rationale of k
.
is to increase the degree of stability and
robustness, as in the typical IMC scheme. A value in
the range 0.11 is advisable on the basis of experience.
Notice that, up to this point, the tuning procedure is
entirely analytical, with the advantages illustrated in
Ref. [13].
In all the cases considered, the obtained IMC reg-
ulator can be approximated with a PID very effec-
tively up to the necessary bandwidth. For this
approximation, an ad hoc numeric procedure is
employed. The idea is to preserve the low-frequency
aspect of the regulators frequency response, with a
certain emphasis on the frequency range around the
cutoff, and its mid-frequency phase lead (when there is
one). With the hypotheses introduced, the IMC reg-
ulator magnitude may have one of the two aspects of
Fig. 2, the low- and high-frequency asymptotic slopes
being 20 dB/dec.
To obtain the PID regulator, the magnitude of the
frequency response of the IMC one is computed, the
presence or absence of the peak is checked, and 20
points are selected as shown in Fig. 2 around the
cutoff frequency, computed by means of the high-
order model and regulator. Then, the frequency
response magnitude of a real PID is fit to these points
with a procedure very similar to the Matlab invfreqs
command, with the constraint that the second PID
pole and its zeros be in the LHP. Thanks to the choice
of .
M
, this simple procedure works effectively in all
the cases of interest. Several alternative procedures
can be found in the literature, see e.g. Refs
[6,22,24,33,35]. Some were tested, with results similar
to those of the simple one proposed. This corroborates
the idea that the structural model improvement, and
the corresponding choice of the control bandwidth,
make also the PID approximation not particularly
critical, provided that the reduction rationale is not to
privilege low-frequency fitting unconditionally, but
rather to preserve (also) the mid- and high-frequency
behaviour of the IMC regulator as much as possible.
As a final remark, notice that the IMC regulator could
be approximated even more effectively by a real PID
cascaded to a first-order filter. This possibility is not
investigated here (apart from a short remark in the
simulation example of Section 6), but may be impor-
tant in the presence of noise.
The proposed method has two design parameters:
k
Q
dictates the limit of the band where Q(s) is con-
sidered an acceptable inverse of M(s), while k
.
deter-
mines the band of the IMC filter, relatively to that
where the inverse is considered reliable. To ease the
application, extensive simulation tests indicate that
two modes can be defined: a simple mode, where
k
Q
is set to 1 and k
.
can range from 0.1 to 0.5, and an
expert mode where recommendations like those
summarised in Table 1 can be employed.
Though a complete treatment of the joint use of k
Q
and k
.
is not reported for brevity, three facts need
stressing.
Also the simple use of the proposed method does
yield an improvement with respect to the FOPDT-
based IMC-PID tuning.
As far as the clarity of the effect of the parameter(s)
is concerned, in some cases the improvement
Fig. 2. PID approximation of the IMC regulator.
Improved IMC-PID Tuning 199
yielded by the simple use of the method exists but
is not particularly evident. These problems are
always avoided using the other parameter, however:
no cases were found in which the closed-loop
responses are not sensitive to at least one parameter
in a clear and sensible way.
The closed-loop responses forecast with the
model (1) are always realistic enough to decide on
the correctness of a specific choice of the para-
meter(s) another improvement with respect to
FOPDT models.
There should be no doubt that the proposed
method can improve performance with respect to the
FOPDT-based IMC-PID, since the dynamics that are
cancelled are more close to the really control-relevant
ones. To compare the robustness properties of the
proposed method with those of the IMC-PID, recall
that in the scheme of Fig. 1 robust stability is guar-
anteed [9] for any additive model error W(s) fulfilling
jW j. C
n
j. j < 1 8., 6
C
n
(s) F(s)Q(s) being the nominal control sensitivity
function. If the FOPDT model used in the IMC-PID
method is a low-frequency approximation of (1), and
if the same IMC filter is used, it is easy to verify that
the nominal control sensitivity functions of the pro-
posed method and of the IMC-PID are almost equal
in the control band, their difference vanishing for
w ! 0. However, in the proposed method the model
error is computed with respect to (1), while in the
IMC-PID it is computed with respect to the FOPDT
model. Hence, the proposed method is more tolerant
to model error than the IMC-PID. This fact is
confirmed by experience, as shown also in the
experimental example of Section 6, but opens another
problem. In the FOPDT case, there is the possibility
of selecting ` based on quantitative model error
information [18]. Things are more complex in the
proposed method, so that precise clues for a case-
specific selection of k
.
and/or k
Q
are not yet available.
The problem is being studied.
Finally, some generality considerations are in order.
Quite intuitively, the method is as general as the set of
characteristics that the model structure is able to
represent. Given a set of dynamics that an auto-
tuner must be capable of dealing with, if the IMC
approach is adopted with an adequately complex
model structure, a correspondingly designed con-
struction of the approximated model inverse, and a
consequent choice of the IMC filter, then, for the
processes that exhibit those dynamics as the control-
relevant ones, the tuning results will be good, the
design parameter(s) will have a clear effect, and nei-
ther the model identification nor the PID approx-
imation procedures will be critical. Employing the
IMC approach in this way is very effective, also to
obtain tuning procedures that are specialised for
certain application domain.
4. Possible Extensions
The proposed idea, besides the procedure sketched
above, can lead to some extensions.
Obviously, it is possible to employ different model
structures, to better tailor the method to different
classes of situations if needed. It would be very
interesting, for example, to employ identification
procedures with structural selection capabilities [18],
possibly with some human intervention. Some pre-
liminary results are available, and the idea seems
promising.
It is also possible to employ different IMC filters,
e.g. introducing lead elements when the closed-loop
bandwidth is higher than the open-loop one. Com-
putations are analogous to those reported, and are
omitted for brevity. As will be shown in the simulation
example, introducing lead-type filters may yield
some advantage, but not as evident as in the FOPDT-
based case.
Research is underway on these extensions, and
results will be presented in future works.
5. A Simulation Example
The process considered is P(s) (1 10s)/
[(1 100s)(1 s)]. The two first order models
M
1
(s) 1/(1 100s) and M
2
(s) 0.1/(1 s) represent
Table 1. Recommendations for the use of design parameters in the expert mode.
Characteristic of model or response Preferred parameter Range Other parameter Value
Dominant delay k
.
0.050.3 k
Q
high (typ. 1)
Overdamped, delay-free (incl. oversh. only) k
Q
0.11.5 k
.
medium (typ. 0.6)
Loosely damped, delay-free k
.
0.050.3 k
Q
low (typ. 0.25)
Undershooting only, delay-free k
.
0.050.3 k
Q
low (typ. 0.25)
Under-and overshooting, delay-free k
.
0.050.6 k
Q
medium (typ. 0.6)
200 A. Leva
it accurately in different frequency ranges, as shown
in Fig. 3.
The IMC-PID rules were applied to M
1
(s) and
M
2
(s) taking Q(s) as the model inverse and (a) F(s)
1/(1 5s), (b) F(s) 1/(1 s) and (c) F(s) (1 10s)/
(1 5s)
2
. The closed-loop responses to a unit load
disturbance step and the regulators magnitude plots
are shown in the first and second row of Fig. 4.
Then, the proposed method was applied to
P(s) taking k
Q
1 and (a) F(s) 1/(1 s)(k
.
0.2),
(b) F(s) 1/(1 0.2s) (k
.
1) and (c) F(s) (1 10s)/
(1 5s)
2
. The closed-loop responses to a unit load
disturbance step and the regulators magnitude plots
are shown in the third row of Fig. 4.
When using the model M
1
(s) and the IMC-PID
rules, the best closed-loop performance in terms of
peak deviation from the set point and settling time
is achieved with the lead-lag filter (c), see the
corresponding curves in the first row of Fig. 4.
However, that regulator has quite a large high-
frequency gain.
On the other hand, when using the IMC-PID rules
with model M
2
(s), the best results are achieved with
the lag-only filter (b), see the corresponding curves in
the second row of Fig. 4. Also, the regulators high-
frequency gain is smaller. Note, however, that it is
very unlikely that the typical identification methods
used in autotuning produce a model like M
2
(s): in fact,
to decide that M
2
(s)is a good model for PID tuning,
knowledge of the entire P(s)is necessary.
The results of the proposed method (third row) are
similar to those obtained with M
2
(s), i.e. the knowl-
edge of P(s) does yield an advantage. Note that the
regulator magnitude is quite small and rolls off at high
frequency, as the PID approximation leads to a very
high-frequency zero, that can be omitted.
To summarise, the shortcomings of the FOPDT
models can adversely affect the tuning results or not,
depending on the identification method: if the model
is accurate in the right bandwidth, the IMC-PID
results are comparable with those of the proposed
method. The problem is that such information is not
available a priori. This should help answering the
frequent question when is a FOPDT model accurate?
or, better, understanding that the question is de facto
of very limited interest in the autotuning context: a
Fig. 3. Process and models in the simulation example.
Fig. 4. Results of the simulation example.
Improved IMC-PID Tuning 201
FOPDT model is always good if well identified, and
bad if identified with the wrong method, but to decide,
extensive process knowledge is needed. In addition, if
an accurate model is used, there is little disadvantage
using a lag-only IMC filter (which saves a design
parameter, by the way): in fact, adopting M
2
(s)
instead of M
1
(s) improves the settling time without
using a lead-type filter. The real problem is that
transients are sluggish when the low-frequency
dynamics of the model are cancelled, and these are not
representative of the dynamics of the process that are
control-relevant for the problem at hand: statements
like load response is sluggish when the required
closed-loop time constant is smaller than that of the
process can be very misleading, because rigorously
speaking there is no such thing as the process time
constant, and that of the model depends on the
identification.
6. Experimental Results
The process considered is a laboratory setup for
temperature control, where a metal plate is heated by
two transistors: one is the control actuator, the other
acts as a load disturbance. The controlled variable is
the plate temperature.
Figure 5a shows an open-loop step experiment and
two identified models: one has the proposed structure
and is
M
1
s
0.1061 81.5s 28.2s
2
e
0.75s
1 192.4s 5180.4s
2
12310s
3
. 7
The other has the FOPDT structure and is
M
2
s
0.106e
2.4s
1 126s
, 8
see figure 5(a). Taking
Qs
1 192.4s 5180.4s
2
12310s
3
0.1061 81.5s 28.2s
2
1 1.5s
, 9
i.e., k
Q
1, and F(s) 1/(1 6s), i.e., k
.
0.25, leads
to the approximating PID regulator
R
1
s 32.5 1
1
32s
4s
1 2s
. 10
Figure 5c and f reports the experimental closed-loop
transients obtained with the proposed method (R
1
),
and Fig. 5b and e those of the IMC-PID rules [25] with
the FOPDT model (R
2
) and a requested closed-loop
dominant time constant of 10 s, i.e.
R
2
s 96.8 1
1
127s
1.2s
1 1.23
. 11
The response to a set point step of amplitude 2
C
and to a load disturbance step of amplitude 30% are
shown. Note the apparent lack of integral action on
the part of R
2
. The integral time of R
1
is one third of
that of R
2
, and the transients of Fig. 5 evidence a
significant response speed increment owing to stron-
ger feedback (see in particular the load disturbance
response, where the output-derivation structure of the
Fig. 5. Experimental tuning results.
202 A. Leva
PID used in the experiments is irrelevant to the ana-
lysis). Notice also the lower high-frequency regulator
gain obtained with the proposed method, and the
corresponding smoother behaviour of the control
signal in the presence of noise. In this example the
process step response appears definitely of the first
order: nonetheless, more extensive process knowledge
is of help.
In Fig. 5d the two PIDs (10) and (11) are compared
in the frequency domain, by reporting the open-loop
Nyquist plots with the (accurate) model (7). Finally,
Fig. 6 shows the magnitude plots of the inverse of the
nominal control sensitivity in the various cases, to
show the improved robustness of the proposed
method, especially in the vicinity of the cutoff
frequency.
7. Concluding Remarks
An extension of the IMC-PID tuning method was
presented. A reasoned choice of the process model
structure, and an empirical quantification of the
bandwidth where this model is reliable from the
synthesis standpoint, allow the IMC procedure to
yield a regulator that can be approximated with a real
PID effectively, and produces better results than the
one obtained with a FOPDT model (the most com-
monly used structure in IMC-based PID tuning), both
in terms of performance and robustness. Simulation
and experimental results have been presented to back
up the proposed approach.
If the IMC approach is used as proposed herein, the
criticality of the model identification method is
reduced, and the clarity and interpretability of the
effects of the IMC design parameter(s) are improved
with respect to any use of the IMC approach where
the model structure is selected based only on the
regulator type. Also these statements are corroborated
by experience.
The presented research is continuing, with three
major goals. The first is to further investigate the role
of design parameters, so as to obtain simple clues for
their selection like in FOPDT-based IMC-PID tuning
(the table proposed for the expert mode is to be
considered a preliminary result). The second is to
further extend the method by employing a set of
possible model structures, among which the choice is
to be made on-line. The third, and more ambitious, is
to employ non-parametric models, so as to minimise
the information loss on the process dynamics.
It is finally worth recalling that here a tuning
principle, not an engineered product, was introduced.
To achieve an industry-ready implementation, a
number of aspects would need addressing that cannot
be treated in a research article. In any case, experience
with a Matlab prototype of the presented procedure
are definitely encouraging.
References
1. Anderson BDO. Control design: moving from theory to
practice. IEEE Control Syst 1993; 13(4): 1625
2. A