Proceedings of the

44th IEEE Conference on Decision and Control, and

ThB05.1
the European Control Conference 2005
Seville, Spain, December 12-15, 2005

Model-based PID tuning for high-order processes:

when to approximate
Antonio Visioli

Abstract— In this paper different methods for the (model- view of the achievable performance and of the ease of use.
based) tuning of Proportional-Integral-Derivative (PID) con- The paper is organized as follows. In Section 2 the different
trollers for high-order processes are analyzed and compared. considered approaches are reviewed and their characteristics
In particular, two approaches in the internal model control
framework are addressed and discussed: (i) the (high-order) are compared. Section 3 is devoted to the presentation of
controller that results from considering the high-order process simulation results regarding processes with different dynam-
model is reduced through a Maclaurin series expansion in ics. These results are discussed in Section 4 and conclusions
order to obtain a PID controller; (ii) the process model is first are drawn in Section 5.
reduced in order to obtain naturally a PID controller (different
techniques are considered for this purpose). Simulation results
regarding different process dynamics are evaluated in order to II. METHODOLOGIES
draw general conclusions.
A. IMC design: generalities
I. INTRODUCTION The internal model control methodology [8] has been
Proportional-Integral-Derivative (PID) controllers are the widely adopted for the purpose of PID controller tuning
most commonly adopted controllers in industrial settings (though, being based on a pole-zero cancellation approach
due to the cost/benefit ratio they are able to achieve. Be- it is not suitable for lag-dominant processes subject to load
cause of their simple structure, they are particularly suited disturbances [9], [10]). Indeed, it provides the user with a
to control processes whose dominant dynamics is of first desirable feature as a tuning parameter that handles the trade-
or second order [1]; however they are employed also for off between robustness and aggressiveness of the controller.
high-order processes, because of the economic advantages In a general form, the IMC control design can be described as
provided by their standardization. However, despite many follows. Consider a standard unity feedback control system
tuning rules having been devised in the past assuming a first- (see Figure 1) in which the (stable) process to be controlled
order plus dead time (FOPDT) or a second-order plus dead is described by the model:
time (SOPDT) process model [2], the case where a high-
order process dynamics (which is assumed to be known) is G(s) = pm (s)pa (s) (1)
explicitly addressed has received a significant attention only
recently [3]-[6]. Actually, it has to be taken into account that where pa (s) is the all-pass portion of the transfer function
in many cases, an apparent time delay is indeed due to the containing all the nonminimmum phase dynamics (pa (0) =
presence of a high-order dynamics [7]. 1). The controller transfer function is chosen as
It is realized that, because of the relative low-order of the f (s)p−1
m (s)
controller, a model reduction has necessarily to be performed. C(s) = (2)
1 − f (s)pa (s)
In this context, two approaches can be followed: (i) design a
model-based high-order controller by considering the (full) in which
high-order dynamics of the process and then reduce the con- 1
f (s) = (3)
troller to a PID form; (ii) reduce first the process model to an (λs + 1)r
appropriate low-order form so that a model-based controller
results directly to be in PID form. Actually, despite the fact is the IMC filter where λ is the adjustable time constant and
that it is obvious that the procedure for the determination of r is an appropriate order so that the controller is realizable.
the process model plays a key role in the controller tuning It has to be noted that the nominal closed-loop transfer
and therefore in the control system performance, this aspect function, i.e. the transfer function from the set-point signal
has been often overlooked in the literature [7]. ysp and the process output y, is
In this paper the two previously-mentioned approaches are pa (s)
thoroughly analyzed and compared in the internal model T (s) = . (4)
(λs + 1)r
control (IMC) framework [8], which has been extensively
adopted for the purpose of PID controller tuning, in order to This makes clear the role of the free design parameter λ
assess their advantages and disadvantages from the point of in selecting the desired closed-loop dynamics (and therefore
in handling the trade-off between robustness and aggressive-
This work was partially supported by MIUR scientifi c research funds
A. Visioli is with Dipartimento di Elettronica per l’Automazione, Uni- ness, as unavoidable mismatches between the true process
versity of Brescia, Italy, [email protected] dynamics and its model have to be taken into account).

0-7803-9568-9/05/$20.00 ©2005 IEEE 7127

d numerator time constants, the process transfer function is in
ysp y the following form:
C(s) G(s)
j (−Tj0 s + 1) −θ0 s
G̃(s) = e (8)
i (τi0 + 1)
Fig. 1. The control scheme.

where Tj0 > 0 and the time constants are ordered according
to their magnitude. Then, a SOPDT transfer function
Obviously, in general, the resulting controller is not in PID k
G(s) = e−θs (9)
form, i.e.: (τ1 s + 1)(τ2 s + 1)
is obtained by applying the half rule, i.e. by setting
Ti s + 1 Td s + 1
C(s) = Kp , (5)
Ti s Tf s + 1 τ1 = τ10 , τ2 = τ20 +
τ30
, (10)
2
if the series (“interacting”) form is considered, or
τ30

θ = θ0 + + τi0 + Tj0 . (11)
1 1 2
C(s) = Kp 1 + + Td s , (6) i≥4 j
Ti s Tf s + 1
It appears that, being the rules (7) based on the final apparent
if the ideal (“non-interacting”) form is implemented; Kp time delay θ, in the first part of the algorithm there is the
is the proportional gain, Ti and Td are the integral and need to guess this final value and to iterate in case at the end
derivative time constants respectively, and Tf is the filter the result is incorrect.
time constant. A PID controller results if the process model Once the SOPDT process model is obtained, the PID param-
has one positive zero and two poles (note that this results if a eters are determined by applying the IMC design procedure
FOPDT transfer function is considered and a first order Padè (and by approximating the delay term as e−θs = 1 − θs) and
approximation is adopted for the delay term [11]), whilst by possibly modifying the value of Ti in order to address
a PI controller results if the plant has a simple first order the case of lag-dominant processes [6] (note that this fact in
dynamics. Thus, if a high order process model is considered, not of concern in the examples presented in Section III). It
this must be reduced to this suitable form before applying results that the PID parameters in (5) are selected as
the IMC design or, alternatively, the resulting high-order τ1
controller has to be subsequently reduced to a PID form. Kp = , Ti = τ1 , Td = τ2 , Tf = 0.01Td . (12)
k(λ + θ)
B. PID tuning with the Skogestad’s half rule Note that the conversion of the tuning rule (12) for the
PID controller in the ideal form is straightforward and
The method proposed by Skogestad in [6] considers a a recommended choice for the desired closed-loop time
process model reduction based on the so-called “half rule”, constant is λ = θ [6].
which states that the largest neglected (denominator) time Summarizing, the method is based on simple, easy to re-
constant is distributed evenly to the effective dead time and member, tuning rules. However, the possible iterations in
the smallest retained time constant. In practice, given a high the model reduction algorithm make the overall procedure
order transfer function, each numerator term (T0 s + 1) with somewhat difficult to automate.
T0 > 0 is first simplified with a denominator term (τ0 s + 1),
τ0 > 0 using the following rules: C. Isaksson and Graebe’s analytical PID design
⎧ The technique proposed by Isaksson and Graebe in [5]
⎪ T0 /τ0 for T0 ≥ τ0 ≥ θ
⎪
⎪ is also based on a suitable process model reduction before
⎪
⎨ T0 /θ for T0 ≥ θ ≥ τ0
T0 s + 1 1 for θ ≥ T0 ≥ τ0 applying the IMC design. The model reduction is performed
≈
τ0 s + 1 ⎪
⎪ T0 /τ0 for τ0 ≥ T0 ≥ 5θ as followa. Let the initial (high-order) process model be
⎪
⎪
⎩ (τ̃0 /τ0 ) def described by the transfer function
(τ̃0 −T0 )s+1 for τ̃0 = min(τ0 , 5θ) ≥ T0
(7) B(s)
G̃(s) = (13)
where θ is the final effective delay (to be determined sub- A(s)
sequently). It has to be noted that τ0 is normally chosen
Then, the numerator and denominator polynomials are con-
as the closest larger denominator time constant (τ0 > T0 ),
sidered separately and the polynomials B1 (s) and A1 (s) that
except when a larger denominator time constant does not
retain only the slowest roots are determined. Subsequently,
exist or there is a smaller denominator time constant closer
the polynomials B2 (s) and A2 (s) that retain the low-order
to T0 ; this is true if the ratio between T0 and the smaller
coefficients are calculated. Finally, the reduced-order model
denominator time constant is less than the ratio between the
is obtained as
larger denominator time constant and T0 and less than 1.6
2 (B1 (s) + B2 (s))
1
at the same time.
G(s) = (14)
2 (A1 (s) + A2 (s))
Once this procedure has been terminated for all the positive 1

7128
By choosing B1 (s) and B2 (s) of first order and A1 (s) and It turns out that the first part of the series expansion contains
A2 (s) of second order and by subsequently applying the IMC a proportional term, an integral term and a derivative term
design (with a first order filter (3)), a PID controller (6) and therefore, if the high-order terms are neglected, a PID
naturally arises. If there are no zeros, two solutions can be controller (6) results (a first order filter can be easily added
exploited: (i) a second order denominator is calculated in the in order to make the controller proper and its time constant
reduction procedure and a second order filter (3) is applied in can be selected sufficiently small so that its dynamics is not
the IMC design, yielding to a PID controller; (ii) a first order significant).
denominator is calculated in the reduction procedure and a Hence, the overall procedure can be easily automated, al-
first order filter (3) is applied in the IMC design, yielding to though it is not based on tuning formulae and its compu-
a PI controller. tational burden is somewhat considerable. However, it has
It has to be noted that, differently from the method described to be stressed that a wrong choice of the design parameter
in subsection II-B, the case of complex conjugate roots is λ can result in the overall control system being unstable
also addressed in [5], but it will not be considered hereafter (see Section IV). Although this can be easily checked before
(see Section III). applying the controller, it can be considered as a major
Summarizing, the Isaksson and Graebe’s method can be drawback of the method.
easily automated, although it is not explicitly based on tuning
formulae. III. SIMULATION RESULTS

D. Model approximation with step response data In order to analyze and compare the different methodolo-
gies, the following processes with high order dynamics have
Usually, for the purpose of PID tuning, a FOPDT or been considered:
SOPDT process model is obtained by means of step response
data. In this context, the least-squares based method proposed (15s + 1)2 (4s + 1)(2s + 1)
G1 (s) = , (17)
in [12] is considered in this paper. The nice feature of this (20s + 1)3 (10s + 1)3 (5s + 1)3 (0.5s + 1)3
technique is that it is capable of providing a SOPDT process
(−0.3s + 1)(0.08s + 1)
model without any iteration and, being based on process G2 (s) = ,
output integrals, it is very robust to measurement noise. (2s + 1)(s + 1)(0.4s + 1)(0.2s + 1)(0.05s + 1)3
(18)
Starting from the identified model, the tuning rule (12)
(−45s + 1)
has been adopted. However, for this purpose, the obtained G3 (s) = ·
SOPDT model must have real poles and no zeros. Thus, if (20s + 1)3 (18s + 1)3 (5s + 1)3
a zero is determined the half rule is then adopted, whilst if (4s + 1)
· ,
complex conjugate poles occur, a FOPDT model (obtained (10s + 1) (16s + 1)(14s + 1)(12s + 1)
2

with the same identification method) is actually employed. (19)

1
In this latter case a PI controller results. G4 (s) = , (20)
It is worth stressing that, given a high-order process model, (s + 1)4
the reduced order transfer function based on step response 1
data can be obtained without the need of time consuming G5 (s) = , (21)
(s + 1)8
and costly experimental results, as a simulation can be
performed [13]. Doing so, the overall procedure can be easily 1
automated, although a relatively significant computational G6 (s) = . (22)
(s + 1)20
power is actually necessary.
The main characteristics of the processes are summarized
E. PID tuning by means of a Maclaurin series expansion in Table I. It has to be noted that transfer functions G1 (s)
and G3 (s) have been taken from [14], G2 (s) from [6]
The methods described in the previous subsections are
and G4 (s) − G6 (s) are representative of typical industrial
based on the reduction of the process model before applying
processes [15], [16].
the IMC design. Conversely, it is possible to apply the IMC
The reduced order models that have been adopted for the
procedure described in subsection II-A by considering the
PI(D) tuning are reported in Table II. Note that two trans-
full process dynamics and then reduce the obtained high-
fer functions might occur for the Isaksson and Graebe’s
order controller to a PID controller form. For this purpose, a
technique, whereas the process dynamics has no zeros, as
Maclaurin series expansion can be employed. The expression
explained in subsection II-C. Indeed, the first one yields to
of the resulting controller can be always written as [3], [4]:
a PID controller (with a second order IMC filter), whilst
r(s) the second one yields to a PI controller (with a first order
C(s) = (15)
s IMC filter). Besides, whereas a FOPDT transfer function is
reported for the step response based method, this means that
and expanding C(s) in a Maclaurin series in s we obtain:

the resulting SOPDT model has complex conjugate poles and
1 r (0) 2 has not therefore been employed (see subsection II-D).
C(s) = r(0) + r (0)s + s + ··· (16)
s 2 In order to make a fair comparison, for each method and for

7129
G1 (s) Minimum phase dynamics
G2 (s) Presence of a nondominant positive zero 1.2

set−point step response

G3 (s) Presence of a dominant positive zero 1

G4 (s) Minimum phase dynamics with a 0.8

small number of coincident poles 0.6 Skogestad
G5 (s) Minimum phase dynamics with a Isaksson and Graebe
step response
0.4
medium number of coincident poles Mclaurin

G6 (s) Minimum phase dynamics with a 0.2

high number of coincident poles 0

0 100 200 300 400 500 600
time
TABLE I
M AIN CHARACTERISTICS OF THE CONSIDERED PROCESSES .

load disturbance step response

0.6

0.4

each process, the value of λ that minimizes the integrated 0.2

absolute error, defined as
∞ 0

IAE = |e(t)|dt, (23) −0.2

0 100 200 300 400 500 600
0 time

have been selected for both the set-point and the load Fig. 2. Optimal responses for G1 (s).
disturbance step responses (i.e. a unit step has been applied
on signals ysp and d separately, see Figure 1). For those 1.2

methods that do not provide the value of the filter time 1

set−point step response

constant Tf explicitly, this has been selected in such a way 0.8

its dynamics is negligible. 0.6

Skogestad
The resulting values of the integrated absolute error and the 0.4 Isaksson and Graebe
step response
corresponding optimal values of λ are reported in Table III. 0.2 Mclaurin

Note again that for the Isaksson and Graebe’s method two 0
0 1 2 3 4 5 6 7 8 9 10
cases (PID and PI control) can emerge, depending on the fact time

that a second order or first order IMC filter respectively has

0.3
been adopted (begin no zeros in the process to be controlled).
load disturbance step response

0.25
Analogously, a PID or a PI controller results from the 0.2

technique based on the step response, depending on the use 0.15

of a SOPDT model or a FOPDT model (the latter in case 0.1

the identified SOPDT model has complex conjugate poles). 0.05

0
In order to evaluate better the results, the phase margin PM
−0.05
and the open loop cutoff frequency ωc for each considered 0 2 4 6 8 10
time
12 14 16 18 20

control system are reported in Table IV. Finally, the resulting Fig. 3. Optimal responses for G2 (s).
(set-point and load) unit step responses are plotted in Figures
2-7. For the sake of clarity, the process responses obtained
with a PI controller resulting from the Isaksson and Graebe’s is true only if an appropriate value of λ is selected. Indeed, a
method are not reported. wrong choice of λ might yield the system to instability. For
example, for system G3 (s), if λ ≤ 6 or λ ≥ 162 the resulting
IV. DISCUSSION closed loop system is unstable and, in any case, if λ ≥ 20 at
From the results obtained it appears that the approach least one of the PID parameter results to be less than zero.
based on the Maclaurin series expansion provides in general Actually, it might happen that a quite narrow range of values
the best performance, both for the set-point following and the for λ is suitable. Despite the fact that an inappropriate value
load disturbance rejection task. This is due to its capability of λ can be easily recognized during the design phase, this
of providing a higher open loop cutoff frequency without de- can be considered as a major drawback of the method, which
creasing the phase margin with respect to the other methods. has been overlooked in the literature. Indeed, this makes
From another point of view, this means that in the set-point the overall design more complicated and, most of all, the
step responses a low rise time is achieved without impairing physical meaning of the filter time constant, which should
the overshoot and in the load disturbance step responses, a handle the trade-off between aggressiveness and robustness
low peak error results without the occurrence of significant and control activity of the control system, is somewhat lost.
oscillations. It has to be also noted that the optimal values of λ are
It turns out that it is better to reduce the model of the significantly different between the considered methodology,
controller than that of the plant, as the approximation in- although it appears, as expected, that in general a higher-
troduced by adopting only the first three terms of the series order filter (i.e. for the Maclaurin series based technique or
expansion is not detrimental in the range of frequencies that when a PID controller is adopted instead of a PI controller
is significant for the considered control system. However this in the Isaksson and Graebe’s method) implies a lower value

7130
Process Skogestad Isaksson and Graebe step response
1 25.5s + 1 1
G1 (s) e−35.5s e−27.38s
(20s + 1)(15s + 1) 2642s2 + 73.25s + 1 44.46s + 1
1 − 0.26s + 1 1
G2 (s) e−0.77s e−0.79s
(2s + 1)(1.2s + 1) 3.21s2 + 3.38s + 1 2.48s2 + 3.17s + 1
1 − 42s + 1 1
G3 (s) e−180s e−127.7s
(30s + 1)(20s + 1) 8564s2 + 115.7s + 1 106.6s + 1
1 1 1 1
G4 (s) e−1.5s e−1.38s
(1.5s + 1)(s + 1) 2.25s2 + 3s + 1 2.5s + 1 2.71s + 1
1 1 1 1
G5 (s) e−5.5s e−3.88s
(1.5s + 1)(s + 1) 14.52s2 + 5.02s + 1 4.51s + 1 4.24s + 1
1 1 1 1
G6 (s) e−17.5s e−12.72s
(1.5s + 1)(s + 1) 95.98s2 + 11.40s + 1 10.70s + 1 7.76s + 1
TABLE II
R ESULTING MODEL REDUCTIONS FOR THE DIFFERENT METHODOLOGIES .

Process Task Skogestad Isaksson and Graebe step response Maclaurin

PID PI PID PI
G1 (s) setpoint 71.38 (9.6) 63.08 (54.4) 73.49 (37.2) 42.07 (4.4)
load 54.45 (0.01) 48.13 (39.1) 60.16 (20.8) 25.35 (2.8)
G2 (s) setpoint 1.852 (0.83) 1.963 (1.22) 1.819 (0.78) 1.783 (0.18)
load 0.869 (0.03) 0.985 (0.37) 0.863 (0.01) 0.986 (0.07)
G3 (s) setpoint 376.5 (120.1) 251.7 (122.3) 301.1 (159.0) 231.6 (11.5)
load 384.3 (115.2) 253.6 (113.0) 307.7 (153.6) 233.5 (10.9)
G4 (s) setpoint 3.133 (0.49) 2.722 (0.87) 4.099 (3.44) 3.998 (2.06) 2.041 (0.41)
load 1.952 (0.001) 1.073 (0.39) 3.183 (2.33) 3.063 (0.96) 0.876 (0.19)
G5 (s) setpoint 11.58 (3.01) 7.776 (2.78) 9.550 (8.83) 9.662 (4.96) 6.561 (0.74)
load 10.95 (2.06) 6.401 (2.34) 8.880 (7.88) 9.030 (4.00) 5.394 (0.61)
G6 (s) setpoint 36.86 (11.2) 23.01 (8.83) 25.84 (24.6) 27.36 (12.2) 19.85 (0.90)
load 36.49 (10.5) 21.88 (8.42) 25.22 (23.8) 26.95 (11.4) 18.81 (0.86)
TABLE III
S UMMARY OF THE OPTIMAL IAE ’ S ( AND CORRESPONDING VALUES OF λ).

Process Task Skogestad Isaksson and Graebe step response Maclaurin

PID PI PID PI
PM ωc PM ωc PM ωc PM ωc PM ωc PM ωc
G1 (s) setpoint 50.76 0.020 77.9 0.019 60.9 0.017 61.6 0.031
load 42.01 0.025 43.3 0.038 47.6 0.227 27.1 0.062
G2 (s) setpoint 58.9 0.658 62.5 0.630 59.1 0.671 58.9 0.733
load 23.0 1.337 21.4 1.404 23.6 1.323 23.5 1.509
G3 (s) setpoint 55.7 3.33 · 10−3 55.8 4.65 · 10−3 62.7 3.72 · 10−3 61.2 4.84 · 10−3
load 55.1 3.39 · 10−3 54.8 4.86 · 10−3 62.1 3.80 · 10−3 61.3 5.02 · 10−3
G4 (s) setpoint 50.7 0.459 53.8 0.544 59.4 0.306 60.4 0.316 58.5 0.707
load 41.1 0.576 17.3 1.157 41.8 0.447 52.6 0.380 12.8 1.487
G5 (s) setpoint 54.2 0.114 56.5 0.162 63.2 0.122 62.3 0.120 64.5 0.175
load 50.0 0.127 54.5 0.193 58.9 0.139 58.0 0.136 62.8 0.232
G6 (s) setpoint 55.4 0.034 56.6 0.051 64.8 0.044 60.4 0.041 61.9 0.058
load 54.6 0.035 56.2 0.053 63.6 0.046 59.3 0.043 62.0 0.061
TABLE IV
S UMMARY OF THE RESULTING PHASE MARGINS PM ( IN DEGREES ) AND CUTOFF FREQUENCIES ωc ( IN RAD / S ).

of λ (and the load disturbance rejection task requires a lower worse performance than the Isaksson and Graebe’s one whilst
value of λ than the set-point following task). no general conclusions can be drawn with respect to the
From the results obtained, it also appears that the Isaksson Skogestad’s method.
and Graebe’s method provides in general better performance
that the Skogestad’s one and, as expected, the PID con- V. CONCLUSIONS
troller is better than the PI controller in the context of the In order for a method to be adopted in the industrial
technique described in subsection II-C. Note, however, that context, its pros and cons should be clearly pointed out.
the tuning rules (12) have been conceived with the aim of For this purpose, in this paper different model based PID
being applicable to a wide range of processes and of being tuning methods for high-order processes have been analyzed
easy to memorize. As for the method based on the step and compared. As a main result, it has been found that the
response data, it can be deduced that in general it provides best performances are obtained by reducing the controller

7131
1.2 1.2

1
set−point step response

set−point step response

1
0.8
0.8
0.6
Skogestad 0.6
0.4 Skogestad
Isaksson and Graebe Isaksson and Graebe
step response 0.4 step response
0.2
Mclaurin Mclaurin
0 0.2

−0.2 0
0 500 1000 1500 0 5 10 15 20 25 30 35 40 45 50
time time
load disturbance step response

load disturbance step response

1 0.8
0.8
0.6
0.6
0.4
0.4

0.2 0.2

0 0

−0.2 −0.2
0 500 1000 1500 0 5 10 15 20 25 30 35 40 45 50
time time
Fig. 4. Optimal responses for G3 (s). Fig. 6. Optimal responses for G5 (s).

1.2
1.2
set−point step response

set−point step response

1 1

0.8 0.8

0.6 Skogestad 0.6

Skogestad
Isaksson and Graebe
0.4 Isaksson and Graebe
step response 0.4
step response
Mclaurin
0.2 Mclaurin
0.2
0
0 2 4 6 8 10 12 14 16 18 20 0
0 50 100 150
time
time

0.6 1
load disturbance step response

load disturbance step response

0.8
0.4
0.6

0.2 0.4

0.2
0
0

−0.2 −0.2
0 5 10 15 20 25 30 0 50 100 150
time time
Fig. 5. Optimal responses for G4 (s). Fig. 7. Optimal responses for G6 (s).

model instead of the process one, although a careful choice [8] M. Morari, E. Zafi riou, Robust process control, Prentice Hall, Engle-
of the IMC filter time constant is needed. The computational wood Cliffs, NJ, 1989.
complexity of the methods considered has also been ad- [9] C. Scali, D. Semino, “Performance of optimal and standard controllers
for disturbance rejection in industrial processes”, Proc. IEEE Int. Conf.
dressed in order to evaluate the ease of their implementation on Ind. Elec., Cont. and Instr. IECON’91, pp. 2033-2038, 1991.
in Distributed Control Systems. [10] F. G. Shinskey, Process Control Systems - Application, Design, and
Tuning, McGraw-Hill, New York, USA, 1996.
R EFERENCES [11] D. E. Rivera, S. Skogestad, M. Morari, “Internal model control. 4.
PID controller design”, Ind. Eng. Chem. Proc. Des. Dev., Vol. 25, pp.
[1] K. J. Åström, T. Hägglund, PID controllers: theory, design and tuning, 252-265, 1986.
ISA Press, Research Triangle Park (NC), 1995. [12] Q.-G. Wang, X. Guo, Y. Zhang, “Direct identifi cation of continuous
[2] A. O’Dwyer, Handbook of PI and PID controller tuning rules, World time delay systems from step response data”, Journal of Process
Scientifi c Press, 2003. Control, Vol. 11, pp. 531-542, 2001.
[3] Y. Lee, S. Park, M. Lee, C. Brosilow, “PID controller tuning for desired [13] B. Huang, “A pragmatic approach towards assessment of control loop
closed-loop responses for SI/SO systems”, AIChE Journal, Vol. 44, performance”, International Journal of Adaptive Control and Signal
No. 1, pp. 106-115, 1998. Processing, Vol. 17, pp. 589-608, 2003.
[4] Y. Lee, S. Park, M. Lee, “PID controller tuning to obtain desired closed [14] L. Wang, W. R. Cluett, From plant data to process control - ideas for
loop responses for cascade control systems”, Industrial Engineering process identification and PID design, Taylor and Francis, London,
and Chemistry Research, Vol. 37, pp. 1859-1865, 1998. Uk, 2000.
[5] A. J. Isaksson, S. F. Graebe, “Analytical PID parameter expressions [15] K. J. Åström, T. Hägglund, “Benchmark systems for PID control”,
for higher order systems”, Automatica, Vol. 35, pp. 1121-1130, 1999. Preprints IFAC Workshop on Digital Control PID‘00, pp. 181-182,
[6] S. Skogestad, “Simple analytic rules for model reduction and PID 2000.
controller tuning”, Journal of Process Control, Vol. 13, pp. 291-309, [16] F. G. Shinskey, “PID-deadtime control of distributed processes”,
2003. Preprints IFAC Workshop on Digital Control PID‘00, pp. 14-18, 2000.
[7] A. Leva, “Autotuning process controller with improved load distur-
bance rejection”, Journal of Process Control, Vol. 15, pp. 223-234,
2005.

7132