You are on page 1of 25

344 Ind. Eng. Chem. Res.

2008, 47, 344-368

A Comparative Study of Recent/Popular PID Tuning Rules for Stable,


First-Order Plus Dead Time, Single-Input Single-Output Processes
M. G. Lin, S. Lakshminarayanan, and G. P. Rangaiah*
Department of Chemical and Biomolecular Engineering, National UniVersity of Singapore,
4 Engineering DriVe, Singapore 117576

Numerous PID tuning rules exist, and new rules continue to be proposed in the literature. The primary objective
of this article is to evaluate comprehensively and compare several recent or popular PID tuning methods
based on the integral of absolute error in the controlled variable, total variation of the manipulated variable
and maximum sensitivity. Several tuning rules such as internal model control, direct synthesis, and optimization-
based methods were considered and evaluated for a range of first-order plus dead time process dynamics
subjected to set-point changes and/or load disturbances via simulation. On the basis of the results,
recommendations were then provided for selecting the optimal tuning rule for industrial processes according
to the desired control objective (servo versus regulatory) and the process dead time to time constant ratio.

1 Introduction
In spite of all of the advances in process control over the
past several decades, the proportional, integral and derivative
(PID) controller remains the most-dominant form of feedback
controller in use today. The transparency of the PID control
mechanism, the availability of a large number of reliable and
cost-effective commercial PID modules, and their widespread Figure 1. Block diagram of a feedback control system.
acceptance by operators are among the reasons for its success.
Koivo and Tanttu1 estimated that there are perhaps only 5-10%
of control loops that cannot be controlled by single-input single-
output (SISO) PID controllers. According to Desborough and
Miller,2 a survey of more than 11 000 controllers in the refining,
chemicals, and pulp and paper industries showed that 97% of
regulatory controllers had the PID structure.
Because the PID controller finds widespread use in the
process industries, considerable amounts of research efforts have Figure 2. Internal model control (IMC) structure.
also been directed toward finding the optimal values for the
PID controller’s three parameters: the proportional gain (Kc),
the integral time (τi), and the derivative time (τd). In a striking
statistic, O’Dwyer3 reported that 293 out of 408 sources of
tuning rules have been recorded since 1992, reflecting the
upsurge of interest in the PID controller in the past decade.
Despite these research efforts, surveys conducted on the state
of the art of industrial control practices reported sobering results.
For instance, Van Overschee and De Moor4 reported that 80%
of PID controllers are badly tuned. About 30% of the PID Figure 3. Rearrangement of the IMC structure.
controllers operate in manual mode. Among the PID controllers methods, we restrict the treatment to FOPDT models in this
that are in auto mode, 25% use default factory settings, implying article because of the above reasons. There are also many
that they have never been tuned. This and other surveys show industrially proven techniques now available for deriving
that the determination of PID controller parameters is a vexing empirical models from plant data.
problem in many industrial applications. The main contribution of this study is that it facilitates the
The first-order plus dead time (FOPDT) model is a good determination of the best tuning rule to compute optimal PID
approximation for many chemical processes that exhibit a controller settings for stable SISO FOPDT processes under
monotonic response to step inputs. Experimental data from different scenarios. Depending on the objective of the control
processes that exhibit higher order and even slightly nonlinear system (set-point tracking or disturbance rejection) and the
behavior are often fitted with FOPDT models to facilitate different kinds of disturbances (deterministic/stochastic) it is
controller design.5 Whereas our analysis can include any linear likely to experience, our study aids the identification of
transfer function model (e.g., higher-order models and processes appropriate PID tuning parameters. Our work is also fairly
with numerator dynamics) and corresponding controller tuning comprehensive in the sense that it includes the control of lag-
dominant, balanced, and delay-dominant stable SISO FOPDT
* To whom correspondence should be addressed. E-mail: chegpr@ systems. We also consider the severity of manipulated variable
nus.edu.sg. Fax: 65-67791936. action and robustness in addition to output performance.
10.1021/ie0704546 CCC: $40.75 © 2008 American Chemical Society
Published on Web 01/09/2008
Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 345

Table 1. Summary of Selected PID Tuning Rules


tuning rule basis proportional gain (Kc) integral time (τi) derivative time (τd) remarks
Ziegler and Nichols7 E 1.2τ/Kθ 2θ θ/2
Rivera et al.9 IMC θ τ + θ/2 τθ/(2τ + θ) λ g 0.1τ
τ+ λ g 0.8θ
2
θ
(
K λ+ )
[ ]
2 θ2 θ
Lee et al.10 IMC τi/K(λ + θ) τ + θ2/2(λ + θ) 1-
2(λ + θ) 3τi

Skogestad11 IMC τ/K(λ + θ) min[τ,4(λ + θ)] 0 λ)θ

( )( ) ( )( )
Chen and Seborg12 DS regulatory
2 2 2
θ θ θ θ τθ θ
2τθ +
2
3λ +
2
- 2τθ +
2
3λ +
2
- 3λ2τθ +
2
3λ + -
2( )
1 2λ - 3λ θ
3 2
2λ - 3λ θ
3 2
2(τ + θ)λ3
K 3
2λ + 3λ2θ +
θ2
2
3λ +
θ
2 ( )
(2τ + θ)θ
( 2τθ +
θ2
2
3λ +)(
θ
2
- )
2λ - 3λ θ
3 2

Sree et al.13 DS 1/K(τ/θ + 0.5) τ + θ/2 0.5θ(τ + 0.1667θ)/(τ + 0.5θ) servo


Huang and Jeng14 CO (0.36θ + 0.76τ)/Kθ τ + 0.47θ 0.47τθ/(0.47θ + τ) θ/τ < 1/3

(0.8194τ + 0.2773θ)/Kτ0.0262θ0.9738 1.0297τ + 0.3484θ 0.4575τ + 0.0302θ


θ (1.0297τ + 0.3484θ)
θ/τ g 1/3

Huang et al.15 CO koτ/Kθ τ aθ series settings


ko ) 0.65,a ) 0.4

Åström and Hägglund16 CO (R1θ + R2τ)/Kθ θ ( R3θ + R4τ


θ + R5τ ) R6τθ/(θ + R7τ)
regulatory
MS ) 1.4

τ θ regulatory
1
(θτ) ]
2.2
Syrcos and Kookos17 CO
K (
0.31 + 0.6
θ ) (
τ 0.777 + 0.45
τ) [
τ 0.44 - 0.56 θ/τ ∈ [2,5]

Balestrino et al.6 compare three PID tuning rules (derived in Gc(s) is the PID controller transfer function and may be
1967, 1981, and 1993) in terms of normalized performance represented as follows:

( )
indices for controlling FOPDT models. They consider the
normalized integral of absolute error (IAE), the normalized 1
Gcideal(s) ) Kc 1 + + τds (2)
integral of squared error (ISE), and also the variability of control τis
action (normalized). However, they do not consider noise and
disturbance effects as we do here. In our work, the following Known as the ideal or parallel controller algorithm, eq 2 is
10 PID tuning rules were selected after a review of the recent unfortunately physically unrealizable in practice because no
and/or popular tuning rules. The Ziegler and Nichols7 tuning device could be constructed that truly differentiates an input
method was included for comparison purposes because it was signal.5 Even if it can be realized, a parallel PID controller has
recently shown to provide the best compromise between the potential problem of giving a very unsteady high-frequency
robustness and performance.8 Because of the popularity of control signal due to noise in the measurement - such noise is
model-based PID tuning methods, a comprehensive study was very often present.
conducted on tuning methods proposed by Rivera et al.,9 Lee Consequently, over the years, several approximations to the
et al.,10 Skogestad,11 and Chen and Seborg12 as they reduce the derivative term have been developed by the vendors. In this
complexity of the PID tuning problem to the specification of a article, the PID controller was implemented in the popular
single adjustable parameter, λ. Recently proposed tuning parallel with filtered deriVatiVe form:

( )
methods by Sree et al.,13 Huang and Jeng,14 Huang et al.,15
Åström and Hägglund,16 and Syrcos and Kookos17 were then 1 τd,parallels
Gc,parallel ) Kc,parallel 1 + + (3)
included for comparison to encompass the latest developments τi,parallels Rτd,parallels + 1
in PID control.
In this equation, the derivative term includes a parameter R,
2. Background known as the derivative filter parameter, which reduces the
sensitivity of control calculations to high-frequency measure-
2.1. Feedback Control Loop. Figure 1 shows a block ment noise. Most commercial PID modules have a preset value
diagram of a SISO feedback control system where ysp is the set of R ) 0.10.
point, y is the controlled output, u is the controller output/ Another popular implementation of PID control is the series
manipulated input, and e is the error ysp - y. The load with filtered deriVatiVe form:
disturbance d enters at the process input, and the feedback signal

( )(
τd,seriess + 1
)
is corrupted by random measurement noise n at the process
1
output. Gp(s) represents the combined dynamics of the stable Gc,series ) Kc,series 1 + (4)
process, the transmitters, and the final control element and is
τi,seriess Rτd,seriess + 1
assumed to be well approximated by a FOPDT transfer function
The relation between controller parameters for the parallel form
-θs (eq 3) and the series form (eq 4) can be obtained through a
Ke term by term matching (for a ) 0) of eqs 3 and 4. This yields
Gp(s) ) (1)
τs + 1 eqs 5-7:
346 Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

( )
Kcparallel ) Kcseries 1 +
τdseries
τiseries
(5)
MS ) max |S(jω)|
ω
(11)

( )
The Nyquist stability criterion states that a closed-loop system
τdseries will be stable if the Nyquist plot of the open-loop transfer
τiparallel ) τiseries 1 + (6) function GcGp(s) does not encircle the critical point (-1,0).
τiseries Geometrically, the maximum sensitivity MS is the inverse of
1 the shortest distance from the Nyquist plot of GcGp(s) to the

( )
τdparallel ) τdseries (7) critical point (-1,0). Thus, as MS decreases, the system becomes
τdseries
more robust. For satisfactory robustness, MS should be in the
1+
τiseries range of 1.2 to 2.5

In the literature, the tuning rules are specified for a particular 3. PID Tuning Rules
form of the PID controller - ideal (parallel) form, series (also The selected 10 tuning rules as applied to FOPDT processes
known as cascade or interacting) form etc. Equations such as are briefly described in this section. They can be classified
5-7 are useful in converting one form to the other when according to the basis for their development, that is, empirical,
feasible. Of the methods considered here (also listed in Table internal model control, direct synthesis, or constrained optimiza-
1), only Skogestad11 and Huang et al.15 use the series form; tion, although overlaps exist. The formulas for the selected
others derive tuning relations are based on the ideal form of tuning rules are summarized in Table 1.
the PID controller. In the case of Skogestad,11 we use only the 3.1. Empirical Methods (E). The tuning method proposed
PI controller, and hence there is no difference between the by Ziegler and Nichols7 is based on the open-loop process step
alternate forms of the controller. response. Because FOPDT processes have a maximum slope
2.2. Controller Performance Metrics. Understandably, there of K/τ at t ) θ for a unit-step input change, these same rules
are several measures for control-loop performance. Integral error can be used to derive a set of PID tuning parameters for a
criteria such as IAE, ITAE, and ISE are typically considered FOPDT process, which gives roughly quarter-decay response.5
(Åström and Hägglund18) and are quite popular.5 For example, 3.2 Internal Model Control (IMC). The IMC method is
the IAE in the controlled variable, based on an assumed process model, G̃p(s) and leads to an
analytical expression for the controller settings related directly
IAE ) ∫0∞ |e(t)|dt (8) to the model parameters. The IMC controller G/c (s) has the
advantage of taking into consideration model uncertainties and
allowing the tradeoffs between system performance and robust-
Whereas the ISE criterion penalizes large errors (resulting in
ness to be considered in a more systematic fashion via the
the most-aggressive settings) and the ITAE criterion penalizes
specification of a single adjustable parameter, λ:
persistent errors (resulting in the most-conservative settings),
the IAE criterion tends to produce moderate settings that are 1
between those for the ISE and ITAE criteria. Hence, in this study, G/c (s) ) (12)
IAE criterion was selected as one of the metrics for controller
G̃p-(s)(λs + 1)r
performance.
where G̃p-(s) contains all of the stable left-half plane zeros, r
As another measure of the controller performance, the
required control effort was computed by calculating the total is a positive integer selected such that G/c (s) is proper, and λ
variation (TV) of the manipulated input, u denotes the desired closed-loop time constant. Rearranging the
IMC system in Figure 2 as in Figure 3, it can be shown that the
∞ standard feedback controller Gc(s) in Figure 3 and the IMC
TV ) ∑
i)0
|u(i + 1) - ui| (9) controller G/c (s) in Figure 2 are related by

G/c (s)
The TV of u(t) is the sum of all of the control moves, both up Gc(s) ) (13)
1 - G/c G̃p(s)
and down. Thus, it is also a good measure of the smoothness
of the manipulated input signal. It is desired that TV be as small Assuming that G̃p(s) is a FOPDT transfer function, Rivera et
as possible. al.9 obtained a set of parallel PID controller settings after
For a control system to remain stable and function effectively, introducing a first-order Padé approximation for the time-delay
it should not be unduly sensitive to small changes in the process term. For sufficient robustness, λ g 0.8θ and λ g 0.1τ are
or to the inaccuracies in the process model. A control system recommended. Expanding Gc(s) in the Maclaurin series in s
that satisfies this requirement is said to be robust. On the basis results in a controller, which contains the proportional, integral,
of the block diagram in Figure 1, the sensitivity function S can and derivative terms, in addition to an infinite number of higher-
be defined as order derivative terms. Because it is impossible to fully
implement such a controller, Lee et al.10 approximated the
1
S(jω) ) (10) parallel PID controller by retaining only the first three
1 + GcGp(jω) terms.
For FOPDT processes, Skogestad11 does not recommend the
The amplitude ratio of S(jω) provides a measure of sensitivity use of derivative action. Hence, by introducing a first-order
of the closed-loop control system to changes in the process. Taylor series approximation for the time-delay term, a set of
Consequently, the maximum sensitivity MS, defined as the parallel PI tuning relations is obtained. It is also well-known
maximum value of |S(jω)| for all of the frequencies, serves as that for lag-dominant processes, the choice of τi ) τ results in
a very useful metric for robustness analysis: a long settling time for load disturbances entering at the process
Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 347

Figure 4. Various perturbations studied.

input. For this reason, Skogestad11 requires that τi be computed term by a first-order Padé approximation, a set of parallel PID
as the lesser of τ and 4(λ + θ). To achieve a fast response with controller settings is obtained. It is apparent from Table 1 that
good robustness, λ ) θ is recommended. Kc, τi, and τd of Chen and Seborg12 can become negative for
3.3. Direct Synthesis (DS). In the DS method, the controller large values of λ. Thus, care should be taken in specifying the
design is based on an assumed process model and a desired desired closed-loop time constant λ.
closed-loop transfer function. As such, it is very similar to the Unlike Chen and Seborg,12 the DS method proposed by
IMC approach and can produce identical controllers if the design Sree et al.13 is based on servo responses and does not have
parameters are specified in a consistent manner. The two an adjustable parameter, λ. The closed-loop transfer function
selected DS methods are outlined below. relating set point ysp to the controlled output y is given by
Unlike the usual DS methods where the PID controller
settings are based on the specification of a closed-loop transfer y GcGp(s)
function for set-point changes, Chen and Seborg12 developed a ) ) [K1q + K2 + K3q2]e0.5q ×
ysp 1 + GcGp(s)

[ ]
modified DS method based on disturbance rejection. Consider
Figure 1 where it is assumed that the disturbance dynamics is e-q
(15)
equal to the process dynamics. Thus, in the following develop- (s + 1)qe0.5q + (K1q + K2 + K3q2)e-0.5q
ment, Gd(s) ) Gp(s); as well, the assumed disturbance dynamics
G̃d(s) is taken to be equal to the assumed process dynamics where K1 ) KcK, K2 ) K1τ/τi, K3 ) K1(τd/τ),  ) θ/τ, and q )
G̃p(s). Because the closed-loop transfer function for load sτ. Removing the e-q term in the numerator from further
disturbance (y/d) is not known a priori and the actual process analysis (because it will only shift the corresponding time axis),
Gp(s) may not be exactly known, it is practical to replace them a first-order Taylor series expansion is introduced to approximate
with the desired closed-loop transfer function for disturbance e0.5q and e-0.5q in the denominator such that the order of the
rejection, (y/d)desired and an assumed process model, G̃p(s), numerator becomes equivalent to that of the denominator.
respectively. This gives the design equation Because the objective of the controller is to achieve (y/ysp)desired
) 1, coefficients of powers of q in the numerator are equated
G̃d(s) 1 with the corresponding coefficients in the denominator to solve
Gc(s) ) - (14)
(y/d)desiredG̃p(s) G̃p(s) for K1, K2, and K3, resulting in a set of parallel PID controller
settings.
As stated before, G̃p(s) ) G̃d(s). Assuming that G̃p(s) is well 3.4. Constrained Optimization (CO). Tuning methods
described by a FOPDT model and approximating the time-delay developed using constrained optimization are mostly based on
348 Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

Figure 5. Effect of the tuning parameter on the performance of controller tuning methods for a unit-step change in the set point: (a) θ/τ ) 0.1, (b) θ/τ )
1.0, and (c) θ/τ ) 2.0.

With reference to Figure 1 for a FOPDT process, the open-


loop transfer function GcGp(s) is

KcK(τiτds2 + τis + 1)e-θs


GcGp(s) ) GcidealG̃p(s) ) )
τis(τs + 1)
p1(p2s2 + p3s + 1)e-θs
(16)
s(τs + 1)

where p1 ) KcK/τi, p2 ) τiτd and p3 ) τi. The optimization


problem can then be formulated as a search for the optimal
parameters, p1, p2 and p3, such that IAE is minimized. Huang
and Jeng14 solved the optimization problem for various values
of θ/τ and derived two sets of parallel PID controller relations;
one for θ/τ < 1/3 and another for θ/τ g 1/3.
For FOPDT processes, Huang and Jeng19 found that the best
achievable GcGp(s) such that IAE is minimized, and this can be
approximated as

ko (1 + aθs) -θs
GcGp(s) ≈ e (17)
θ (1 + µθs)s
Figure 6. Legend for Figures 5, 7-25, and 27-32.
where the denominator transfer function (1 + µθs) is imple-
mented to make the PID controller realizable; ko and a denote
the model parameters returned from the minimization search.
the minimization of certain integral of error criterion subject to On the basis of the above findings, Huang et al.15 further
desired constraints. A unique advantage of this design approach developed correlations between ko and a with gain and phase
is that the tuning methods developed are obtained via direct margins and proposed a model-based design method that results
optimization of the selected objective. Such a desired charac- in a series with a filtered derivative PID controller with two
teristic is not shared by many other classical design methods. adjustable parameters ko and a, which are closely related to
In addition, the constrained optimization approach also serves robustness and performance of the closed-loop control system.
to complement classical model-based design methods, as The default values of ko ) 0.65 and a ) 0.4 for a gain margin
demonstrated by Huang et al.15 and Åström and Hägglund.16 of 2.7 and a phase margin of 65° are recommended.
Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 349

Figure 7. Responses of controller tunings with similar IAE values for the set-point change and θ/τ ) 0.1: (a) Rivera et al.9 (IAE ) 2.0) and (b) Skogestad11
(IAE ) 3.0).

Figure 8. Performance of controller tuning methods for a unit-step change in the set point: θ/τ ) 0.1.
Panagopoulos et al.20 proposed a tuning method based on the generalized parallel PID controller settings by varying the
optimization of load disturbance rejection with constraints on specified design parameter MS. For satisfactory performance
robustness to model uncertainties. For processes well described with excellent robustness, MS ) 1.4 is recommended. For this
by a FOPDT model, Åström and Hägglund16 derived a set of method, the parameters of the ideal parallel PID algorithm are
350 Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

Figure 9. Performance of controller tuning methods for a unit-step change in the set point and output corrupted by measurement noise (nj ) 0, σ2n ) 0.052):
θ/τ ) 0.1.

Figure 10. Performance of controller tuning methods for a unit-step change in the set point: θ/τ ) 1.
provided in terms of parameters R1 to R7 (Table 1). The Ri values Syrcos and Kookos17 formulated a nonlinear programming
for various values of maximum sensitivity (MS) are tabulated problem for a PID controller design solely in the time domain.
by Åström and Hägglund.16 For example, if MS ) 1.4 is desired, With no approximations introduced and solved using direct
they are given by: R1 ) 0.168, R2 ) 0.460, R3 ) 0.363, R4 ) optimization for a number of delay-dominant FOPDT processes
0.871, R5 ) 0.111, R6 ) 1.70, and R7 ) 4.37. with θ/τ ∈ [2,5], subjected to a unit-step load disturbance d, a
Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 351

Figure 11. Performance of controller tuning methods for a unit-step change in the set point and output corrupted by measurement noise (nj ) 0, σ2n ) 0.052):
θ/τ ) 1.

Figure 12. Performance of controller tuning methods for a unit-step change in the set point: θ/τ ) 2.

set of parallel PID controller settings for delay-dominant Furthermore, the PID controller was implemented in the widely
processes were derived. popular parallel with filtered deriVatiVe form with the derivative
filter parameter specified as R ) 0.1:

( )
4. Simulation and Comparison Strategy
1 τds
Gcparallel(s) ) Kc 1 + + (19)
For this study, the following stable FOPDT process model τis τd
was used: s+1
10

1 To provide a holistic comparison of the various tuning


Gp(s) ) e-θs (18)
10s + 1 methods, the closed-loop system is subjected to both set-point
352 Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

Figure 13. Performance of controller tuning methods for a unit-step change in the set point and output corrupted by measurement noise (nj ) 0, σ2n ) 0.052):
θ/τ ) 2.

Table 2. Summary of Recommended Tuning Rules in the Order of θ/τ ) 1.0 is representative of a balanced process, and θ/τ )
Preference for Servo Control 2.0 is representative of a delay-dominant process.
lag-dominant delay-significant Type 2. For FOPDT processes, an important factor that
inputs (θ/τ ) 0.1) (θ/τ ) 1 and 2) affects the performance and robustness of the closed-loop system
1. Huang and Jeng14 1. Huang and Jeng14 is the θ/τ ratio. Termed as the controllability index by
step 2. Huang et al.15 2. Lee et al.10 Deshpande and Ash,21 the θ/τ ratio can be interpreted to
3. Lee et al.10
represent the real or apparent time delay of the closed-loop
1. Skogestad11 1. Skogestad11 system.22 To explore the effects of the θ/τ ratio on the various
step (noisy) 2. Huang et al.15 2. Huang and Jeng14
3. Lee et al.10 3. Lee et al.10
tuning methods, the time delay θ was varied from 1 to 20 such
that θ/τ ∈ [0.1,2].
changes and load disturbances: (a) for the servo problem, the To facilitate the simulations and comparisons described above,
set point ysp is modeled as a unit-step input in the absence of an integrated MATLAB/SIMULINK program was developed; this
measurement noise, n, and is then repeated in the presence of code can be made available upon request. For the ease of cross-
measurement noise with mean, nj ) 0, and variance, σ2n ) referencing, all of the simulations were carried out for 200 units
0.052; and (b) for the regulatory problem, the load disturbance of time using the Dormand-Prince (4,5) variable step algorithm
d is first modeled as a unit-step input, followed by a random as the ordinary differential equation solver.
input with mean dh ) 0 and variance σ2d ) 1; the latter is also
known as a stochastic load disturbance. Similar to the servo 5. Results and Discussion
problem, the regulatory problem is first studied in the absence
of measurement noise n and then repeated in the presence of Using the comparison strategy described in Section 4, the
measurement noise with mean nj ) 0 and variance σ2n ) 0.052 results are presented separately for the two problems: servo
(Figure 4). control and regulatory control. Nonetheless, before these two
All of the simulations carried out in this study can be problems are addressed, it is useful to discuss the curves
categorized into two distinct types. observed for the model-based tuning methods as a result of
Type 1. To explore the limits of model-based tuning methods varying λ.
with a single adjustable parameter λ, the recommended values 5.1. The Nature of Model-Based Tuning Methods. Figure
for λ in Section 3 were not utilized. Instead, λ was adjusted 5 illustrates the variation in the performance of the selected
incrementally from the minimum value of λ that gives closed- controller tuning methods having λ or MS as an adjustable
loop system stability. On the other hand, the upper limit for λ parameter, for the servo control of three representative pro-
is of secondary importance because the closed-loop system tends cesses: lag-dominant (θ/τ ) 0.1), balanced (θ/τ ) 1), and delay-
toward better robustness as λ increases. The same approach is dominant (θ/τ ) 2). For the model-based methods with a single
applied the tuning method of Åström and Hägglund,16 where adjustable parameter λ, the marker x denotes λ f 0 and the
the adjustable parameter is MS. To cover the various FOPDT marker O denotes the recommended value for λ, if any. The
processes, the methodology was repeated for three different θ/τ legend for all of the plots in this article, unless otherwise
ratios: θ/τ ) 0.1 is representative of a lag-dominant process, specified, is as shown in Figure 6.
Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 353

Figure 14. Performance of controller tuning methods for a unit-step load disturbance: θ/τ ) 0.1.

Figure 15. Performance of controller tuning methods for a unit-step load disturbance and output corrupted by measurement noise (nj ) 0, σ2n ) 0.052): θ/τ
) 0.1.
The curves shown in Figure 5 have the following implications. achieved by a particular tuning method, an alternative tuning
(i) For each tuning method with a single adjustable parameter, may be tried because a more aggressive setting (smaller λ) will
there exists an optimum value of λ for which a minimum IAE only cause the output performance to deteriorate.
value can be achieved. However, this does not mean that the (ii) Further, as illustrated in Figures 5 and 7, it is possible to
minimum achievable IAE value for each tuning method is the achieve a desired level of output performance with two different
same. Hence, if satisfactory output performance cannot be values of λ, such that both settings give a stable closed-loop
354 Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

Figure 16. Performance of controller tuning methods for stochastic disturbances (dh ) 0, σ2d ) 1): θ/τ ) 0.1.

Figure 17. Performance of controller tuning methods for stochastic disturbances (dh ) 0, σ2d ) 1) and output corrupted by measurement noise (nj ) 0, σ2n
) 0.052): θ/τ ) 0.1.
response with similar IAE values. Consequently, the selection (iii) Whereas it is safe to conclude that increasing λ will lead
of an appropriate adjustable parameter should be based on an to a more-robust control system, the conclusion that overly
output performance related metric (i.e., IAE, ISE) and at least decreasing λ will lead to an unsatisfactory performance is
one non-output performance related metric, which could be a subjective. Figure 7 shows that decreasing λ will lead to a faster
control effort metric (i.e., TV) or a robustness metric (i.e., MS). and more oscillatory response; because the controller gain Kc
Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 355

Figure 18. Performance of controller tuning methods for a unit-step load disturbance: θ/τ ) 1.

Figure 19. Performance of controller tuning methods for a unit-step load disturbance and corrupted by measurement noise (nj ) 0, σ2n ) 0.052): θ/τ ) 1.

increases as λ decreases, this translates to greater control action and MS are desirable properties, the optimal tuning rule/s can
and, consequently, a greater amount of control effort required. be located in the lower-left portion of the individual plots. For
If rapid responses are desired, λ can be decreased at the expense the tuning method of Huang et al.,15 default values of ko ) 0.65
of robustness and greater control effort. and a ) 0.4 were used in this study.
5.2. Servo Control. Because TV and MS are plotted on the y 5.2.1. Servo Control of Lag-Dominant FOPDT Processes.
axis against IAE on the x axis and minimal values of IAE, TV, From Figure 8, it is apparent that Rivera et al.,9 Lee et al.,10
356 Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

Figure 20. Performance of controller tuning methods for stochastic disturbances (dh ) 0, σ2d ) 1): θ/τ )1.

Figure 21. Performance of controller tuning methods for stochastic disturbances (dh ) 0, σ2d ) 1) and corrupted by measurement noise (nj ) 0, σ2n ) 0.052):
θ/τ ) 1.

Huang and Jeng,14 and Huang et al.15 all provide excellent set- in set point. For more-robust tuning, Huang et al.15 or Lee et
point responses with sufficient robustness for the control of lag- al.10 with larger λ values can be considered as alternatives.
dominant processes in the absence of measurement noise. This If the control effort required is a priority in the selection of
is expected for Huang and Jeng14 because the tuning method is an optimal tuning method, Skogestad11 can be considered at
based on the global minimization of IAE for a unit-step change the expense of achievable maximum output performance.
Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 357

Figure 22. Performance of controller tuning methods for a unit-step load disturbance: θ/τ ) 2.

Figure 23. Performance of controller tuning methods for a unit-step load disturbance and corrupted by measurement noise (nj ) 0, σ2n ) 0.052): θ/τ ) 2.

However, this is discouraged, given that the minimum IAE able IAE, more-strenuous control action would be required and
/
achievable for Skogestad11 (JSkogestad ≈ 2.5) is about 1.8 times hence a greater TV compared to a PI controller.
/
that of Huang and Jeng14 (JHuang&Jeng ≈ 1.4) (Figure 8). Performance presented in Figure 9 resulted when random
Because the use of derivative action improves the responsive- measurement noise n with mean nj ) 0 and variance σ2n ) 0.052
nessof the control system, thus improving the minimum achiev- was introduced into the process output. In terms of optimal
358 Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

Figure 24. Performance of controller tuning methods for stochastic disturbances (dh ) 0, σ2d ) 1): θ/τ ) 2.

Figure 25. Performance of controller tuning methods for stochastic disturbances (dh ) 0, σ2d ) 1) and corrupted by measurement noise (nj ) 0, σ2n ) 0.052):
θ/τ ) 2.

output performance with minimum control effort and maximum should not be implemented if the system is corrupted by
robustness, it is evident that Skogestad11 is the superior tuning substantial measurement noise.
method for the servo control of noisy lag-dominant processes, 5.2.2. Servo Control of Balanced and Delay-Dominant
/
where the difference between JSkogestad (∼10.5) and FOPDT Processes. For the servo control of balanced and delay-
/
JLee(∼10.1) is only about 4%, but the control effort required dominant processes, collectively known as delay-significant
by Lee et al.10 is about 18 times that required by Skogestad.11 processes, similar conclusions can be drawn for both the
The contrast of results between Figures 8 and 9 shows that, for noiseless and noisy cases. From Figures 10-13, it is apparent
the servo control of lag-dominant processes, derivative action that the method of Huang and Jeng14 provides the best servo
Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 359

5.3. Regulatory Control. In this section, the performance


of the various controller tuning methods for regulatory control
is first evaluated for a unit-step load disturbance first in the
absence and later in the presence of measurement noise. Because
random load disturbances are frequently encountered in the
process industries, the load disturbance input is subsequently
remodeled as a stochastic input with a mean dh ) 0 and a
variance σ2d ) 1, and the evaluation was repeated.
5.3.1. Regulatory Control of Lag-Dominant FOPDT
Processes. From Figure 14, it is apparent that for sufficiently
robust controller tuning (MS < 2), Åström and Hägglund16
followed by Chen and Seborg,12 are the superior choices for
the regulatory control of lag-dominant processes in the absence
of measurement noise. This can be expected because both tuning
Figure 26. Schematic of a typical industrial fired heater. methods were designed for disturbance rejection. Consequently,
one would also expect Åström and Hägglund16 and Chen and
Table 3. Summary of Recommended Tuning Rules in the Order of Seborg12 to give favorable results when applied to the regulatory
Preference for Regulatory Control
control of balanced and delay-dominant processes. Although
lag-dominant balanced delay-dominant Ziegler and Nichols7 delivers the best achievable output
inputs (θ/τ ) 0.1) (θ/τ ) 1) (θ/τ ) 2)
performance, it suffers from poor robustness (MS > 4) and thus,
step 1. Åström and 1. Åström and 1. Lee et al.10 is not highly recommended. If the use of Ziegler and Nichols7
Hägglund16 Hägglund16
is necessary, the process model has to be sufficiently accurate
1. Skogestad11 1. Skogestad11 1. Skogestad11 and invariant in nature.
step
2. Åström and 2. Åström and 2. Lee et al.10
(noisy)
Hägglund16 Hägglund16 When random measurement noise n with mean nj ) 0 and
1. Rivera et al.9 1. Åström and 1. Rivera et al.9
variance σ2n ) 0.052 was introduced into the lag-dominant
random Hägglund16 process output; the performance of various controllers is as
2. Huang et al.15 2. Lee et al.10 shown in Figure 15. Evidently, Skogestad11 exhibits superior
1. Skogestad11 1. Skogestad11 1. Skogestad11 output performance with a minimum control effort required and
random 2. Rivera et al.9 2. Åström and 2. Rivera et al.9 good robustness (MS ≈ 2) for the regulatory control of noisy
(noisy) Hägglund16 lag-dominant processes. A comparative study of Rivera et al.,9
3. Huang et al.15 3. Lee et al.10 Lee et al.,10 Chen and Seborg,12 and Skogestad11 for the
Table 4. Optimal Tuning Rules Identified for the Fired Heater regulatory control of FOPDT processes has also been conducted
by Foley et al.23 Although the input disturbance was modeled
step Huang and Jeng14
servo control
step (noisy) Skogestad11 as a random walk and different performance metrics were used,
the relative performances of the four tuning rules were es-
step Åström and Hägglund16
step (noisy) Skogestad11 sentially similar to the results obtained in Figures 14 and 15.
regulatory control
random Rivera et al.9 For the regulatory control of stochastically disturbed lag-
random (noisy) Skogestad11 dominant processes not corrupted by measurement noise,
responses for processes with significant apparent time delays Åström and Hägglund16 and Chen and Seborg12 cease to be the
(θ/τ g 1), regardless of the effects of measurement noise. For optimal tuning rules (Figure 16). Instead, with a maximum
a more-robust system, Lee et al.10 can be considered as it leads output performance corresponding to the minimum IAE achiev-
to relatively lower values of TV and MS at the expense of slightly able and robustness of MS ≈ 1.8, Rivera et al.9 is recommended
higher IAE values. for the regulatory control of stochastically disturbed lag-
Although the use of derivative action improves the servo dominant processes in the absence of measurement noise.
response in terms of the minimum IAE achievable for delay- Alternatives such as Huang et al.15 and Huang and Jeng14 can
significant processes, its use for noisy processes is not neces- also be considered.
sarily warranted. From Figure 13, the minimum achievable IAE Figure 17 shows that Skogestad11 continues to be the preferred
/
by Skogestad11 (JSkogestad ≈ 45) is about 30% more than that tuning rule for the control of noisy lag-dominant processes
/
by Huang and Jeng14 (JHuang&Jeng ≈ 32), whereas the control despite the random nature of the load input. Nonetheless, if the
effort required by the latter is about 25 times more than that control effort required is not an important consideration, Lee
required by the former. Hence, for noisy delay-significant et al.10 can be considered as the alternative tuning rule, though
processes, it is recommended that Skogestad11 be used first to the improvement in output performance is only about 4%-
MS < 2 /
determine whether satisfactory output performance can be (IAE Skoges t ad ≈ 8.8, JLee ≈ 8.5), whereas the increase in
attained. If not, Huang and Jeng14 or Lee et al.10 can be control effort required is about 10 times. Similar conclusions
considered to improve the output performance. that have been previously drawn for the servo problem verify
The inferior minimum IAE achievable using a Skogestad11 the disincentive in the use of derivative action for noisy lag-
PI tuning algorithm compared to PID tuning algorithms is dominant processes.
indicative of the benefits of derivative action on systems with 5.3.2. Regulatory Control of Balanced FOPDT Processes.
significant apparent time delays. It is also noteworthy that For the regulatory control of noise-free balanced processes, Chen
although Sree et al.13 was designed for set-point tracking and Seborg12 exhibits superior performance in dealing with both
whereas Chen and Seborg12 was designed for disturbance step and stochastic load disturbances (Figures 18 and 20). This
rejection; the latter outperforms the former in the servo control was expected because Chen and Seborg12 was designed specif-
of delay-significant processes. Recommended tuning rules for ically for the purpose of disturbance rejection. Nevertheless, at
servo control are summarized in Table 2. optimal IAE values, the tuning rule is mostly nonrobust, that
360 Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

Figure 27. Performance of controller tuning methods for a unit-step change in the set point: θ/τ ) 0.262.

Figure 28. Performance of controller tuning methods for a unit-step change in the set point and output corrupted by measurement noise (nj ) 0, σ2n ) 0.052):
θ/τ ) 0.262.
is, when J* ≈ 10.3, MS ≈ 3.0 in Figure 18 and when J* ≈ 7.2, the deteriorating minimum IAE achievable. Åström and Häg-
MS ≈ 2.4 in Figure 20. Hence, Åström and Hägglund16 should glund16 can be considered as the alternative tuning rule for robust
be considered as the default tuning rule for a necessarily robust and improved output performance if the control effort required
controller (MS < 2), exhibiting relative superior output perfor- is of secondary importance.
mance. For the regulatory control of noisy balanced processes 5.3.3. Regulatory Control of Delay-Dominant FOPDT
(Figures 19 and 21), as the apparent time-delay increases, Processes. For the regulatory control of noise-free delay-
Skogestad11 is no longer the superior tuning rule because of dominant processes subjected to unit-step load disturbances
Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 361

Figure 29. Performance of controller tuning methods for a unit-step load disturbance: θ/τ ) 0.262.

Figure 30. Performance of controller tuning methods for a unit-step load disturbance and output corrupted by measurement noise (nj ) 0, σ2n ) 0.052): θ/τ
) 0.262.
(Figure 22), Lee et al.10 exhibits both superior performance and but is insufficiently robust (MS ≈ 2.5). For a sufficiently robust
robustness, whereas for the regulatory control of noise-free controller tuning, Rivera et al.9 should be considered instead.
delay-dominant processes subjected to stochastic load distur- Similar to previous conclusions on lag-dominant and balanced
bances (Figure 24), Ziegler and Nichols7 outperforms all of the noisy processes, it is recommended that Skogestad11 be used
recent tuning methods in terms of maximum achievable IAE first to determine whether satisfactory output performance can
362 Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

Figure 31. Performance of controller tuning methods for a stochastic disturbance (dh ) 0, σ2d ) 1): θ/τ ) 0.262.

Figure 32. Performance of controller tuning methods for a stochastic disturbance (dh ) 0, σ2d ) 1) and output corrupted by measurement noise (nj ) 0, σ2n
) 0.052): θ/τ ) 0.262.

be attained before considering Lee et al.10 for unit-step load 22 and 23 were also similar to the results obtained in the
disturbances or Rivera et al.9 for stochastic load disturbances comparative study conducted by Foley et al.23
(Figures 23 and 25). The relative performances of Rivera et Recommended tuning rules for regulatory control under
al.,9 Lee et al.,10 Chen and Seborg,12 and Skogestad11 in Figures various situations are summarized in Table 3. A comparison of
Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 363

Figure 33. Legend for the subsequent figures in this section (Figures 34-40).

Figure 34. Robustness of controller tuning methods as θ/τ increases.

PI tuning rules conducted by Foley et al.24 indicates thatSkoges- dinlet variable. The following transfer function has been reported
tad11 is a well-rounded tuning rule for the control of lag- by Marlin25 for this process
dominant systems.
5.4. Case Study: Fired Heater. This section serves to verify Tout(s) 0.1e-1.1s
Gp(s) ) ) (20)
the validity of recommendations in Tables 2 and 3 by using FFuel(s) 4.2s + 1
them to predict the optimal tuning rules for the feedback control
of a typical industrial process: a fired heater shown in Figure With the θ/τ known, the optimal tuning rules can be identified
26 where the temperature of the feed outlet Tout is the controlled using the recommended tuning rules for the servo and regulatory
output variable and the fuel flow rate FFuel is the manipulate- controls listed in Tables 2 and 3. Because θ/τ ) 0.262 and is
364 Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

Figure 35. Performance of controller tuning methods for a unit-step change in the set point as θ/τ increases.

Figure 36. Performance of controller tuning methods for a unit-step load disturbance as θ/τ increases.

,1, the optimal tuning rule with respect to each of the six based on recommendations were congruent with the results
perturbation types is identified by considering the recommended obtained. Because the identified optimal tuning rules turn out
tuning rules for the lag-dominant processes, and these are to be the best based on the above comparisons, the determination
summarized in Table 4. of an optimal PID tuning rule for the control of a FOPDT
Figures 27-32 illustrate the actual simulation results obtained process can be achieved using Tables 2 and 3 and the θ/τ ratio
as outlined in Section 4. The optimal tuning rules identified of the process.
Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 365

Figure 37. Performance of controller tuning methods for a stochastic load disturbance (dh ) 0, σ2d ) 1) as θ/τ increases.

Figure 38. Performance of controller tuning methods for a unit-step change in the set point and output, corrupted by measurement noise (nj ) 0, σ2n )
0.052), as θ/τ increases.
5.5. Effects of the Controllability Index. To investigate the main objective is to generalize the trends of IAE, TV, and MS
effects of the controllability index (i.e., the ratio of the apparent for the respective tuning rules as θ/τ increases. Because MS is
time delay to the time constant in the FOPDT model of a calculated directly from the closed-loop transfer functions and
process), the IAE, TV, and MS for the various tuning methods is independent of the process inputs, one MS versus θ/τ plot is
were plotted as a function of θ/τ for each of the six different sufficient. For the tuning methods proposed by Rivera et al.,9
types of external perturbations considered in this study. The Skogestad,11 and Åström and Hägglund,16 the values of the
366 Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

Figure 39. Performance of controller tuning methods for a unit-step load disturbance and output corrupted by measurement noise (nj ) 0, σ2n ) 0.052), as
θ/τ increases.

Figure 40. Performance of controller tuning methods for a stochastic load disturbance (dh ) 0, σ2d ) 1) and output corrupted by measurement noise (nj )
0, σ2n ) 0.052), as θ/τ increases.

adjustable parameter recommended by the respective authors From Figure 34, the robustness of the various tuning methods,
were used. On the other hand, Lee et al.10 and Chen and Seborg12 with the exception of Ziegler and Nichols,7 generally remains
were excluded in this study because of the lack of such constant as θ/τ increases. The decreasing MS trend for Ziegler
recommended values. The legend for all of the plots in this and Nichols7 can be attributed to the smaller proportional gain
section is as shown in Figure 33. and larger integral time as θ/τ increases. Hence, the control
Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 367

system tends to be less oscillatory and more stable as θ/τ Gd(s) ) disturbance transfer function (assumed equal to Gp(s)
increases. Nevertheless, Ziegler and Nichols7 tuning remains in this work)
insufficiently robust with MS > 2. Gp(s) ) process transfer function
With reference to Figures 35-40, nearly identical trends of Gp-(s) ) portion of Gp(s) with left half plane zeros
the effects of the controllability index on the performance of IAE ) integral of the absolute error
controllers tuned by different tuning rules are observed for all IMC ) internal model control
of the six different types of external perturbations considered. ISE ) integral of the squared error
For FOPDT processes with θ/τ < 0.2, Skogestad11 is comparable ITAE ) integral of the time weighted absolute error
in output performance to the rest of the PID tuning methods J* ) minimum IAE
but requires significantly less control effort. For noisy systems, ko ) correlation parameter in Huang et al.15
the control effort required is even less. Again, this illustrates
K ) process steady-state gain
the disincentive in the use of derivative action for the control
Kc ) proportional gain of controller
of lag-dominant and/or noisy processes. Consequently, Skoges-
MS ) maximum sensitivity
tad10 should be considered as the default tuning rule for such
processes. CO ) constrained optimization
As θ/τ increases, Skogestad11 is increasingly inferior in output PI ) proportional and integral
performance compared to the rest of the PID tuning methods PID ) proportional, integral and derivative
(with the exception of Ziegler and Nichols7). Furthermore, the s ) Laplace transform variable
control effort required using PID tuning algorithms diminishes S(jω) ) sensitivity function
as θ/τ increases. This relatively enhanced performance of PID SISO ) single input single output
control systems is illustrative of the benefits of implementing TV ) total variation of manipulated input u
derivative action on delay-significant processes. In general, the u ) manipulated input
IAE is observed to be increasing as θ/τ increases, whereas the y ) controlled output
TV decreases as θ/τ increases. More than just illustrating the ysp ) set point
tradeoff that exists between the output performance and the θ ) process dead time or time delay
control effort required, it also indicates the increasing difficulty λ ) tuning parameter in IMC and DS methods
of controlling processes with larger apparent time delays. τ ) process time constant
τi ) integral time
6. Conclusions τd ) derivative time
This article will conclude with specific recommendations on R ) derivative filter parameter
optimal tuning rules for controlling industrial process control Ri ) correlation parameter in Åström and Hägglund16
loops that are affected by set-point changes, load disturbances, ω ) angular frequency
and the corruption of feedback signals by measurement noise. µ ) filter parameter in Huang et al.15
For the general servo control problem, if the system is
substantially noise-free, it is recommended that Huang and Literature Cited
Jeng14 be used for rapid servo responses with satisfactory robust
(1) Koivo, H. N.; Tanttu, J. T. Tuning of PID Controllers: Survey of
stability. If a more-robust setting is desired, Lee et al.10 is a SISO and MIMO Techniques, Proceedings of the IFAC Intelligent Tuning
favorable alternative to consider. On the other hand, for a noisy and AdaptiVe Control Symposium (Singapore); 1991, 758.
system, Skogestad11 should first be considered to determine (2) Desborough, L.; Miller, R. Increasing Customer Value of Industrial
whether satisfactory output performance can be attained, before Control Performance Monitoring - Honeywell’s Experience, Sixth Inter-
national Conference on Chemical Process Control. AIChE Symp. Ser. 2002,
considering Huang and Jeng14 or Lee et al.10 For the general
326 (98).
regulatory problem, if the system is substantially noise-free, (3) O’Dwyer, A. Handbook of PI and PID Controller Tuning Rules;
Åström and Hägglund16 is recommended for the control of lag- Imperial College Press: London, 2003.
dominant and balanced processes, whereas Lee et al.10 is (4) Van Overschee, P.; De Moor, B. RaPID: The End of Heuristic PID
recommended for the control of delay-dominant processes. On Tuning. Preprints of Proceedings of PID ’00: IFAC Workshop on Digital
the other hand, for noisy systems, Skogestad11 should first be Control (Terrassa, Spain); 2000, 687-692.
(5) Seborg, D. E.; Edgar, T. F.; Mellichamp, D. A. Process Dynamics
considered to determine whether satisfactory output performance and Control, Second Edition; John Wiley and Sons: New York, 2004.
could be attained, before considering Åström and Hägglund16 (6) Balestrino, A.; Landi, A.; Medaglia, M; Satler, M. Perfor-
or Lee et al.10 It is to be noted that we consider only the mance Indices and Tuning in Process Control. Proceedings of the 14th
PI controller for Skogestad’s tuning, whereas all of the Mediterrenean Conference on Control and Automation (Ancona, Italy);
other considered controllers are the PID type. Thus, it is no 2006.
(7) Ziegler, J. G.; Nichols, N. B. Optimum Settings for Automatic
surprise that Skogestad’s tuning appears attractive for noisy Controllers. Trans. ASME 1942, 64, 759.
systems and that it results in less input usage compared to PID (8) Tan, W.; Liu, J.; Chen, T.; Marquez, H. J. Comparison of Some
controllers. Well-Known PID Tuning Formulas; Comput. Chem. Eng. 2006, 30, 1416-
1423.
(9) Rivera, D. E.; Morari, M.; Skogestad, S. Internal Model Control:
Nomenclature 4. PID Controller Design. Ind. Eng. Chem. Res. Process Des. DeV. 1986,
25, 252.
a ) correlation parameter in Huang et al.15
(10) Lee, Y.; Lee, M.; Park, S.; Brosilow, C. PID Controller Tuning
d ) load disturbance for Desired Closed-Loop Responses for SI/SO Systems. AIChE J. 1998,
DS ) direct synthesis 44 (1), 106.
e ) error (11) Skogestad, S. Simple Analytic Rules for Model Reduction and PID
FOPDT ) first-order plus dead time Controller Tuning. J. Process Control 2003, 13 (4), 291-309.
(12) Chen, D.; Seborg, D. E. PI/PID Controller Design Based on Direct
Gc(s) ) feedback controller transfer function Synthesis and Disturbance Rejection. Ind. Eng. Chem. Res. 2002, 41,
G/c (s) ) IMC controller transfer function 4807.
368 Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

(13) Sree, R. P.; Srinivas, M. N.; Chidambaram, M. A Simple Method (20) Panagopoulos H.; Åström, K. J.; Hägglund T. Design of PID
of Tuning PID Controllers for Stable and Unstable FOPDT Systems. Controllers Based on Constrained Optimization. IEE Proceedings: Control
Comput. Chem. Eng. 2004, 28, 2201. Theory and Applications; 2002, 149 (1), 32-40.
(14) Huang, H. P.; Jeng, J. C. Identification for Monitoring and Auto- (21) Deshpande, P. B.; Ash, R. H. Elements of Computer Process
Tuning of PID Controllers. J. Chem. Eng. Japan. 2003, 36, 284. Control with AdVanced Control Applications; Instrument Society of
America, Research Triangle Park: NC, 1981.
(15) Huang, H. P.; Jeng, J. C.; Luo, K. Y. Auto-Tune Using Single- (22) Åström, K. J.; Hang, C. C.; Persson, P.; Ho, W. K. Towards
Run Relay Feedback Test and Model-Based Controller Design. J. Process Intelligent PID Control. Automatica 1992, 28 (1), 1-9.
Control 2005, 15, 713. (23) Foley, M. W.; Julien, R. H.; Copeland, B. R. A Comparison of
(16) Åström, K. J.; Hägglund, T. Revisiting the Ziegler-Nichols Step PID Controller Tuning Methods. Canadian J. Chem. Eng. 2005, 83, 712.
Response Method for PID Control. J. Process Control 2004, 14, 635- (24) Foley, M. W.; Ramharack, N. R.; Copeland, B. R. Comparison of
650. PI Controller Tuning Methods. Ind. Eng. Chem. Res. 2005, 44, 6741.
(25) Marlin T. E. Process Control: Designing Processes and Control
(17) Syrcos, G.; Kookos, I. K. PID Controller Tuning Using Mathemati- Systems for Dynamic Performance, Second Edition; McGraw-Hill Higher
cal Programming. Chem. Eng. Process. 2005, 44, 41. Education: New York, 2000.
(18) Åström, K. J.; Hägglund, T. The Future of PID Control. Control
Engineering Practice 2001, 9, 1163-1175. ReceiVed for reView March 28, 2007
ReVised manuscript receiVed September 21, 2007
(19) Huang, H. P.; Jeng, J. C. Monitoring and Assessment of Control Accepted September 28, 2007
Performance for Single Loop Systems. Ind. Eng. Chem. Res. 2002, 41,
1297. IE0704546

You might also like