Sunteți pe pagina 1din 22

ARTICLE IN PRESS

Control Engineering Practice 16 (2008) 407–428


www.elsevier.com/locate/conengprac

Dead-time compensators: A survey


Julio E. Normey-Ricoa,, Eduardo. F. Camachob
a
Depto. de Automac- ão e Sistemas, Universidade Federal de Santa Catarina, 88040-900, Florianópolis, SC, Brazil
b
Depto. de Ingenierı´a de Sistemas y Automática, Universidad de Sevilla, Sevilla 41092, Spain
Received 9 October 2006; accepted 25 May 2007
Available online 12 July 2007

Abstract

This paper presents a review of the main dead-time compensators (DTC) described in literature. The paper analyses the basic Smith
predictor (SP) showing its advantages and drawbacks. DTC structures designed to improve closed-loop characteristics and to control
unstable systems are described. The paper concludes with some recommendations for designing dead-time compensator controllers.
r 2007 Elsevier Ltd. All rights reserved.

Keywords: Dead-time systems; Predictors; Robustness; Dead-time compensators; Process control

1. Introduction in the process industry (Huzmezan, Gough, Dumont, &


Kovac, 2002; Normey-Rico, Bordons, & Camacho, 1997)
Many processes in industry, as well as in other areas, but also in other fields such as robotics (Normey-Rico,
exhibit dead times in their dynamic behaviour. Dead times Gómez-Ortega, & Camacho, 1999) and internet congestion
are mainly caused by mass, energy or information control (Mascolo, 2006). The first DTC structure, the
transportation phenomena, but they can also be caused Smith predictor (SP), was presented at the end of the 1950s
by processing time or by the accumulation of time-lags in a (Smith, 1957) to improve the performance of classical
number of simple dynamic systems connected in series. controllers (PI or PID controllers) for plants with dead
Processes with significant dead times are difficult to time. It is one of the most popular dead-time compensating
control using standard feedback controllers mainly because methods and most widely used algorithm for dead-time
of the effect of the perturbations is not felt until a compensation in industry.
considerable time has elapsed and furthermore, the effect Over the past 25 years, numerous extensions and
of the control action takes some time to be felt in the modifications of the SP have been proposed in order to:
controlled variable. The control action that is applied (a) improve the regulatory capabilities of the SP for
based on the actual error, tries to correct a situation that measurable or unmeasurable disturbances; (b) to allow its
originated some time before. These difficulties can also be use with unstable plants; (c) to improve the robustness or
explained in the frequency domain: the dead time (d) to facilitate the tuning in industrial applications.
introduces an extra decrease in the system’s phase which Practical stability of SPs was analysed in Palmor (1980)
make the process more difficult to control. and Palmor and Halevi (1983) showing that if the primary
When the process has a significant dead time, the controller is not properly tuned, the SP could be unstable
performance of the closed-loop system can be improved when a small mismatch in the delay is considered, in spite
by using a predictor structure. These predictor based of having high gain and phase margins for the ideal system.
controllers are known as dead-time compensators (DTC) The properties of the SP were analysed in Jerome and Ray
and have been applied to many engineering fields, mainly (1986) and in Garcia and Morari (1984) and Morari and
Zafiriou (1989) using some ideas derived from internal
Corresponding author. Fax: +55 48 3721 9934. model control (IMC). Different tunings of the parameters
E-mail addresses: julio@das.ufsc.br (J.E. Normey-Rico), of the SP are proposed in Santacesaria and Scattolini
eduardo@cartuja.us.es (E.F. Camacho). (1993) and Lee, Wang, and Tan (1996) in order to improve

0967-0661/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2007.05.006
ARTICLE IN PRESS
408 J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428

robustness. An analysis of the asymmetry of the dead-time- Problems such as the appropriate selection of the discrete
estimation error effect on performance and stability is model and sampling time are discussed.
made in Limon-Marruedo, Normey-Rico, Pagano, and Next section presents the SP while Sections 3 and 4 are
Aracil (1999), Ingimundarson and Hägglund (2001), and dedicated to analyse the DTCs for stable and unstable
Pagano, Normey-Rico, and Franco (2001). Simple tuning processes, respectively. The discrete implementation is
rules for the SP are given in Hagglung (1996) while two- presented in Section 5. The paper ends with some
degree-of-freedom (2DOF) structures has been introduced recommendations and conclusions.
in Normey-Rico and Camacho (2002). An SP with a
modified fast model is presented in Watanabe and Ito 2. The SP and its properties
(1981) to improve the disturbance rejection capabilities.
This structure can also be used to control unstable plants. The structure of the SP, shown in Fig. 1, can be divided
A modified structure with a filtered predictor is proposed in into two parts: the primary controller CðsÞ and the predictor
Normey-Rico et al. (1997) and is used to control a process structure. The predictor is composed of a model of the
with variable dead time in Normey-Rico, Bordons, plant without dead time (G n ðsÞ), also known in literature as
Berenguel, and Camacho (1998). Other structures, includ- the fast model, and a model of the dead time eLn s . Thus,
ing a disturbance-observer in the DTC, are presented in the complete process model is Pn ðsÞ ¼ G n ðsÞeLn s . The fast
Zhong and Normey-Rico (2002) and also analysed in model Gn ðsÞ is used to compute an open-loop prediction.
Zhong and Mirkin (2002) and Zhong and Li (2003). To consider the modelling errors, the difference between
DTC structures for integrative and unstable processes the output of the process and the model including dead
have been analysed in several papers. Aström, Hang, and time is added to the open-loop prediction, as can be seen in
Lim (1994) introduced a modified SP for integrative the scheme of Fig. 1. If there are no modelling errors or
processes based on a different structure. A simple DTC disturbances, the error between the current process output
for integrative processes based on the estimation of the and the model output will be null and the predictor output
disturbance was presented in Mataušek and Micić (1996) signal yp ðtÞ will be the dead-time-free output of the plant.
and Mataušek and Micić (1999). Modified versions of this Under these conditions, CðsÞ can be tuned, at least in the
structure were presented in Kwak, Whan, and Lee (2001), nominal case, as if the plant had no dead time. Some
Kaya (2003), Chien, Peng, and Liu (2002), and Hang, fundamental characteristics of the SP must be analysed
Wang, and Yang (2003) where more complex algorithms when considering perfect modelling (Jerome & Ray, 1986),
and tuning rules are described. 2DOF control strategies that is, when PðsÞ ¼ Pn ðsÞ; GðsÞ ¼ G n ðsÞ and L ¼ Ln .
for integrating processes were proposed in Zhong and
Normey-Rico (2002) and Zhong and Li (2003) based on the 2.1. Nominal properties of the SP
disturbance observer approach. Other solutions using a
similar approach for the unstable process case were The SP structure for the nominal case (no modelling
presented in Liu, Cai, Gu, and Zhang (2005) and Lu, errors) has the following fundamental properties:
Yang, Wang, and Zheng (2005).
Because of implementation problems, only the discrete
2.1.1. Dead-time compensation and prediction
versions of the dead-time compensators are used in
The dead time is eliminated from the closed-loop
practice. Some of the particular properties of the digital
characteristic equation. From Fig. 1, it is easy to see that
version of the SP are discussed in Palmor and Halevi
if qðtÞ ¼ 0 and GðsÞeLs ¼ G n ðsÞeLn s , the error signal ep ðtÞ
(1990), Guo, Wang, and Shieh (2000), Torrico and
is zero. Under this condition yp ðtÞ ¼ yðt ^ þ Ln Þ and the
Normey-Rico (2005), and Normey-Rico and Camacho
characteristic equation is
(2007).
Previous reviews of the SP and its modifications can be 1 þ CðsÞG n ðsÞ ¼ 0. (1)
found in Palmor (1996), Normey-Rico and Camacho
(1998, 1999c) for monovariable systems. The extension of
the SP for multivariable systems with multiple delays is
q (t)
discussed in Alevisakis and Seborg (1973), Ogunnaike and
Ray (1979), Ogunnaike, Lemaire, Morari, and Ray (1983) r(t) + + y (t)
C (s) P (s)
and also in Palmor and Halevi (1983), Bhaya and Desoer
_
(1985), Jerome and Ray (1986), Feng (1991), and Rao and
Chidambaram (2006). y^ (t + Ln) ^y (t)
+
This survey deals with the analysis, design and tuning of Gn (s) e-Lns
DTCs for stable and unstable processes. Some of the most _
important modifications presented in literature are de- yp (t) +
ep (t)
scribed. Special attention is paid to the simple tuning
+
procedures of the DTCs for industrial processes. Further-
more, the discrete implementation of DTCs is analysed. Fig. 1. Smith predictor controller.
ARTICLE IN PRESS
J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428 409

The feedback signal yp ðtÞ produced by the predictor in q (t)


e (t) u (t)
Fig. 1 anticipates the system output for changes in the set- r (t) + + y (t)
point, although this is not the case for disturbances1: C (s)
P (s)
_ + _ +
^ þ Ln Þ þ Pn ðsÞ½qðtÞ  qðt þ Ln Þ.
yp ðtÞ ¼ yðt
Note that for slow changes in the disturbance qðtÞ ’ +
Gn (s) Pn (s) _
^ þ Ln Þ but if
qðt þ Ln Þ and yp ðtÞ is a good prediction of yðt
qðtÞ changes rapidly then the disturbance cannot be C'(s)
eliminated from the feedback signal yp ðtÞ.
Fig. 2. Equivalent control structure of the Smith predictor.
2.1.2. Performance limitations of the SP
The structure of the SP implicitly factorises the plant
into two parts: Gn ðsÞ that, in some cases, can be invertible
and eLn s that is non-invertible due to the dead time. Using  the poles of PðsÞ cannot be eliminated from the
this idea and considering that an ‘‘ideal’’ controller (one disturbance rejection transfer function, except for a
with infinity gain) could be applied, it follows (see Fig. 2): pole at s ¼ 0. Eq. (3) clearly shows this, even in the ideal
case (CðsÞ ! 1), Pn ðsÞ appears in the expression of yðtÞ,
CðsÞ
C 0 ðsÞ ¼ ¼ ðGn ðsÞÞ1 . (2) but when Pn ðsÞ has a pole at s ¼ 0, Pn ðsÞ½1  eLn s  has
1 þ CðsÞG n ðsÞ not. Note that ½1  eLn s  is zero for s ¼ 0, thus the root
Considering PðsÞ ¼ Pn ðsÞ, the following ‘‘ideal’’ output is of the numerator cancels the root of the denominator.
obtained: In the non-ideal case, note that Pn ðsÞ is a factor in
expression (5) and its denominator will appear in the
yðtÞ ¼ rðt  Ln Þ þ Pn ðsÞ½qðtÞ  qðt  Ln Þ. (3)
denominator of YQðsÞðsÞ
even when CðsÞ cancels the plant
Note that the ‘‘ideal’’ transfer function between the
poles. This has a fundamental consequence: if the poles
reference and the output is a simple delay. Although this
of PðsÞ are slower than the defined closed-loop poles of
ideal controller cannot be used in practice, it shows what
H r ðsÞ, they will dominate the disturbance rejection
could be achieved with the SP and gives an upper limit
response, and thus, the slow transients cannot be
of performance for the closed-loop. Even in the ideal case,
cancelled. In practice this is only a problem when the
if a disturbance is applied at t ¼ 0, it is necessary to wait
dead time is small, because in other cases the closed-loop
until t ¼ 2Ln to note the effect of the controller on the
poles are, in general, very close to the open-loop ones, as
output.
can be seen in this paper.
2.2. Reference tracking and disturbance rejection
2.3. Robustness
To understand how the SP performs at the set-point and
In the real case there are modelling errors, thus
disturbance rejection responses, the closed-loop transfer
Pn ðsÞaPðsÞ. Consider a family of plants PðsÞ such that
functions of the structure in Fig. 1 are computed in the
nominal case. When the model of the plant is perfect PðsÞ ¼ Pn ðsÞ½1 þ dPðsÞ ¼ Pn ðsÞ þ DPðsÞ
(PðsÞ ¼ Pn ðsÞ) and
Y ðsÞ CðsÞPn ðsÞ jdPðjoÞjpdPðoÞ; o40
H r ðsÞ ¼ ¼ , (4)
RðsÞ 1 þ CðsÞG n ðsÞ
for all plants in the family. If CðsÞ stabilises G n ðsÞ, the
 
Y ðsÞ CðsÞPn ðsÞ condition for the robust stability of the SP closed loop is
H q ðsÞ ¼ ¼ Pn ðsÞ 1  . (5) that, for all frequencies
QðsÞ 1 þ CðsÞG n ðsÞ
Using the previous transfer functions it must be noted j1 þ CðjoÞG n ðjoÞj
dPðoÞodPðoÞ ¼ . (6)
that jCðjoÞGn ðjoÞj
Function dPðoÞ is the upper bound of the multiplicative
 if CðsÞ is tuned to define the disturbance rejection modelling error of the plant to guarantee stability and it
dynamics then it is not possible to attain a desired set- can be used as a measure of the controller robustness. Note
point response. This is a common problem of all one- that: (i) dPðoÞ is a rational function because the predictor
degree-of-freedom structures, not just a drawback of the has eliminated the dead time. This, in general, allows a
SP. Note that as CðsÞ is the only degree of freedom, it is better trade-off between robustness and performance to be
not possible to arbitrarily define YQðsÞ
ðsÞ
and YRðsÞ
ðsÞ
. achieved than in the PID case where dPðoÞ is dead-time
dependant; (ii) if CðsÞ is not appropriately chosen, small
1
The notation X ðsÞuðtÞ indicates the inverse Laplace transform of uncertainties could drive the system to instability, as
X ðsÞUðsÞ. pointed out by Palmor (1980).
ARTICLE IN PRESS
410 J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428

Because of the important effect of errors in dead-time q (t)


estimation on robustness, it is usual to define the delay + + + y (t)
r (t)
margin of the system Dm as the largest variation in the dead C (s) P (s)
_
time that can occur in the process PðsÞ before the closed- _ +
loop system becomes unstable. Several tuning rules have
been defined using the idea of dead-time sensitivity. For
Gn (s)-Pn (s)
example, in Morari and Zafiriou (1989) the gain of CðsÞ (a
PI controller) is defined using the maximum dead-time Ceq (s)
error DLmax while in Palmor and Blau (1994) the same gain
is related to Dm . Similar tuning procedures are analysed in Fig. 3. Equivalent structure of the Smith predictor.
Sections 3 and 4 as this is an important property of the
dead-time compensator structures. must be stable. In this case C eq ðsÞ is
CðsÞ
C eq ðsÞ ¼
2.4. Shortcomings of the SP 1 þ CðsÞðG n ðsÞ  Pn ðsÞÞ
that is, the primary controller has a feedback block
2.4.1. SP for general unstable plants HðsÞ ¼ G n ðsÞ  Pn ðsÞ. Computing the static gain of HðsÞ
As pointed out in the previous analysis, the poles of PðsÞ for the SP, it follows:
cannot be eliminated from the disturbance rejection
transfer function (5) except for a pole at s ¼ 0. This has lim HðsÞ ¼ Ln G e ð0Þa0, (8)
s!0
a very important consequence: the SP cannot be used
that is, even if CðsÞ has integral action, the equivalent
with general unstable plants that possess pole(s) with
controller C eq ðsÞ will not have the pole at s ¼ 0 as desired
ReðsÞ40.
and, in consequence, will not reject the step disturbances.
Eq. (4) shows that the reference tracking can be obtained
Because of these problems, over the last twenty years
even in the unstable case, as CðsÞ can be chosen to stabilise
several authors have proposed special tuning procedures or
G n ðsÞ. However, the transfer function YQðsÞ ðsÞ
always has
modified versions of the SP. In Sections 3 and 4 some of
unstable poles.
these problems are addressed.

2.4.2. SP for integrative processes 2.5. When to use a DTC


Integrative processes are a special case of unstable
plants. In this case, the SP gives a stable closed-loop if the DTC structures are more complex and require more
controller is properly implemented. This point is analysed knowledge for tuning than traditional PIDs. Is the more
with details in Sections 4.1.3 and 5.4. Consider that CðsÞ complex design and tuning when compared to PID
stabilises Gn ðsÞ and gives a closed-loop with unitary gain, compensated for by better performance and robustness? In
thus which situations the DTC offers better performance? There is
CðsÞG n ðsÞ N e ðsÞ a tendency to think that DTC should be used when the dead
¼ , time of the process is dominant, (i.e. when the normalised
1 þ CðsÞG n ðsÞ De ðsÞ L
dead-time t ¼ LþT 40:5). However, not many papers
where De ðsÞ has all its roots with ReðsÞo0 and analyse the performance and robustness of PID and DTC
N e ð0Þ ¼ De ð0Þ. The disturbance to the output transfer for dead-time processes to determine when DTC must be
function is then used (Ingimundarson & Hägglund, 2002; Normey-Rico &
  Camacho, 2007; Rivera, Morari, & Skogestad, 1986).
N e ðsÞ Ln s
H q ðsÞ ¼ Pn ðsÞ 1  e (7) In Ingimundarson and Hägglund (2002) the integrated
De ðsÞ
absolute error (IAE) is compared for PI, PID and simple
which has a root at s ¼ 0 both in the numerator and DTC for stable and integrative processes when an input
denominator. The pole at s ¼ 0 does not appear in H q ðsÞ step disturbance is applied and a certain robustness is
when properly implemented, avoiding closed-loop instabi- imposed. The conclusions of the analysis are that DTC
lity. offer better performance than PI for T=L in the interval
Assume now a step disturbance and that Pn ðsÞ ¼ PesðsÞ, ½0:1; 5 while the performance of a PID is better than a
Ln s
where Pe ðsÞ is stable and Pn ðsÞ ¼ G e ðsÞ e s . To reject the DTC for T=L on ½1; 10. However, these results were
Y ðsÞ
disturbance, the gain of QðsÞ must be zero. This gain cannot obtained for a fixed value of Dm and the performance of
be computed directly because the cancellation causes an the DTC improves when smaller values of Dm are
indetermination. The analysis can be made using the considered. The analysis concludes that the advantages of
equivalent structure shown in Fig. 3. The internal model the DTC are more significant for integrative processes.
principle states that in order to reject a step disturbance the A different analysis is presented in Normey-Rico and
equivalent cascade controller, C eq ðsÞ in Fig. 3, must have Camacho (2007) where it is shown that the improvement
integral action (a pole at s ¼ 0) and the closed-loop system obtained with a DTC has more to do with the error in the
ARTICLE IN PRESS
J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428 411

estimation of the dead time than with the absolute value of and the effect of the disturbance can be eliminated from
the dead time. The PID is considered as an SP with a dead- the output of the process independently of the type of
time model computed using a Padé approximation. The disturbance if C ff ðsÞ exists such that
conclusion is that the PID can be tuned to offer almost Pq ðsÞ
the same trade-off between performance and robustness as C ff ðsÞ ¼ .
PðsÞ
the SP when the maximum relative dead-time estimation
error is greater than 80%. Finally, for small modelling Consider the plant and load disturbance transfer functions
errors, the advantages of the DTC are more appreciable PðsÞ and Pq ðsÞ defined as
when the dead time is dominant.
PðsÞ ¼ GðsÞeLs ; Pq ðsÞ ¼ Gq ðsÞeLq s .
The SP performs better than the PID controller when the
dead time is dominant and well known. In general, the There are two situations:
improvement in the set-point tracking is more noticeable
than in the disturbance rejection response.  LpLq . In this case the controller is
Gq ðsÞ
3. Improvements and modifications of the SP for stable C ff ðsÞ ¼ eðLq LÞs .
GðsÞ
plants G ðsÞ
q
If GðsÞ can be computed the disturbance is eliminated
This section presents modifications to the SP that have from the output. Otherwise, a pseudo inverse of GðsÞ can
been proposed in literature suitable for open-loop stable be computed
processes. C ff ðsÞPðsÞ ¼ Pq ðsÞX ðsÞ.
Y ðsÞ
The final QðsÞ is
3.1. Improving disturbance rejection
Y ðsÞ
¼ eLq s G q ðsÞ½1  X ðsÞ, (9)
The regulatory capabilities of the DTC for measurable QðsÞ
disturbances can be improved introducing a feed-forward where 1  X ðsÞ has zero static gain and the fastest
action. The formulation is almost the same as the one used achievable response.
in classical controllers and has been incorporated in the SP  L4Lq . In this case is not possible to compute the inverse
structure in Palmor and Powers (1985). When the of eðLq LÞs and the feed-forward controller is given by
disturbances are not measurable the same idea can be used
but an estimation of qðtÞ is required. These two solutions G q ðsÞ
C ff ðsÞ ¼
are presented in the following sections. GðsÞ
and the final transfer function is
3.1.1. Measured disturbances: the DTC with a feed-forward Y ðsÞ
Fig. 4 shows a block diagram of a DTC with a feed- ¼ eLq s G q ðsÞ½1  X ðsÞeðLLq Þs . (10)
QðsÞ
forward using the IMC structure. In this controller the
block Pnq ðsÞ represents the model for Pq ðsÞ. The structure Note that even in this case the solution is better than the
used can represent input disturbances (when Pq ðsÞ ¼ PðsÞ), one obtained when the feed-forward is not used. To
output disturbances (when Pq ðsÞ ¼ 1) and a general show this, compare the transfer function YQðsÞ ðsÞ
in Eq. (10)
disturbance (when Pq ðsÞaPðsÞ and Pq ðsÞa1). In the ideal with the one without the feed-forward
case, when Pnq ðsÞ ¼ Pq ðsÞ and Pn ðsÞ ¼ PðsÞ, the transfer  
Y ðsÞ Lq s CðsÞG n ðsÞ Ls
function YQðsÞ
ðsÞ
is QðsÞ
¼e G q ðsÞ 1 
1 þ CðsÞG n ðsÞ
e . (11)
Y ðsÞ
¼ ½Pq ðsÞ  C ff ðsÞPðsÞ The first term on the right-hand side of the two
QðsÞ equations is the same: eLq s G q ðsÞ. The second term on
the right-hand side (the term between brackets) is
different. First note that the dead time of this term in
q (t)
Eq. (10) is L  Lq while in Eq. (11) it is L. That is, the
Pq (s) dead time has been reduced in Lq . Secondly, the
u1 (t)
Cff (s) dynamics of this term in Eq. (10) is given by X ðsÞ while
r (t) u2 (t) Pnq (s) CðsÞG n ðsÞ
+ _ + y (t) in Eq. (11) it is given by 1þCðsÞG . In general X ðsÞ can be
+ + n ðsÞ
_ C (s) P (s) CðsÞG n ðsÞ
_ chosen to have a faster response than
+ 1þCðsÞGn ðsÞ.
_ +
Gn (s) + +
Pn (s) It is clear that in the two cases analysed, the advantage of
this solution is less important when Lq ! 0.
The previous structure cannot be used when disturbance
Fig. 4. Control structure of the Smith predictor with a feed-forward. is not measurable. However, the idea can be used to
ARTICLE IN PRESS
412 J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428

improve the controller using an estimation of the distur- The robust stability of the closed-loop system is defined by
bance qðtÞ. In this case, the controller is composed of a 1
disturbance observer and a feed-forward controller. One dPðoÞodPðoÞ ¼ 8o 2 ½0; 1. (16)
jV ðjoÞj
advantage of this approach is that the controller can easily be
tuned to reject other types of disturbances and not only steps. Expressions (14–16) show that the controller has decoupled
the set-point tracking and the disturbance rejection
3.1.2. The disturbance-observer-based DTC responses and that only V ðsÞ defines the robust stability.
For clarity, the following analysis will consider the Thus, V ðsÞ is tuned for a compromise between robustness
system model Y ðsÞ ¼ PðsÞðUðsÞ þ QðsÞÞ where an input and disturbance rejection response and F ðsÞ for a desired
disturbance is used. However, the formulation of the set-point tracking behaviour. To reject the disturbances in
controller can be made for the general case (Zhong & the steady state V ð0Þ ¼ 1.
Normey-Rico, 2002). The disturbance can be computed The structure in Fig. 5a is not appropriate for
ideally as implementation so a modified scheme must be used, as is
shown in Fig. 5b. In this new structure it is clear that if
QðsÞ ¼ P1 ðsÞY ðsÞ  UðsÞ G1
n ðsÞ has poles on the right-hand side of the s-plane, V ðsÞ
and then used in a feed-forward controller. This ideal has to be designed to eliminate these poles from the
solution is not realisable as P1 ðsÞ contains a term eLs , thus controller. This condition is achieved if V ðsÞ has the same
a delayed disturbance is computed: zeros as Gn ðsÞ on the right-hand side of the s-plane.
Furthermore, for the implementation, V ðsÞ must guarantee
eLs QðsÞ ¼ G 1 ðsÞY ðsÞ  eLs UðsÞ. (12) that GVnðsÞ
ðsÞ is proper. Thus, the relative degree of V ðsÞ must be
^
The estimated delayed disturbance QðsÞ can be computed same as the relative degree of G n ðsÞ.
using the model Gn ðsÞ and a filter V ðsÞ that gives a proper This structure also shows that the equivalent controller
and stable V ðsÞG 1 C eq ðsÞ has integral action (1  V ðsÞeLn s ¼ 0 for s ¼ 0),
n ðsÞ
thus the closed-loop system tracks step changes at the
^ ¼ V ðsÞ½G1 ðsÞY ðsÞ  eLs UðsÞ
QðsÞ (13) reference. The settling time and overshoot of the set-point
n

and allows a feed-forward action to be implemented to response can be modified with an appropriate choice of
design the disturbance rejection characteristics of the F ðsÞ which must be proper.
closed-loop system. To reject the disturbances of the form s1m in the steady
The complete controller must also include another state the numerator of
degree-of-freedom to define the set-point response, giving Pn ðsÞ½1  V ðsÞeLs , (17)
the 2DOF-DTC based on a disturbance observer as shown
in Fig. 5a. must have m roots in s ¼ 0. This implies
8
The nominal transfer functions between the reference, > ð1  V ðsÞeLs Þjs¼0 ¼ 0;
>
>
the disturbance and the output of this system are given by >
> d
>
> Ls
Y ðsÞ < dsð1  V ðsÞe Þjs¼0 ¼ 0;
>
¼ F ðsÞGn ðsÞeLs , ð14Þ .. (18)
RðsÞ >
> .
>
>
Y ðsÞ > dm1
>
¼ Pn ðsÞð1  V ðsÞeLs Þ. ð15Þ >
>
: m1 ð1  V ðsÞeLs Þjs¼0 ¼ 0;
QðsÞ ds

q (t) where m is the order of the disturbance (1 for steps, 2 for


r (t) + + y (t) ramps, etc.).
+ u (t)
F (s) P (s) Consider, for instance, a FOPDT model of the process
_ 1
and a step disturbance. As m ¼ 1 a simple V ðsÞ ¼ 1þT vs
will
_ +
e-Ls Gn-1 (s) be enough to satisfy the conditions. However, in the case
where a ramp disturbance is considered (m ¼ 2) a second-
order V ðsÞ must be used. Moreover, 1  V ðsÞeLn s must
V (s) have two roots at s ¼ 0 which implies
q (t) ½1  V ðsÞeLn s js¼0 ¼ 0 ! V ð0Þ ¼ 1,
r (t) u (t)
F (s) Gn (s) + V(s) + + y (t)  
V (s)
P (s) d½1  V ðsÞeLn s  dV ðsÞ
_ Gn (s) [1-V(s) e
-Ls
]
ds  ¼ 0 ! ds  ¼ Ln .
s¼0 s¼0

The simplest V ðsÞ is


Fig. 5. Control structure of 2DOF-DTC with a disturbance observer and 1 þ bs
a feed-forward action: (a) structure for analysis; (b) structure for V ðsÞ ¼ .
implementation. ð1 þ sT v Þ2
ARTICLE IN PRESS
J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428 413

To satisfy the previous conditions, b ¼ 2T v þ Ln and T v is 3.1.3. The SP with a modified fast model
the tuning parameter. T v is tuning for a trade-off between As has been pointed out in the original structure
performance and robustness. Note that dPðoÞ for this proposed by Smith, the disturbance rejection properties
controller is defined by of the closed-loop system cannot be arbitrarily defined as
  the open-loop poles are also closed-loop poles of the
ð1 þ sT v Þ2 
dPðoÞ ¼  . transfer function YQðsÞ
ðsÞ
. To avoid this problem a modified
1 þ sb 
dead-time-free model (G m ðsÞ) can be chosen, G m ðsÞa
Thus, this approach allows a more general solution that Pn ðsÞeLn s in such a way that HðsÞ ¼ G m ðsÞ  Pn ðsÞ does
can be applied to every process model and every type of not have the poles of Pn ðsÞ (Watanabe & Ito, 1981). This
disturbance. This controller also allows independent tuning N
scheme is shown in Fig. 7. Consider that Pn ðsÞ ¼ Dpp eLn s .
of V ðsÞ and F ðsÞ. Furthermore, this structure can be used
for other types of disturbances, for example sinusoidal A simple way to obtain this condition is using Gm ðsÞ
signals. If qðtÞ ¼ sinðo0 tÞ, then the condition that V ðsÞ must N m ðsÞ
satisfy is Gm ðsÞ ¼ ; N m ðsÞaN p ðsÞ
Dp ðsÞ
ð1  V ðsÞeLs Þjs¼jo0 ¼ 0. and selecting N m ðsÞ in such a way that the following
The following example illustrates the design of the conditions are verified:
controller for the case of ramp disturbances. Hðsj Þ ¼ 0; sj ¼ 0; sj =Dp ðsj Þ ¼ 0. (19)
5s
e When these conditions are satisfied the poles of Pn ðsÞ are
Example. Consider the process PðsÞ ¼ ð1þsÞ 3 and the model
e6s
Pn ðsÞ ¼ ð1þ2sÞ. The closed-loop system must follow step not poles of YQðsÞ
ðsÞ
. For an off-set free closed-loop system CðsÞ
changes at the set-point with a settling time similar to the
open-loop one and it also has to reject ramp disturbances. q (t)
r (t) +
Thus + + y (t)
F (s) C (s) P (s)
ð6 þ 2T v Þs þ 1 _ _ +
V ðsÞ ¼ ; F ðsÞ ¼ 1
ðT v s þ 1Þ2
are used. Fig. 6 shows the effect of T v on the responses. Gm (s)-Pn (s)
v (t)
Note that for T v ¼ 2 the closed-loop system has a faster H (s)
response but some oscillations appear because of the effect
of the model uncertainties. Fig. 7. 2DOF-DTC with a modified fast model.

case Tv = 3
1
y
y and r

0
-1 r
-2
0 50 100 150 200 250 300 350 400 450 500

5
u and q

0 u
-5 q

0 50 100 150 200 250 300 350 400 450 500


case Tv = 2
1
y and r

0 y
-1 r
-2
0 50 100 150 200 250 300 350 400 450 500

5
u and q

0 u
-5 q

0 50 100 150 200 250 300 350 400 450 500


time

Fig. 6. Plant output, reference, disturbance and control action for the 2DOF disturbance-observer-based DTC for T v ¼ 3 and 2.
ARTICLE IN PRESS
414 J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428

has integral action. Note that CðsÞ must be tuned to q (t)


‘‘control’’ the new fast model G m ðsÞ. In general, this tuning r (t) y (t)
+ + +
is more complicated than the original case because to F (s) C (s) P (s)
_ _ +
satisfy the conditions in HðsÞ the fast model normally
includes zeros on the right-hand side of the s-plane.
For example, if Gn (s)-Pn (s)
v (t)
H (s)
eLs 1 þ sT 1
Pn ðsÞ ¼ ; G m ðsÞ ¼
1 þ Ts 1 þ sT Fig. 8. Structure of the two degree of freedom Smith predictor.

and T 1 ¼ ð1  eL=T ÞT, Gm ðsÞ has a positive zero that goes


to zero if LbT (T 1 ! 1).
A direct consequence of the new fast model and primary 3.1.4.2. Tuning of the 2DOF-SP for FOPDT processes. The
controller is a negative effect on robustness. If CðsÞ is a PI dynamic behaviour of many processes in industry can be
tuned for a pair of fast poles in 1=T 0 , the upper bound of approximated by a first-order plus a dead time (FOPDT)
the multiplicative uncertainty dPðoÞ is given by model
 
ð1 þ joT 0 Þ2 
dPðoÞ ¼    K p eLs
1 þ joT i  Pn ðsÞ ¼ .
1 þ Ts
and as T 0 5T i , dPðoÞ will be less than one at medium
frequencies, thus, the closed-loop system will have poor The tuning procedure for this case can be summarised as
robustness. follows:
Furthermore, it is important to note that the elimination
K
of the open-loop poles from YQðsÞ ðsÞ
only has an important  p
choose Gn ðsÞ ¼ 1þTs ;
effect if the poles of the term in brackets in the transfer  choose CðsÞ ¼ kc ð1 þ sT1 i Þ. By making T i ¼ T, the
function characteristic equation is
 
Y ðsÞ CðsÞPn ðsÞ kc K p
¼ Pn ðsÞ 1  1 þ CðsÞG n ðsÞ ¼ 1 þ ;
QðsÞ 1 þ CðsÞG m ðsÞ sT i
are faster than the open-loop ones. This is not the case
when the process has a dominant dead time, because slow  define a closed-loop time constant T 0 and compute
responses are necessary to achieve robust performance of kc ¼ KTp Ti 0 ;
the closed-loop system. Furthermore, when the dead time is  choose F ðsÞ ¼ 1þsT
1þsT 1 .
0

dominant the contribution of the open-loop poles to the


closed-loop response will be small, thus their elimination With this choice, the closed-loop transfer functions are
will contribute with a small increment in the speed of the given by
transients.  
Y ðsÞ eLs Y ðsÞ K p eLs eLs
¼ ; ¼ 1 (20)
3.1.4. The 2DOF SP RðsÞ 1 þ T 1 s QðsÞ 1 þ Ts 1 þ T 0s
When the primary controller of the SP is tuned to
accelerate the closed-loop disturbance rejection response, which shows that the nominal set-point response (defined
the set-point tracking response deteriorates. This effect is by parameter T 1 ) has been decoupled from the disturbance
caused by the position of some of the zeros of the reference response (defined by parameter T 0 ). As a measure of
to the output closed-loop transfer function which are controller robustness:
introduced by the primary controller. As these zeros do not dPðoÞ ¼ j1 þ joT 0 j (21)
appear in the disturbance rejection transfer function, their
effect can be attenuated using a reference filter, as shown in is used, showing that T 0 also defines the robust stability
the scheme of Fig. 8. and can be chosen as a good compromise between
robustness and disturbance rejection. Note also that T 0
3.1.4.1. General tuning procedure. The tuning procedure and T 1 have clear physical meaning and can easily be used
consists of defining the primary controller CðsÞ to achieve by practitioners. It is clear that small values of T 0 will
the highest bandwidth for the nominal disturbance cause faster disturbance responses and, at the same time,
response maintaining the robustness condition given by less robust behaviour of the closed-loop.
expression (6). For this computation, a nominal model The proposed controller only has five parameters
Pn ðsÞ and an estimation of the modelling errors dPðsÞ are (K p ; T; L; T 1 and T 0 ) to be tuned, the same as a 2DOF
necessary. After this, as a second step, F ðsÞ can be defined PID. The controller offers both simplicity and performance
to improve the closed-loop set-point response. in the control of dead-time processes.
ARTICLE IN PRESS
J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428 415

3.2. Robustness q (t)


r (t) + y (t)
+
3.2.1. Robust tuning C (s) P (s)
_ +
For processes with dominant dead time, robust tuning
y^ (t + Ln) ^y (t)
of the 2DOF-DTC can be obtained taking only dead-time-
+
estimation errors into account. Note that under these Gn (s) e-Lns
circumstances, the error in the dead time dominates _
the other sources of errors (Ingimundarson & Hägglund, yp (t) +
Fr (s) ep (t)
2001)
+

dPðoÞ ¼ j1 þ joT 0 j4j1  eDLjo j 8o. (22) Fig. 9. Structure of the filtered SP.

When the maximum value of the delay estimation error


DLmax is known, a simple tuning rule for a robust stable periodic disturbance with a period approximately equal to
controller can be defined by choosing T 0 in such a way that minðLn ; LÞ (Normey-Rico et al., 1997).
the right part of the inequality (22) intersects the left part A simple solution to this problem is to use a filter F r ðsÞ
for DL ¼ DLmax . This is approximately obtained for with unitary static gain (F r ð0Þ ¼ 1), as shown in Fig. 9
T 0 ¼ 0:69DLmax . Note that for a defined T 0 this relation- (Normey-Rico et al., 1997). The filter should be designed
ship allows the maximum admissible error in the dead-time to attenuate the oscillations in the plant output especially
estimation (DLmax ) to be computed. Although this tuning at the frequency where the uncertainty errors are impor-
rule does not consider errors in the estimation of the gain tant. The robust stability condition for the filtered SP
and equivalent time constant of the process, it introduces a (FSP) is
small error when the dead time is dominant (Ingimundar-
son & Hägglund, 2001). j1 þ CðjoÞGn ðjoÞj
In practice, robust stability is not sufficient and robust dPðoÞodPFSP ðoÞ ¼ (24)
jCðjoÞGn ðjoÞF r ðjoÞj
performance is desired. In this case, a greater value of T 0
must be used in order to obtain a lower oscillatory response 8o40, thus
when the model differs from the real process. To define the
robust performance for the disturbance response a simple dPSP ðoÞ
dPFSP ðoÞ ¼ . (25)
rule can be used. For example, a factor of 0.5 can be jF r ðjoÞj
applied to the robust stability condition That is, if the filter F r ðsÞ is low pass it can be used to
j1  eDLjo
j improve the robustness of the system at the desired
o0:5 8o. (23) frequency region. Eq. (25) shows that dPFSP ðoÞ can be
dPðoÞ
increased arbitrarily at medium and high frequencies using
Using the same procedure, the condition for the robust an adequate F r ðsÞ.
performance is T 0 ¼ 1:7DLmax . A rule of thumb for tuning Note that if there are no disturbances (qðtÞ ¼ 0), the
the reference filter is to choose T 1 2 ½T; T 0 . The PI tuning nominal performance of the closed-loop system is not
is in this case similar to the one proposed in Morari and modified by the inclusion of F r ðsÞ. It must be noted,
Zafiriou (1989) using the IMC approach. however, that the nominal disturbance to the output
closed-loop transfer function is modified by the inclusion
of the filter. In general, after the inclusion of the filter, the
3.2.2. A simple robust solution: the filtered SP response will be slower than the nominal one. Note that the
Dead-time errors can drive the SP to instability. A new YQðsÞ
ðsÞ
is
simple analysis of the block diagram of the SP in Fig. 1  
shows that when dead-time errors are considered, the Y ðsÞ Ls F r ðsÞCðsÞG n ðsÞ Ls
¼ Gn ðsÞe 1 e
errors between the real and the predicted outputs are fed QðsÞ 1 þ CðsÞG n ðsÞ
back to the controller in a ‘‘periodic’’ way. Note that after
so F r ðsÞ will produce slower transients than the ones
the change in the set-point is applied at t ¼ t0 , the error
obtained with the SP.
between the real output and the predicted one
^ The properties of the FSP suggest the following two-step
(ep ðtÞ ¼ yðtÞ  yðtÞ) will be zero until the instant t ¼ t0 þ x
general procedure for tuning the controller:
where x ¼ minðLn ; LÞ. This signal is then fed back to the
controller and its reaction will be perceived at ep ðtÞ after
only x seconds. This error may cause closed-loop  compute CðsÞ and F ðsÞ in order to obtain the desired
instability but on the other hand, if this error is not fed closed-loop performance for the nominal plant.
back to the controller, the disturbances will not be rejected.  estimate the uncertainty bound of the plant, and
Thus, the effect of the dead-time-estimation error can be compute the filter F r ðsÞ in order to obtain robust
interpreted as the addition of the nominal response plus a stability or robust performance.
ARTICLE IN PRESS
416 J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428

q (t) examples of these types of processes that can be found in


u (t)
+
y (t)
industry.
r (t) +
Qr (s) P (s) As has been pointed out the original structure of the SP
_ + cannot be used to control unstable plants because: (i) if the
+ process has a pole p such that ReðpÞ40, the SP results in an
_ internally unstable closed-loop system; (ii) for the parti-
Pn (s)
cular case of integrative processes, the SP does not reject
step disturbances at the plant input. In this case, internal
Qd (s) instability can be avoided by proper implementation of the
controller.
Different approaches can be considered to solve these
Fig. 10. Structure of the two-degree-of-freedom IMC.
problems:
3.3. IMC interpretation of the DTC
 a DTC with a modified fast model instead of the one
The SP can be considered as a particular case of the IMC used in the SP (DTC-MFM);
(Morari & Zafiriou, 1989). The modified versions of the SP  modified SP structures incorporating feed-forward
presented in the previous sections such as the FSP, FPPI, controller ideas;
the 2DOF-DTC and the disturbance-observer-based DTC  special solutions for integrative processes.
can also be represented as a 2DOF IMC.
A 2DOF IMC structure is shown in Fig. 10. In the IMC All of these approaches eliminate the open-loop unstable
Qd ðsÞ defines the robust stability and the disturbance poles from the transfer function YQðsÞðsÞ
and guarantee the
rejection and Qr ðsÞ defines the set-point response. In the rejection of step disturbances in the steady state.
IMC approach the tuning of Qd ðsÞ and Qr ðsÞ is done using Because of the importance in practice, special attention
optimal procedures considering a defined type of input is paid to some solutions that are only valid for integrative
signals rðtÞ and qðtÞ. processes. As in the case of stable processes, 2DOF
controllers are considered and the tuning take into account
3.4. Predictive PI the robustness of the closed-loop system.

In the dominant-dead-time case the closed-loop response 4.1. Using a modified fast model
of the 2DOFSP will also be dominated by the dead time
and it is not necessary to speed it up. A simple tuning can The idea used in the stable case to eliminate the slow open-
be used loop poles of the process from the transfer function YQðsÞ
ðsÞ
can
T 0 ¼ T 1 ¼ T ) K c ¼ 1=K p ; T i ¼ T. be used here. Consider the structure in Fig. 7, where Pn ðsÞ is
the plant model and G m ðsÞ is the dead-time-free model. As in
This controller is called predictive-PI (PPI) (Hagglung, the stable case, Gm ðsÞ can be chosen with the same
1996). Note that the simple tuning rule (only three denominator as Gn ðsÞ and with a modified numerator
parameters are tuned) eliminates the zero of the reference
filter and gives a fixed robustness index. N m ðsÞ
Gm ðsÞ ¼
The idea of the FSP can be applied to the PPI to obtain Dp ðsÞ
better robustness and to reduce the number of tuning N m ðsÞ is designed in order to eliminate the unstable poles of
parameters. In this case, filter F r ðsÞ can be defined as
1
Dp ðsÞ in YQðsÞ
ðsÞ
, this is equivalent to HðsÞjs¼p ¼ ½Gm ðsÞ 
F r ðsÞ ¼ 1þ0:5Ls . As shown in Normey-Rico et al. (1997)
Pn ðsÞjs¼p ¼ 0 for all p such that Dp ðpÞ ¼ 0; ReðpÞ40.
T f ¼ Ln =2 gives a good solution for dead-time errors of up
To reject steps in qðtÞ an integral action is used in CðsÞ
to 30%. This controller, denominated filtered PPI (FPPI)
and N m ðsÞ is such that HðsÞjs¼0 ¼ 0.
has been used in several DTC structures in literature
The complete model of the process is then used for the
(Ingimundarson & Hägglund, 2001; Normey-Rico et al.,
tuning of CðsÞ and the reference filter F ðsÞ. The implemen-
1997; Rao & Chidambaram, 2006). Also, the same idea is
tation of HðsÞ cannot be achieved by the cancellation of
used in model predictive control to improve the robustness
unstable roots.
of the controller for dead-time processes (Normey-Rico &
Camacho, 1999b, 2000).
4.1.1. The DTC-MFM for the integrative case
Consider that the process can be described by the model
4. DTCs for unstable plants
Pn ðsÞ ¼ GesðsÞ eLn s where G e ðsÞ is a stable transfer function.
In this case G m ðsÞ can be chosen as
There are some cases where the process exhibits
integrative or unstable open-loop behaviour and a dead NðsÞ
Gm ðsÞ ¼ Ge ðsÞ ,
time. Some batch chemical reactors and heating boilers are s
ARTICLE IN PRESS
J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428 417

where NðsÞ will be tuned to obtain where A; B; C are matrices of proper dimensions and
  Pn ðsÞ ¼ CðsI  AÞ1 BeLn s the output signal of the inner
NðsÞ  eLn s
lim G e ðsÞ ¼ 0. (26) loop (vðtÞ) is obtained as
s!0 s Z 0  Z Ln 
The more simple solution for this problem is NðsÞ ¼ vðtÞ ¼ CeALn eAt Buðt þ tÞ dt  CeAt B dt uðtÞ.
1  Ln s, thus Ln 0
(29)
1  Ln s
G m ðsÞ ¼ G e ðsÞ (27) Because of the finite limits of the integrals, the right-hand
s
will be used in the controller. The tuning will consider the side of Eq. (29) has no singularities. Thus, the implementa-
faster disturbance rejection closed-loop response maintain- tion is made without the pole-zero cancellation (Palmor,
ing the robustness conditions. In a second step, F ðsÞ is 1996).
designed to improve the set-point response. This solution considers that the controller will be
implemented in analogous equipments. This is not the
case in practice, as the real time implementation is carried
4.1.2. The DTC-MFM for plants with one RHP pole
out using digital discrete platforms. This point will be
If a simple unstable process model is considered with
analysed in detail in Section 5.
only one right-half plane pole (RHP)
G e ðsÞeLn s 4.1.4. Simple tuning for the IPDT model
Pn ðsÞ ¼ ; T40
Ts  1 As in the stable case, because of the practical impor-
thus, tance, tuning rules will be derived for the simple-
Ln s
integrative-plus-dead-time model (IPDT) Pn ðsÞ ¼ K v es
T ms þ 1
G m ðsÞ ¼ G e ðsÞ which is the simple model most used in industry to describe
Ts  1 integrative processes. For this case, the fast model and the
gives an HðsÞ with a zero at s ¼ 0. T m is computed in primary controller are
order to obtain HðsÞ ¼ Gm ðsÞ  Pn ðsÞ without a pole at  
s ¼ 1=T. As K v ð1  Ln sÞ 1
Gm ðsÞ ¼ ; CðsÞ ¼ kc 1 þ
s sT i
1
HðsÞ ¼ G e ðsÞ ½T m s þ 1  eLn s  and, for the closed-loop transfer function between the
Ts  1
disturbance and the output, a double pole in s ¼ 1=T 0 is
the cancellation of the root at s ¼ 1=T is obtained if
defined. This gives the following parameters:
½T m s þ 1  eLn s js¼1=T ¼ 0.
This condition is equivalent to
 T i ¼ 2T 0 þ Ln .
 kc ¼ K 2T 0 þLn
ðT þL Þ2
.
v 0 n
T m ¼ Tð1  eLn =T Þ
which gives a fast model G m ðsÞ with a pole and a zero To decouple the disturbance rejection and set-point
located on the RHP (note that T m o0). responses, a filter is defined as
ð1 þ sT 0 Þ2
4.1.3. Implementation issues F ðsÞ ¼ .
ð1 þ sT i Þð1 þ sT 1 Þ
The proposed controller must be implemented to
guaranty a stable inner loop HðsÞ ¼ G m ðsÞ  Gn ðsÞeLn s , With this choice the closed-loop transfer functions are
otherwise the scheme shown in Fig. 7 is not internally given by
stable. Note that there is a pole-zero cancellation at the Y ðsÞ eLn s
unstable poles in HðsÞ. This problem arises in the case of ¼ ,
RðsÞ 1 þ T 1 s
integrative and unstable plants. In the stable case, analysed
 
in Section 3, only stable roots are cancelled which do not Y ðsÞ K v eLn s eLn s ð1 þ sðLn þ 2T 0 ÞÞ
cause instability. ¼ 1 (30)
QðsÞ s ð1 þ T 0 sÞ2
This pole-zero cancellation cannot be eliminated by the
use of polynomial division because the numerator of HðsÞ and
is a non-rational expression. A solution for this problem  
 ð1 þ joT 0 Þ2 
consists of the derivation of a non-dynamical form of the dPðoÞ ¼  (31)
1 þ joðLn þ 2T 0 Þ
inner loop (Palmor, 1996). If a state representation of Pn ðsÞ
is used that is, the parameter T 1 defines the set-point response and
T 0 defines the disturbance response as well as the robust
dx
¼ AxðtÞ þ Buðt  Ln Þ, stability. The same conclusions about the compromise
dt between performance and robustness of the controller
yðtÞ ¼ CxðtÞ, ð28Þ derived for stable processes are valid for this case.
ARTICLE IN PRESS
418 J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428

The proposed controller has four parameters (K v ; Ln ; T 1 102


and T 0 ), and when a model of the process is obtained
experimentally, only two parameters (T 0 and T 1 ) need to
be tuned.
When the real process has high-order dynamics, a 101
K v eLn s
second-order model can be used: Pn ðsÞ ¼ sð1þsTÞ . In this

magnitude
case G m ðsÞ ¼ Ksð1þsTÞ
v ð1Ln sÞ
and CðsÞ can be chosen as a PID
100
controller. The same tuning procedure as in the previous
case allows a closed-loop system with a double pole at s ¼
1=T 0 to be obtained.
10-1
norm-error
4.1.5. Robust tuning for the IPDT model norm-dP (T´0=1)
norm-dP (T´0=3)
A similar study to the stable case is done here, again norm-dP (T´0=5)
considering the multiplicative representation of the model- 10-2
10-2 10-1 100 101 102
ling errors. The robust stability of the controller is defined
normalised frequency
by the condition
  Fig. 11. Analysis of expression (34) for different values of T 00 ; d ¼ 0:2 and
 ð1 þ joT 0 Þ2  T 0u ¼ 0:05.
jdPðjoÞjodPðoÞ ¼    8o. (32)
1 þ joðLn þ 2T 0 Þ
Three types of uncertainties will be considered: (i) errors in
estimating the dead time; (ii) errors in the velocity gain K v low frequencies is computed as
and (iii) unmodelled dynamics (that is GðsÞa1=s). The real
dPn0 ¼ lim jdPn ðoÞj ¼ dL þ T 0u .
plant is then given by o!0

K Thus, the minimum value b0 and the maximum value dL þ


PðsÞ ¼ eLs T 0u can be used to tune the robustness of the controller. The
sð1 þ sT u Þ
robust condition is given by
and the model used in the control algorithm is given by
Pn ðsÞ ¼ K v s1 eLn s . The error in the velocity gain only ðT 00 Þ2
affects the curve at very low frequencies where dPðoÞ has b0 4dL þ T 0u ! 4dL þ T 0u
ð1 þ 2T 00 Þ
very high values. Therefore, the error in the velocity gain
will be neglected in the following analysis. The multi- or equivalent
plicative error with a normalised frequency on ¼ Ln o is
T 20
1 b¼ 4L  Ln þ T u .
dPðjon Þ ¼ ejon dL  1. Ln þ 2T 0
1 þ jon T 0u
For dominant-dead-time processes the maximum dead-
Then, Eq. (32) is written as time-estimation error is used to tune the robust condition
   
 1   ð1 þ jon T 00 Þ2 
 e jon dL
1 o  b ¼ DLmax ) T 0 ¼ b½1 þ ð1 þ Ln =bÞ1=2 .
1 þ jo T 0   1 þ jo ð1 þ 2T 0 Þ, (33)
n u n 0
Thus, it follows:
where T 00 ¼ T 0 =Ln , or what is equivalent
     h pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffii
 1 1   ð1 þ jon T 00 Þ2  T 0 ¼ DLmax 1 þ 1 þ Ln =DLmax . (35)
 jon dL 
1 o  
jo 1 þ jo T 0 e jo ½1 þ jo ð1 þ 2T 0 
Þ
8o,
n n u n n 0
If the controller must be chosen for robust performance
(34)
specifications, b should be chosen greater than DLmax . As a
where both sides of the equation are divided by jon . The general tuning rule, the value of b0 can be between 1:5dPn0
analysis of the expression on the right-hand side of Eq. (34) and 2dPn0 if a small overshoot time response is desired.
shows that it is a monotonic decreasing function of the When the unmodelled dynamics impose harder restrictions
frequency and has a constant value at high frequencies on the parameters of the controller than the dead-time
b0 ¼ ðT 00 Þ2 =ð1 þ 2T 00 Þ. Fig. 11 shows the shape of this error the complete expression b ¼ DL þ DT u must be used.
expression for different values of T 00 . This is the case when T e is not negligible if compared to Le
The modelling error divided by jon (jdPn j ¼ jdP=jon j) is and the variations on the high-order dynamics of the
shown in the same figure. Note that this function also has a process are important. As in the stable case, T 1 can be
monotonic decreasing characteristic with the frequency in tuned so as to attenuate the peaks in the step response.
the region where it can intersect the normalised limit of Bigger values of T 1 allows slower responses and smaller
robustness. The constant value assumed by this function at values of the overshoot to be obtained. It is difficult to
ARTICLE IN PRESS
J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428 419

obtain a general rule for tuning T 1 , specially when poor and the pole-zero cancellation in the equivalent controller
information about uncertainties is available. In practice,
V ðsÞ
manual tuning is normally needed and T 1 equal to T 0 is a C eq ðsÞ ¼
good starting point. G n ðsÞ½1  V ðsÞeLn s 
must be eliminated. This can be achieved by using F ðsÞ
having zeros in the same position as the poles of Gn ðsÞ on
4.2. Rejecting disturbances with feed-forward action
the right-hand side of the s-plane. Moreover, F ðsÞ and
C eq ðsÞ must be proper. This implies that the relative degree
When the disturbances can be measured, the feed-
of V ðsÞ must be rd . Furthermore, C eq ðsÞ must be
forward action can also be used to improve the regulatory
implemented with the same procedure as the one explained
capabilities of the DTC for unstable processes. The
in the DTC with a modified fast model.
analysis of this case is similar to the stable one and will
The final tuning takes into account a compromise
not be addressed here. Note that the unstable poles of the
between performance and robustness using an appropriate
plant impose new constraints on the computation of the
tuning of the degree of freedom in F ðsÞ and V ðsÞ. The
pseudo inverse of the process model. However, even for
simple case of lower order models is presented next.
unmeasurable disturbances including a feed-forward action
that uses an estimation of the disturbances, it is possible to
reject them in a DTC when applied to an unstable process. 4.2.2. Tuning for low-order integrative models
Ln s
Because of its importance in practice, some special First consider the model Pn ðsÞ ¼ K v es with a step
structures based on this idea have been presented in disturbance, then:
literature to reject step disturbances when integrator Kv
processes are considered. Gn ðsÞ ¼ ; rd ¼ 1.
s
A simple filter can be used in the set-point controller
s
4.2.1. The disturbance observer approach F ðsÞ ¼ K v ð1þsT 1Þ
, giving
The disturbance observer approach presented in the
Y ðsÞ eLn s
stable case (see Fig. 5) can also be used for unstable ¼ .
N ðsÞ RðsÞ 1 þ sT 1
processes. In this case Gn ðsÞ ¼ DppðsÞ where N p ðsÞ and Dp ðsÞ
The speed of the set-point response is defined by T 1 .
have no common roots and it is assumed that Dp ðsÞ has, at
For this case m ¼ 2, thus, a second-order V ðsÞ with
least, one pole on the right-hand side of the s-plane.
relative degree 1 is necessary. Moreover, 1  V ðsÞeLn s
The set-point response can be defined using F ðsÞ as in the
must have two roots at s ¼ 0 which gives
stable case. The conditions in V ðsÞ are defined by the type
of process and the type of disturbance. 1 þ ð2T 0 þ Ln Þs
V ðsÞ ¼
ð1 þ sT 0 Þ2
 For stability conditions the numerator of 1  V ðsÞeLs and T 0 is the tuning parameter. T 0 is used to obtain an
must cancel the unstable roots of Dp ðsÞ. adequate compromise between performance and robust-
 For a disturbance of type sm11 the numerator of 1  ness because the robust stability condition is given by
V ðsÞeLs must have m roots in s ¼ 0, where  
 1 
m ¼ m1 þ m2 . m1 is the order of the disturbance (1 for jdPðjoÞjo  8o.
steps, 2 for ramps, etc.) and m2 is the number of effective V ðjoÞ
integrators of the plant (the number of roots in s ¼ 0 of The transfer function YQðsÞ
ðsÞ
is
Dp ðsÞ). Thus, this structure can be also used for double-  
integrator processes. Y ðsÞ 1 þ ð2T 0 þ Ln Þs Ln s
¼ Pn ðsÞ 1  e (36)
 For a sinusoidal disturbance of frequency o0 , the QðsÞ ð1 þ sT 0 Þ2
numerator of 1  V ðsÞeLs must include roots at s ¼ this given the same solution as that obtained with the dead-
jo0 and s ¼ jo0 . time compensator based on the modified fast model. Note
that in this case the tuning is simple because the tuning of
For the implementation, V ðsÞ must guarantee that GVnðsÞðsÞ is V ðsÞ does not depend on the tuning of F ðsÞ.
proper. Thus, the order of the controller will depend on the K v eLn s
For a second-order model Pn ðsÞ ¼ sð1þsTÞ and a step
relative degree rd of G n ðsÞ. Furthermore, because of the Kv
disturbance, the fast model is Gn ðsÞ ¼ sð1þsTÞ and rd ¼ 2.
unstable poles of Pn ðsÞ, the implementation is made with
the structure in Fig. 5b with an equivalent filter F eq ðsÞ and For the set-point response a second-order filter is used
cascade controller C eq ðsÞ. Filter F ðsÞ must be chosen so as sð1 þ sTÞ Y ðsÞ eLn s
to obtain a proper and stable equivalent reference filter F ðsÞ ¼ 2
) ¼ ,
K v ð1 þ T 1 Þ RðsÞ ð1 þ sT 1 Þ2
F ðsÞG n ðsÞ where T 1 defines the settling time of this response. For the
F eq ðsÞ ¼
V ðsÞ disturbance rejection, V ðsÞ is a third-order filter with
ARTICLE IN PRESS
420 J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428

dV ðsÞ considered the difference between the model and process


relative degree 2 and must verify: V ð0Þ ¼ 1, ds js¼0 ¼ Ln .
This gives outputs in steady state is an estimation of the amplitude of
1 þ ms qðtÞ. Thus, using MðsÞ this estimated value is introduced in
V ðsÞ ¼ ; m ¼ 3T 0 þ Ln . the loop to eliminate the effect of the disturbance. As MðsÞ
ð1 þ sT 0 Þ3
acts as a feedback controller, it must be tuned to achieve
This last solution will give better results when the process closed-loop stability and also an adequate transient of the
exhibits high-order dynamics because the second-order disturbance rejection response. Moreover, the primary
model allows the model uncertainties to be reduced (Zhong controller CðsÞ is used to define the set-point response of
& Normey-Rico, 2002). the controller.
If the model is equal to the process, the closed-loop
4.2.3. Tuning for the first-order unstable model transfer function between RðsÞ and Y ðsÞ is
Ln s
Consider model Pn ðsÞ ¼ Ke Ts1 . For step disturbance Y ðsÞ CðsÞPn ðsÞ
rejection ½1  eLn s V ðsÞ should have a zero at s ¼ 0 and ¼ ¼ C eq ðsÞPn ðsÞ, (37)
RðsÞ 1 þ CðsÞG n ðsÞ
another one at s ¼ 1=T (V ð0Þ ¼ 1 and V ð1=TÞeLn =T ¼ 1).
This gives where Pn ðsÞ and G n ðsÞ are, respectively, the process model
1 þ sa and dead-time-free process model and C eq ðsÞ is the
V ðsÞ ¼ equivalent controller in Fig. 12. Note that CðsÞ can be
ð1 þ sT 0 Þ2 computed to obtain a desired set-point response and it is
with a such that such that
1 þ a=T lim C eq ðsÞPn ðsÞ ¼ 1.
eLn =T ¼ 1 ) a ¼ ½ð1 þ T 0 =TÞ2 eLn =T  1T, s!0
ð1 þ T 0 =TÞ2
For the disturbances, it follows:
where T 0 is the tuning parameter as in the previous cases.
The nominal closed-loop transfer function is Y ðsÞ Pn ðsÞ
¼ ½1  C eq ðsÞPn ðsÞ ; (38)
QðsÞ 1 þ MðsÞPn ðsÞ
Y ðsÞ 1
¼ eLn s thus, if MðsÞ is stable and stabilises Pn ðsÞ
RðsÞ 1 þ sT 1
being T 1 the desired closed-loop time constant, F ðsÞ is Pn ðsÞ
given by 1 þ MðsÞPn ðsÞ
Ts  1 is stable and as
F ðsÞ ¼
Kð1 þ sT 1 Þ lim½1  C eq ðsÞPn ðsÞ ¼ 0
s!0
and the equivalent reference filter
the transfer function in Eq. (38) has all the poles with
1 ð1 þ sT 0 Þ2 ReðsÞo0 and a zero at s ¼ 0. This guarantees the rejection
F eq ðsÞ ¼ ¼ .
V ðsÞðT 1 s þ 1Þ ð1 þ sT 1 Þð1 þ saÞ of step disturbances.
The condition for robust stability is
 
4.3. Special solutions for integrative processes 1 þ MðjoÞPn ðjoÞ
jdPðjoÞjodPðoÞ ¼  .
C eq ðjoÞ þ MðjoÞ 
In the case of integrative processes, the principal
objective in designing a modified SP is the rejection of Several authors have used this idea to control integrative
step disturbances. Some solutions have been presented with processes and also double integrative processes with a dead
this objective (Aström et al., 1994; Kaya, 2004; Mataušek time (Chien et al., 2002; Hang et al., 2003; Kaya, 2003;
& Micić, 1996, 1999). The simplest controller, proposed in Kwak et al., 2001). In these works, CðsÞ and MðsÞ are
Mataušek and Micić (1996), can be seen in Fig. 12. This defined as P, PI, PD or PID controllers. Because of the
structure can be interpreted as a disturbance observer practical importance, in the next section the analysis and
DTC. The idea here is that when step disturbances are tuning of the controller is presented for the simplest case.

q (t) 4.3.1. The simplest case


r (t) Ceq (s)
The most simple solution for this problem is to use
+ Ln s
+ +
C (s)
u (t) +
P (s)
y (t) model Pn ðsÞ ¼ K v es and a simple gain in each controller
_ _ + (Mataušek & Micić, 1996)
+
Gn (s)
M (s)
_ CðsÞ ¼ K c ; MðsÞ ¼ K 0 .
Pn (s) In this case
Y ðsÞ eLn s 1
Fig. 12. Structure of the modified Smith predictor with feed-forward of ¼ ; Tr ¼ (39)
the estimation of q. RðsÞ 1 þ T r s K cK v
ARTICLE IN PRESS
J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428 421

that is, T r can be used to define the speed of the set-point dPn ðon Þ, is plotted for different cases in Fig. 13. Thus, it is
response. possible to tune parameter T r using an estimation of the
To complete the tuning of the controller, K 0 is computed uncertainties. Note that the shapes of the curves in Fig. 13
imposing a phase margin fm ¼ 60 of the inner loop 1 þ are very similar to the one obtained in the DTC with the
MðsÞPn ðsÞ ¼ 0 giving (Mataušek & Micić, 1996) modified fast model (see Fig. 11). Thus, using the
1 maximum value of the normalised modelling error (dPn0 ),
K0 ¼ . obtained in Section 4.1.5 a parameter b can be defined for
2K v Ln
robust stability (b ¼ DLmax þ T u ) or for robust perfor-
With these parameters, in the nominal case mance (b ¼ 2ðDLmax þ T u Þ) and the tuning of T r as
  
Y ðsÞ eLn s K v eLn s b
¼ 1 (40) T r4 ,
QðsÞ 1 þ sT r s þ K 0 K v eLn s 1  ðb=2Ln Þ
that is, as expected, a stable transfer function with zero static where T u is again, in this case, the equivalent time constant
gain. It is important to note that in this solution parameter of the unmodelled dynamics. Finally, the controller gain K c
T r which defines the set-point response also affects the is K c ¼ T r1K v .
disturbance rejection performance and the closed-loop This controller presents the same advantages of the FPPI
robustness. For this controller, dPðoÞ is given by presented in the stable case, as just as a PID controller it
has only three tuning parameters: K v , K c and Ln . In
jjoðjo þ K c K v Þð2Ln jo þ ejoLn Þj manual tuning, when the evaluation of the modelling error
dPðoÞ ¼ . (41)
j  ð2Ln K v K c þ 1Þo2 þ joK v K c j is not possible the practical rule is to choose T r similar to
As in Section 4.1.5, this equation is divided by jo. This the equivalent time constant of the non-integrative part of
bound depends almost entirely on T r the process. The disadvantage of this controller, if
compared to the 2DOF-DTC analysed in previous
jðjoT r þ 1Þð2Ln jo þ ejoLn Þj sections, is that T r also affects the disturbance rejection
dPn ðoÞ ¼ 8o. (42)
j  ð2Ln þ T r Þo2 þ joj response and the robustness of the controller and therefore
Using a normalised frequency on ¼ Ln o and T 0r ¼ T r =Ln : it does not allow the set-point and disturbance response to
be decoupled. Furthermore, this controller cannot improve
jðjon T 0r þ 1Þð2jon þ ejon Þj its disturbance rejection when the plant has very long dead
dPn ðon Þ ¼ 8o. (43)
j  ð2 þ T 0r Þo2n þ jon j times, as is shown in Normey-Rico and Camacho (1999a).
The controller presented in this section uses both the
For low frequencies the bound behaves as
simplest model and the simplest controllers CðsÞ and MðsÞ.
1 Using high-order models and/or controllers, the results
dPn ðon Þ ffi ; o!0
on obtained in this section can be improved. Several authors
and for high frequencies: have proposed solutions using these more complex models
(Chien et al., 2002; Hang et al., 2003; Kaya, 2003; Kwak et
2T 0r al., 2001). The following section presents a brief resume of
dPn ðon Þ ffi ; o ! 1,
2 þ T 0r these methods.

102 4.3.2. The improved solutions


Ln s
dPn (T´r =0.1) Maintaining the simplest model Pn ¼ K v es and using a PI
dPn (T´r =0.3) controller in CðsÞ and a PD controller in MðsÞ the closed-
dPn (T´r =0.6) loop performance can be improved. The tuning of CðsÞ and
MðsÞ can follow several approaches. For example, CðsÞ and
101 MðsÞ can be, respectively, tuned for a desired gain margin and
magnitude

a phase margin of loops 1 þ CðsÞPn ðsÞ and 1 þ MðsÞPn ðsÞ.


Moreover, the set-point response can be improved if a
reference filter is included in the controller. This solution
obviously increases the design and tuning complexity.
100
In several cases, the behaviour of the real process can
K v eLn s
better be represented using a SOIPDT model Pn ¼ sð1þsTÞ
(Kaya, 2004). This model will give better results than the
one based on the first-order model only when T is not
10-1 negligible when compared to Ln . However, it must be noted
10-2 10-1 100 101 102 that using a second-order model it is possible to obtain a
normalised frequency
small modelling error and therefore improve the robustness
Fig. 13. jdPn ðon Þj for the Mataušek and Micić’s DTC for different values of the controller. In this case, MðsÞ is a PD controller while
of T 0r . CðsÞ is a PI or a PID with set-point weighting. The tuning
ARTICLE IN PRESS
422 J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428

can be done, for example, using pole-zero cancellation and performance and robustness as the disturbance observer.
the same frequency approach as in the simplest case. Tuning However, a poor performance is expected in the case of
can be improved if the pole-zero cancellation is avoided in dominant dead-time. Also, it is important to note that
MðsÞ and also if a 2DOF-PID is used in CðsÞ. Moreover, when the controller has to detuned to improve the
other procedures can be used to tune CðsÞ and MðsÞ. robustness, the differences between the real F ðsÞ and the
Another solution to this problem can be obtained using ideal one are attenuated. This solution, obtained by low
the idea presented in Liu et al. (2005). In this work the order approximation of the non-rational controller could
authors use a slightly different structure that includes a be also applied to the other cases analysed in this paper,
stabilising controller, as it is shown in Fig. 14. This scheme but as mentioned, in practice the controller will be
can be used with general unstable processes. It uses the implemented in a discrete platform where these approx-
controller gain K c to stabilise the open loop unstable dead- imations are unnecessary.
time free model G n ðsÞ. After defining K c the set-point The following example illustrates some of the aspects
tracking controller G c ðsÞ is tuned to improve the set-point discussed in this section.
response. The disturbance rejection response is defined Ls
e
with F ðsÞ. The nominal transfer function between the Example. Consider the process P ¼ sð1þ10sÞð1þsÞð1þ0:5sÞð1þ0:25sÞ
disturbance (q) and the control action (f) generated by F is with two different dead times. In case 1: L ¼ L1 ¼ 5, and in
case 2: L ¼ L2 ¼ 30. The first-order models are defined as
F ðsÞPn ðsÞ
C q ðsÞ ¼ .
1 þ F ðsÞPn ðsÞ e16:8s e41:8s
Pn1 ¼ ; Pn2 ¼
Ideally the faster achievable response is obtained when s s
C q ðsÞ ¼ eLs (note that the dead-time cannot be elimi- and the second-order models as
nated). In the real case, C q ðsÞ has a numerator and
e6:66s e31:66s
denominator such that Pn3 ¼ ; Pn4 ¼ .
sð1 þ 10:14sÞ sð1 þ 10:14sÞ
1 C q ðsÞ
F ðsÞ ¼ As can be seen, only in case 2 L4T. A dead-time
Pn ðsÞ 1  C q ðsÞ
estimating error of 20% and a gain estimating error of
is proper and does not cancel the unstable poles of Pn ðsÞ. 5% are also considered.
For the FOPDT unstable case, if C q ðsÞ is defined as
Two different tunings for CðsÞ and MðsÞ are considered.
1 þ as
C q ðsÞ ¼ eLs
ð1 þ lsÞ2  DTC1 uses first-order models and:
the estimator F ðsÞ is Tds þ 1
CðsÞ ¼ K c ; MðsÞ ¼ K 0
ðTs  1Þð1 þ asÞ 0:1T d s þ 1
F ðsÞ ¼ fm ¼ 60 ; Am ¼ 2 are used for tuning MðsÞ (Mataušek &
K½ð1 þ slÞ2  ð1 þ asÞeLs 
Micić, 1999) obtaining
and to avoid the zero at s ¼ 1=T, a must verify
0:8
a ¼ T½eL=T ð1 þ l=TÞ2  1, T d ¼ 0:45L; K0 ¼ .
LK v
where l can be used as a tuning parameter. The pole-zero K c is tuned using the robustness analysis of the simplest
cancellation of the root at s ¼ 1=T may cause the unstable case considering b ¼ L  Le þ T u  T e . Note that in
operation of F ðsÞ and cannot be removed directly. Thus, this case the error in the estimation of T u is of the same
the authors suggest to use an approximate transfer order as the error in the dead time. It is assumed that
function of the controller, based on the Maclauring T u  T e ¼ 1:5 for the tuning of the DTC1. Thus, using
expansion formula (Liu et al., 2005). Without considering the formula for robust performance gives b ¼ 2ð1 þ
this approximation this controller achieves the same 1:5Þ ¼ 5 in case 1 and b ¼ 2ð6 þ 1:5Þ ¼ 15 in case 2.

q (t)

r (t) + + u (t) + y (t)


Gc (s) P (s)
_ _
+

Kc f (t) +
F (s)
_ _
Gn (s) e-Ls
+

Fig. 14. Disturbance observer with stabilising controller.


ARTICLE IN PRESS
J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428 423

 The second controller, DTC2, uses the second-order 1.05. As can be seen, the DTC2 has a better closed-loop
models and response. It is possible to see here the effect of the time
T cs þ 1 Tds þ 1 constant T of the second-order model, that has a non-
CðsÞ ¼ K c ; MðsÞ ¼ K 0 negligible value when compared to Ln .
0:1T c s þ 1 0:1T d s þ 1
A new situation is simulated for case 2 (L ¼ L2 ¼ 30): a
and a tuning based on a pole-zero cancellation (Kaya, step change (from 0 to 2) of the set-point at t ¼ 10 and a
2004). Thus, T c ¼ T and T d ¼ T. Gains K c and K 0 are 0:05 step disturbance at t ¼ 500. Initially, the DTC1 was
chosen, respectively, for a desired fm of the loop tuned with b ¼ 15 giving T d ¼ 18:81, K 0 ¼ 0:019,
transfer functions CðsÞPn ðsÞ and MðsÞPn ðsÞ K c ¼ 0:0547. The obtained disturbance rejection response
Kc K v eLn s was two times superior to the one with the DTC2, that
CðsÞPn ðsÞ ¼ , offers, in this case a slow response with the proposed
0:1Ts þ 1 s
tuning parameters: T c ¼ T d ¼ 10:14, K c ¼ K 0 ¼ 0:0158.
K0 K v eLn s Thus, b ¼ 30 was used in the DTC1 giving T d ¼ 18:81,
MðsÞPn ðsÞ ¼ .
0:1Ts þ 1 s K 0 ¼ 0:019, K c ¼ 0:0214. This case is shown in Fig. 16 for
Note that the solution is the same for the two loops. If the nominal dead time and gain and for the real dead time,
the filter of the derivative action is neglected the loop L ¼ 36, and the real gain, 1.05. Note that the advantages of
transfer functions are the same as the one obtained when the DTC2 disappear and DTC1 has better performance.
Pn ¼ Ksv and MðsÞ ¼ K 0 . Thus, using fm ¼ 60 to tune This is the effect of the dominant dead time.
the gain, the obtained K c ¼ K 0 is
1
Kc ¼ K0 ¼ . 5. Discrete DTCs
2K v Ln
Although DTC appeared in the mid fifties, their
The closed-loop behaviour of DTC1 and DTC2 is first implementation with the analogue control equipments
compared for case 1. The simulation considers a step was very difficult and DTCs were not used in industry
change (from 0 to 2) of the set-point at t ¼ 10 and a 0:05 until digital controls took over in the eighties. Moreover,
step disturbance at t ¼ 150. The closed-loop responses the dead-time compensation ideas can be applied areas
obtained with the DTC1 (T d ¼ 7:56, K 0 ¼ 0:0476, K c ¼ other than the traditional continuous processes as, for
0:172) and the DTC2 (T c ¼ T d ¼ 10:14, K c ¼ K 0 ¼ 0:075) instance, to dynamic systems directly described in the
are shown in Fig. 15 without considering modelling errors discrete time domain. In these cases, the design of the
and when the real dead time is L ¼ 6 and the real gain is controller is carried out using only discrete time tools.

case: L=5, K=1


2
r, y

1 y (DTC1)
y (DTC2)
0
0 50 100 150 200 250 300 350

1.5
u (DTC1)
1
u (DTC2)
u

0.5
0
-0.5
0 50 100 150 200 250 300 350
case: L=6, K=1.05
2
r, y

1 y (DTC1)
y (DTC2)
0
0 50 100 150 200 250 300 350

1.5
u (DTC1)
1
u (DTC2)
u

0.5
0
-0.5
0 50 100 150 200 250 300 350
time

Fig. 15. Response of the DTC1 (IPDT model and P/PD controller) and the DTC2 (SOIPDT model and PD/PD controller) for case 1, L ¼ 5.
ARTICLE IN PRESS
424 J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428

case: L=30, K=1


2

r, y
1 y (DTC1)
0 y (DTC2)
-1
0 100 200 300 400 500 600 700 800 900 1000

0.6 u (DTC1)
0.4 u (DTC2)
u

0.2
0
-0.2
0 100 200 300 400 500 600 700 800 900 1000
case: L=36, K=1.05
2
r, y

1 y (DTC1)
0 y (DTC2)
-1
0 100 200 300 400 500 600 700 800 900 1000

0.6 u (DTC1)
0.4 u (DTC2)
u

0.2
0
-0.2
0 100 200 300 400 500 600 700 800 900 1000
time

Fig. 16. Response of the DTC1 (solid) and the DTC2 (dashed) for case 2, L ¼ 30.

A discrete synthesis procedure can be used when a DTC r ( kTs) q (kTs) y (kTs)
has to be applied to a discrete time system. On the other + +
F (z) C (z) P (z)
hand, the design of a discrete DTC for a continuous _ +
process can be made using two different approaches:
S (z)
 computing a discretisation of the continuous controller,
yp (kTs)
or +
Fr (z)
 applying a direct discrete design method using a discrete +
representation of the process.
Fig. 17. Unified representation of the discrete DTCs.

The performance of the controller will depend not only on


the sampling time choice but also on the discretisation
blocks CðzÞ and F ðzÞ represent the 2DOF primary
method used to transform the original differential equation
controller.
(of the plant or of the controller) into a discrete model
This structure is a unified dead-time compensator—
(Guo et al., 2000). In fact, the most important character-
UDTC. The UDTC structure can be used to compute a
istics of the continuous DTC can be maintained in the
controller taking into account the ideas previously
discrete case if the sampling time is very small compared to
discussed aiming at increasing robustness, coping with
the dominant time constant of the process. However, in
unstable plants, improving the disturbance rejection
practice, the real time implementation of the controller
properties, or decoupling the set-point and disturbance
may have time constraints and the sampling time must be
responses.
chosen greater than the ideal theoretical value affecting the
The closed-loop behaviour of the UDTC is defined by
closed-loop performance and robustness of the system,
the transfer functions from the reference to the output
especially in the case of dead-time systems.
(H r ðzÞ) and from the disturbance to the output (H q ðzÞ)
F ðzÞCðzÞPðzÞ
5.1. General properties of the discrete DTC H r ðzÞ ¼ ,
1 þ CðzÞ½PðzÞF r ðzÞ þ SðzÞ
For the analysis of the properties of the discrete DTC a
PðzÞ½1 þ CðzÞSðzÞ
unified representation of DTCs for analysis in the discrete H q ðzÞ ¼ .
1 þ CðzÞ½PðzÞF r ðzÞ þ SðzÞ
domain is presented in Normey-Rico and Camacho (2007)
using a general block diagram, as shown in Fig. 17, where As in the continuous case the robustness is analysed using
blocks F r ðzÞ and SðzÞ represent the predictor structure and the multiplicative modelling error represented in the
ARTICLE IN PRESS
J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428 425

frequency domain 8o 2 ½0; p


Pðejo Þ  Pn ðejo Þ 100
dPðjoÞ ¼ pdPðoÞ,
Pn ðejo Þ

j1 þ Cðejo Þ½Pn ðejo ÞF r ðejo Þ þ Sðejo Þj


dPðoÞodPðjoÞ ¼ .

magnitude
jG n ðejo ÞCðejo ÞF r ðejo Þj 10-1

(44)
Note that F r ðzÞ only affects H q ðzÞ and the robustness
condition, while F ðzÞ only affects H r ðzÞ. Thus, the analysis 10-2
and tuning in the discrete domain can follow the same
procedure as used in the continuous case. |δ P| (ZOH)
|δ P| (Tustin)
|δ P| (forward)
5.2. Discretisation of the continuous DTC 10-3
10-1 100 101
frequency
The simplest way to implement continuous DTC on
digital platforms is to use a discretisation procedure. As the Fig. 18. Modelling error for the TUSTIN, FORWARD and ZOH.
DTC is a model based controller the discretisation of
the model plays an important role in the final controller. As as in the real process a ZOH block is used to compute the
the control signal is applied to the process through a zero discrete control actions.
order hold (ZOH) the best method for obtaining the
discrete model of the process is a ZOH-based discretisa-
tion. As shown in Normey-Rico and Camacho (2007) if the A discrete bound for dPðjoÞ can be computed:
model is not appropriate chosen the extra modelling error
introduced in the discretisation of the controller, can dPðjoÞpdPs ðejoTs Þ 8o 2 ½0; p=T s .
produce instability in some cases.
In this case a representation of the form:

5.2.1. Robustness analysis Ps ðejoT s Þ  Pn ðejoT s Þ


In the continuous process case, the frequency domain dPs ðejoTs ÞX ; i ¼ 1; 2; 3
Pn ðejoT s Þ
representation of the process plus the sampler and the
hold is is obtained and dPs ðejoTs Þ is used for the robust tuning of
1  ejoT s the discrete controller.
Ps ðjoÞ ¼ PðjoÞ 8o 2 ½0; p=T s . It is interesting to compute the sampling time T s in order
joT s
to obtain a simple representation of dPs ðejoTs Þ using the
The nominal model (Pn ðzÞ) will depend on the method used ZOH discretisation procedure, both for Pn ðsÞ and PðsÞ. In
to compute the controller. Thus, the modelling error is this case, if T s is adequately chosen, only a small error is
obtained as introduced in the approximation. An equivalent discrete
model Ps ðzÞ is computed for each process model PðsÞ in the
1  ejoT s
PðjoÞ  Pn ðejoT s Þ family of continuous processes, and a nominal model Pn ðzÞ
joT s
dPðjoÞ ¼ 8o 2 ½0; p=T s , such that Ps ðzÞ ¼ Pn ðzÞ½1 þ dPs ðzÞ. This model approxima-
Pn ðejoT s Þ tion will depend on T s . For the particular case of dead-time
which clearly depends on the modelling error of PðsÞ, on systems, where the dead-time-estimation error (DL) is the
the nominal model selection and on the sampling time T s . most important model uncertainty, the sampling time
Note that, for a given sampling time, important choice should be related to DL. The maximum value of the
variations in the modelling error can be observed depend- modelling error of a dead-time system dPðjoÞ is obtained
ing on the discretisation procedure. The following example close to the frequency p=DLmax . Thus, to capture the
illustrate this result. dynamics of the modelling error in this bandwidth, the
4s
sampling time has to be defined as
Example. Consider the process PðsÞ ¼ e1þs. The modelling
error dPðjoÞ is analysed using these models: Pn1 ðzÞ p=T s Xp=DLmax ) T s pDLmax .
obtained with Tustin’s method (TUSTIN), Pn2 ðzÞ obtained
with Euler’s forward method (FORWARD) and Pn3 ðzÞ In Torrico and Normey-Rico (2005) the effect of T s on the
obtained with the ‘‘zoh’’ method (ZOH) (see Fig. 18). As performance is analysed and a rule of thumb is given to
can be seen the modelling errors introduced by Euler’s and avoid performance deterioration with the choice of T s . For
Tustin’s discretisation are greater than the one obtained example, for DLmax 2 ½0217%, T s ¼ DLmax and for
with the ZOH method. Note that this result was expected DLmax 2 ½18245%, T s ¼ DLmax =2.
ARTICLE IN PRESS
426 J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428

5.3. Direct discrete design take place. If a discrete implementation is used the pole-
zero cancellation in:
The second approach to design discrete DTCs is based 1
on a discrete model of the continuous process. From a HðzÞ ¼ ½N m ðzÞ  N p ðzÞzd n 
Dp ðzÞ
synthesis point of view, there is no difference between the
two solutions if it is assumed that they use the same process is solved by directly dividing the polynomials ½zd N m ðzÞ 
model, including the uncertainties, and the same control N p ðzÞ and Dp ðzÞ.
structure. However, there is a fundamental difference in the
computation of the nominal model and on the model of the 6. Some design recommendations
uncertainties, this will depend on the sampling time and on
the discretisation procedure. The SP is a simple and effective controller for dead-time
processes. It can offer better responses than a PID
controller, mainly in the case where the dead time is
5.3.1. Dahlin algorithm: the lambda controller
dominant and well known. These advantages are more
The discrete dead-time compensation scheme can be also
evident when high-order models are used. In general,
derived using a direct synthesis approach which can be
the improvement of the performance compensates for the
easily interpreted as an IMC controller. A particular case is
increase in the tuning complexity. However, it has some
the popular dead-time compensation algorithm proposed
limitations that must be taken into account:
by Eric Dahlin in 1968 (Dahlin, 1968) for FOPDT models.
The algorithm is known today as the Lambda Tuning or
Dahlin controller. It was mainly used in the pulp and paper  it cannot be used with integrative and unstable
industry in the 1970s and was made possible by the processes;
increasing use of mini-computer based control systems.  the disturbance rejection response cannot be faster than
The algorithm was developed in the discrete domain but it the open-loop one. This can be important when the dead
can also be used in the continuous time domain. time is non-dominant.
The algorithm considers the plant to be a FOPDT model
and makes the closed loop behave like a FOPDT system The performance and robustness of the DTCs can be
with unit static gain. That is, if the desired close-loop is compared to the traditional PID controller. When the dead
1 time is estimated with a small error, the DTC clearly allows
given by Pd ¼ ð1lÞz
1lz1
zd and the plant is given by better performance. However, when the DTC has to be
1
PðzÞ ¼ Kð1aÞz
1az1
zd , the direct synthesis controller is detuned to maintain robustness, the advantages of the use
of this type of controller are less.
1  az1 1l Two aspects must be considered in the DTC design: (a)
CðzÞ ¼ .
Kð1  aÞ 1  lz  ð1  lÞzðdþ1Þ
1
model selection; (b) DTC structure and (b) tuning and
Note that the Dahlin controller corresponds to a SP with a practical implementation.
1
ð1lÞð1az Þ
PI primary controller C SP ¼ Kð1aÞð1z 1 Þ where l adjust the
 The selection of the nominal model depends on the
controller gain. relationship between T, the time constant representing
the dynamics of dead-time-free process (or the non-
5.4. Implementation issues integrative part of the process) and L. If the process
exhibits a dominant dead time, that is LbT, the simple
Flexibility, low cost and simplicity in the implementation FOPDT or IPDT models can be used. In other cases,
of complex control algorithms are some of the advantages when T is non-negligible, better results will be obtained
of discrete controllers. For the particular case of the DTC, using a SOPDT or SOIPDT. Unstable processes with a
the digital implementation of the controller has some dead time are not as common as the integrative one in
special advantages related to the polynomial representation practice. The simple unstable first-order model can be
of dead time. generally used in this case.
As has been pointed out in Sections 3 and 4, there are  The tuning of the controller for all cases is simpler with
some situations where is necessary to cancel the roots of the structure of the disturbance observer approach that
some polynomials and pseudo-polynomials for the im- also gives the best results. The analogous implementation
plementation of the DTC. This is particulary important to of this controller is, however, more complicated than the
avoid internal instability when controlling unstable dead- FPPI or Mataušek and Micić controllers. Thus, if the
time processes. controller must be implemented on an analogous plat-
Consider, for instance, a DTC with a modified fast form, these latter controllers can be used. Finally, when a
model controlling an unstable plant. The implementation discrete platform is used, the disturbance observer
of the minor loop must be made to have a stable HðsÞ ¼ approach is again the best option. Moreover, the distur-
G m ðsÞ  G n ðsÞeLs in order to obtain internal stability. No bance observer approach is easy to analyse and tune
‘‘pole-zero’’ cancellation at the unstable poles of G n ðsÞ must because of the decoupling properties of its structure.
ARTICLE IN PRESS
J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428 427

 The use of the DTC with a modified fast model to Chien, I.-L., Peng, S. C., & Liu, J. H. (2002). Simple control method for
improve the disturbance rejection is only recommended integrating processes with long deadtime. Journal of Process Control,
when the dead time is not dominant, because in other 12(3), 391–404.
Dahlin, E. B. (1968). Designing and tuning digital controllers. Instruments
cases the disturbance rejection is not appreciably and Control Systems, 41, 77–83.
improved. Feng, W. (1991). On practical stability of linear multivariable feedback
 The feed-forward controller should be used when the systems with time delays. Automatica, 27, 389–394.
disturbances are measurable. It is a very efficient method Garcia, C. E., & Morari, M. (1984). Internal model control 1: A unified
to minimise the effect of the disturbances on the process review and some new results. Industrial & Engineering Chemistry
Process Design and Development, 21, 308–316.
output. Guo, S. M., Wang, W., & Shieh, L. S. (2000). Discretisation of two degree-
of-freedom controller and system with state, and output delays. IEE
The analysis and design of a discrete DTC based on a Proceedings Control Theory and Applications, 147(1), 87–96.
Hagglung, T. (1996). An industrial dead-time compensating PI controller.
discrete model of the process use the same concepts and Control Engineering Practice, 4(6), 749–756.
procedures as in the continuous case, but special attention Hang, C. C., Wang, Q.-G., & Yang, X.-P. (2003). A modified Smith
must be paid to the modelling error introduced by the predictor for a process with an integrator and long dead time.
discretisation when the design procedure is obtained from Industrial & Engineering Chemistry Research, 42, 484–489.
the discretisation of a continuous controller. Huzmezan, M., Gough, W. A., Dumont, G. A., & Kovac, S. (2002). Time
delay integrating systems: A challenge for process control industries.
A safe procedure to design the controller should have, as A practical solution. Control Engineering Practice, 10(10), 1153–1161.
a first step, an analysis of the modelling error. An adequate Ingimundarson, A., & Hägglund, T. (2001). Robust tuning procedures of
bound for this error should be computed before the dead-time compensating controllers. Control Engineering Practice, 9,
controller tuning. 1195–1208.
Ingimundarson, A., & Hägglund, T. (2002). Performance comparison
between pid and dead-time compensating controllers. Journal of
7. Conclusions Process Control, 12, 887–895.
Jerome, N. F., & Ray, W. H. (1986). High performance multivariable
control strategies for systems having time delays. AIChE Journal,
DTCs are effective for controlling dead-time processes as 32(6), 914–931.
they eliminate the effect of the dead time in the nominal Kaya, I. (2003). Obtaining controller parameters for a new PI-PD
set-point response. Furthermore, a good trade-off between Smith predictor using auto tuning. Journal of Process Control, 13,
robustness and performance can be obtained by appro- 465–472.
Kaya, I. (2004). Two-degree-of-freedom IMC structure and controller
priate tuning of the primary controller and predictor.
design for integrating processes based on gain and phase-margin
Several structures and tuning rules have been analysed specifications. IEE Proceedings Control Theory and Applications,
for controlling stable, integrative and unstable processes 151(4), 401–407.
with a dead time. Some of the approaches arrive at similar Kwak, H., Whan, S., & Lee, I.-B. (2001). Modified Smith predictor for
results (and in some cases at the same result) if the same integrating processes: Comparisons and proposition. Industrial &
Engineering Chemistry Research, 40, 1500–1506.
process model and the same set of specifications are
Lee, T. H., Wang, Q. G., & Tan, K. K. (1996). Robust Smith predictor
defined. It has been shown that: (a) the DTCs originally controller for uncertain delay systems. AIChE Journal, 42(4),
designed to control stable dead-time processes can be 1033–1040.
modified to control integrative and unstable plants; (b) in Limon-Marruedo, D., Normey-Rico, J. E., Pagano, D. J., & Aracil, J.
the cases where the plant can be modelled by simple (1999). Stability of saturated dead-time compensating PI controllers
for uncertain dead-time systems. In Proceedings of ECC99. Germany.
models, a good compromise between performance and
Liu, T., Cai, Y. Z., Gu, D. Y., & Zhang, W. D. (2005). New modified
robustness can be obtained using simple controllers and Smith predictor scheme for integrating and unstable processes with
tuning rules. time delay. IEE Proceedings Control Theory and Applications, 152(2),
238–246.
Lu, X., Yang, Y.-S., Wang, Q.-G., & Zheng, W.-X. (2005). A double two-
Acknowledgement degree-of-freedom control scheme for improved control of unstable
delay processes. Journal of Process Control, 15(5), 605–614.
Mascolo, S. (2006). Modeling the internet congestion control using a smith
Partially support by CAPES-BRASIL Contract BEXO
controller with input shaping. Control Engineering Practice, 14(4),
0828/05-0 and CICYT Contract DPI 2005-4568. 425–435.
Mataušek, M. R., & Micić, A. D. (1996). A modified Smith predictor for
controlling a process with a integrator and long dead time. IEEE
References Transactions on Automatic Control, 41(8), 1199–1203.
Mataušek, M. R., & Micić, A. D. (1999). On the modified Smith predictor
Alevisakis, G., & Seborg, D. (1973). An extension of the Smith predictor for controlling a process with a integrator and long dead-time. IEEE
to multivariable linear systems containing time delays. International Transactions on Automatic Control, 44(8), 1603–1606.
Journal of Control, 3, 541–551. Morari, M., & Zafiriou, E. (1989). Robust process control. Englewood
Aström, K. J., Hang, C. C., & Lim, B. C. (1994). A new Smith predictor Cliffs, NJ: Prentice-Hall.
for controlling a process with a integrator and long dead time. IEEE Normey-Rico, J. E., Bordons, C., Berenguel, M., & Camacho, E. F.
Transactions on Automatic Control, 39(2), 343–345. (1998). A robust adaptive dead-time compensator with application to a
Bhaya, A., & Desoer, C. A. (1985). Controlling plants with delay. solar collector field. In Proceedings of IFAC-workshop on linear time
International Journal on Control, 41, 813–830. delay systems (LDTS’98) (pp. 105–110). Grenoble, France.
ARTICLE IN PRESS
428 J.E. Normey-Rico, E.F. Camacho / Control Engineering Practice 16 (2008) 407–428

Normey-Rico, J. E., Bordons, C., & Camacho, E. F. (1997). Improving Palmor, Z. J. (1996). The control handbook. Time delay compensation:
the robustness of dead-time compensating PI controllers. Control Smith predictor and its modifications. Boca Raton and London: CRC
Engineering Practice, 5(6), 801–810. Press and IEEE Press.
Normey-Rico, J. E., & Camacho, E. F. (1998). Dead-time compensators: Palmor, Z. J., & Blau, M. (1994). An auto tuner for Smith dead-time
A unified approach. In Proceedings of IFAC-LDTS’98 (pp. 141–146). compensator. International Journal of Control, 60, 117–135.
Grenoble, France. Palmor, Z. J., & Halevi, Y. (1983). On the design and proper-
Normey-Rico, J. E., & Camacho, E. F. (1999a). Robust tuning of dead- ties of multivariable dead time compensators. Automatica, 19,
time compensators for processes with an integrator and long dead 255–264.
time. IEEE Transactions on Automatic Control, 44(8), 1597–1603. Palmor, Z. J., & Halevi, Y. (1990). Robustness properties of sampled-data
Normey-Rico, J. E., & Camacho, E. F. (1999b). Robustness effects of a systems with dead time compensators. Automatica, 26, 637–640.
prefilter in Smith predictor based generalised predictive controller. IEE Palmor, Z. J., & Powers, D. V. (1985). Improved dead time compensator
Proceedings Control Theory and Applications, 146, 179–185. controllers. AIChE Journal, 31, 215–221.
Normey-Rico, J. E., & Camacho, E. F. (1999c). Smith predictor and Rao, A. S., & Chidambaram, M. (2006). Smith delay compensator for
modifications: A comparative study. In Proceedings of ECC99. multivariable non-square systems with multiple time delays. Computers
Germany. and Chemical Engineering, 30, 1243–1255.
Normey-Rico, J. E., & Camacho, E. F. (2000). Multivariable generalised Rivera, D. E., Morari, M., & Skogestad, S. (1986). Internal model control
predictive controller based on the Smith predictor. IEE Proceedings 4. PID controller design. Industrial & Engineering Chemistry Process
Control Theory and Applications, 147, 538–546. Design and Development, 25, 252–265.
Normey-Rico, J. E., & Camacho, E. F. (2002). A unified approach to Santacesaria, C., & Scattolini, R. (1993). Easy tuning of Smith predictor in
design dead-time compensators for stable and integrative processes presence of delay uncertainty. Automatica, 29, 1595–1597.
with dead-time. IEEE Transactions on Automatic Control, 47(2), Smith, O. J. M. (1957). Closed control of loops with dead time. Chemical
299–305. Engineering Progress, 53, 217–219.
Normey-Rico, J. E., & Camacho, E. F. (2007). Control of dead-time Torrico, B. C., & Normey-Rico, J. E. (2005). 2DOF discrete dead-time
processes. Berlin: Springer. compensators for stable and integrative processes with dead time.
Normey-Rico, J. E., Gómez-Ortega, J., & Camacho, E. F. (1999). A Smith Journal of Process Control, 15, 341–352.
predictor based generalized predictive controller for mobile robot path Watanabe, K., & Ito, M. (1981). A process-model control for linear
tracking. Control Engineering Practice, 7, 729–740. systems with delay. IEEE Transactions on Automatic Control, 26(6),
Ogunnaike, B. A., Lemaire, J. P., Morari, M., & Ray, W. H. (1983). 1261–1269.
Advanced multivariable control of a pilot scale distillation column. Zhong, Q.-C., & Li, H. X. (2003). Control of integral processes with dead
AIChE Journal, 29(4), 632–642. time. Part 3: Dead-beat disturbance response. IEEE Transactions on
Ogunnaike, B. A., & Ray, W. H. (1979). Multivariable controller design for Automatic Control, 48(1), 153–159.
linear systems having multiple time delays. AIChE Journal, 25, 1043–1060. Zhong, Q.-C., & Mirkin, L. (2002). Control of integral processes with
Pagano, D., Normey-Rico, J. E., & Franco, A. L. D. (2001). Stability dead time. Part 2: Quantitative analysis. IEE Proceedings Control
analysis of a modified Smith predictor for integrative plants with Theory and Applications, 149(4), 291–296.
dead-time and saturations. In Proceedings of the 40th IEEE-CDC Zhong, Q.-C., & Normey-Rico, J. E. (2002). Control of integral processes
(pp. 1855–1860). Orlando, FL. with dead time. Part 1: Disturbance observer-based 2DOF control
Palmor, Z. J. (1980). Stability properties of Smith dead time compensator scheme. Control Theory and Applications. IEE Proceedings, 149(4),
controller. International Journal of Control, 32, 937–949. 285–290.

S-ar putea să vă placă și