You are on page 1of 10

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 12, NO.

1, JANUARY 2004 123

Identification-Based PID Control Tuning for Power


Station Processes
Steve Glickman, Roland Kulessky, and Gregory Nudelman

Abstract—This paper presents a new approach for PID-con- method (see [5]–[17] and [10], [11, references therein]). The
troller tuning based on the requirements of power station DD approach reflects the industrial control system practice for
processes. It uses a novel optimization procedure based on four which the PID-control is currently most often implemented.
power process-oriented criteria. A new approach for transfer
function identification of process models is also proposed as a Actually, this approach requires a two-stage design procedure:
robustness basis for the PID-control design. The tuning of PID the first stage is process identification while the second one is
controllers implemented on Israel Electric Corporation (IEC) controller optimization using the identified model so obtained.
steam power plant processes provides sufficiently good settings for Main features of the modern DD are as follows.
these controller parameters and illustrates industrial applications As a rule, FOPTD process model is used [4], [6], [10], [12],
of this approach.
[14]. Limit cycle analysis [14], [15], or time response approxi-
Index Terms—Identification, optimization, performance limits, mation [21], [22] identifies the model parameters. The second-
PID-control, temperature and flow control. order plus time delay (SOPTD) model is identified by “the char-
acteristic area approach” [7], frequency response analysis [13]
I. INTRODUCTION or minimization of the prediction error mean square [9], [11].
Using models of higher orders [16] are not typical. For uncer-
A. Methodologies of PID-Control Tuning
tainty models, creation of a procedure of random changing of

A number of methodologies exist for PID-controller tuning


(see, for example, [1]–[24]). Three main methodologies
are dealt with here: the Internal Model Control (IMC), the Quan-
the frequency response data is proposed in [18] but it is not used
in power station practice.
Robustness characteristics are gain and phase margins [5],
titative Feedback Theory (QFT), and the “Direct Design” (DD) [9], [10], [12], [18], or sensitivity functions [11], [18]. In the
approach. realm of PID-controller optimization they are constrained by
The IMC approach [1] is essentially an open loop design in empirical formulas, not by uncertainty models analysis.
which PID-control parameters depend mainly on the nominal The best optimization criterion in the case of the FOPTD
process model which leaves the robustness filter as the only tun- process model is the minimum integral of absolute error multi-
able parameter. This filter time constant is calculated to satisfy plied by time (ITAE) [6]. The minimum integral of error squared
the system’s largest bandwidth and the performance specifica- (ISE) is applied to analytical approaches [7], [8]. The minimum
tion for the worst case of the closed loop sensitivity function integral of absolute error (IAE) is also transformed into the cri-
[2]. Because power station processes generally include a time terion for maximum PID-integrating if the control error is posi-
delay, the full IMC-controller has to contain the approximation tive [9], [11]. The procedures for performance and robustness
of transcendental functions since it is not exactly the PID-order. calculation of well-known PID formulas are proposed in de-
In other words, PID-controller is an approximation [2]. We note tail in [10], [12] for the case of the FOPTD process. Optimiza-
that the IMC approach is widely used in practice (for example, tion procedures are based on both methods of variation calculus,
[19], [21], [22]) and is based on a simple first-order plus time mathematical programming including the MATLAB optimiza-
delay (FOPTD) process model. tion toolbox, an equivalence to the Smith predictor idea [17]
The QFT approach [3] is a closed loop-shaping procedure in and exhaustive search [7]–[11]. Such characteristics as the over-
which the order of the controller depends on plant uncertainties shoot , the settling time and the rise time are applied
and performance specifications. Also PID-controller may be an usually to the time response analysis of a PID-closed loop after
approximation of the designed control. The Horowitz and IMC its optimization. The required is used as a criterion in [16].
methods may yield similar controllers [2]. Three main parameters of a PID-controller are usually opti-
The approach pioneered by [4] named here as the DD is mized: proportional gain , reset and differentiating gain .
a PID-closed loop optimization based on knowledge of the However, the fourth parameter must be also optimized: the filter
process transfer function. There have been many studies of this time constant of the differentiating element (see (1) below).
This filter is needed not only to reduce noise influence [11] but
also to prevent overactivity and mechanical impact on the final
Manuscript received March 24, 2003. Manuscript received in final form Au- control element. In the above approach, this filter is either absent
gust 28, 2003. Recommended by Associate Editor K. Schlacher. [6], [7], [9], [14], [16] or predetermined (not optimized) [10],
The authors are with the Generation and Transmission GroupThe Israel [11] by the ratio . In addition, this ratio often is limited
Electric Corporation, Ltd., Haifa 31000 Israel (e-mail: uhi10@iec.co.il;
kulesskyr@iec.co.il; nudelmang@iec.co.il). by the existing equipment of various companies (Metso/MAX-
Digital Object Identifier 10.1109/TCST.2003.821955 Control, Bailey etc.) by . In order to prevent
1063-6536/04$20.00 © 2004 IEEE
124 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 12, NO. 1, JANUARY 2004

end element overactivity, this ratio can be , as one


can see from application examples (Section VI).
Summing up the preceding, main disadvantages of the above
works are as follows:
None of the works [4]–[7], [9]–[14], [16] give a way for un-
certainty models identification, which can be used for power sta-
tion processes. Only three main parameters of a PID-controller
, , are usually optimized that sufficiently limits the degree
of freedom of PID control for power station processes (see, for
example [6], [7], [9]–[11], [14], [16]). The most of the works Fig. 1. Block diagram for the tuned control loop.
(see, for example, [4], [6], [10], [12], [14]) are oriented on the
simple FOPTD model, which can not perform well for power The Control System Tuning Program (CSTP) [33] based on
station processes. these principles was developed at IEC and has been effectively
used as the tuning tool.
B. Existing Practice of Power Station PID-Control This paper is organized as follows. The typical PID-control
There exist a number of effective computer packages, for the loop is described in Section II. The proposed identification ap-
tuning of PID-control loops (see, for example, [19]–[24]). They proach is presented in Section III. The presented PID-control
are based on the IMC approach (LALTS01 [21], WES-Tune optimization is discussed in Section IV and V. Industrial appli-
[22], INTUNE [19] etc.) and the DD approach (ExperTune [20], cation examples are given in Section VI. Some of the conclu-
Tune Wizard [23], Protuner System Analyzer [24]). Unfortu- sions can be seen in Section VII.
nately, in most power station applications PI-controllers are
usually used instead of the more suitable PID-controllers. This II. PID-CONTROL LOOP DESCRIPTION
phenomenon can be explained by the fact that specific features We consider the following typical SISO control loop structure
of power station processes and control equipment are not fully (Fig. 1). The process model and the controller are represented
taken into consideration by the existing PID control function- by transfer functions , , respectively, where four pa-
ality, namely: rameters of PID-controller, , , , , are optimized simul-
Both PI- and PID- controllers robustness has to be based on taneously according to the controller transfer function
their optimization which requires the most accurate nominal
model of a power process. However, the packages [21]–[23] (1)
are usually oriented on very approximate models, as a rule the
FOPTD model which are often inadequate for these processes. The signals in Fig. 1 are set point , process output ,
Dissimilar power station processes require different optimiza- error signal , loop disturbance , power unit disturbance
tion criteria. In spite of this, the same optimization criterion (depending on the unit mode), measurement noise and
is applied to all processes [19]–[24]. PID-controller parame- controller output .
ters are also often limited by control valve speed constraints, The disturbance is the internal signal defining the loop
which can significantly restrict strong and robust tuning param- operation mode (for example, it is the network frequency de-
eters. The PID control optimization by known tuning packages viation [30] in the control loop of a power unit load where
[17]–[24] are not fulfilled under these constraints. ). The disturbance is the external signal in-
Ignoring these requirements causes unrealistic PID-control troduced to this loop, which determines the transfer function
design and PI-control becomes more effective. for the unit steady-state mode. For example, may be
steam pressure or/and its temperature deviations for a certain
C. Novel Contributions of This Paper steady-state load.
The novel contributions are as follows: The controller acts through a control valve ,
a) A new way for model family identification of power pro- the output of which should usually be in a normalized scale
cesses and the criteria (both in time and frequency do- according to practical requirements. On this basis,
mains) for searching the nominal model and uncertainty parameters of the structure (Fig. 1) are normalized such that
models from this family. , , also belong to the same scale . Hence, the
b) Power process-oriented criteria for optimal PID-control, structure (Fig. 1) is correct under the following assumptions
including restrictions for executing element (control
(2)
valve) of power stations.
c) A new way for simultaneous calculation of set parameters (3)
, , , among them the optimal parameters of PID
control. Here, there is an upper bound on the valve speed denoted by
d) Tuning procedure allowing effective usage of the degrees and measured in units of %/s. To prevent overactivity in a
of freedom of a PID-control (its four parameters , , , control valve, (2) and (3) have to be carried out.
) for robust controller design and improvement of the Remark 1: A prefilter is usually used to reduce the effect
power station behavior. of measurement noises. Without loss of generality, such filter
GLICKMAN et al.: IDENTIFICATION-BASED PID CONTROL TUNING 125

can be included into which describes a process. Then Usually, transfer functions of power station processes are of
will be a noise after this prefilter. In addition, set point filters the third order at the most [30]. Because of this reason and as-
(SP-filters) are sometimes used for the overshoot reduction of suming that the band corresponds to the identified
the time response to . This problem however is beyond the process, several transfer functions (TF) of the first-, second-,
scope of this paper. and third-order (with and without zeros) are computed. Then
Remark 2: The scheme in Fig. 1 describes PID-control under one may expect the true TF to fall somewhere between these
(2)–(3), where a control system operates in linear regions for all computed values.
of its parameters. Of course, if only one parameter reaches its
limit a closed control system degenerates into an open control B. Model Family Identification Statement
loop.
The method [30], [31] is used for this model family identifi-
In the framework in which the proposed tuning procedure is
cation.
developed, the concept is as follows:
Let a process be described by the identification experiment
a) in the linear region, the PID-control tuning is fulfilled data set . As a rule, the closed loop experiment is
within the proposed procedure; preferable when is the step function
b) if a controller output reaches a limit, the standard anti- where . We emphasize that the process is assumed
windup control begins to operate. to contain a control valve.
The anti-windup control is not dealt with as it is beyond the Without loss of generality, expected values of , are:
scope of this paper. for and ,
for . In the case of a linear stable time-invariant process
is the recovery time. Therefore, the usage of finite is only
III. PROCESS IDENTIFICATION
the approximation in continuous systems.
A. About Identification Method Due to the noises, nonlinearities, inaccurate measurements,
small deviations of , and other factors a fixed linear time-
The transfer function identification (TFI) technique is used
invariant identified process model will never exactly represent
very widely in power system applications [25]–[31]. Of the two
the data obtained from the process. Furthermore,
main methods, ARMA TFI and Prony TFI [28], only the first
such a model can forecast the process behavior with different
one is directly related to the TFI problem.
accuracy on different parts of these data. What is meant by this
According to ARMA TFI, a pseudorandom binary sequence
is that choosing different subsets from data and ful-
(PRBS) or many uncorrelated PRBS are usually applied to the
filling identification we will obtain a family of linear time-in-
input of a system operating in closed loop (closed loop identi-
variant models. The most accurate of them may be taken, as the
fication) and a SIMO or a MIMO process is identified by this
nominal one while the others will define the process uncertain-
method using response data [25]–[27]. However, even if iden-
ties related to modeling errors.
tification of a system operating in closed loop is theoretically
Suppose that the data set is broken down into
possible, the estimates obtained may be in practice very poor,
-data subsets , , , ,
due to the effect of the MIMO feedback terms in the input signal
where: , ,
[29]. From these considerations, the individual SISO identifica-
is the expected value of the bracketed function.
tion procedure for each link transfer function (for each equation)
Applying the method [31] to the subsets
of a model may be more accurate [25], [30]. In addition, deter-
a family of linear time-invariant models may be ob-
ministic test signals can be used (see Theorem 2 in [31])
tained. For each above subset, types of models, namely
instead of the PRBS signals which are mandatory for MIMO
, are represented by their transfer
processes. This leads both to a less time-consuming identifica-
functions.
tion procedure and to a safer identification experiment.
Let this model family is calculated in the frame of the fol-
In this paper the identification method [30], [31] is used
lowing general type
without changes. Below only main principles of this method
are introduced.
The following two-stage procedure for TFI is proposed in
(4)
[31]: the first stage is frequency response (FR) identification of
a process, while the second one is TF computation using the fre-
quency response obtained from the first stage. The main reason The following five model types are suggested here to identify
for this is to provide the maximum of TFI accuracy (see The- for
orem in [30]) through the optimal limiting the bandwidth in the
frequency response identification by .
There are several methods for TF computation (see [30])
using the frequency response data in which problems of (5)
TF order determination and of parameters estimation are
solved simultaneously. The convergence is problematic in
these methods. In [30], [31] the above problems are solved
separately. where the FOPTD-model is .
126 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 12, NO. 1, JANUARY 2004

Taken into account that a SISO-process is de- D. Nominal Model Identification


scribed by the following model family Let defines the maximum identification quality in the
time domain

...............................................
(6)
...............................................
(11)

For further usage, the quality index is introduced Now we select such model set , ,
, which satisfies

(7) (12)

In our experience, is usually chosen.


where is the model time response. Clearly, Next we select the nominal model from
and the ideal identification corresponds model set , which satisfies the following
to .
(13)
The functional (7) estimates identification accuracy in the
time domain. The correct usage of an identified model in the fre-
quency domain requires calculation of a frequency bandwidth, According to (12) and (13), the nominal model describes a
in which a model correctly describes frequency properties of a process in the widest bandwidth and, simultaneously, with the
process [32]. The following theorem is applied to this problem. permissible accuracy in the time domain.

E. Additive Uncertainties Estimation


C. Theorem 1
Let is defined by
Suppose that a stable process is described by its frequency
response . Assume also that , . (14)
Next we introduce the complex number
According to (14), defines the maximum identifica-
(8) tion accuracy in the time domain by using the data subset with
number .
Now we select the model set , , which
Then the phase shift between the frequency responses of a satisfies
process and its model is equal to
(15)
(9)
The best model , identified by the data
subset with number satisfies
where

(10) (16)

Models give the closest approach to the identified


is the argument of the complex number ;
process using data subsets ,
, are Fourier transforms;

..............
(17)
..............
Proof: See Appendix I.
The expression (9) relates an error in time domain between a We note that the nominal model be-
process and its model to the error in frequency longs to
domain between a process and its model . Keeping with (17), additive uncertainties can be introduced
Let be the upper frequency bound of the model to model errors in high-frequency dynamics. Denoting additive
, which corresponds to a certain computed by (9). uncertainties as we have
Then will determine the model bandwidth upper
limit on the stage of controller optimization. (18)
GLICKMAN et al.: IDENTIFICATION-BASED PID CONTROL TUNING 127

In addition, allowable fluctuations in the steady-


state mode caused by the noises should be met

(21)

Here is the standard deviation of the bracketed function


and is its permissible upper bound.

C. Control Loop Performance Specification


Following [6], [10] we use the ITAE criterion. Then the index
quality for the controller adjusting is applied to the control loop
with the nominal process model

(22)
Fig. 2. Typical PID-controller response u(t).

where is a step response of the closed loop system to the unit


where , .
step. Obviously, and for the ideal tuning
.
IV. PID-CONTROL TUNING PROBLEM Because industrial thermal processes are not stationary ones,
A. About Control Loop Operation Mode their transfer function (parameters or/and structure) change with
time (usually, during months). Therefore phase/gain margins
The primary design goal may be to achieve good rejection may be rejected with relation to an initial tuning of the PID con-
of load disturbances [5], [11]. However, detailed assumptions troller. So we impose the following restrictions (for the nominal
can not be made about . Sometimes, PID-controllers are tuned model) for the phase margin
providing good response to a step as a possible dis-
turbance [11]. More often good response to the unit step change
(23)
in demand is the tuning goal [7], [8], [12], [13],
[16], [17], [20]–[24]. There are several reasons for this.
First, the set point step is equivalent to and the gain margin
. This disturbance spectrum is wider
than because is a low-pass link [14]. (24)
Hence is a sufficiently strong disturbance in
comparison with the unit step and is therefore a of the tuned system. Here, required lower bounds are denoted
reasonable replacement for the step. At the same time, the as , .
response to set point step must be optimized as well. Finally, These are some processes, for which such indexes as or
quality estimates such as , , are more convenient for or are more important than . Then the index quality
operating staff when applied to set point step response.
Hence, we use a good response to the step demand (25)
as the PID-control tuning goal.
can be introduced instead of (22) where is equal to either
B. Control Valve Performance Limits or or . The minimizing of is the optimization purpose
Fig. 2 shows a typical PID-controller response to a set for certain processes.
point step change. Now the following constraints have to take If nonstationary properties of a particular power station
place in accordance with Fig. 2: process are not negligible, it is reasonable to introduce the
minimum of the phase margin as the stability guarantee
for all uncertainty models of a tuned system
(19)
(26)

Usually is measured in the digital form. Then ac-


cording to (3) we have D. PID-Control Tuning Problem
PID-controller is tuned for the nominal process model
(20) through optimal adjusting all four parameters in (1). Addition-
ally, this controller has to ensure required phase margin for all
where is the sampling interval and . plant uncertainty models.
128 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 12, NO. 1, JANUARY 2004

Now the tuning problem can be formulated as an optimization 1) Lower Bound : The bandwidth lower bound can be
problem: approximately estimated using the identified FOPTD model
Find controller parameters , , , that simultaneously from the family (6)
maximize (22) or minimize (25) i.e.,
(31)
(27)
This model does not need to be the nominal one. Considering
that the open loop is optimal according to [8] one can prove (see
under restrictions (19)–(21), (23), (24) for the nominal model Appendix III) that
and the constraint (26) for , , .
Remark 3: Checking of (19), (20) requires modeling of the (32)
tuned control loop with real values of and . This is fulfilled
on the final step of the proposed tuning procedure. 2) Upper Bound : The estimate of is equal to
and limits the frequency region by the condi-
V. TUNING PROCEDURE tion of a nominal model correct usage (Theorem 1)
Before the tuning procedure will be discussed, permissible (33)
domains of controller parameters and a possible frequency re-
gion are defined.
It is important to define these domains with a safety margin, C. Permissible Domain Of
as we deal with a wide variety of power stations processes. The
above domains are estimated “approximately” however PID op- The ratio is not a constant in the presented procedure.
timal parameters inside of them are calculated accurately. We emphasize again that the upper bound is defined by
mechanical effects in a final control element, which is usually
A. Permissible Domain of Controller Parameters a control valve. These effects are indirectly prevented by limits
(19) and (20). Generally, we have
Two principal conditions (28) and (29) give the relationship
between , , , , , ( is the crossover
frequency). (34)
1) Theorem 2: Let be the crossover frequency of the
control loop (Fig. 1). Then the following condition is met for
It is clear that is a nonlinear function of and
the PID-control
which determine and in (19), (20). Therefore,
can be approximately defined only for certain operating mode
(28) parameters. For this approximate calculation we consider that
, , , . These
parameters are typical for different power stations control loop
tuning. As one proves, (see Appendix IV).
where
The lower bound of is chosen as be-
Real
cause it approximately corresponds to the extreme case of a
Imag
PI-control.
Proof: See Appendix II. D. Permissible Domain of
2) Consequence: Zeroing and in (28) we define the
corresponding condition for PI-controller The constraint (21) can be used for an approximate calcula-
tion of the permissible domain :

(29) (35)

A process which is a low-pass link [14] significantly filters


As one can see from (29), PI-controller parameters and fluctuations so effectively that they will influence control
are calculated separately. valve behavior only. Then can be approximately com-
puted in the open loop where
B. Permissible Domain Of
Usage of (28), (29) requires to define a domain of (36)

(30) Here the permissible domain of is known from (34).


Now is represented by the Parseval’s theorem and ,
where , are upper and lower bounds of . are computed according to known values , .
GLICKMAN et al.: IDENTIFICATION-BASED PID CONTROL TUNING 129

TABLE I
PARAMATERS OF PROCESS TRANSFER FUNCTIONS

E. Tuning Procedure
Suppose that , are known, values of , ,
, and are chosen and the certain function in (25) is
given. The proposed tuning procedure consists of the following
Fig. 3. Time responses of the FOPTD model G (s)-(a) and the optimal
main stages fulfilled in both frequency and time domains: modelG s ( )-(b).
a) Data acquisition from identification ex-
periment and selecting -data subsets , perheater temperature set point is broken down into 6-data sub-
. sets , . Applying the above identi-
b) Applying the method [31] to the above data and transfer fication method to the subsets the family
functions calculation of the nominal model and was calculated (see Table I). The nominal model is described
of uncertainty model family. by because it and only it satisfies (12) and (13) where
c) Calculating permissible domains of , , . and . Other
d) Assembling sets of parameters , , , for PID-con- models satisfy only (15), (16) and, therefore, describe uncer-
troller (or , for PI-controller) by calculating equation tainty models.
systems (28) or (29) in permissible domains of , For example, time responses (Fig. 3), , , illustrate
, . two identification results for where model transfer func-
e) Searching the optimal values , , , or , that tions and values of , are as follows:
satisfy the optimization criterion (27) under constraints
(24), (26) and (19)–(21). We recall that checking (19),
(20) requires to simulate the designed closed loop with (37)
real values of and .
f) Inserting the optimal parameters , , , into a tuned (38)
controller and checking control loop performance.
Remark 4: This tuning procedure is not iterative and calcu- As one can see from (38) and Fig. 3, the FOPTD-model does
lates PID optimal parameters from their permissible set through not perform well for this process.
exhaustive search. Because the permissible parameter set is as- 2) PID Controller Tuning: The initial PI-controller param-
sembled by calculating equation systems (28) or (29) in per- eters are , . Applying the tuning procedure
missible domains of , , , it is nonempty and this 5.5 to the identified process, two optimal controllers were cal-
procedure is favorable. culated: PI with parameters , and PID-con-
We note that the tuning procedure in [16] is closely related to troller with parameters , , ,
the proposed here. It includes a definition of model family and . We note that . Requirements for the
calculation of parameters sets , , . These sets can simul- optimal loop operation were as follows:
taneously place characteristic polynomial roots into the desired
a) Minimum overshoot of the control loop step response as
region in the complex plane that guarantees a specified settling
the optimization criterion.
time of the responses.
b)

VI. INDUSTRIAL APPLICATION EXAMPLES


Two industrial application examples given below illustrate
the proposed tuning procedure. These examples deal with a
Westinghouse DCS that does not limit the inserted ratio .
The superheater temperature control loop response are shown
CSTP [33] was used as the tuning tool.
in Fig. 4. Because of a relatively low level of noise and suffi-
ciently high value of , PID-controller provides the highest
A. Superheater Temperature Control on a 140 MW Unit quality.
1) Process Identification: The data set , where 3) Some Conclusions: The following conclusions follow
is the superheater steam temperature and is the desu- from this example:
130 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 12, NO. 1, JANUARY 2004

Fig. 4. Plot of time responses for superheater temperature control of 140MW Fig. 6. Plot of time responses for main air flow control of 140 MW unit.
unit.

tification. In addition, Fig. 3 also demonstrates the high


identification accuracy (time responses of process and its
model coincide practically). These practical results val-
idate the high identification accuracy which is theoreti-
cally proved in [30], [31] (see Section III-A).

B. Main Air Flow Process of 140 MW Unit


1) Process Identification: The data set is
broken down into 7-data subsets , ,
. Below we touch upon usage of the FOPTD model
only.
Time responses (Fig. 5) illustrate two identification results for
Fig. 5. Time responses of the FOPTD model G (s)-(a) and the nominal the following model transfer functions and values of ,
model G s ( )-(b).

a) The optimal PID controller sufficiently improves the


system dynamical behavior in comparison with the (39)
optimal and initial PI controllers (see Fig. 4).
b) All methods in which (for example, [6], [7], [9], (40)
[14], [16]) can result in, at best, only the above PI op-
timal control because (for the PID optimal con- As one can see from (40) and Fig. 5, the FOPTD-model de-
trol) should be equal to 2.5. The same results can be scribes this process more accurate in the time domain. However,
reached by methods [10], [11] where is predeter- this description holds within the bandwidth of 0.105 rad/s. At
mined (not optimized) or by using the equipment where the same time, the model providing the closed accuracy
. is valid within the bandwidth of 0.25 rad/s. So the model
c) Additional deterioration of dynamic behavior based on is preferred over its counterpart according to (12), (13).
methods [4], [6], [10], [12], [14]) may take place due to 2) PID Controller Tuning: The initial PI-controller has the
the FORTD model, which does not perform well for the parameters , . PID-control parameters after
superheater process [see Fig. 3(a)]. its tuning: , , , ,
d) Table I gives parameters , , , of transfer . Actually we have PI-optimal controller because
functions (4), (5). In particular, first row corresponds to . Requirements for optimal loop opera-
, second, third, fourth rows—to , , tion were as follows:
, fifth, sixth rows—to , . Differences
a) Maximum of the control loop step response as the
between parameter values belonged to the same model
optimization criterion.
type are not significant. Between different model types,
b)
parameters are distinguished sufficiently.
e) The superheater PID control is tuned under constraints
(19), (20), which take into account possible superheater
stresses. According to the proposed tuning procedure
(Section V-E) the calculation of optimal controller
parameters includes checking (19), (20). Hence, the im- The main air flow control loop responses are shown in Fig. 6.
proved dynamic behavior does not stress the superheater. Because of a relatively low level of noise PID-controller pro-
f) Identification accuracy of the process nominal model is vides sufficiently high quality. However, due to small dy-
equal to 99.17% in relation to 100% for the ideal iden- namics, improvement is less than in the previous example.
GLICKMAN et al.: IDENTIFICATION-BASED PID CONTROL TUNING 131

3) Identification Accuracy: Identification accuracy of the Next we substitute into (44) algebraic forms of complex num-
process nominal model is equal to 97.25% in relation to 100% bers
for the ideal identification. In addition, Fig. 5 illustrates the
high identification accuracy (time responses of process and its
model coincide practically). These practical results validate the
high identification accuracy which is theoretically proved in (45)
[30], [31] (see Section III-A). Equating separately real and imaginary parts in both left- and
righ-hand sides of (44) we arrive at (28).
VII. CONCLUSION
A new approach to the PID-control tuning for power station APPENDIX III
processes is presented. Its main principles are as follows: A. Appendix 3
A novel procedure for highly accurate identification of the Let be so much small that the phase shift
process model family is presented. This family contains both the so that a frequency response deviation
nominal and uncertainty models describing a process by transfer caused by the delay link in (31) is negligible. Then the
functions up to the third order with and without zeros. The ro- simple FOPTD process model (31) can be approximately
bustness parameters, gain and phase margins, are constrained by described by in this low frequency
using these uncertainty models in the tuning procedure. domain.
Four parameters of PID-controller, , , , , are opti- It is known [8], the second-order optimal closed control loop
mized simultaneously according to its transfer function (1). The is represented by the transfer function of Butterworth’s filter
optimization is fulfilled in the frame of the robust theory ap- . This loop structure is optimal from
proach using four power process-oriented criteria considering both a few deterministic and statistics optimization criteria point
also the control valve performance limits. of view [8]. Considering that we get
Main stages of the proposed tuning procedure are data acqui- which determines the lowest loop bandwidth determined
sition for process identification, calculation of model transfer by this process inertia only.
functions, calculating permissible domains of certain param-
eters, composing sets of controller parameters , , , , APPENDIX IV
searching the optimal parameters , , , , checking tuned
control loop performances. A. Appendix 4
Theoretical principles of this approach have been successfully Computing the controller (1) response to the error signal
applied for PID/PI-control loop tuning in IEC power stations. we have

APPENDIX I
A. Appendix 1 (46)
First we introduce functions Expression (46) can be used for checking (19) and (20) ful-
fillment.
(41) Zeroing in (46) and substituting and into the
(42) second inequality of (19) we find

where , are defined in comments to (10). Cal- (47)


culating where and
we have: Taking from the interval which corresponds
to the highest permissible we have
(43) (48)

Here, is from (42). Substituting (43) and , Now substituting and into (20) we arrive at
into (41) and taking into account (8) we find after conversions
the (10). Since we arrive (49)
at (9).
where the sampling period usually belong to interval s.
From (48), (49) it is follows that . Analysis of the
APPENDIX II
first inequality in (19) may reduce in certain cases. For
A. Appendix 2 calculation the (46) can be used.
According to the Nyquist stability criterion and the require-
ment (23), the following condition is held ACKNOWLEDGMENT

(44) The authors wish to thank D. Kohn and M. Bachar for their
help in planning and fulfilling this work at the Israel Electric
where . Corporation.
132 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 12, NO. 1, JANUARY 2004

REFERENCES [28] J. R. Smith, J. F. Hauer, and D. J. Trudnowski, “Transfer function iden-


tification in power system applications,” IEEE Trans. Power Syst., vol.
[1] M. Morari and E. Zafiriou, Robust Process Control. Englewood Cliffs,
8, pp. 1282–1289, Aug. 1993.
NJ: Prentice-Hall, 1988. [29] I. Gustavsson, L. Ljung, and T. Soderstrom, “Identification of processes
[2] C. G. Beril, J. T. Rockman, D. R. Lewin, and P. O. Gutman, “A com-
in closed loop—Identifiability and accuracy aspects,” Automatica, vol.
parison between the horowitz and IMC design methods for a chemical 13, pp. 59–75, 1977.
process control problem,” in ECC’91 Europ. Contr. Conf., Grenoble, [30] Y. Hain, R. Kulessky, and G. Nudelman, “Identification—Based power
France, 1991, pp. 292–297.
unit model for load-frequency control purposes,” IEEE Trans. Power
[3] I. Horowitz and M. Sidi, “Synthesis of feedback systems with large plant Syst., vol. 15, pp. 1313–1321, 2000.
ignorance for prescribed time-domain tolerances,” Int. J. Contr., vol. 2,
[31] R. Kulessky, G. Nudelman, and Y. Hain, “Thermal power plant dynamics
no. 16, pp. 287–309, 1972. identification,” in 1999 Amer. Contr. Conf., San Diego, CA, 1999, pp.
[4] J. G. Ziegler and N. B. Nichols, “Optimum settings for automatic con-
843–847.
trollers,” Trans. ASME, vol. 64, pp. 759–768, 1942. [32] R. Kulessky and G. Nudelman, “Power boiler control loops optimal
[5] K. J. Åström and T. Hägglund, “Automatic tuning of simple regulators
tuning,” in 9th Mediterranean IEEE Conf., Tel-Aviv, Israel, 1998, pp.
with specifications on phase and amplitude margins,” Automatica, vol. 484–488.
20, pp. 645–651, 1984. [33] , (2000) CSTP: User’s Guide. Israel Electric Corporation, Haifa.
[6] A. M. Lopez, J. A. Miller, C. L. Smith, and P. W. Murrill, “A comparison
[Online]. Available: http://www.ardan-pic.co.il/CSTPsite/
of controller,” Control Eng., pp. 73–76, December 1967.
[7] Y. Nishikawa, N. Sannomiya, T. Ohta, and H. Tanaka, “A method for
auto-tuning of PID control parameters,” Automatica, vol. 20, no. 1, pp.
321–332, 1984.
[8] G. C. Newton Jr, L. A. Gould, and J. F. Kaiser, Analytical Design of
Linear Feedback Controls. New York: Wiley, 1957. Steve Glickman was born in the United States in
[9] T. S. Schei, “Automatic tuning of PID controllers based on transfer func- 1947. He received the B.Sc. degree in chemistry,
tion estimation,” Automatica, vol. 30, no. 12, pp. 1983–1989, 1994. math, and physics from Stevens Institute of Tech-
[10] W. K. Ho, O. P. Gan, E. B. Tay, and E. L. Ang, “Performance and gain nology, NJ.
and phase margins of well-known PID tuning formulas,” IEEE Transac- He worked for a major manufacturer of industrial
tions on Control Systems Technology, vol. 4, no. 4, pp. 473–477, 1996. control and computer systems, Leeds & Northrup, PA
[11] H. Panagopoulos, PID Control. Design, Extension, Application. Lund: for 16 years where he was a field systems startup en-
Lund Institute of Technology, 2000. gineer, field project manager, as well as a systems
[12] A. P. Swanda and D. E. Seborg, “Controller performance assessment programmer. In 1985, he immigrated to Israel and
based on setpoint response data,” in 1999 Amer. Contr. Conf., San Diego, began working for the Israel Electric Corporation at
CA, pp. 3863–3867. the Rutenberg Power Station where he worked for 17
[13] Q. G. Wang, T. H. Lee, H. W. Fung, Q. Bi, and Y. Zhang, “PID tuning years. He was directly involved in the startup of the station’s four generating
for improved performance,” IEEE Trans. Contr. Technol., vol. 7, pp. 0
units (total capacity 2250 MW ). At Rutenberg, he worked as Manager of the
457–465, 1999. Controls Department, and later, as Plant Manager. He is presently Assistant to
[14] Z. J. Palmor, Y. Halevi, and N. Krasney, “Automatic tuning of decen- the Manager of the IEC Generating Division in charge of electrical matters, con-
tralized PID controllers for TITO processes,” Automatica, vol. 31, no. trols, and computerization.
7, pp. 1001–1010, 1995.
[15] L. Grouzman, R. Kulessky, and E. Zeheb, “Identification-based power
station models for purposes of robust control design,” in Proc. 2003
Amer. Control Conf., Denver, CO, pp. 2288–2293.
[16] J. Ackermann and D. Kaesbauer, “Design of robust PID controllers,” in Roland Kulessky was born in Russia in 1937. He re-
2001 Europ. Contr. Conf., Porto, Sept. 2001. ceived the M.S. and Ph.D. degrees, all in electrical en-
[17] S. Glickman, R. Kulessky, and G. Nudelman, “PID control design for gineering, from the Ural Politechnic Institute (UPI),
power station processes based on time—Delay compensation—Smith Russia, in 1959 and 1967, respectively. He received
predictor,” in MMAR 2002 IEEE Conf., Poland, 2002, pp. 1163–1168. the Dr.Sc degree in electrical engineering from the
[18] S. P. Bhattacharyya, H. Chapellat, and L. H. Keel, Robust Control. The Moscow Energy Institute, Russia, in 1988.
Parametric Approach. Englewood Cliffs, NJ: Prentice-Hall, 1992. Until 1991, he worked in Russia: the Ural
[19] INTUNE: Process Loop Monitoring, Tuning, & Diagnostic Software. Turbine Plant (1959–1961), Electro-Project Institute
Presented by ControlSoft, Inc., Cleveland, Ohio, USA. [Online]. Avail- (1961–1967), the UPI (1967–1991). Since 1989
able: http//www.controlsoftinc.com he has been Full Professor at the UPI. After
[20] ExperTune: PID Tuning, Analysis and Simulation. ExperTune Inc, Hu- immigrating to Israel, he joined the Control System
bertus, WI. [Online]. Available: http//www.expertune.com Department of the Israel Electric Corporation, where he has been working as
[21] Bailey infi90: Loop Tuning System LALTS01. Instruction. Cleveland, a Senior Controls Specialist. His research interests are digital control systems
OH: Controlsoft Inc, 1992. optimization, process identification, variable bandwidth control, and amplitude
[22] (1995) WES-Tune. User’s Guide, Document UO-8105. Westinghouse quantization theory.
Electric Corp., Pittsburgh, PA. [Online]. Available: http//www.westing-
housepc.com
[23] TuneWizard: PID Controller Tuning, Process Diagnostic, Loop Simu-
lation. CONTROLSERVE. [Online]. Available: http//www.tunewizard.
com
[24] Protuner: Loop Tuning, Analysis, Simulation. TECHMATION, Scotts-
dale, AZ. [Online]. Available: http//www.protuner.com Gregory Nudeman was born in Russia in 1958. He received the M.S. degree in
[25] G. Pellegrinetti and J. Bentsman, “Nonlinear control oriented boiler control system engineering from the Moscow Institute of Railway Engineering,
modeling—A benchmark problem for controller design,” IEEE Trans. Russia, in 1980. Until 1991, he had worked in Russia as a Control Field En-
Contr. Syst. Technol., vol. 4, pp. 57–63, Jan. 1996. gineer with the Start Up Company (1981–1991). There he dealt mainly with
[26] P. K. Chawdry and B. M. Hogg, “Identification of boiler models,” IEEE optimization of power unit control systems configuration and their performance
Proc., vol. 136, pp. 261–271, Sept. 1989. optimization in their main regimes. SInce 1992, he has worked in Israel as a Con-
[27] E. Swidenbank and B. M. Hogg, “Application of system identification trols Engineer with the Haifa Power Station of the Israel Electric Corporation.
techniques to modeling a turbogenerator,” IEEE Proc., vol. 136, pp. His research interests are thermal power process identification, load–frequency
113–120, May 1989. coordination control, industrial control system tuning, and autotuning problems.

You might also like