Sie sind auf Seite 1von 9

Journal of Process Control 15 (2005) 883–891

www.elsevier.com/locate/jprocont

Analysis and control of a nonlinear boiler-turbine unit


a,*,1
Wen Tan , Horacio J. Marquez b, Tongwen Chen b, Jizhen Liu a,1

a
Department of Automation, North China Electric Power University, Zhuxinzhuang, Dewai, Beijing 102206, PR China
b
Department of Electrical and Computer Engineering, University of Alberta, Edmonton, AB, Canada T6G 2V4

Received 23 September 2004; received in revised form 22 December 2004; accepted 18 March 2005

Abstract

A distance measure is proposed via the gap metric in this paper, and the concept is applied to a boiler-turbine unit to analyze its
dynamics. It is shown that the unit shows severe nonlinearity, but the nonlinearity can be avoided by careful choice of the operating
range. A single linear controller can be designed to work in such an operating range. It is also shown that the controller constraint is
another source of the nonlinearity, which can be compensated using anti-windup techniques. Simulation results are given to verify
the conclusions.
 2005 Elsevier Ltd. All rights reserved.

Keywords: Boiler-turbine unit; Nonlinearity measure; Gap metric; Anti-windup bumpless transfer techniques; Controller constraints

1. Introduction sponse to electricity demands from a power grid or


network.
There are two types of configurations in the electricity It is generally accepted that a boiler-turbine unit is a
generation using drum boilers and steam turbines: highly nonlinear and strongly coupled complex system.
However, there is no definite quantification of the com-
(1) A single boiler is used to generate steam that is plexity of a unit. Specifically, how nonlinear is it? can a
directly fed to a single turbine. This configuration linear controller be used to cover the whole operating
is usually referred to as a boiler-turbine unit. range? These are fundamental issues in the control sys-
(2) A header is used to accommodate all the steam tem design for a boiler-turbine unit. Without a thorough
produced from several boilers, and the steam is understanding of the nonlinearity, the operating range
then distributed to several turbines through the and performance of a linear controller cannot be
header. The steam can be used to generate electric- guaranteed.
ity as well as other purposes. This configuration is One way to approach this problem is to study the
commonly used in industrial utility plants. nonlinearity of a unit. The nonlinearity measure at-
tracted much attention in the past years, and several def-
Boiler-turbine units are nowadays preferred over initions and computation methods were proposed, see
header systems, because they can achieve quick re- the survey paper [1] and the references cited. Roughly
speaking, a nonlinearity measure can be regarded as
*
Corresponding author. Tel.: +86 10 8079 8549; fax: +86 10 8079 the ÔdistanceÕ between a nonlinear system and a class
8458. of feasible linear systems [2,1]. The first nonlinearity
E-mail addresses: wtan@ieee.org (W. Tan), marquez@ece.ualber- measure defined in [3] is
ta.ca (H.J. Marquez), tchen@ece.ualberta.ca (T. Chen).
1
Sponsored by Specialized Research Fund for the Doctoral v :¼ inf kN  Lk; ð1Þ
Program of Higher Education (20020079007), China. L2K

0959-1524/$ - see front matter  2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jprocont.2005.03.007
884 W. Tan et al. / Journal of Process Control 15 (2005) 883–891

where N is the nonlinear system considered, and the inf- 0 6 ui 6 1ði ¼ 1; 2; 3Þ;
imum is taken over all the linear operators L in the fea- 0.007 6 u_ 1 6 0.007;
sible set K. The norm here can be any appropriate norm, ð4Þ
2 6 u_ 2 6 0.02;
such as L2 or L1. A larger v means that the system is
ÔmoreÕ nonlinear; and in this case, a linear control may 0.05 6 u_ 3 6 0.05.
not achieve good global performance. The unit model has been widely studied in the litera-
The available nonlinearity measures have the follow- ture using various control techniques, e.g.,
ing limitations:
(1) Linear control [5–7];
(1) Almost all the examples found in the literature are (2) Gain scheduling control [8];
for SISO systems. The computation is rather diffi- (3) Nonlinear control [9];
cult for MIMO systems. (4) Intelligent control [10–14].
(2) The available nonlinearity measures are not appli-
cable for integrating and unstable systems, since Though robustness was considered in some cases, the
distances between unstable systems cannot be overall performance of the designed controllers cannot
measured in terms of standard norms. be guaranteed without a quantification of nonlinearity.
We will show in this paper that the unit does have severe
In this paper, we will propose to compute the distance nonlinearity, however, by careful choice of the operating
between a nonlinear system and a (fixed) linear system range, a single linear control can work well for the de-
based on the gap metric. The distance measure is closely sired operating range.
related to nonlinearity measures in that a nonlinearity
measure such as the one in (1) computes the distance be-
2. Gap metric and distance measure
tween a specified nonlinear system and any of the feasi-
ble linear systems. So the distance measure is the basis
Let P 1 ¼ N 1 M 1 1
1 and P 2 ¼ N 2 M 2 be normalized co-
for computing nonlinearity measures.
prime factorizations of two linear systems P1 and P2.
We will then apply the measure to analyze the nonlin-
Then the gap between the two systems can be defined
earity of a boiler-turbine unit. The dynamics of the unit
by [15,16]
is given by [4]
8 9=8 dðP 1 ; P 2 Þ ¼ maxf~
dðP 1 ; P 2 Þ; ~
dðP 2 ; P 1 Þg; ð5Þ
>
> x_ ¼ 0.0018u2 x1 þ 0.9u1  0.15u3 ;
> 1
> where ~ dðP 1 ; P 2 Þ is the directed gap and can be computed
>
> 9=8
>
> x_ 2 ¼ ð0.073u2  0.016Þx1  0.1x2 ; by
>
>
<     
x_ 3 ¼ ð141u3  ð1.1u2  0.19Þx1 Þ=85;  M1 M2 
ð2Þ dðP 1 ; P 2 Þ ¼ inf 
~  Q ð6Þ
>
> y 1 ¼ x1 ; Q2H 1  N N  .
>
> 1 2 1
>
>
>
> y 2 ¼ x2 ; For any two linear systems, the gap is bounded as
>
>
:
y 3 ¼ 0.05ð0.13073x3 þ 100acs þ qe =9  67.975Þ; 0 6 dðP 1 ; P 2 Þ 6 1. ð7Þ
where state variables x1, x2, and x3 denote drum pres- Analysis of process control systems using the gap
sure (kg/cm2), electric output (MW), and fluid density metric has been an interesting topic due to its simplicity
(kg/m3), respectively. The inputs, u1, u2, and u3 are the for control-relevant measure, see [17,18] for the first
valve positions for fuel flow, steam control, and feed- application of the gap metric in the process community.
water flow, respectively. The output y3 is the drum water The gap can be regarded as the ÔdistanceÕ between two
level (m) and acs and qe are steam quality and evapora- linear systems, and it is a generalization of the conven-
tion rate (kg/s), respectively and are given by tional distance expressed by the 1-norm. An important
feature of the gap metric is that it is applicable not only
ð1  0.001538x3 Þð0.8x1  25.6Þ
acs ¼ ; to stable systems, but also to integrating and unstable
x3 ð1.0394  0.0012304x1 Þ systems. For instant, consider the following two linear
qe ¼ ð0.854u2  0.147Þx1 þ 45.59u1  2.514u3  2.096. systems
ð3Þ 1 1
P1 ¼ ; P2 ¼ . ð8Þ
s s þ 0.1
The model is based on the basic conservation laws,
and the parameters were estimated from the data mea- The distance in the 1-norm sense is infinite, since
sured from the Synvendska Kraft AB Plant in Malmo, kP 1  P 2 k1 ¼ 1; ð9Þ
Sweden. The plant is oil-fired and the rated power is but the distance in the gap sense is finite,
160 MW. Due to actuator limitations, the control inputs
are subject to the following constraints: dðP 1 ; P 2 Þ ¼ 0.0995. ð10Þ
W. Tan et al. / Journal of Process Control 15 (2005) 883–891 885

The Bell and Astrom model in (2) contains an inte- The distance between a nonlinear system and a
grating action in the dynamics of y3—the drum level, nominal linear system can help determine the oper-
so the distance computation in the 1-norm sense pro- ating range of a linear controller designed using
vides no useful information, while the gap can be used the nominal model.
to measure the distance.
Recently, the gap metric was extended to nonlinear The last statement is based on the following result.
systems, see, for example, [19–22]. Among the various Let P be a linear system, and K be a stabilizing control-
generalizations, the differential gap proposed in [19] is ler of P. Let
relatively simple. Given two nonlinear systems N1 and   1
 I 
bP ;K :¼   .
N2, the differential gap is defined by 1
 K ðI þ PKÞ ½I P  ð15Þ
1
dd ðN 1 ; N 2 Þ :¼ maxf~
dd ðN 1 ; N 2 Þ; ~
dd ðN 2 ; N 1 Þg; ð11Þ
Then we have:
where
Proposition 1. Suppose the feedback system with the pair
~
dd ðN 1 ; N 2 Þ :¼ sup inf dðLr1 N 1 ; Lr2 N 2 Þ; ð12Þ
r2
(P, K) is stable. Let
r1
P :¼ fP D : dðP ; P D Þ < cg. ð16Þ
where Lri N i ði ¼ 1; 2Þ denote the linearization of Ni along
trajectory ri. then the feedback system with the pair (PD,K) is stable for
The gap between two nonlinear systems is closely re- all P D 2 P if and only if
lated to the nonlinearity measure. Replacing the norm in c 6 bP ; K . ð17Þ
(1) with the gap, we can define a nonlinearity measure as
the gap between a nonlinear system and a linear system. The proposition is directly obtained from Theorem
For example, with the differential gap, we can define a 17.3(a) in [16] for the case that there is no uncertainty
nonlinearity measure as in the controller K. It shows that a linear controller K
can stabilize all the linear systems that have distances
vd :¼ inf dd ðN ; LÞ ¼ inf sup dðLr N ; LÞ; ð13Þ
L2K L2K r to P less than bP,K. Thus a distance measure between
a nonlinear system and a nominal linear system can
where LrN is the linearization of N along trajectory r.
determine the operating range of a linear controller de-
In the sequel, we will consider the following distance
signed based on the nominal model. In contrast, a non-
measure
linearity measure such as vd can help determine whether
vg :¼ sup dðLr0 N ; LÞ; ð14Þ a linear controller is enough for the whole operating
r0 range, and if not, where the operating points should
where Lr0 N is the linearization of N at the operating be selected.
point r0. The difference between LrN and Lr0 N is that Obviously vd and vg require that the nonlinear system
the linearization along a trajectory is usually time vary- considered be differentiable. So they are not applicable
ing, while at an operating point it is time-invariant. to the quantification of nondifferentiable nonlinearities
While vd is more appropriate for quantifying the nonlin- such as saturation. The gap defined in [20] might be used
earity of a system and vg only reflects the nonlinear instead for such a purpose. However, the computation
dynamics near an operating point, there are several of such a gap is quite difficult and we will not discuss
advantages in using vg: it here.
We often need to add weights in controller design. In
(1) The computation of vg is simple. vd involves the this case, we can define the weighted gap d(W2P1W1,
gap between a linear time-varying system LrN W2P2W1), and the nonlinearity measure can be defined
and a linear time-invariant system L. There is no by
efficient method to compute it. On the contrary, vgw :¼ inf sup dðW 2 ðLr0 N ÞW 1 ; W 2 LW 1 Þ; ð18Þ
L2K r0
the gap between two linear systems can be easily
computed [23]. where W1 and W2 are some suitable weights determining
(2) In practical industrial processes, normal operation the desired performance of the closed loop as in loop-
is confined to the neighborhood of the equilibrium shaping H1 design [24]. The weights can be chosen fol-
points; so vg is a reflection of the nonlinear dynam- lowing the guidelines discussed in [25]:
ics due to the operating point change, as long as
the change is slow. (1) W2 is usually chosen as a constant, reflecting the
(3) Linear controller design is usually based on a nom- relative importance of the outputs to be con-
inal linear model, either from linearization of a trolled. It is often set to an identity matrix if the
nonlinear model or identification from real data. model is well scaled.
886 W. Tan et al. / Journal of Process Control 15 (2005) 883–891
2 3
(2) In general W1 has the form W1 = WpWaWg. 1 0 0
• Wp contains dynamic shaping. Integral action for 6 7
C¼4 0 1 0 5;
low frequency performance; phase-advance for
0.0063 0 0.0047
reducing the roll-off rates at crossover; and 2 3
phase-lag to increase the roll-off rates at high fre- 0 0 0
6 7
quencies should all be placed in Wp if desired. D¼4 0 0 0 5. ð20Þ
• Wa is a constant that aligns the singular values at 0.253 0.512 0.014
a desired bandwidth (optional). This is effectively
a constant decoupler and should not be used if To analyze the nonlinearity of the unit, we will com-
the plant is ill-conditioned. pute the measure vg. Since the system contains three
• Wg is an additional gain matrix to provide con- states, and the combination of the three states will result
trol over actuator usage (optional). in many operating points, the computation of vg is com-
plicated. For simplicity, we will fix L as the linearized
It should be noted that the weights can affect the model at operating point #4, and we will compute for
numerical value of the gap between two systems, which the following three cases:
is the main drawback of the gap metric. However, our
experience with the gap metric shows that the gap com- (1) Drum pressure y1 = 108, electric output y2 varies
puted with reasonable weights is in the same scale as the from 20 to 160, and drum level y3 varies from
gap computed without weights, so weights will not affect 0.5 to 0.5. vg is shown in Fig. 1.
the degree of the nonlinearity of a system. (2) Drum pressure y1 varies from 60 to 160, electric
output y2 = 66.65, and drum level y3 varies from
0.5 to 0.5. vg is shown in Fig. 2.
3. Analysis of nonlinear dynamics (3) Drum pressure y1 varies from 60 to 160, electric
output y2 varies from 20 to 160, and drum level
Some typical operating points of the Bell and Astrom y3 = 0. vg is shown in Fig. 3.
model (2) are shown in Table 1.
The linear control design for the unit found in the lit-
erature usually takes the linearized model at opera-
ting point #4 as the nominal model. The linearized 0.7

model is 0.6

( 0.5
d_x ¼ Adx þ B du
P 0 :¼ ð19Þ 0.4

dy ¼ Cdx þ D du
vg

0.3

0.2
with
2 3 0.1

0.0025 0 0 0
6 7 –0.5
A ¼ 4 0.0694 0.1 0 5;
0.0067 0 0 0
2 3 Drume Level (m) 140 160
0.9 0.349 0.15 80 100 120
6 7 0.5 40 60
B¼4 0 14.155 0 5; 20
Electric Output (MW)

0 1.398 1.659 Fig. 1. Distance measure vg: case 1.

Table 1
Typical operating points of Bell and Astrom model
#1 #2 #3 #4 #5 #6 #7
x01 75.60 86.40 97.20 108 118.8 129.6 140.4
x02 15.27 36.65 50.52 66.65 85.06 105.8 128.9
x03 299.6 342.4 385.2 428 470.8 513.6 556.4
u01 0.156 0.209 0.271 0.34 0.418 0.505 0.6
u02 0.483 0.552 0.621 0.69 0.759 0.828 0.897
u03 0.183 0.256 0.340 0.433 0.543 0.663 0.793
y 03 0.97 0.65 0.32 0 0.32 0.64 0.98
W. Tan et al. / Journal of Process Control 15 (2005) 883–891 887

(1) For fixed pressure operation (Fig. 1), there is a


0.9 maximum electric output. The linearized models
0.8 at the region with large drum levels have a small
0.7
distance to the nominal model.
0.6
0.5
(2) Different combination of drum pressure and drum
level can result in the same electric outputs (Fig.
vg

0.4
0.3 2). The linearized models at the region with small
0.2
drum pressures and large drum levels have a small
0.1
0.0
distance to the nominal model.
–0.5 (3) For fixed drum level operation (Fig. 3), only the
linearized models at the region with small electric
0 outputs have a small distance to the nominal
160
Drume Level (m) 120
140 model.
100
0.5
60
80 (4) Larger values of vg occur at the region with large
Steam Pressure (kg/cm2)
drum pressures, large electric outputs, and small
Fig. 2. Distance measure vg: case 2. drum levels; The maximum is about 0.9 (Fig. 2).

To evaluate the nonlinearity of the unit, we need to


compute vg by taking L as linearized models at other
operating points. It is found that the maximum of vg is
always larger than 0.7, so we conclude that the Bell
0.8 and Astrom model in (2) shows severe nonlinearity.
0.6
The plots of vg show us how to ÔavoidÕ the nonlinear
dynamics due to the operating point change. Since a boi-
0.4
vg

ler-turbine unit must follow the electricity demand from


0.2 the grid, and the same amount of electric output can be
0.0 obtained with different combination of drum pressure
160
140
120
and drum level, the operating range of the unit should
100
120
be carefully chosen. For instant, if a linear controller
80
100 60
is designed at operating point #4, then we should avoid
Steam Pressure (kg/cm2) 80 40 Electric Output (MW)
driving the unit to operating points that have large drum
60 20 pressures, large electric outputs, and small drum levels,
Fig. 3. Distance measure vg: case 3.
since at these operating points the dynamics are quite
different from the nominal one. A good operation is to
increase the drum pressure and drum level as the electric
Note that the combination of y1, y2, and y3 can result output is increased. In this case the dynamics at the
in impractical operating points, e.g., u2 > 1. At these operating region will not be far away from the nominal,
operating points the corresponding gaps to the nominal thus stability of the system can be guaranteed. The typ-
point cannot be computed, and there are no correspond- ical operating points shown in Table 2 are those that sat-
ing values in the figures shown. For example, in Fig. 3, isfy the good operation requirements.
the operating point ðy 01 ¼ 80; y 02 ¼ 100; y 03 ¼ 0Þ is impos- To verify the argument, we compute the gaps be-
sible. In fact at this point u02 ¼ 1.209, out of the scope of tween the linearized models at the typical operating
the controller operating range. points (Table 2). It is clear that the gaps between them
From Figs. 1–3 we arrive at the following conclu- are small, the largest being between models at operating
sions: point #1 and #7. If the nominal model is taken at #4,

Table 2
Gaps between linearized models at typical operating points
#1 #2 #3 #4 #5 #6 #7
#1 0 0.074 0.139 0.195 0.245 0.289 0.329
#2 0.074 0 0.066 0.123 0.174 0.219 0.261
#3 0.139 0.066 0 0.058 0.110 0.157 0.199
#4 0.195 0.123 0.058 0 0.053 0.10 0.142
#5 0.245 0.174 0.110 0.053 0 0.048 0.091
#6 0.289 0.219 0.157 0.10 0.048 0 0.043
#7 0.329 0.261 0.199 0.142 0.091 0.043 0
888 W. Tan et al. / Journal of Process Control 15 (2005) 883–891

then by proposition 1, as long as the designed linear con- constraints at the design stage, and the effects of the con-
troller K satisfies bP,K > 0.195, then the controller can straints are compensated after the controller design.
guarantee the closed-loop stability at other operating Such techniques are called anti-windup bumpless trans-
points. fer (AWBT) in the literature. For a single-loop PI con-
troller, the technique is known as anti-reset, and one
of the possible AWBT configurations is shown in Fig.
4. Linear controller design and controller constraints 4, which was proposed in [27]. Here the PI controller
has the form KðsÞ ¼ kð1 þ s1I sÞ, and the compensation
As shown in the previous section, the Bell and As- parameter sr can be chosen equal to k for a good com-
trom model in (2) shows severe nonlinearity, however, pensation effect.
careful choice of operation can avoid the nonlinearity. The AWBT technique for a general controller in the
Thus a linear controller is enough for the unit in the state space form is more difficult to implement than a
operating range. PI controller. Considering the practical implementation
To show this, we will design a linear controller at oper- issue, we simplify the final loop-shaping H1 controller
ating point #4 via the loop shaping H1 approach [24] as using the PID reduction procedure proposed in [7].
in [7]. We choose the pre- and post-compensators as After eliminating some small terms, we finally get the
2 3 following linear controller:
5 þ 1s 0 0 2 3
6
W 1 ¼ W a4 0 1 þ 1s
7
0 5; W 2 ¼ I; ð21Þ 0.0485 þ 0.0012
s
0 1.2091 þ 0.0486
s
6 7
0 0 5 þ 5s KðsÞ ¼ 4 0 0.0197 þ 0.0045
s
0 5.
0 0 7.2548 þ 0.2914
s
where Wa is a constant that aligns the singular values of
the model at 0.001 rad/s. It is effectively a constant ð23Þ
decoupler, and given by The singular value plots of the original H1 controller
2 3
0.001095 0.00373 0.02136 and the reduced PI controller are shown in Fig. 5. It can
6 7 be shown that they are close at the low frequencies but
W a ¼ 4 0.004273 0.007065 0 5. ð22Þ
different at the high frequencies, so the PI controller will
0.000374 0.00595 0.1281 have similar performance as the H1 controller, but the
Since we are using weights, we need to compute vgw robustness will degrade. In fact, bW 2 P 0 W 1 ;K for the PI con-
for the model to proceed. The procedure and the plots troller is 0.27, smaller than that for the H1 controller.
are similar to those in the previous section, hence omit- However, the PI controller can still guarantee the stabil-
ted here for brevity. For example, the weighted gaps be- ity of the system at other operating points, as compared
tween the linearized models at the typical operating with the values in Table 3.
points and the nominal point are shown in Table 3. It While the compensation is effective, methods that can
is in the same scale as the corresponding row in Table directly consider the constraints are desired. Model pre-
2. bW 2 P 0 W 1 ; K for the final H1 controller is 0.42, so by dictive control [28] can be an effective one in dealing
Proposition 1, it can guarantee the stability of the sys- with the constraints. However, they are out of the scope
tem at other operating points. of the paper.
Despite the nonlinear dynamics due to the operating
point change, there is another source of nonlinearity in
K(s) d
the unit—the controller constraints in (4). It is well-
known that the constraints have great impacts on the r y
k G(s)
performance of the closed-loop systems, e.g., causing _ +
large overshoot and even instability [26]. The constraint _
k 1
effects are not considered in [7]. Here we slightly modify τI _ s
the pre-compensator to take the effect of the constraints
1
into consideration. τr
The nonlinearity measure given by vd or vg cannot
deal with such constraints, since they are not differentia-
ble. In fact, most of the linear design methods ignore the Fig. 4. Anti-windup technique for a PI controller.

Table 3
Weighted gaps between linearized models at typical operating points and the nominal point
#1 #2 #3 #4 #5 #6 #7
#4 0.199 0.125 0.0596 0 0.0538 0.103 0.147
W. Tan et al. / Journal of Process Control 15 (2005) 883–891 889

Singular Values 125


80

y1 (kg/cm2)
120

115
60
110

40 105
0 100 200 300 400 500 600
Singular Values (dB)

140
20
120

y2 (MW)
100
0
80

60
–20 0 100 200 300 400 500 600
0.6
–40 0.4

y3 (m)
0.2
–60 0

–0.2
0 100 200 300 400 500 600
–80 Time (sec.)
10–4 10–3 10–2 10–1 100 101 102
Frequency (rad/sec)
Fig. 6. Responses of system from nominal operating point to a ÔnearÕ
Fig. 5. Singular value plots of the H1 controller and the reduced PI operating point.
controller (solid lines: H1 controller; dashed lines: PI controller).

160
5. Simulation results
y1 (kg/cm2)

140

120
To show the effects of the nonlinear dynamics and the
controller constraints on the controlled system, we give 100
0 100 200 300 400 500 600
some simulations in this section. We consider the follow- 100
ing cases: 90
y2 (MW)

80

(1) Nonlinear dynamics. We first show the effects of 70

the nonlinear dynamics on the controlled system, 60


0 100 200 300 400 500 600
without the controller constraints in (4). 200
(a) From the nominal point to another operating 100
y3 (m)

point that is ÔcloseÕ. At t = 100, drum pressure


0
y1 is increased from 108 to 120, electric output
y2 increased from 66.65 to 120, and drum level –100
0 100 200 300 400 500 600
y3 kept at 0. The weighted gap between the Time (sec.)

linearized model at the final operating point Fig. 7. Responses of system from nominal operating point to a ÔfarÕ
and the nominal model is 0.154. The system operating point.
responses are shown in Fig. 6. It is clear that
the system can enter the final operating point
in a stable fashion due to a close distance. (c) If one wants to increase the drum pressure
(b) From the nominal point to another operating from 108 to 150 without causing instability,
point that is ÔfarÕ. At t = 100, drum pressure one method is to increase the drum level at
y1 is increased from 108 to 150, electric output the same time. For instant, at t = 100 increas-
y2 increased from 66.65 to 90, and drum level ing y3 from 0 to 0.5, then the weighted gap
y3 kept at 0. The weighted gap between the between the linearized models at the final oper-
linearized model at the final operating point ating point and the nominal model is only
and the nominal model is 0.727, larger than 0.183, so the system can enter the final operat-
bW 2 P 0 W 1 ;K . The system responses are shown in ing point in a stable fashion, see Fig. 8.
Fig. 7. It can be observed that drum pressure (2) Controller constraints. To show the effect of the
and electric output are stable while drum level controller constraints on the controlled system,
is obviously out of the operating range and thus we consider the case (a) again, with the controller
can be regarded as unstable. The reason is that constraints in (4) in effect. The system responses
the distance between the linearized models at are shown in Fig. 9, where the solid lines are
the final and the nominal operating points is responses with AWBT compensation, and the
large. dashed lines are those without compensation.
890 W. Tan et al. / Journal of Process Control 15 (2005) 883–891

160 technique as discussed in the previous section. It


is clear that the controller constraints cause larger
y1 (kg/cm2)

140

120
overshoot and larger settling time without com-
pensation. The controller with AWBT compensa-
100
0 100 200 300 400 500 600 tion achieves good performance.
100 (3) A large operating point change. To show that the
90 linear controller can operate well in the careful
y2 (MW)

80
chosen operating range, we consider the operating
70
point change from #1 to #7 at t = 100. Drum
60
0 100 200 300 400 500 600 pressure increases from 75.6 to 140, electric output
1 from 15.3 to 128, and drum level from 0.97 to
0.5 0.98. The changes are large. However, since the
y3 (m)

0
weighted gap between the linearized models at
the two operating points is 0.335, it can be
–0.5
0 100 200 300 400 500 600 expected that a well designed linear controller
Time (sec.)
can fulfill the task. The system responses for the
Fig. 8. Responses of system from nominal operating point to another designed linear controller with AWBT compensa-
operating point. tion are shown in Fig. 10.

130
y1 (kg/cm2)

120

110
200
100
0 100 200 300 400 500 600 700 800 900 1000 150
y1 (kg/cm2)

140
100
120
y2 (MW)

100 50
0 100 200 300 400 500 600 700 800 900 1000
80 150
60
0 100 200 300 400 500 600 700 800 900 1000 100
y2 (MW)

0.2
50
0.1
y3 (m)

0 0
0 100 200 300 400 500 600 700 800 900 1000
–0.1 1
–0.2 0.5
0 100 200 300 400 500 600 700 800 900 1000
a 0
y3 (m)

Time (sec.)
–0.5
1
1
0.8 0 100 200 300 400 500 600 700 800 900 1000
a Time (sec.)
0.6
u1

0.4
1
0.2
0 100 200 300 400 500 600 700 800 900 1000
u1

1 0.5
0.9

0.8 0
u2

0 100 200 300 400 500 600 700 800 900 1000
0.7
1
0.6
0 100 200 300 400 500 600 700 800 900 1000 0.8
u2

1 0.6

0.4

0.5 0.2
u3

0 100 200 300 400 500 600 700 800 900 1000
1
0
0 100 200 300 400 500 600 700 800 900 1000
b Time (sec.)
u3

0.5

Fig. 9. Responses of system: with vs. without anti-windup.


0
b 0 100 200 300 400 500 600 700 800 900 1000
Time (sec.)
Since the final controller consists of 4 SISO PI
controllers, it is easy to implement the AWBT Fig. 10. Responses of system from operating point #1 to #7.
W. Tan et al. / Journal of Process Control 15 (2005) 883–891 891

6. Conclusions [10] R. Dimeo, K.Y. Lee, Boiler-turbine control system design using a
genetic algorithm, IEEE Trans. Energy Convers. 10 (4) (1995)
752–759.
The nonlinearity of a boiler-turbine unit was ana- [11] A.B. Abdennour, K.Y. Lee, An autonomous control system for
lyzed in this paper. It was shown that the unit exhibits boiler-turbine units, IEEE Trans. Energy Convers. 11 (2) (1996)
severe nonlinearities, and choice of a good operating 401–406.
range can help avoid the nonlinearity. A linear control- [12] F.A. Alturki, A.B. Abdennour, Neuro-fuzzy control of a steam
ler was designed via loop-shaping H1 approach, and boiler-turbine unit, in: Proc. 1999 IEEE Int. Conf. Control Appl.,
Hawaii, 1999, pp. 1050–1055.
with anti-windup compensation the single linear con- [13] A.B. Abdennour, An intelligent supervisory system for drum type
troller works well in the desired operating range. It boilers during severe disturbance, Electrical Power Engergy Syst.
should be noted that the analysis is for the specified boi- 22 (2000) 381–387.
ler-turbine unit. General conclusions on the operation [14] U.-C. Moon, K.Y. Lee, A boiler-turbine system control using a
and control of a boiler-turbine unit should be made with fuzzy auto-regressive moving average (FARMA) model, IEEE
Trans. Energy Convers. 18 (1) (2003) 142–148.
caution. [15] T.T. Georgiou, M.C. Smith, Optimal robustness in the gap
metric, IEEE Trans. Automat. Contr. 35 (1990) 673–686.
[16] K. Zhou, J.C. Doyle, Essentials of Robust Control, Prentice Hall,
1998.
References [17] Y. Samyudia, M. Green, P.L. Lee, I.T. Cameron, A new approach
to decentralized control design, Chem. Eng. Sci. 50 (1995) 1695–
[1] M. Nikolaou, P. Misra, Linear control of nonlinear processes: 1706.
recent developments and future directions, Comput. Chem. Eng. [18] P.L. Lee, H. Li, I.T. Cameron, Decentralized control design for
27 (2003) 1043–1059. nonlinear multi-unit plants: a gap metric approach, Chem. Eng.
[2] M. Nikolaou, When is nonlinear dynamic modeling necessary? in: Sci. 55 (2000) 3743–3758.
Proc. of American Control Conf., San Francisco, 1993, pp. 1460– [19] T.T. Georgiou, Differential stability and robust control of
1464. nonlinear systems, Math. Control Signals Syst. 10 (3) (1993)
[3] C.A. Desoer, Y.T. Wang, Foundations of feedback theory for 289–306.
nonlinear dynamical systems, IEEE Trans. Circ. Syst. 27 (2) [20] T.T. Georgiou, M.C. Smith, Robustness analysis of nonlinear
(1980) 104–123. feedback systems: an input–output approach, IEEE Trans.
[4] R.D. Bell, K.J. Astrom, Dynamic models for boiler-turbine- Automat. Contr. 43 (9) (1997) 1200–1221.
alternator units: data logs and parameter estimation for a 160 [21] M.R. James, M.C. Smith, G. Vinnicombe, Gap metrics, repre-
MW unit, Tech. Rep. Report LUTFD2/(TFRT-3192)/1-137, sentations, and nonlinear robust stability, in: Proc. 39th IEEE
Department of Automatic Control, Lund Institute of Technology, Conf. Decision Contr., Sydney, Australia, 2000, pp. 2936–2941.
Lund, Sweden, 1987. [22] W. Brian, M. French, Coprime factorization and gap metric for
[5] W.H. Kwon, S.W. Kim, P.G. Park, On the multivariable robust nonlinear systems, in: Proc. 42nd IEEE Conf. Decision Contr.,
control of a boiler-turbine system, in: IFAC Symposium on Power Maui, Hawaii USA, 2003, pp. 4694–4699.
Systems and Power Plant Control, Seoul, Korea, 1989, pp. 219– [23] T.T. Georgiou, On the computation of the gap metric, Syst.
223. Contr. Lett. 11 (1988) 253–257.
[6] C.-S. Hwang, D.-W. Kim, A design of robust two-degree of [24] D.C. McFarlane, K. Glover, Robust Controller Design Using
frwdom boiler-turbine control system using H1 optimization Normalized Coprime Factorization Description, Springer-Verlag,
method, in: Proc. SICEÕ95, Sappero, 1995, pp. 1263–1268. 1990.
[7] W. Tan, Y.G. Niu, J.Z. Liu, H1 control for a boiler-turbine unit, [25] S. Skogestad, I. Postlethwaite, Multivariable Feedback Control:
in: Proc. IEEE Conf. on Control Applications, Hawaii, USA, Analysis and Design, Wiley, 1996.
1999, pp. 807–810. [26] P.J. Campo, M. Morari, Robust control of processes subject to
[8] P.C. Chen, J.S. Shamma, Gain-scheduled l1-optimal control for saturation nonlinearities, Comput. Chem. Eng. 14 (4/5) (1990)
boiler-turbine dynamics with actuator saturation, J. Process 343–358.
Control 14 (2004) 263–277. [27] H.A. Fertik, C.W. Ross, Direct digital control algorithm with
[9] F. Fang, J. Liu, W. Tan, Nonlinear internal model control for anti-windup feature, ISA Trans. 6 (4) (1967) 317–328.
boiler-turbine coordinated systems of power unit (in chinese), [28] J.M. Maciejowski, Predictive Control with Constraints, Prentice
Proc. CSEE 24 (4) (2004) 195–199. Hall, Harlow, England, 2002.

Das könnte Ihnen auch gefallen