Sei sulla pagina 1di 22

1

CHAPTER 6: Decentralized Control (Multi-loop control)


Introduction The first step of designing the decentralized is to determine the control structure and control configuration.
Decentralized control is always conducted system, i.e. number of MV equal number of CV. on square

The process consists of N MV's and M CV's, therefore, there are three cases: If N = M (Square system), then control loop configuration, i.e. input output pairing, should be determined. There are several possible configurations of control loops. The number of different loop configuration increase rapidly with N: For N = 3 For N = 4 For N = 5 we have 3! = 6 different loop configurations we have 4! = 24 different loop configurations we have 5! = 120 different loop configurations

If N > M, therefore, we need to extract the best M MV's to be used with the M CV's. This is called control structure design. Having determined the best structure, we need to go back to step 1 and determine the loop configuration. The remaining r = N M inputs can be used in split-range or left for emergencies. If N < M, then there are r = M N control variables can not be controlled. In this case, the r controlled variables that have the lowest priority should be taken out of the control objective list or controlled through override scheme. For the remaining N CV's, the loop configuration should be determined, i.e. step 1. Non-square RGA is useful to role out some outputs.

1. Control loop configuration Once all MV's and CV's are determined, we need to decide how they are going to be interconnected through control loops. This means what output measurement will actuate a given MV, or what MV will be used to regulate a given output measurement. There are large numbers of loop pairing, choosing the "best" configuration is a critical task. Various criteria can be used to select the best pairing: Use plant experience and physical reasoning, qualitative method Use a quantitative method The two common quantitative methods are: RGA method, which determine the control configuration that yield control loops with minimum interaction. SVD method 1.1 Loop pairing using the RGA Now we will consider how the RGA may be used as a guide for selection of input/output pairs that lead to minimum interaction among control loops. The interpretation of the values of the RGA can be classified according to the following categories: 1. ij = 1, indicates that open loop gain between yi and uj is identical to the closed-loop gain. Loop i will not be subject to retaliatory actions from other control loops when they are closed. Thus, uj can control yi without interference from other control loop. Pairing recommendation: Pairing yi and u will therefore be ideal. 2. ij = 0, indicates that open-loop gain between yi and uj is zero. This means uj has no direct influence on yi. Pairing recommendation: Do not pair yi with uj. 3. 0 < ij < 1, indicating the open-loop gain between yi and uj is smaller than the closed-loop gain. Since the closed-loop gain is the sum of the open-loop gain and the retaliatory effect from the other loops, the loops are definitely interacting. Pairing recommendation: if possible avoid pairing yi with uj whenever ij = 0.5.

4. ij > 1, indicating that the open-loop gain between yi and uj is larger than the closed-loop gain. The loops interact, and the retaliatory effect from the other loops acts in opposition to the main effect of u j and yi. Pairing recommendation: where possible, do not pair yi with uj if ij takes a very high value, e.g. >25. 5. ij < 0, indicating that open-loop and closed-loop gains between yi and uj have opposite signs. The loops interact, and the retaliatory effect from the other loops is not only in opposition to the main effect, but also the more dominant of the two effects. Pairing recommendation: avoid pairing yi with uj. The foregoing discussion leads to the following rule: RGA RULE A: pair input and output variables that have positive RGA elements and closets to one.
NIEDERLINSKY INDEX Even though pairing Rule A is usually sufficient in most cases; it does not consider the stability of the resulting control structure. Therefore, it is necessary to check the stability of the resulted control structure. This can be according to the Niederlinsky theorem. Consider the n n multivariable system whose input and output variables have been paired as follows: y1 m1 , y 2 m2 ,K , y n mn , resulting in a transfer function model of the form:

y = Gu
In this model, each element of G, gii, is rational and open loop stable. Furthermore, assume there are no individual feedback controllers with integral action and each controller is stable when the other n-1 loops are open. When all loops are closed, the system will be unstable for all possible values of controller parameters (Structurally monatomic unstable), if the Niederlinsky index, N defined in the following equation is negative.

G K = N = n n g ii K ii i =1 SS i =1

RULE B: any pairing is unacceptable if it leads to a control system for which the Niederlinsky index is negative.

1.1.2 Examples Example 1: Consider the mixer process, which can be modeled at steady state as follows:
F = F1 + F2

x=

F1 + F2 F

F1 , x 1 F2, x2

Mixer
Figure 1: Mixer example

F, x

The outputs are F, x The inputs are F1, F2 Create the steady state transfer function (i.e., linearize):

F =1 F1 F =1 F2
F x (1 x ) F (1 x ) = 2 = = 2 F1 F 2 F F

x F1 xF x = = 2 = F2 F 2 F F

Therefore:

1 F x = 1 x F

1 F1 x F2 F

The operating condition is F = 200 mole/h, x = 0.6

1000 1000 K = 3 2
0.6 0.4 = 0.4 0.6
The RGA recommend pairing F with F1 and x with F2.
FC FT

F1, x1 F2, x2

Mixer
CC CT

F, x

Figure 2: Mixer under feedback control

Because all relative gains are close to 0.5, the control loop interaction will be serious. Example 2: The relative gain for a 4X4 refinery distillation column is given as follows:

0.150 0.08 0.164 0.931 0.011 0.429 0.286 1.154 = 0.135 3.314 0.27 1.19 2.03 0.90 1.919 0.215 The recommended pairing is, y1-u1, y2-u4, y3-u2, y4-u3

Example 3: Consider the following system: 5 3 K = 1 1 1 1 3 1 1 1 1 3

10 4.5 4.5 = 1 4.5 4.5 1 4.5 4.5 The recommended pairing is 1-1/2-2/3-3. According to the Niederlisnky rule: K = det( K ) = 0.145

K ii = ( 3 )( 3 )( 3 ) = 27
i =1

5 1 1

Therefore;
N= K <0

K ii
i =1

This leads to unstable configuration. As a remedy, change the pairing to 1-1/2-3/3-2, gives: 10 4.5 4.5 = 1 4.5 4.5 1 4.5 4.5

Which corresponds to: 5 3 K = 1 1 1 1 3 1

1 1 1 3

Computing the Niederlinski index gives:


N = 4/45
stable looping

1.1.3 Shortcomings The above RGA pairing method ignores process dynamics. It has been shown that if the transfer function has very large time delay or time constant relative to the others, steady state RGA analysis provide an incorrect recommendation.
Example: Consider the following transfer function:

2e s 1.5e s y1 10s + 1 s + 1 u1 y = s u 2 2 1.5e s 2e s + 1 10s + 1

2 1.5 K = 1.5 2 0.64 0.36 = 0.36 0.64


The recommended pairing is 1-1/2-2. However, the off-diagonal elements indicates that y1 responds ten times faster to u2 than u1 because their relative time constant.

Computer simulation indicates that the opposite pairing is better performance. Show SIMLNK simulation

1.2 Singular value decomposition 1.2.1 Definition: The singular value decomposition of a matrix K results in three component matrices as follows:
K = U V T

where
K: is an nxn matrix U: is an nxn orhtonormal matrix, with its column is called left singular vector V: is an mxm orhtonormal matrix, with its column is called right singular vector : is nxm diagonal matrix of scalars called the singular values that are organized in decending order.

1.2.2 Physical interpretation: K is the steady state gain matrix, contains the sensitivity of each measured variable (sensor) to change in the manipulated variable. It is very important that elements of K be scaled.
K=

y u

A good physical scaling should give:

K=
U

%sensor span %range of MV

= U1:U2: Un provides the most appropriate coordinate for viewing the process sensor. The first column indicates the easiest sensor direction in which the system can be changed by the MV. = V1:V2: Vm provides the most appropriate coordinate for viewing the MV. The first column of VT indicates the combination of control action that has the most effect on the system. = diag(1, 2, m) provide ideal decoupled gain of the open-loop process. The ratio of the maximum singular value to the minimum singular value (max/min) is the condition number.

10

The condition number is a measure of the difficulty of the decoupled multivariable control problem. Large condition number indicates that it is difficult or impossible to accomplish all the control objectives.

1.2.3 Example Consider the mixer of two different temperature streams:


F1 , Th
Hot water

FT

TT

F2 , Tc
Cold water Figure 3: Mixing thermal streams

The linearized model is given as follows:


Tm F1 F = K F 2 m

For an operating condition of F1 = 10 gpm, F2 = 20 gpm, Th = 100oF, Tc = 65 oF, the steady state gain matrix have the following numerical values:

0.7778 0.3889 K = 1.000 1.000


Which decomposes to:

0.276 0.961 U = 0.961 0.276 0.809 0.587 V = 0.587 0.809 0 1.453 = 0.803 0

11

1.4531 = 1.7 0.803

Assume the feed changed by the amount shown in the circle. The effect on the outputs is shown by the ellipse, which indicates the following: The major effect, which corresponds to the first column U1, increases both outputs but with more emphasis to Fm. The minor effect, which corresponds to the first column U2, decreases Tm and increases Fm. The second effect is minor compared to the first one because 2 < 1. So we can conclude that most of the input effect goes toward Fm (major axis).
major Fm [0.276,0.961]
[-0.961,0.276]

F1

1U1

2U2
F2

Tm

Figure 4: Operating Ellipse 1

For the give values for the outputs shown in the circle below, the backward effect of the matrix gain is shown by the ellipse:

12
F1 best ombination 1/1V1 1/2V2 Fm

F2

K-1

Tm

Figure 5: Operating Ellipse 2

The figure indicates: The best effect, which corresponds to [V1]T, indicates that the best effect on the outputs comes from the combination of 80% of F1 on the positive direction and 60% of F2 in the negative direction. Suppose now that the operating condition is changed to: F1 = 100 gpm, F2= 150 gpm, Th=100 oF, Tc=65 oF

The gain matrix and its decomposition is:

0.084 0.056 K = 1.00 1.00 0.014 0.999 U = 0.999 0.014 0.708 0.706 V = 0.706 0.708 0 1.414 = 0.099 0
= 14.28 First we note that the condition number is at least ten times larger, which indicates that the second operating condition becoming ill-condition.

13

F1

Fm

F2

Tm

Figure 6: Operating Ellipse 3

By inspecting the operating ellipse we can observe that most of the input effort will be directed towards the major axis (Fm), i.e. the area of operation will be along the major axis. Comparing the two operating ellipse (Figure 4 & 6), we notice that the first operating ellipse has a clear two degree of freedom, while the second operating ellipse has one degree of freedom which along the major axis. Therefore, it might be difficult to control the system in the direction of the minor axis (Tm). 1.2.4 Pairing method: 1. Connect the output associated with largest element of the vector U1 with the input that associated with the largest element of the vector V1. 2. Repeat the same procedure for the other columns of U and V. Example 1: a 2x2 system
0.2756 0.9612 Tm U = 0.9612 0.2756 Fm 0.8091 0.5877 F1 V = 0.5877 0.8091 F2

14

Example 2: a 3x3 system 0.595 0.627 0.502 T1 U = 0.583 0.767 0.669 T2 0.553 0.134 0.823 T3 0.869 0.084 0.488 R V = 0.454 0.718 Q 0.527 0 . 197 0.846 0 . 495 S R T1 S T2 Q T3 Using SVD for Decoupling Does not require inversion of a model. Using SVD for selecting appropriate location of a sensor Distillation columns, tubular rector. Using SVD control structure design Selection of appropriate MV is the one with smaller condition number. Smaller CN indicates that area of operation will be relatively equal (or reasonable) for both directions (major and minor).

15

1.3 Comparison between SVD and RGA RGA problem it is likely that an RGA pairing results in a high condition system because RGA by itself does not monitor the system ill-conditioning. For high condition number, the success of multivariable control system is unlikely. RGA is not sensitive to scaling. SVD problem it addresses only the open-loop nature of the process. RGA can detect with negative numbers those controller pairing that would change sign if some other control loop were to be put in manual. SVD is sensitive to scaling.
Conclusion: it is a good practice to consider both SVD and RGA analysis when final decision is to be made.

16

2. Strategies for reducing interaction When interaction is a problem, alternative strategies are available: Detune one or more feedback loop. Select different MV or CV. Consider decoupling controllers Consider a multivariable control scheme. 2.1 Change of variable The idea is to select new MV or CV to introduce zeros in the steady state gain matrix. Reconsider the mixer example, Let Then,
F=m x = F1/m m = F1 + F2

Choose u1 F1,

u2 m

F =0 F1

F =1 m
x 1 = F1 m
x F1 = m m 2

0 F x = 1 m

1 F1 F1 m m2

For any value of F1 and m, the RGA will be:

17

0 1 = 1 0
This indicates total decoupling. The resulting block diagram:
_ F1+F2 +
FC CC CT

F1

F1 , x 1 F2 , x 2 Mixer

FT

F, x

Figure 7: Control of a Mixer

Remark: Although this method has a great payoff, but it requires insight, intuition, and much analytical study.

2.2 Decoupling method The design objective is to reduce control loop interaction by adding additional controllers called decouplers to the multi-loop control configuration. It can provide two benefits: Interaction is eliminated and consequently the stability of the closed loop system is determined by the stability characteristic of the individual single loops. A set point change of one controlled variable has no effect on the other controlled variable. 2.2.1 Design procedure Reconsider the 2X2 interacting system:

y1 = g11u1 + g12u2 y2 = g 21u1 + g 22u2

18

One-way decoupling (partial decoupling): Suppose we want to eliminate the effect of u2 on y1, then we introduce the decoupling element D1 as shown in the figure.

sp1

_ +

gc1

++

u1

g11 g12

++

y1

D1 sp2 gc2 u2

g21 g22
+ +

+_

y2

Figure 8: Multivariable system under one-way decoupler

Therefore,
y1 = g11 (u1 + D1u 2 ) + g12 u 2

y1 = g11u1 + ( g11 D1 + g12 )u 2

To eliminate the effect of u2

( g11D1 + g12 ) = 0

D1 =

g12 g11

Two-way decoupling (complete decoupling): In the same fashion we can obtain: D2 = g 21 g 22

The full decoupled system looks as follows:

19

sp1

_ +

gc1 D2 D1

+ +

u1

g11 g12 g21

++

y1

sp2

+_

gc2

+ +

u2

g22

+ +

y2

Figure 9: Multivariable system under tow-way decoupler

Alternative design procedure:


y1 g11 = g y2 21 g12 u1 g 22 u 2

We introduce a decoupling matrix D:


y1 g11 = g y 2 21 g12 d11 g 22 d 21 d12 u1 b11 b12 u1 d 22 u 2 b21 b22 u 2

Where,
b11 b12 d11 g11 + d 21 g12 b = d g + d g b 21 22 11 21 21 22 d12 g11 + d 22 g12 d12 g 21 + d 22 g 22

Note that for partial decoupling we want either b12 or b21 equal zero, for complete decoupling we want b12 = b21 = 0. For simplicity set d11 = d22 = 1, then
b12 = d12g11 + d22g12 = 0 b21 = d11g21 + d21g22 = 0 d12 = - g12/g11 d21 = - g21/g22

20

1 D= g 21 g 22

g12 g11 1

Overall closed-loop transfer function: The overall closed-loop transfer function will be as follows (Case 1): y1 = g c1 ( g11 g12 g 21 g 22 ) sp y 1 + g c1 ( g11 g12 g 21 g 22 ) 1 g c 2 ( g 22 g12 g 21 g11 ) sp y 1 + g c 2 ( g 22 g12 g 21 g11 ) 2

y2 =

(compare this with that given in Appendix of CH6a)


Example: physically unrealizable de-coupler Suppose a process has the following transfer function:

5e 5 s 2e 4 s 4 s + 1 8s + 1 G p (s) = 3e 3s 6e 3s 12s + 1 10s + 1


The de-coupler design results in: 0.25( 4 s + 1)e s D12 = 8s + 1 Thus, the de-coupler is unrealizable because of es. There are other alternative approaches to resolve the problem, one of which is using static de-coupler:
D12 = lim g12 k = 12 s 0 g11 k11

21

D21 = lim

g 21 k = 21 s 0 g 22 k 22

Advantages of static de-coupler: Always realizable Require less process information Easy to implement Disadvantages of static de-coupler: Control loop interaction still exists during transient conditions.

2.2.2 Shortcomings of De-coupling De-couplers require perfect knowledge of the model, since this requirement is rarely justified, the de-couplers offer only partial decoupling with some weak interaction still persists. Chemical processes are nonlinear and non-stationary in nature, therefore, even if the decoupling is initially perfect, as the desired operating condition change, the decoupling deteriorates. De-couplers are conceptually similar to feed-forward controllers.

For large multivariable process, decoupling design becomes more complicated.

22 Case 1:

y1 = g11u1 + g12 u 2
From the block diagram:
sp sp u1 = g c1y1 + g c 2 d1y 2

(1)

sp sp u 2 = g c 2 y 2 + g c1d 2 y1

Inserting the definition of u1 and u2 into Eq. (1) gives:


sp sp sp sp y1 = g11 g c1y1 + g11 g c 2 d1y 2 + g12 g c 2 y 2 + g12 g c1d 2 y1

Noting that

d1 = - g12/g11 d2 = - g21/g22
The last equation becomes:
sp y1 = g11 g c1y1 g11 g c 2

g12 sp g sp sp y 2 + g12 g c 2 y 2 g12 g c1 21 y1 g11 g 22

y1 = g c1 ( g11 g12

g 21 sp )y1 g 22

sp y1 = g c1 ( g11 g12 g 21 g 22 )( y1 y1 )

sp y1 (1 + g c1 ( g11 g12 g 21 g 22 ) ) = g c1 ( g11 g12 g 21 g 22 ) y1

y1 =

g c1 ( g11 g12 g 21 g 22 ) sp y1 (1 + g c1 ( g11 g12 g 21 g 22 ) )

Potrebbero piacerti anche