Professional Documents
Culture Documents
Robust MIMO PID Controllers Tuning Based On Complex Real Ratio of The Characteristic Matriz Eigenvalues
Robust MIMO PID Controllers Tuning Based On Complex Real Ratio of The Characteristic Matriz Eigenvalues
www.elsevier.com/locate/ces
Received 13 September 2005; received in revised form 7 February 2006; accepted 10 February 2006
Available online 6 March 2006
Abstract
Robust MIMO PID controllers tuning based on complex/real ratio of the characteristic matrix eigenvalues is proposed. It is showed that this
tuning criterion is equivalent to H∞ optimal control. Under the proposed criterion, the tuning problem is stated as an optimization problem, in
which the complex/real ratio of the characteristic matrix eigenvalues, a Lyapunov quadratic index, and the spectral abscissa were simultaneously
minimized. Proposed criterion was applied to the multivariate controls of a distillation column, and a non-linear chemical reactor, both reported
in the literature.
䉷 2006 Elsevier Ltd. All rights reserved.
∗ Corresponding author. Tel.: +52 229 934 5701. García-Alvarado et al. (2005) stated that any linear or lin-
E-mail address: miguelg@itver.edu.mx (M.A. García-Alvarado). earized plant with multivariate PID control action may be
0009-2509/$ - see front matter 䉷 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ces.2006.02.015
I.I. Ruiz-López et al. / Chemical Engineering Science 61 (2006) 4332 – 4340 4333
then the characteristic matrix of the system is Corollary. If the system is at the boundary of stability (→0),
−1 1/(2 s 2 + 2s + 1)∞ → ∞.
( + chd) 0 (a − cfd) c
A= . (9)
0 Ic −gd 0
Theorem 2. For a high order stable system with transfer func-
By the other hand, it is demonstrated (Doyle et al., 1989) that tion equal to
one of the best robustness criterion is the standard H∞ control
problem, which state that the infinite norm of the close loop b0 s m + b1 s m−1 + · · · + bm
G(s) = for n > m, (17)
matrix transfer of the function T(s) defined as a0 s n + a1 s n−1 + · · · + an
T(s)∞ = sup (max |T(i)|) (10)
∀∈R+ G(s)∞ → ∞ ⇒ min → 0 where min = min{j }
must be kept lower than a given real positive number. In order ∀j = 1, 2, . . . , k
to assure robust performance and nominal performance of the
control system, sufficient conditions are represented by (Bao where k = n/2 if n is even, and k = (n − 1)/2 if n is non-even.
et al., 1999)
Proof. Eq. (17) may be written as
T(s)WT (s)∞ < 1, (11)
G(s)
S(s)WS (s)∞ < 1, (12) (s − p1 ) · · · (s − pm )
= ,
where S(s) is the sensitivity function of the system, WT (s) (0 s + 1)(21 s 2 + 21 1 s + 1) · · · (2k s 2 + 2k k s + 1)
and WS (s) are weighting functions, and is the optimization (18)
4334 I.I. Ruiz-López et al. / Chemical Engineering Science 61 (2006) 4332 – 4340
where 0 = 0 if n is even. The modulus (applying complex It is well known that (a0 could be introduced without lost of
number properties) is generality)
1 |In+c s − A| = a0 s n+c + a1 s n+c−1 + · · · + an+c . (27)
|G(i)| = 2 + p12 · · · 2 + pm
2 ···
2 20 + 1
From the definition of singular values decomposition and matrix
1 scalar product properties,
···
1 + 22 21 (221 − 1) + 4 41
[T1 (s)] = UT T1 (s)V
1 1
··· . (19) = UT D Adj(In+c s − A)
1 + 22 2n/2 (22k − 1) + 4 4k |In+c s − A|
× (B4 s + B3 )V. (28)
By the corollary of Theorem 1, if the minimum of the k
oscillation indexes min → 0, the system is at the stability That is, any of the r singular values of T1 (s) has polynomial
boundary and exist at least one in which G(s) → ∞. (27) as denominator (the characteristic equation), and therefore
by Theorem 2,
Theorem 3. For a multivariate close loop transfer matrix of a
T1 (s)∞ → ∞ ⇒ min → 0.
stable system T(s),
T(s)∞ → ∞ ⇒ min → 0. (20) Theorem 3 states that a system at stability boundary may be
detected by two equivalent ways
Proof. Eqs. (2), (7) and (8) may be written as
T(s)∞ → ∞ or min → 0.
y = Dx∗ + e, (21)
Therefore a control with robust stability may be obtained by
dx∗ d dyd
= Ax∗ + B1 + B2 + B 3 yd + B 4 , (22)
dt dt dt T(s)∞ < 1 or min > 2 .
where A is defined in Eq. (9), and,
Finally it is evident that
−1
( + chd) 0 b − cfe
B1 = , |In+c s − A| = a0 s n+c + a1 s n+c−1 + · · · + an+c
0 Ic −ge
−1 = a0 (s − 1 )(s − 2 ) · · · (s − n+c )
( + chd) 0 −che
B2 = , = (0 s + 1)(21 s 2 + 21 1 s + 1)
0 Ic 0
· · · (2k s 2 + 2k k s + 1), (29)
−1
( + chd) 0 cf
B3 = , where i are the eigenvalues of A. For those complex eigen-
0 Ic g
−1 values associated to min (or real eigenvalues if min 1),
( + chd) 0 ch
B4 = , D = [d 0 ].
0 Ic 0 1 − 2min
min
i,j = − ± i. (30)
Taking Laplace transform of Eqs. (21) and (22),
y(s) = T1 (s)yd (s) + T2 (s)(s), The complex/real ratio (in the case of min < 1) of the roots is
where
Im(i,j ) 1 − min
2
index defined by García-Alvarado et al. (2005) is considered. constrains and the optimization algorithm were programmed
This index is in Matlab 7.0. Differential equations of process dynamic were
∞ solved with fourth-order Runge–Kutta method, using discrete
IL = (yT Qy) dt = x∗T (0)Px∗ (0) PID algorithms. Final performance of control system was addi-
0
tionally evaluated with the quadratic performance index of the
= [D−1 y(0)]T PD−1 y(0), (32) error
tf
where D−1 D = In+c and P is the solution of
Ie = [(yd − y)T Q(yd − y) + uT Ru] dt (40)
AT P + PA = −DT QD. (33) 0
calculated from simulation results by trapezoidal rule.
Eq. (21) is valid only when
lim x∗T (t)Px∗ (t)
= 0. (34) 4. MIMO PID control of a distillation column
t→∞
Additionally, it is convenient subject the spectral abscissa (), Consider the distillation column model studied by Bao et al.
(1999) and Tan et al. (2002a,b):
= max[Re(i )] ∀i = 1, 2, . . . , n + c.
(35)
−33.89 32.63
y1 (s) (98.02s+1)(0.42s+1) (99.6s+1)(0.35s+1)
Then the tuning method consists in keep, =
y2 (s) −18.85 34.84
(75.43s+1)(0.30s+1) (110.5s+1)(0.03s+1)
IL < 1 , max < 2 , < − 3 . (36)
u (s)
× e−0.01s 1 , (41)
From Eq. (31), the maximum complex/real ratio (max ) is u2 (s)
max = max[|Im(i )/Re(i )|] ∀i = 1, 2, . . . , n + c. (37) where outputs 1 and 2 are temperatures of trays 21 and 7 and
inputs 1 and 2 are the liquid reflux and vapor boilup, respec-
All the values involved in Eq. (36) are scalars, and there- tively.
fore any optimization problem planted with these, represents State-space representation of Eq. (41) may be written in the
a lower computational effort than matrix inequality approach. standard form shown in Eqs. (1) and (2) with x ∈ R 12×1 (vari-
An optimal robust control may be obtained by solving the fol- ables 9–12 were introduce in order to considerate the delay),
lowing problem. ∈ ∅, u ∈ R 2×1 , y ∈ R 2×1 , and the non-zero element of the
Find f, g, h such that matrices as 11 = 98.02, 22 = 0.42, 33 = 99.6, 44 = 0.35,
IL and max → Min (38) 55 =75.43, 66 =0.3, 77 =110.5, 88 =0.03, 99 =11.11 =1,
10.10 =12.12 =5×10−5 , aii =−1, ∀i=1, . . . , 8, a12 =−33.89,
subject to a29 = a41 = a69 = a8.11 = a9.10 = a11.12 = 1, a34 = 32.63,
a56 = −18.85, a78 = 34.8, a10.9 = a12.11 = −1, a10.10 = a12.12 =
< 0. (39)
−0.01, c10.1 = c12.2 = 1, and d11 = d13 = d25 = d27 = 1. In
In the practice, this problem may be solved in sequence. the original example (Bao et al., 1999) the delay of Eq. (41)
First the minimization problem (38) is solved with IL as target. is not included until the robustness test. We include this terms
Then, the result obtained for IL may be used as constriction with a small value because the need of auxiliary equations to
for resolving problem (38) with max as target. The process is represent the transport delay.
repeated until a combined minimum is reached. At the optimal Lyapunov quadratic index (IL ) was evaluated with Q = I2
tuning, the infinite norm was evaluated with Eq. (23). and y(0)=1. Furthermore, the weighted quadratic performance
Two applications of the criterion are presented. In Section index of error (Ie ) was evaluated with Q = I2 , R = 0, and
4 a MIMO PID control of a distillation column is described. tf = 25 min for a simultaneous step set point change in both
This distillation column was represented using an empirical channels. The optimization details are described in Appendix B.
dynamic model reported in the literature (Bao et al., 1999; Tan Results are summarized in Table 1, and the dynamic responses
et al., 2002a,b). In Section 5 a MIMO PID control of a CSTR of outputs 1 and 2 are plotted in Figs. 1 and 2.
reported by Chen et al. (2002) is considered. The CSTR was In order to test robustness, we supposed the model is per-
represented by a mechanistic dynamic mode obtained from heat turbed as suggested by Bao et al. (1999) and Tan et al. (2002a,b),
and mass balances. and is represented by
Matrix operations were solved with Matlab 7.0 routines. Op- −33.89 32.63
timization problems required an algorithm that could manip- y1 (s) (98.01s+1)(0.43s+1) (98.5s+1)(0.33s+1)
=
y2 (s) −18.85 34.84
ulate the target function and the implicit constrain defined by (76s+1)(0.31s+1) (109.5s+1)(0.025s+1)
Eqs. (38) and (39). We found that the classical algorithm pro-
u1 (s)
posed by Box (1965) with modifications defined in Appendix × e−0.1s . (42)
A, assures the requirements, because it does not use functions u2 (s)
derivatives, it can manipulate discontinuous functions, and it For this last model, the weighted quadratic performance in-
can introduce implicit constrains. In order to avoid local op- dex of error (Ie ) was evaluated in the same way as for the nom-
timums, different initial guesses were used. Target functions, inal one. The weighted quadratic performance index of error
4336 I.I. Ruiz-López et al. / Chemical Engineering Science 61 (2006) 4332 – 4340
Table 1
Results obtained for the distillation column
Parameters Bao et al. (1999) Tan et al. (2002a,b) Proposed criteria P Proposed criteria PID
Fig. 1. Process output 1 for a simultaneous set point change in both channels Fig. 3. Process output 1 for a simultaneous set point change in both channels
(nominal plant). (perturbed plant).
Fig. 2. Process output 2 for a simultaneous set point change in both channels Fig. 4. Process output 2 for a simultaneous set point change in both channels
(nominal plant). (perturbed plant).
(Ie ) is shown in Table 1, and the dynamic responses of outputs best performance in the control of the nominal plant, as ex-
1 and 2 are plotted in Figs. 3 and 4. pected from the smallest Lyapunov index (IL ). However, the
As it can be seen from the weighted quadratic performance difference is minimal, because the Bao et al. (1999), Tan et al.
index of error (Ie ), the proposed PID controller achieved the (2002a,b) and proposed criteria are practically the same. The
I.I. Ruiz-López et al. / Chemical Engineering Science 61 (2006) 4332 – 4340 4337
Table 2
Simulation schedule for the CSTR
t (min) Action t (min) Action
1 CA = CA − 0.05 8 T =T −5
2 CA = CA + 0.05 9 CA0 = CA0 − 0.2
3 T =T −5 10 CA0 = CA0 + 0.4
4 T =T +5 11 CA0 = CA0 − 0.2
5 CA = CA + 0.05 11 T0 = T0 − 10
6 CA = CA − 0.05 12 T0 = T0 + 20
7 T =T +5 13 T0 = T0 − 10
characteristics. This robust stability may be predicted from the with Q = I2 and y(0) = 1. Furthermore, the weighted quadratic
max value. Nevertheless, this time the proposed P controller performance index of error (Ie ) was evaluated with Q = I2 ,
achieved the best performance. R = 0, and tf = 15 min for the simulation schedule shown in
From the results presented in Table 1, it can be observed Table 2 . The optimization details are described in Appendix B.
that tuning control parameters that minimize the max value Results are summarized in Table 3, and the dynamic re-
effectively reduces the H∞ -norm of the sensitivity matrix sponses of reactor concentration and temperature are plotted in
and the closed loop transfer matrix. The control tuned with Figs. 7 and 8.
proposed criterion yields comparable singular value plots In this particular example, the optimization procedure yields
(Figs. 5 and 6) to those obtained with the other more compli- a max value of zero, indicating all roots were real; therefore
cated tuning methods, showing similar nominal performance, making the minimization of the spectral abscissa () the only
disturbance rejection, and robust stability properties in fre- way to reach an acceptable level of robustness. For a linear
quency domain. system, the Lyapunov index (IL ) and spectral abscissa () are
4338 I.I. Ruiz-López et al. / Chemical Engineering Science 61 (2006) 4332 – 4340
Table 3
Results obtained for the CSTR
6. Conclusions
Notation
Fig. 7. Dynamic behavior of the controlled CSTR for the exit reactant
concentration. a, b, c, d, e, dynamic
parameters matrices
A dynamic characteristic matrix
Bi for i = 1, 2, . . . , 4 composite matrices of dynamic
parameters
D output mapping matrix
f, g, h PID parameters matrices
I L , Ie quadratic indexes performance
(Lyapunov and error)
P Lyapunov matrix. Symmetric and
positive defined
Q, R weigh matrices, symmetric and pos-
itive defined
S(s) sensitivity function matrix
Ti (s) for i = 1, 2 close loop transfer matrices
U, V orthonormal matrices ∈ R r×r
u, x, y control, state, and output matrices
ui , upd integral, proportional and derivative
control variables
Fig. 8. Dynamic behavior of the controlled CSTR for the exit temperature. yd set point matrix
Greek letters
both measures of the decay rate of the states. In this case, even spectral abscissa
when the CSTR is a non-linear process, the minimization of damping factor
these values produced control parameters that achieved a better matrix A eigenvalues
performance for both, set point changes and load disturbances, input matrix
than the reported by Chen et al. (2002), as it can be seen from time constants
the lower weighted quadratic performance index of error (Ie ). complex/real ratio of matrix A
Finally, as expected from the spectral abscissa values, proposed eigenvalues
I.I. Ruiz-López et al. / Chemical Engineering Science 61 (2006) 4332 – 4340 4339
2. Eq. (A.19) in the original complex method is equal to In the case of min IL
xn corr = 0.5(xn + xc ). f (x) = IL , g1 (x) = , g2 (x) = max ,
The modifications introduced are the result of our experience −∞ < g1 (x) < 0, 0 < g2 (x) < ∞.
in the use of original complex algorithm. When the centroid is
invalid or the most far point the algorithm is cycled. In the case of min max
−100 0000 < f11 < 0, 0 < f22 < 100 000. References
−100 000 < f12 , g12 , f21 , g21 < 100 000.