Article

pubs.acs.org/IECR

Centralized PI/PID Controller Design for Multivariable Processes

Yuling Shen,*,† Youxian Sun,† and Wei Xu‡
†
State Key Laboratory of Industrial Control Technology, Zhejiang University, Hangzhou 310027, China
‡
Shanghai Electric Group Co., Ltd. Central Academe, Shanghai 20070, China

ABSTRACT: A novel centralized controller design method is proposed for multivariable systems, whether square or nonsquare
processes. First, the relationship between equivalent transfer function (ETF) and the pseudo-inverse of multivariable transfer
matrix is derived. Second, the relative normalized gain array (RNGA)-based ETF parametrization method is extended to the
nonsquare processes. Finally, a centralized proportional integral/proportional integral derivative (PI/PID) multivariable
controller is obtained from the Maclaurin expansion. The effectiveness of the proposed approach is verified by analysis of
several multivariable industrial processes; better overall performance is demonstrated compared with other centralized control
methods.

1. INTRODUCTION As the decoupler or controller is designed based on a model

It is known that not all the processes are square and processes inverse, it may result in a very complicated structure. In contrast,
with unequal number of inputs and outputs are frequently the centralized controller is composed of PI/PID controllers, and
encountered in the chemical industry. The traditional approach each of them is designed independently. The resulting controller
for a nonsquare multiple input−multiple output (MIMO) is simple and easy for application. Above all, the presented
process is to square up or square down the original system into approach is suitable for both square and nonsquare processes.
square process, which is realized by the addition or removal of Since nonsquare systems with more outputs than inputs are
inputs or outputs. It proves to be costly and difficult to get generally not desirable, the nonsquare systems with more
satisfactory performance.1 inputs than outputs still often arise in the chemical process.
For multivariable square systems, Davison has formulated a Therefore, this article mainly focuses on the centralized
centralized PI controller design method based on the steady controller design for multivariable processes with control inputs
state gain matrix.2,3 Sarma and Chidambaram extended it for not less than outputs.
multivariable nonsquare systems.4 For MIMO nonsquare systems
with multiple time delay, the Smith delay compensator is used to 2. PRELIMINARIES
enhance the control performance, and the controller is composed
A general multivariable control system is depicted as in Figure 1,
of a static decoupler and a decentralized controller.5,6
where G(s) is an n × m (n ≤ m) process transfer function,
Many equivalent transfer function (ETF) related methods
described by
can be found in the literature. An ETF-based centralized
controller design is simple and easy to implement in engineering ⎡ g (s )
processes,7,8 but it is not concerned with MIMO nonsquare g12(s) ... g1n(s) ⎤
⎢ 11 ⎥
systems, in which the difficulty lies in the ETF model ⎢ g (s ) g22(s) ... g2n(s)⎥
parametrization. Jin et al. developed the application of ETF into G(s) = ⎢ 21 ⎥
nonsquare systems.9 To improve the availability of the ETF ⎢ ... ... ... ... ⎥
models, model reduction is involved, and then the neighborhood ⎢ g (s ) gn2(s) ... gnn(s)⎥⎦
⎣ n1 (1)
search-assisted particle swarm optimization (NPSO) algorithm is
used to design the internal model control-propotional integral and GC(s)is a m × n multivariable centralized PI/PID controller,
derivative (IMC-PID) controller. However, it is a partial decoupling represented by
method technically. For square processes (n = m), the parameters
of the ETF model can be easily determined by a relative ⎡ g (s ) gc,12(s) ... gc,1n(s)⎤
normalized gain array (RNGA)-based method. When n < m, the ⎢ c,11 ⎥
inverse operation of matrix, which is involved in this method, does ⎢ g (s ) gc,22(s) ... gc,2n(s)⎥
not exist any more. In this paper, the RNGA-based ETF GC(s) = ⎢ c,21 ⎥
parametrization method is extended to nonsquare processes. In ⎢ ... ... ... ... ⎥
⎢ ⎥
addition to this, the ETF for process transfer function matrix with ⎣ gc, n1(s) gc, n2(s) ... gc, nn(s)⎦
(2)
zero terms is also discussed.
The proposed method is based on the concept of ideal
decoupling. The traditional decoupling techniques include a Received: January 27, 2014
decoupler and a controller, or an integrated controller with Revised: May 23, 2014
the aforementioned two functions.10−15 Satisfactory control Accepted: May 27, 2014
results for square systems are available from both of them. Published: May 27, 2014

© 2014 American Chemical Society 10439 dx.doi.org/10.1021/ie501541s | Ind. Eng. Chem. Res. 2014, 53, 10439−10447
Industrial & Engineering Chemistry Research Article

Figure 1. Block diagram of multivariable feedback structure.

2.1. Equivalent Transfer Function (ETF) Matrix. Let Table 1. Average Residence Times of FOPDT and SOPDT
Ĝ (s) be the equivalent transfer function (ETF) matrix of Models
G(s), i.e.
g(s) τar
⎡1/g ̂ (s) 1/g12̂ (s) ... 1/g1̂ m (s) ⎤
⎢ 11 ⎥ FOPDT k
e−θs
θ+τ
⎢ ̂ (s ) ̂ (s) ... 1/g2̂ m (s)⎥
τs + 1
̂ s) = ⎢1/g21
G(
1/g22
⎥ k θ+2τ
⎢ ... ... ... ... ⎥ (τs + 1)2
e−θs
⎢1/g ̂ (s) 1/g2̂ n (s) ... 1/gnm̂ (s)⎥⎦
⎣ 2n (3) k θ+τ1+τ2
SOPDT e−θs (τ1 ≠ τ2)
(τ1s + 1)(τ2s + 1)
where ĝij(s) is the equivalent transfer function of gij(s) when all
the other loops are closed under control. For simplicity, the kωn2 ξ
e−θs (ξ < 1) θ + 2ω
form of first order plus dead time (FOPDT) model is adopted s 2 + 2ξωns + ωn2 n

for ĝij(s), that is,

kiĵ ̂ and calculated by16
giĵ (s) = e−θijs
τiĵ s + 1 (4) Λ = K ⊗ K +T (8)
where kiĵ , τ̂ij and θ̂ij are the steady-state gain, time constant, and where K = [gij(0)]n×m and K+ is the generalized inverse matrix
dead time, respectively. of K. Similar to GRGA, the generalized relative normalized gain
2.2. Parameterization for ETF. Define the normalized array (GRNGA) can be defined as
steady-state gain matrix of G(s), calculated as
ΛN = KN ⊙ K̂ N (9)
KN = K ⊙ TAR (5) 17
and calculated by
where K = [gij(0)]n×m, TAR = [τar.ij]n×m, and τar.ij is the average
resident time of gij(s). The calculating formulas for τar.ij are ΛN = KN ⊗ K+NT (10)
given in Table 1. Correspondingly, the normalized steady-state Substituting eq 5 and 6 into eq 9, it is finally obtained as
gain matrix of Ĝ (s) is defined as
ΛN = Λ ⊙ Γ (11)
̂
K̂ N = K̂ ⊙ TAR (6) Where the relative average resident time array (RARTA), Γ, is
where K̂ = [ĝij(0)]n×m, T̂ AR = [τ̂ar.ij]n×m and τ̂ar.ij is the average expressed as
resident time of ĝij(s). ̂ ⊙ TAR
Γ = [γij]n × m = TAR
As is known, the generalized relative gain array (GRGA) for (12)
multivariable process is defined as Then, the RARTA can be obtained from eq 11 as
Λ = K ⊙ K̂ (7) Γ = ΛN ⊙ Λ (13)

10440 dx.doi.org/10.1021/ie501541s | Ind. Eng. Chem. Res. 2014, 53, 10439−10447

Industrial & Engineering Chemistry Research Article

From eq 8 and eq 12, the ETF parameter matrix can be From eq 21, the gain from uj(s) to yi(s) can be taken as 1/g+ji (s).
determined as By definition, the RDGA can be obtained as
K̂ = K ⊙ Λ (14) ⎡ g (s) g (s) ... g (s)⎤
⎢ 11 12 1m ⎥
and ⎢ g (s) g (s) ... g (s)⎥
Λ(s) =⎢ 21 22 2m ⎥
̂ = TAR ⊗ Γ
TAR (15) ⎢ ... ... ... ... ⎥
⎢ g (s) g (s) ... g (s)⎥
According to Table 1, the average resident time matrix of ĝij(s) ⎣ n1 n2 nm ⎦
is calculated as ⎡1/g +(s) 1/g + (s) ... 1/g + (s)⎤
⎢ 11 21 m1 ⎥
̂ = T̂ + Θ̂
TAR (16) ⎢1/g + (s) 1/g + (s) ... 1/g + (s)⎥
⊙ ⎢ 12 22 m2 ⎥
where T̂ = [τ̂ij]n×m and Θ̂ = [θ̂ij]n×m. Generally, the dead time of ⎢ ... ... ... ... ⎥
ETF is determined by ⎢ + + + ⎥
⎣1/g1n(s) 1/g2n(s) ... 1/gmn(s)⎦
Θ̂ = Θ ⊗ Γ (17)
=G(s) ⊗ G+T(s) (25)
Then, the time constant of ETF is calculated as
In contrast with eq 20, the following relationship is derived as
̂ − Θ̂
T̂ = TAR (18)
Ĝ (s) = G+T(s) (26)
Remark 1: When gij(s) = 0, there is a situation, in which zero
is divided by zero. In this case, the ETF of gij(s) is modeled as
3. MULTIVARIABLE PI/PID CONTROLLER DESIGN
giĵ (s) = kiĵ (19) Theorem 1. The ideal control objective for multivariable
process G(s) is equivalent to that for multiple single processes
2.3. Relationship between G(s) and Ĝ (s). As for ĝij(s), that is,
Nonsquare Multivariable Processe, The relative dynamic gain
array (RDGA) can be defined as18 I 1
GC(s) G(s) = ⇔ gc, ij(s)gjî (s) =
s s (27)
Λ(s) = G(s) ⊗ Ĝ (s) where GC(s) = [gc,ij(s)]m×n is the multivariable controller for G(s).
⎡ g (s ) g (s ) ... g1m(s) ⎤ Proof: In ideal control,19 it is hoped that the multivariable
⎢ 11 12 ⎥ process is decoupled into
⎢ g (s ) g (s ) ... g2m(s)⎥
= ⎢ 21 22 ⎥ I
⎢ ... ... ... ... ⎥ GC(s) G(s) =
s (28)
⎢ g (s ) g (s ) ... gnm(s)⎥⎦
⎣ n1 n2 Multiplying eq 26 by eq 28 gives
⎡1/g ̂ (s) 1/g12̂ (s) ... 1/g1̂ m (s) ⎤ T I
⎢ 11 ⎥ GC(s) = Ĝ (s)
⎢1/g ̂ (s) s (29)
1/g22 ̂ (s) ... 1/g2̂ m (s)⎥
⊗⎢ 21 ⎥ Then, eq 29 can be further expanded as
⎢ ... ... ... ... ⎥
⎢1/g ̂ (s) 1/g2̂ n (s) ... 1/gnm̂ (s)⎥⎦ ⎡ g (s) g (s) ··· gc,1n(s) ⎤
⎣ 2n (20) ⎢ c,11 c,12 ⎥
⎢ g (s) g (s) ··· gc,2n(s) ⎥
In ideal control, the following relationship holds for control ⎢ c,21 c,22 ⎥
variable and output variable as ⎢ ⋮ ⋮ ⋱ ⋮ ⎥
⎢ ⎥
U(s) = G+(s) Y(s) (21) ⎢ g (s) g (s) ··· gc, mn(s)⎥⎦
⎣ c, m1 c, m2
where
⎡ 1 1 1 1 1 1⎤
U(s) = [u1(s), u 2(s), ···, um(s)] T ⎢ · · ··· · ⎥
(22)
⎢ g11 ̂ (s) s g21 ̂ (s ) s gn̂ 1(s) s ⎥
⎢ 1 1 1 1 1 1⎥
Y(s) = [y1(s), y2 (s), ···, yn (s)]T (23) ⎢ · · ··· · ⎥
+ = ⎢ g12 ̂ (s) s g22̂ (s) s gn̂ 2 (s) s ⎥
and G (s) is the generalized inverse matrix of G(s), which is ⎢ ⎥
calculated as ⎢ ⋮ ⋮ ⋱ ⋮ ⎥
⎢ 1 1 1 1 1 1⎥
⎡ g + (s ) g + (s ) ··· g1+n(s) ⎤ ⎢ · · ··· · ⎥
⎢ 11 12 ⎥ ⎣⎢ g1̂ m (s) s g2̂ m (s) s ̂ (s) s ⎥⎦
gnm (30)
⎢ g + (s ) g + (s ) ··· g2+n(s) ⎥
G+(s) = ⎢ 21 22 ⎥ According to one-to-one correspondence rule, the following
⎢ ⋮ ⋮ ⋱ ⋮ ⎥ relationship exists
⎢ ⎥
⎢⎣ g (s) g (s)
+ + +
··· gmn (s)⎥⎦ 1
m1 m2 gc, ij(s)gjî (s) =
s (31)
= G̅ ′(s)(G(s)G̅ ′(s))−1 (24) for i = 1,2,···,m and j = 1,2,···,n.
10441 dx.doi.org/10.1021/ie501541s | Ind. Eng. Chem. Res. 2014, 53, 10439−10447
Industrial & Engineering Chemistry Research Article

Figure 2. Step response of Example 1.

Lemma 1. The multivariable process control problem with Substituting eq 35 into eq 36, the controller can be further
the goal of represented as
⎡ k e−l1s ⎤ kα , j/gjî (0) ⎡ g ̂ji′(0) ⎛ g ̂ ′ 2(0)
⎢ α ,1 ⎥ gc, ij(s) = ⎢ 1−s 2⎜
+s 2 2
ji
⎢ s ⎥ s ⎢ gjî (0) ⎜
⎢ ⎥ ⎣ ⎝ g ̂ji (0)
⎢ kα ,2 e−l2s ⎥ ⎤
g ̂ji″(0) ⎞
GC(s)G(s) = ⎢ s
⎥ ⎟ + ···⎥
⎢ ⎥ − ⎥
⎢ ⎥ gjî (0) ⎟⎠
⋱ ⎦ (37)
⎢ −lns ⎥
⎢ kα , n e ⎥ (1) The standard PI controller form is
⎢⎣ s ⎥⎦ (32) kI, ij
gc, ij(s) = kP, ij +
can be solved by designing the single loop controllers so as to s (38)

kα , j e−ljs where kP,ij and kI,ij are the controller parameters. Comparing
gc, ij(s) gjî (s) = eq 37 with eq 38, the controller parameters are derived as
s (33)
⎧ k = k g ̂ ′(0)/g ̂ 2(0)
where lj = min {θ̂ji, I = 1,2···,n} ; kα,j, j = 1,2···,n are the ⎪ P, ij α , j ji ji
⎨
regulation parameters and 0 < kα,j ≤ 1. Equation 33 can be ⎪ kI, ij = kα , j/gjî (0)
rewritten as ⎩ (39)
1 When the ETF elements take the model form of FOPDT, it
gc, ij(s) = F (s ) follows that
s (34)

where g jî ′(0)

= −τar,̂ ji + l j
Fji(s) = kα , j e−ljs /gjî (s) = kα , j/gjî (s) gjî (0) (40)
(35)
Therefore, the PI controller is designed as
Applying the Maclaurin series expansion to the above equation,
the controller can be expressed as ⎧ k = k (τ ̂ − l )/k ̂
⎪ P, ij α , j ar, ji j ji
1 ⎨
gc, ij(s) = [Fji(0) + sF′ji(0) + s 2F″ji(0) + ···] ⎪ kI, ij = kα , j/kjî
s (36) ⎩ (41)

10442 dx.doi.org/10.1021/ie501541s | Ind. Eng. Chem. Res. 2014, 53, 10439−10447

Industrial & Engineering Chemistry Research Article

Figure 3. Manipulated variable responses of Example 1.

(2) The standard PID controller form is where ISEyi−ri = ∫ ∞ 2 ∞

0 [1 − yi] and ISEyi−rj = ∫ 0 [0 − yi] (j ≠ i).
2

kI, ij The detailed steps to design the multivariable PI/PID

gc, ij(s) = kP, ij + + kD, ijs controllers are given as follows:
s (42)
Step 1: obtain the average resident time matrix TAR from Table 1
where kP,ij, kI,ij and kD,ij are the controller parameters. Step 2: calculate GRGA, GRNGA, and RARTA by eq 8,
Comparing eq 37 with eq 42, the controller parameters are eq 10. and eq 13
derived as Step 3: determine the ETF parameters K̂ , T̂ AR and T̂ by
eq 14, eq 15, and 18
⎧ k = k g ̂ ′(0)/g ̂ 2(0) Step 4: design the PI/PID controller according to eqs 41 or
⎪ P, ij α , j ji ji
⎪ eqs 45
⎨ kI, ij = kα , j/gjî (0) Remark 2: kα,j is tuned online to get a good compromise
⎪ between the features of fast track and small overshoot.
⎪ k = k (2g ̂ ′ 2(0) − g ̂ (0)g ̂ ″(0))/g ̂ 2(0)
⎩ D, ij α ,j ji ji ji ji (43)
4. CASE STUDY
For ETF in the form of FOPDT, it is obtained that
Example 1. Consider the shell control problem (2 × 3)20
gji″̂ (0)
= (τar,̂ ji − l j)2 + τji2̂ ⎡ 4.05 e−81s 1.77 e−84s 5.88 e−81s ⎤
gjî (0) (44) ⎢ ⎥
⎢ 50s + 1 60s + 1 50s + 1 ⎥
G(s) = ⎢ ⎥
Substituting eq 40 and eq 44 into eq 43, the PID controller is −54s
5.72 e−42s 6.9 e−45s ⎥
designed as ⎢ 5.39 e
⎣ 50s + 1 60s + 1 40s + 1 ⎦
⎧ k = k τ /k ̂
⎪ P, ij α , j ar, ji ji
⎪ Table 2. ISE Values of Centralized Controller for Example 1
⎨ kI, ij = kα , j/kjî
⎪ ISE values
⎪ k = k ((τ ̂ − l )2 − τ 2̂ )/k ̂
⎩ D, ij α ,j ar, ji j ji ji (45) method step in y1 y2 sum of ISE
Jin r1 132.68 73.47 206.15
To evaluate the control system performance, the ISE r2 15.10 69.84 84.94
performance index is introduced, which is calculated as follows Davison r1 115.79 6.41 122.20
n m r2 0.36 81.40 81.76
ISE = ∑ ∑ ISE y − r i j
proposed r1 111.89 1.61 113.50
i=1 j=1 (46) r2 0.84 73.75 74.59

10443 dx.doi.org/10.1021/ie501541s | Ind. Eng. Chem. Res. 2014, 53, 10439−10447

Industrial & Engineering Chemistry Research Article

Figure 4. Step response of Example 1: Perturbed case.

Table 3. ISE Values of Centralized Controller for Example 1 Table 4. ISE Values of Centralized Controller for Crude
with Perturbation Distillation Column Example
ISE values ISE values
method step in y1 y2 sum of ISE method step in y1 y2 y3 y4 sum of ISE
Jin r1 146.50 108.04 254.54 Davison r1 2.27 0.23 0.19 0.20 2.89
r2 24.12 102.37 126.49 r2 1.00 6.18 3.34 0.95 11.47
Davison r1 210.68 32.88 243.56 r3 0.09 0.10 3.80 0.34 4.33
r2 1.82 90.72 92.54 r4 0.07 0.09 0.20 2.10 2.46
proposed r1 142.85 3.21 146.06 Tanttu r1 11.87 1.48 1.48 0.02 14.85
r2 2.65 82.90 85.55 r2 0.21 17.20 0.35 0.03 17.79
r3 0.12 0.31 17.45 0.12 18.00
r4 0.26 0.97 1.18 14.87 17.28
The RGA, RNGA, and RARTA are calculated for Example 2 as proposed r1 5.47 0.12 0.1 0.15 5.84
⎡ 0.3203 − 0.5946 1.2744 ⎤ r2 0.15 4.94 0.16 0.16 5.41
Λ=⎢ ⎥, r3 0.04 0.01 5.83 0.34 6.22
⎣−0.0170 1.5733 − 0.5563⎦ r4 0.02 0.07 0.15 5.26 5.5
⎡ 0.6662 − 0.6248 0.9585 ⎤
ΛN = ⎢ ⎥,
⎣−0.3174 1.6153 − 0.2978 ⎦
⎡ 2.0803 1.0507 0.7522 ⎤ Then, the ETF parameter matrices are determined by eq 14
Γ=⎢ ⎥
⎣18.6639 1.0267 0.5354 ⎦ and eq 15, respectively.

Chart 1

10444 dx.doi.org/10.1021/ie501541s | Ind. Eng. Chem. Res. 2014, 53, 10439−10447

Industrial & Engineering Chemistry Research Article

Figure 5. Step response of Example 2.

⎡12.64 − 2.9766 4.6140 ⎤ The simulation results of the other two centralized control
K̂ = ⎢ ⎥, methods3,9 are compared in Figure 2 and Figure 3, and ISE
⎣−316.9045 3.6356 − 12.4033⎦
values are given in Table 2. It is shown that the proposed
⎡ 272.5140 151.2953 98.5330 ⎤ approach gets the best control performance and less
̂ =⎢
TAR ⎥ interactions between loops. The manipulated variable curves
⎣1941.0417 104.7211 45.5088 ⎦
are relatively smooth, which is valuable for the control
Set kα,1 = 0.008, kα,2 = 0.012 and solve the eqs 41; the technique when it is applied in practice, because the abrupt
multivariable controller is designed as change of control signal is undesirable for the actuator.
To test the robustness of the proposed method, we mismatch
⎡ 0.0006326 0.00003787 ⎤ the process model by increasing all six steady-state gains, six
⎢ 0.0658 + − 0.06712 − ⎥ time constants, and six time delays by a factor of 1.2, separately.
⎢ s s ⎥
⎢ 0.002688 0.003301 ⎥ Meanwhile, all the controllers are kept the same as before. The
GC = ⎢− 0.2907 − 0.2033 + ⎥ comparison results are shown in Figure 4, and the ISE values
⎢ s s ⎥ are given in Table 3. It shows that under such model
⎢ 0.001734 0.0009675 ⎥
⎢ 0.1291 + − 0.02072 − ⎥ mismatches, the deterioration is reasonable compared with the
⎣ s s ⎦ size of the perturbation.
10445 dx.doi.org/10.1021/ie501541s | Ind. Eng. Chem. Res. 2014, 53, 10439−10447
Industrial & Engineering Chemistry Research Article

Example 2. Consider the crude distillation process (4 × 5)21 ⎡ 2.3206 3.3755 0 0 3.4874 ⎤
⎢ ⎥
=⎢ ⎥
3.1616 13.2277 0 0 0
⎡ ̂
TAR
⎢ 3.2647 ⎥
⎢ 3.8(16s + 1) 2.9 e−6s 4.7487 8.8377 0 0
⎢ ⎥
⎢ 140s 2 + 14s + 1 10s + 1 ⎣ 0.1727 0.2512 6.3957 2.9873 3.7839 ⎦
⎢
⎢ 3.9(4.5s + 1) 6.3
⎢ 96s 2 + 14s + 1 ⎡ 0.1170 2.1097 0 1.9708 ⎤
0
20s + 1 ⎢ ⎥
G(s) = ⎢ 2.9156 13.2277 0 0 0
⎢ 3.8(0.8s + 1) 6.1(12s + 1) e−s T̂ = ⎢ ⎥
⎢ 3.0461 ⎥
⎢ 2 ⎢
4.5011 6.8517 0 0
⎥
⎢ 23s + 13s + 1 337s 2 + 34s + 1
⎣ 0.1685 0.2460 5.6090 1.9915 2.2072 ⎦
⎢ − 1.62(5.3s + 1) e−s − 1.53(3.1s + 1)
⎢ 2 Set kα,1 = kα,2 = kα,3 = kα,4 = 0.1 and solve the eqs 45; the
⎣ 13s + 13s + 1 5.1s 2 + 7.1s + 1
multivariable controller is designed as seen in Chart 1.
− 0.73(− 16s + 1) e−4s⎤ It can be seen from Figure 5 and Table 4 that the proposed
0 0 ⎥
150s 2 + 20s + 1 ⎥ method offers satisfactory control performance compared with
⎥ the other two controllers.3,22 Thus, the proposed method is still
16s e−2s ⎥ effective for high-dimension processes. Especially, simple design
0 0
(5s + 1)(14s + 1) ⎥ is one of its biggest advantages.
⎥
3.4e−2s 22s e−2s ⎥
0 ⎥ 5. CONCLUSION
6.9s + 1 (5s + 1)(10s + 1) ⎥
In this work, a novel multivariable centralized controller design
− 1.3(7.6s + 1) − 0.6 e−s 0.32(− 9.1s + 1)e−s ⎥ method is proposed, which is effective for both square and
⎥
4.7s 2 + 7.1s + 1 2s + 1 12s 2 + 15s + 1 ⎦ nonsquare processes. The RNGA-based ETF parametrization
method is extended to all multivariable processes. The multivariable
It is observed that g25(s) and g35(s) are the processes with PI/PID controller is determined by Maclaurin expansion. Each
controller is designed independently for corresponding ETF. The
derivative element. Sinceg25(0) = g35(0) = 0, so we consider major advantage of the proposed approach is that it can achieve
them as disturbance terms and then the process transfer satisfactory performance with simple control structure, which is
demonstrated by two simulation examples.

■
function is simplified into
AUTHOR INFORMATION
⎡ 3.8 e−3.2s 2.9 e−6s
⎢ Corresponding Author
⎢ 0.17s + 1 10s + 1 *Tel./Fax.: +86-21-2602 7776. E-mail: shenyuling2011@gmail.
⎢ −1.07s com.
⎢ 3.9 e 6.3
⎢ 12.68s + 1 20s + 1 Notes
G(s) = ⎢ The authors declare no competing financial interest.

■
−0.826s
⎢ 3.8 e 6.1 e−1.37s
⎢ 11.51s + 1 24.91s + 1 ACKNOWLEDGMENTS
⎢
⎢ −1.62 e s We are very grateful to the editors and anonymous reviewers
− 0.166
−1.53 e−0.0978s
⎢ for their valuable comments and suggestions to help improve
⎣ 6.66s + 1 4.58s + 1 our paper. This work is supported by the Key Program of
⎤ National Natural Science Foundation of China (No. 61333007)
−0.73e−19.8s ⎥ and the Major Program of National Natural Science Foundation
0 0 ⎥ of China (No. 61290321).

■
25.73s + 1 ⎥
0 0 0 ⎥
⎥ REFERENCES
3.4 e − 2s
⎥ (1) Loh, E. J.; Chiu, M. S. Robust decentralized control of non-square
0 0 ⎥ systems. Chem. Eng. Commun. 1997, 158, 157−180.
6.9s + 1 ⎥ (2) Davison, E. Some properties of minimum phase systems and
−10.5s ⎥ “squared-down” systems. IEEE Trans. Autom. Control 1984, 28, 221−
−1.3 e−0.0575 −0.6 e−s 0.32 e
⎥ 222.
0.41s + 1 2s + 1 14.70s + 1 ⎥ (3) Davison, E. J. Multivariable tuning regulators: The feed forward
⎦
and robust control of a general servomechanism problem. IEEE Trans.
Autom. Control 1976, 21, 35−47.
Following the above design procedure, the ETF parameter (4) Sarma, K. L. N.; Chidambaram, M. Centralized PI/PID
controllers for nonsquare systems with RHP zeros. J. Indian Inst. Sci.
matrices are determined as
2005, 85, 201−204.
(5) Seshagiri Rao, A.; Chidambaram, M. Smith delay compensator for
⎡ 2.2574 − 3.6466 − 322.3574 − 1.7078 − 6.5227 ⎤
⎢ ⎥ multivariable non-square systems with multiple time delays. Comput.
K̂ = ⎢
− 4.9692 3.5297 − 3.5289 2.0383 12.8340 ⎥ Chem. Eng. 2006, 30, 1243−1255.
⎢ 42.8616 − 69.2379 3.4019 − 1.5883 15.6073 ⎥ (6) Chen, J.; He, Z.-F.; Qi, X. A new control method for MIMO first
⎢ ⎥ order time delay non-square systems. J. Process Control 2011, 21, 538−
⎣16.5500 − 26.7346 − 2363.3397 − 0.6072 6.0264 ⎦ 546.

10446 dx.doi.org/10.1021/ie501541s | Ind. Eng. Chem. Res. 2014, 53, 10439−10447

Industrial & Engineering Chemistry Research Article

(7) Vu, T. N. L.; Lee, M. Independent design of multiloop PI/PID

controllers for interacting multivariable processes. J. Process Control
2010, 20, 922−933.
(8) Shen, Y.; Cai, W.-J.; Li, S. Multivariable process control:
Decentralized, decoupling, or sparse? Ind. Eng. Chem. Res. 2009, 49,
761−771.
(9) Jin, Q. B.; Hao, F.; Wang, Q. A multivariable IMC-PID method
for non-square large time delay systems using NPSO algorithm. J.
Process Control 2013, 23, 649−663.
(10) Wallter, M.; Wallter, J. B.; Wallter, K. V. Decoupling revisited.
Ind. Eng. Chem. Res. 2003, 42, 4575−4577.
(11) Wang, Q.-G.; Zhang, Y.; Chiu, M.-S. Decoupling internal model
control for multivariable systems with multiple time delays. Chem. Eng.
Sci. 2002, 57, 115−124.
(12) Linnemann, A.; Wang, Q.-G. Block decoupling with stability by
unity output feedback−Solution and performance limitations.
Automatica (UK) 1993, 29, 735−744.
(13) Liu, T.; Zhang, W. D.; Gao, F. Analytical decoupling strategy
using a unity feedback control structure for MIMO processes with
time delays. J. Process Control 2007, 17, 173−186.
(14) Cai, W.-J.; Ni, W.; He, M.-J.; Ni, C.-Y. Normalized decoupling−
A new approach for MIMO process control system design. Ind. Eng.
Chem. Res. 2008, 47, 7347−7356.
(15) Shen, Y. L.; Cai, W.-J.; Li, S. Y. Normalized decoupling control
for high dimensional MIMO processes with application to room
temperature control of HVAC systems. Control Eng. Pract. 2010, 18,
652−664.
(16) Chang, Jin-Wen; YU, Cheng-Ching The relative gain for non-
square multivariable systems. Chemical Eng. Sci. 1990, 45, 1309−1323.
(17) Luo, Y.; Liu, H.; Cai, W.-J. Jia, Z.; Jia, L.; Song, R. Control
configuration selection based on RNGA for non-square multivariable
processes. IEEE 31st Chineses Control Conference (CCC), 2012, 4722−
4727.
(18) Huang, H.-P.; Ohshima, M.; Hashimoto, I. Dynamic interaction
and multiloop control system design. J. Process Control 1994, 4, 15−27.
(19) Grosdidier, P.; Morari, M. Closed-loop properties from steady-
state gain information. Ind. Eng. Chem. Res. 1985, 24, 221−235.
(20) Valchos, C.; Williams, D.; Gomm, J. B. Solution to the shell
control problem using genetically tuned PID controllers. Control Eng.
Pract. 2002, 10, 151−163.
(21) Muske, K.; Young, J.; Grosdidier, P.; Tani, S. Crude unit
product quality control. Comput. Chem. Eng. 1991, 15, 629−638.
(22) Tanttu, J. T.; Lieslehto, J. A comparative study of some
multivariable PI controller tuning methods. Intelligent Tuning and
Adaptive Control; Pergamon Press: London, 1991; pp 357−362.

10447 dx.doi.org/10.1021/ie501541s | Ind. Eng. Chem. Res. 2014, 53, 10439−10447