Journal of Process Control 23 (2013) 649–663

Contents lists available at SciVerse ScienceDirect

Journal of Process Control

journal homepage: www.elsevier.com/locate/jprocont

A multivariable IMC-PID method for non-square large time delay

systems using NPSO algorithm
Q.B. Jin, F. Hao ∗ , Q. Wang
Institute of Automation, Beijing University of Chemical Technology, Beisanhuan East Road 15, Chaoyang District, Beijing 100029, PR China

a r t i c l e i n f o a b s t r a c t

Article history: Multiple time delays and strong interactions among different loops are the main problems in the design of
Received 4 September 2012 multivariable controller for non-square systems. In this paper, the concept of effective open-loop transfer
Received in revised form 5 December 2012 function (EOTF) is extended to non-square systems. By applying the internal model control (IMC) method,
Accepted 23 February 2013
the controllers with equivalent models are designed. For practical applications, the NPSO algorithm is
used to obtain the parameters of the incremental PID with first-order lag filter. This new method does
Keywords:
not only avoid the complex computation caused by the procedure of decoupling first and then designing
Non-square system
controllers but also employs the advantages of IMC-PID’s suitable for large time delay systems and strong
IMC-PID
EOTF
robustness. Simulation is carried out to demonstrate the effectiveness of the proposed method; also
NPSO significant performance improvement has been achieved with the proposed method compared with
PID with first-order lag filter other methods.
© 2013 Elsevier Ltd. All rights reserved.

1. Introduction shorter response time compared to the method given by Tanttu

and Lieslehto [4] in 1991. However, as a static decoupling method,
In order to meet the high product quality and energy inte- tuning parameters must be implemented for better dynamic per-
gration requirements of the process industries, the multi-input formances. Magni and Scattolini [5] applied the model predictive
multi-output (MIMO) process has become widespread in modern control (MPC) algorithm to control a non-square nonlinear model of
industry. Under this circumstance, there has been more and more a continuous fermenter, which guarantees the tracking of asymp-
attention paid to the control issue of the multivariable system in totically constant reference signals by forcing integral action on
both industry and academia. The MIMO systems are often classified the error variables. However the MPC algorithm is computationally
into two types according to the number of input/output variables, intensive because the norm objective functions are usually need to
i.e. square systems (with equal number of inputs and outputs) and be calculated [6]. Chen et al. [7] came up with a method based on
non-square systems (with unequal number of inputs and outputs). IMC algorithm to design Smith compensator for non-square sys-
Due to the special model structure of non-square systems, most tems with transfer function elements consisting of first order plus
control methods for square systems cannot be applied to non- dead time (FOPDT), and simulation result shows the method can
square systems directly. The traditional methods for non-square achieve good decoupling and control performance.
processes are generally realized by adding or deleting variables to In recent years, the concept of EOTF has been gradually
obtain square system matrix [1]. But adding variables might incur introduced to the design of multi-loop controllers for MIMO sys-
unnecessary control cost, while deleting variables will cause loss of tems [8–10]. The main idea is to reasonably transform the design
system information, even leading to the system becoming uncon- of multi-loop controller to a set of independent single-loop con-
trollable. trollers. The method can solve the interaction problem in MIMO
With the development of advanced control theories, many systems and avoid the complex procedures of decoupling first
design approaches have been presented. Sarma and Chidambaram and designing controllers afterwards on the decoupled systems.
[2] proposed the centralized PI controllers design method by Therefore the control performance can be improved significantly
extending Davison’s method [3] of calculating the pseudo inverse eliminating the possible error caused by complex computation. In
of steady-state gain matrix to non-square systems, which has a literature [9], Vu and Lee designed a multi-loop PI/PID controller for
square systems based on the concept of EOTF and showed favorable
control performance. However, the method is confined to square
systems. In this paper, the application of EOTF is extended to the
∗ Corresponding author at: Mailbox #382, Beisanhuan East Road 15, Chaoyang
control of non-square systems as follows. The non-square system
District, Beijing 100029, PR China. Tel.: +86 13426282946.
model is first transformed into equivalent single-loops on the basis
E-mail addresses: Catherinehaofeng@126.com, Fionahao@126.com (F. Hao).

0959-1524/$ – see front matter © 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.jprocont.2013.02.007
650 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663

of EOTF, meanwhile the IMC method is brought in to design the

single-loop controllers. Then the PID controller which are simple
to implement and widely used can be obtained through the NPSO
algorithm.
The particle swarm optimization (PSO) algorithm was first
introduced by Kennedy and Eberhart [11]. It has the advantage of
excellent global search ability while nothing to do with the initial
parameters. However, the PSO has the weakness of converging to
undesired local solution or premature convergence [12]. In 1984, Fig. 2. Transformation between feedback controller and internal model controller.

Pan [13] improved the LJ algorithm proposed by Luus and Jaakola

[14] and proposed a New LJ (NLJ) algorithm. This optimize method of the feedback controller Gc can be obtained as long as the transfer
is completely based on random search, but the identification results function of the internal model controller is known exactly, thus the
largely depend on the selection of initial values. In literature [15], parameters of PID controller can be obtained through further math-
a new NPSO algorithm was proposed in which the search results ematical process. However in terms of MIMO systems, the above
of the PSO algorithm were used as the initial values of the NLJ procedure needs a series of complicated mathematical derivations.
algorithm. As a whole, the NPSO algorithm can make better opti- It seems even more difficult when it comes to non-square systems
mization for its combination of the global search capability of PSO whose transfer function matrixes are asymmetric.
and the exactly local optimization of NLJ.
In this paper, the idea of applying NPSO algorithm to obtain the 2.2. The incremental PID with first-order lag filter structure
parameters of each incremental PID controller is first put forward.
The proposed method has not only the advantages of traditional The control law of traditional PID controllers [19] can be
PID control method, but also the ones of IMC method, i.e. good anti- described as:
interference ability, little interactivity and strong robustness. The   t 
1 T de(t)
organization of the paper is as follows: Section 1 reviews the devel- u(t) = Kp e(t) + e(t)dt + d (1)
opment of the control methods for non-square systems in recent Ti 0
dt
years; Section 2 gives details of the theories about the IMC-PID where u(t) is the process input, e(t) is the deviation between
and the derivation of the incremental PID with first-order lag fil- the set-point and the output, means e(t) = r (t) − y (t). Kp , Ti , Td are
ter; Section 3 explains the multivariable IMC-PID controller design known as factors of proportionality, integration time and deriva-
method for non-square large time delay systems based on EOTF tive time respectively. In order to determine the parameters of
using NPSO algorithm; Section 4 presents two simulation exam- desired PID controller using numerical search algorithm, the tradi-
ples of fat systems (number of inputs is more than outputs) to tional integral and differential items cannot be used directly, thus
demonstrate the effectiveness of the proposed method. Finally, the following approximations are involved:
conclusions are drawn in Section 5.
 t 
k

k

2. Preliminaries t = kT, (k = 0, 1, 2, . . .), e(t)dt = Te(j) = T e(j),

0
j=0 j=0
2.1. Basic theories of the IMC-PID de(t) e(k) − e(k − 1)
=
dt T
In 1982, Garcia and Morari [16] proposed the IMC method on the
basis of model algorithm control (MAC) and dynamic matrix control Hence the position PID is derived:
(DMC). The structure of IMC is shown in Fig. 1. Where R is the set-

k
point, U is the manipulated variable, Gp and Gm are the transfer u(k) = Kp e(k) + Ki e(j) + Kd [e(k) − e(k − 1)] (2)
function matrix of process and its model respectively, GIMC is the
j=0
traditional internal model controller, Yp and Ym are the output of the
process and the internal model respectively, d denotes the external where the integral coefficient and the differential coefficient are
disturbance. The form of the internal model controller [17,18] can defined as Ki = Kp T/Ti and Kd = Kp Td /T, respectively.
−1
be expressed as GIMC (s) = Gm− (s)f (s), where Gm– (s) is the minimum Note from (2) that full-output may lead to heavy computational
phase part of the model. f (s) represents the filter with the form effort for the reason that each of its output is closely linked to the
as f (s) = 1/(1+s) , where  is large sufficiently to guarantee the whole past states, also the total sum of e(k) has to be raised [20].
controller proper and realizable.  is the filter time constant. To avoid this problem, the incremental PID is employed. A formula
For the convenience of field implementation, it is desirable can be obtained from (2) that:
to transform the controllers to PID forms which are widely used
in industry fields. Note from Fig. 2 that the equivalence relation

k−1
u(k − 1) = Kp e(k − 1) + Ki e(j) + Kd [e(k − 1) − e(k − 2)] (3)
between the feedback controller and the internal model controller
j=0
is derived as Gc = GIMC /(1 −vGm GIMC ). In theory, the transfer function
The incremental PID is derived from (2) and (3) as:

u(k) = u(k) − u(k − 1) = Kp [e(k) − e(k − 1)]

+ Ki e(k) + Kd [e(k) − 2e(k − 1) + e(k − 2)] (4)

It is known to all that the derivative element can improve the

dynamic performance of the control system, but it will lead to
strong sensitivity to disturbance. In formula (2), there exists ud (k) =
(Kp (Td /T ))[e(k) − e(k − 1)] = Kd [e(k) − e(k − 1)]. Given e(k) is the
step function, the output values are worked out as ud (0) = Kd ,
Fig. 1. Structure of IMC method. ud (1) = ud (2) = · · · = 0, therefore the controller only has output in
 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663 651

matrix of Gm with the ith row jth column (1 ≤ i ≤ m, 1 ≤ j ≤ n)

element removed, gm(i) is the ith row vector of Gm with the jth col-
−jR
umn element removed. gm(j) is the jth column vector of Gm with
−iC
the ith row element removed. r , y are the set-point and output
−i −i
variables with the ith loop’s ri , yi dropped from r, y, respectively. u
−j

Fig. 3. PID controller with the first-order lag filter structure.

is the manipulated variables with uj dropped from u. Q represents
−ji
the controller when the jth row ith column element is dropped from
the first period with an amplitude of Kd = Kp Td /T, while the outputs Q.
of the later periods are all zeros. That means the effect of the deriva- In Fig. 4, equations can be obtained that:
tive element’s output is confined in the first period, hence the actual
control effect is too weak to control error in advance for those sys- gm(j) uj + Gm u = y (12)
−iC −ij −j −i
tems with large time constants. On the other hand, Kd is normally a
large value, and it may cause harmful to actuators if the changes in gm(ij) uj + gm(i) u = yi (13)
ud are either too much or too fast. Taking above into consideration, −jR −j

a first-order filter Gf (s) = 1/(1 + Tf (s)) is employed to compose the u= r −Qy (14)
incremental PID controller, and then the shortcomings of derivative −j −i −ji −i
element can be compensated as the result. The position PID with Let r = 0, substituting (12) into (14), rearranging leads to:
the first-order lag filter can be obtained from the block diagram −i
shown in Fig. 3 by discretization as follows: −1
  u = − Q (I + Gm Q ) gm(j) uj (15)
−j −ji −ij −ji −iC
K Kd s
U(s) = Kp + i + E(s) = Up (s) + Ui (s) + Ud (s) (5)
s 1 + Tf s Then substituting (15) to (13), the relation between yi and uj can
be written as:
u(k) = up (k) + ui (k) + ud (k) (6)
yi = gm(ij) uj + gm(i) u = [gm(ij) − gm(i) Q (I + Gm Q )−1 gm(j) ]uj (16)
The up (k) and ui (k) are the same as those in tradition PID form, while −jR −j −jR −ji −ij −ji −iC
ud (k) is expressed as:
Note from (16) that the relation between yi and uj depends on not
Kp Td s
Ud (s) = E(s) (7) only the transfer function gm(ij) of the ith row jth column, but also
1 + Tf s the coupling terms among other loops.
Considering the inverse Laplace transform of (7): Rewriting (16) leads to the simplified form as below:

dud (t) de(t) yi = [gm(ij) − gm(i) Gm −1 Gm Q (I + Gm Q )−1 gm(j) ]uj (17)

ud (t) + Tf = Kp Td (8) −jR −ij −ij −ji −ij −ji −iC
dt dt
Using ud (k) + Tf ((ud (k) − ud (k − 1))/T ) = Kp Td ((e(k) − e(k − 1))/T ) On the premise that the frequencies are lower than the cross-over
to approximate the derivative element and simplifying (8) leads frequency [9], perfect control approximation can be considered as:
to:
Gm Q (I + Gm Q )−1 = I (18)
Tf Kp Td −ij −ji −ij −ji
ud (k) = u (k − 1) + [e(k) − e(k − 1)] (9)
Tf + T d Tf + T
For square systems, Gm −1 in (17) can be obtained directly, but that’s
For a straightforward view in searching for parameters of the con- −ij

trollers, a factor is assumed as ˛ = Tf /(Tf + T), (˛ < 1), therefore (9) is not the case for non-square systems. As a result, the Moore-Penrose
rearranged as: pseudo inverse [21] of Gm has to be employed as substitution. Here
−ij

ud (k) = Kd (1 − ˛)[e(k) − e(k − 1)] + ˛ud (k − 1) (10) the following lemma is considered.

It can be seen easily from (10) that the effect of derivative element Lemma 1. Assume that A is an m × n matrix (m < n), the necessary
decreases in a ˛ related ratio since the first period, and therewith and sufficient condition of existing Moore-Penrose pseudo inverse A*
also the sensitivity to disturbance. Thus the control performances is rank(A) = m, then the Moore-Penrose pseudo inverse of A can be
−1
can be improved significantly. Referring to the derivation of (4), the expressed as A∗ = AH (AAH ) , where AH is the Hermitian matrix of A.
incremental PID with first-order lag filter structure is obtained as −1
follows: Thus the pseudo inverse Gm ∗ = Gm H (Gm Gm H ) is used instead
−ij −ij −ij −ij

u(k) = Kp [e(k) − e(k − 1)] + Ki e(k) + Kd (1 − ˛)[e(k) − 2e(k − 1) of G−1

m and (17) can be expressed as:
−ij
+ e(k − 2)] + ˛[uD (k − 1) − uD (k − 2)] (11)
−1
yi = [gm(ij) − gm(i) (Gm H (Gm Gm H ) Gm ) Q (I + Gm Q )−1 gm(j) ]uj (19)
−jR −ij −ij −ij −ij −ji −ij −ji −iC

3. The controller design Substituting (18) to (19) leads to:

−1
yi = (gm(ij) − gm(i) Gm H (Gm Gm H ) gm(j) )uj = gem(j) uj (20)
3.1. The concept of the equivalent model
−jR −ij −ij −ij −iC

The derivation of the equivalent model based on EOTF [9] is where

given in this subsection. Fig. 4 shows the block diagram of the open- −1
gem(j) = gm(ij) − gm(i) Gm H (Gm Gm H ) gm(j) (21)
loop equivalent MIMO model, where Gm is the transfer function −jR −ij −ij −ij −iC
−ij
652 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663

Fig. 4. Block diagram of the open-loop equivalent non-square model.

It can be known from (21) that each equivalent model transfer func- of gem(j) , Gemc (s) represents the equivalent model internal model
tion gem(j) is attainable with the model’s transfer function matrix controller:
itself only, while no pre-knowledge of other loops is required.
⎡g ⎤
emc1 (s) 0
⎢ ⎥
3.2. Design of the equivalent model controller ⎢ .. .. .. ⎥
⎢ . . . ⎥
⎢ ⎥
Applying the IMC method, the single-loop internal model con- ⎢ 0 gemc(m) (s) ⎥
troller can be obtained based on the equivalent model of the Gemc (s) = ⎢
⎢
⎥
⎥ (23)
⎢ 0 · · · gemc(m+1) (s) ⎥
non-square system. The Gp (s) in Fig. 5 indicates the transfer func- ⎢ ⎥
tion matrix of the m × n non-square system which has n inputs and ⎢ .. .. ⎥
⎣ . ··· .
⎦
m outputs written as:
⎡ ⎤ 0 ··· gemc(n) (s) n×m
g11 (s) g12 (s) ··· g1n (s)
⎢ g (s) g (s) . . . g (s) ⎥
⎢ 21 22 2n ⎥ Note from Section 2 that the internal model controllers are given as
Gp (s) = ⎢ ⎥ −1
(22) GIMC (s) = Gm− (s)f (s), then each element of non-square equivalent
⎢ . . . . ⎥
⎣ .. .. .. .. ⎦ model can be expressed as:

gm1 (s) gm2 (s) ··· gmn (s) −1

gemc(j) = gem(j)− f (s) (24)
where gij (s) = gij0 (s)e−ij s ,
(i ≤ m, j ≤ n) is the transfer function
between the jth input and ith output, gij0 (s) is strictly proper and where gem(j)− is the minimum phase part of equivalent model gem(j) ,
stable; e−ij s is the time-delay term and ij0 ≥ 0 is the constant of −1
gem(j)− is the inverse of gem(j)− .
time-delay. Gem (s) is the non-square equivalent model composed

3.3. Determining parameters of the IMC-PID controller using

NPSO algorithm

In this paper, the input and output values of the controllers

are collected for the purpose to convert the transfer functions of
the equivalent model to PIssD forms through the random search
algorithm. That means only with the collected values, the wanted
parameters of the incremental PID controllers with first-order lag
filter structure can be easily obtained by running the NPSO pro-
Fig. 5. Internal model controller with equivalent model for non-square system. gram.
 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663 653

good estimate for any optimization problem with a determined

performance. While to PSO algorithm, the characteristic of each
particle depends on a velocity vector − →vi = (vi,1 , vi,2 , . . . , vi,d ), a
position vector − →xi = (xi,1 , xi,2 , . . . , xi,d ) and its experience vector
−
→
pi = (pi,1 , pi,2 , . . . , pi,d ). The updating equations are presented as
follows:
Fig. 6. The structure of the PID controller for non-square system.

vi,j (t + 1) = ω · vi,j (t) + c1 · ˇ1 · (Pi,j − xi,j (t)) + c2 · ˇ2 · (Pg,j − xi,j (t))

(26)
3.3.1. The new method to collect input/output values by multiple
set-points change
The structure of the PID controller for non-square system are
shown in Fig. 6, where the controller CPID can be expressed as: xi,j (t + 1) = xi,j (t) + vi,j (t + 1) (27)
⎡ c (s) 0
⎤
ii
where vi,j (t + 1) and vi,j (t) represents the current and previous
⎢ . ⎥
⎢ . .. .. ⎥ velocity of a design variable respectively, while xi,j (t + 1) and xi,j (t)
⎢ . . . ⎥ represents the corresponding position. The acceleration parame-
⎢ ⎥
⎢ 0 cmm (s) ⎥ ters c1 and c2 are set to constant value. ˇ1 and ˇ2 are two uniformly
CPID (s) = ⎢
⎢ 0 ··· c
⎥
⎥ (25) distributed random, ω is inertial weight which controls the influ-
⎢ (m+1)m (s) ⎥
⎢ ⎥ ence of previous velocity on the new velocity. Pi,j is the best fitness
⎢ . .. ⎥ of the ith particle so far and Pg,j is the global design of the swarm in
⎣ .. ··· .
⎦
current generation.
0 ··· cnm (s) n×m This optimization searching algorithm was extended to deter-
mine the parameters for the PID controllers in this paper. When
To obtain the parameters of the PID controller, a serious of
the unit step change is made to ri , the sub-PID controller is written
numerical searching work has to be done based on the input/output
as:
values. Hence a novel method for collecting values by multiple
set-points changing is concluded according to the input/output ⎡ (i) ⎤
c11 (s) 0
signal structure of non-square PID controller shown in Fig. 7. In
⎢ ⎥
terms of an m × n non-square system, the input/output values of ⎢ .. ⎥
⎢ . ⎥
every element of the controller CPID can be obtained in the case ⎢ . . ⎥
⎢ .. (i)
cii (s) . ⎥
that the set-point changes are made from r1 to rm in sequence. ⎢ . ⎥
Assume that the unit step change in ri , the input value of the cor- ⎢ ⎥
⎢ .. ⎥
responding controller equals to the error between the set-point (i) ⎢ . ⎥
CPID (s) = ⎢ ⎥ (28)
(i)
value and the output value of the system, means ei = ri − yi ,
(i)
⎢ 0 cmm (s) ⎥
(i)
⎢ ⎥
while the controller’s output value equals to the process’s input ⎢ ⎥
(i) ⎢ 0 ··· c(m+1)m (s) ⎥
(i)
ui . For the controller’s input/output values whose set-point val- ⎢ ⎥
⎢ ⎥
(i)
ues are rk = 0 k(1 < k ≤ m, k =
/ i), the inputs of these controllers ⎢ . .. ⎥
(i) (i) ⎣ .. ··· . ⎦
are ek = 0 − yk , while the jth (j =
/ i) output value is the same as
(i)
the input of the system uj .
(i) 0 ··· cnm (s) n×m

(i)
3.3.2. The NPSO algorithm where the elements of CPID (s) are composed of the parameters
(i) (i) (i)
As mentioned in the introductory section, the NPSO algo- Kp , Ki , ˛(i) The detailed steps to obtain the values are
Kd ,
rithm [15] is a combination of the NLJ algorithm [13] and the presented as follows:
PSO algorithm [11]. As to NLJ algorithm, as long as the search
range and the search density are large enough, there is a very (i) (i) (i)
Step 1: Make the parameters Kp , Ki , Kd , ˛(i) into an array
x to be determined. Then initialize the population, assign con-
stant values of inertial weight ω, acceleration parameters c1 and
c2 , respectively.
Step 2: For each particle x, evaluate the fitness function Ê =
t 2
fitness(x̂i,j ) = 0 [u(t) − û(t)] dt where u(t) indicates the internal
model controller outputs and û(t) indicates the sub-PID controller
outputs.
Step 3: Find the personal best location (x̂p ) and the global best loca-
tion (x̂g ); if current fitness evaluation is better than the previous,
then let the current replace the previous.
Step 4: Update the velocity of each particle and the coordinate
position.
Step 5: Determine whether the iteration number k meet the set
value or not, if not, then jump to Step 7.
Step 6: Start NLJ algorithm. Set the x̂g as the initial value of NLJ,
means a0 = x̂g .
Step 7: Referring to the current optimization to determine the next
Fig. 7. Input/output signal structure of non-square PID controller. range of search ri (k) = Hai [k−1] vk−1 , where H is a constant value
654 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663

which is generally taken as 2, while the v value depends on the

rate .
⎧
⎪
⎪
v0 = 1 v1 = 1
⎪
⎪ ⎧    
⎪
⎨ ⎪ ai
[k−1]
ai
[k−1]
1
⎪
⎨ vk−1 ϕk max [k−1]
≤  and max
[k−1]
≥

ai (0) ai (0)
⎪
⎪ v =     k≥2
⎪
⎪
k
⎪
⎪
[k−1] [k−1]
⎪
⎩ ⎩ 1 max
ai
>  or max
ai
<
1
[k−1]
ai (0)
[k−1]
ai (0) 

ϕk is convergence coefficient which is usually taken from the fol-

lowing empirical formula:

1.34 1.29
ϕk = 0.981k or ϕk = 0.979k

Step 8: Generate p groups of n + 1 random numbers randki to get the

[k] [k]
current optimal solution ai = ai (j)|min[J(ai (j))] as the initial value
of the next iteration. The updating forum is as follows:

(j) (j−1)
ak (i) = ak (i) + randki × r (j) (i)

where k = 1, . . . , p indicates the array and j = 1, . . . , m indicates

the iteration number. So the sub-PID controller output û(t) and the
performance assessment value ÊN are obtained.
Step 9: Return to Step 7 until the ÊN meets the termination condi-
tion, along with the best estimated â is acquired.
Step 10: Choose the superior result from Ê and ÊN by the crite-
rion E = min{Ê, ÊN } and use this result to denote the corresponding
parameters x.
Step 11: If E ≤ ε (ε is infinitesimal positive number), exit the itera-
tion; otherwise, return to Step 2.

3.3.3. Obtaining the final PID controller through sub-PID

controllers
As stated above, for the unit-step changes in ri , processing the
collected input/output values through NPSO algorithm can lead to
(i)
the every elements of sub-PID controller CPID (s) in (28). To obtain Fig. 8. The flowchart of the proposed method.
the final CPID (s) through these sub-PID controllers, the control
actions of the sub-PID controllers when unit-step changes were ∞ ∞
as ISEy2 −r2 = 0 (r2 − y2 )2 dt and ISEy1 −r2 = 0 (ȳ1 − y1 )2 dt. On the
made to different ri had to be synthesized. Hence, the elements in
whole, the ISE value of the controlled system is expressed as:
(25) can be given as:
Ki Kd s ISEy−r = ISEy1 −r1 + ISEy2 −r1 + ISEy1 −r2 + ISEy2 −r2 (30)
c(s) = Kp + + (29)
s 1 + Tf s There is an inverse relationship between the ISEy–r and the per-
m (i) m (i) m (i)
formance of the system, i.e. the smaller the value, the better the
where Kp = K , Ki =
i=1 p
K , Kd =
i=1 i
K . On the other
i=1 d performance.
hand, because ˛ is a ratiocoefficient, it should be processed accord-
m
ing to the formula ˛ = i=1
˛(i) /m. Then referring to ˛ = Tf /(Tf +T) 4. Simulation study
and the given sampling time T, the Tf for c (s) could be obtained
consequently. Example 1 The full Shell process is a multivariable (5 inputs and
The proposed method can be expressed straightforwardly by the 7 outputs) heavy oil fractionator with strong interactions and large
flowchart in Fig. 8. In addition, two well-known examples in pro- dead times [23]. In this paper, the subset of key elements of the Shell
cess industries are given in Section 4 where the simulation results standard control problem (3 inputs and 2 outputs) is introduced,
are presented showing good performance to demonstrate the effec- where the outputs are top end point (y1 ) and side end point (y2 ),
tiveness of the proposed method. while the manipulated variables are top draw (u1 ), side draw (u2 ),
and bottoms reflux duty (u3 ). Since this proposed method is signif-
3.3.4. The performance criterion icant for large time delay processes, the time delays are increased
To evaluate the performance of closed-loop system quantita- to three times, thus the transfer function matrix of the process is
tively, the idea in literature [22] is referred to. Here the integral written as:
square error (ISE) values are calculated with ȳ1 , ȳ2 denote the cor- ⎡ ⎤
responding steady state outputs when the set-point change in the 4.05e−81s 1.77e−84s 5.88e−81s
⎢ 50s + 1 ⎥
other. In case the step change in r1 , the corresponding∞ ISE val- ⎢ 60s + 1 50s + 1 ⎥
ues to y1 and interaction response y2 are ISEy1 −r1 = 0 (r1 − y1 )2 dt Gp (s) = ⎢ −54s 5.72e−42s ⎥ (31)
∞ ⎣ 5.39e 6.9e−45s ⎦
2
and ISEy2 −r1 = 0 (ȳ2 − y2 ) dt, respectively. In a similar way, as to
the step change in r2 , the corresponding ISE values are presented 50s + 1 60s + 1 40s + 1
 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663 655

Table 1 Table 2
The results of NPSO algorithm when r1 = 1, r2 = 0. The results of NPSO algorithm when r1 = 0, r2 = 1.
(1) (1) (1) (2) (2) (2)
Kp Ki Kd ˛(1) Kp Ki Kd ˛(2)
(1) (2)
c11 0.0118 0.0008 10.1021 0.9794 c11 0.0193 0.0017 0.6079 0.9890
(1) (2)
c22 0.1592 0.0018 0.2443 0.4390 c22 0.0095 0.0009 0.0413 0.9544
(1) (2)
c32 0.0062 0.00008 0.3755 0.9973 c32 0.0015 0.00001 0.8423 0.8719

Table 3
The equivalent model is derived according to (21), and through
The final results of the PID parameters.
model reduction method can the low order model be obtained:
Kp Ki Kd Tf
0.6843 × (345.875s + 1)e−7.7077s
gem1 = , c11 0.0311 0.0025 10.7100 436.0380
(321.53s + 1)(52.259s + 1) c22 0.1687 0.0027 0.2856 16.0795
c32 0.0077 0.00009 1.2178 100.0336
3.5533e−37.938s −2.7179e−69.447s
gem2 = , gem3 =
62.449s + 1 65.969s + 1
To achieve the minimum ISE value, the filter time constants are change in r1 , while those subfigures with the Y-axis u1 (2), u2 (2),
set as 1 = 125, 2 = 20, 3 = 357.1429. Then the internal model u3 (2) show the comparison when a unit step change in r2 .
controller are obtained according to (24): For this simulation example, the sampling time was chosen as
⎡ ⎤ T = 7 s. Synthesizing the results in Tables 1 and 2, the PID controller
(321.53s + 1)(52.259s + 1)
0 parameters Kp , Ki , Kd , ˛ are obtained shown in Table 3.
0.6843(345.875s + 1)(125s + 1)
⎢ ⎥ Fig. 10 shows the manipulated variable behavior for a unit step
Gemc (s) = ⎢
⎢ 62.449s + 1 ⎥
0 ⎥ change in r1 , r2 at t = 0, respectively. Fig. 11 shows the correspond-
⎣ 3.5533(20s + 1) ⎦ ing servo responses. Sarma K.L.N.’s method [2] and Chen’s method
65.969s + 1
0 − [24] are also considered for a fair comparison. The ISE values of all
2.7179(357.1429s + 1)
the comparative methods are tabulated in Table 4. Fig. 12 shows
After that, the method for collecting input/output values of the the manipulated variable behavior for a unit step perturbation in
controller is used, meanwhile the NPSO algorithm is employed to d when r1 = 0, r2 = 0 and the corresponding responses are shown in
search for the optimal parameters. The sub-PID controllers’ param- Fig. 13. It can be seen that the proposed method can offer satisfac-
eters were tabulated in Tables 1 and 2 with a unit step change in r1 , tory dynamic performance and anti-interference ability.
r2 , respectively. In Fig. 9, the subfigures with the Y-axis u1 (1), u2 (1), To demonstrate the robust performance of the proposed
u3 (1) show the comparison curves between the internal model con- method, a perturbation uncertainty of +20% in each process gain
troller outputs and the sub-PID controller outputs when a unit step while −20% in the time constant and time delay were inserted

−3
x 10
0.3 5

0.25
0
0.2
u1(1)

u1(2)

−5
0.15
−10
0.1

0.05 −15
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 80
Time(s) Time(s)

0 0.7

0.6
−0.1
0.5
u2(1)

u2(2)

−0.2
0.4
−0.3
0.3

−0.4 0.2
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 80
Time(s) Time(s)

0.15 −0.06

0.1 −0.08
u3(1)

u3(2)

0.05 −0.1

0 −0.12 The internal model controller outputs

The sub−PID controller outputs
−0.05 −0.14
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 80
Time(s) Time(s)

Fig. 9. The comparison of the internal model controller outputs and the sub-PID controller outputs.
656 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663

Manipulated variable responses for a step in r1 Manipulated variable responses for a step in r1 Manipulated variable responses for a step in r1
1.2 0 0.4

1
−0.1 0.2
0.8

0.6
u1

u2

u3
−0.2 0
0.4

0.2
−0.3 −0.2
0

−0.2 −0.4 −0.4

0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000
Time(s) Time(s) Time(s)

Manipulated variable responses for a step in r2 Manipulated variable responses for a step in r2 Manipulated variable responses for a step in r2
0.2 0.4 0.4

0.3
0.2
0
0.2

0.1 0
u1

u2

u3
−0.2
0 −0.2
−0.1 Proposed
−0.4
−0.4 Chen
−0.2
Sarma K L N
−0.6 −0.3 −0.6
0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000
Time(s) Time(s) Time(s)

Fig. 10. Manipulated variable responses for a unit step in r when model matches.

into (31), whereas the controller parameters remain the same Example 2. The crude distillation process [2] which has 5
as those when the model matches. As shown in Fig. 14, the inputs and 4 outputs is considered. As to this process, the out-
proposed method can still offer good performance when model puts are naphtha/kerosene cutpoint (y1 ), kerosene/light gas oil
mismatches which means the higher robustness level of the design (LGO) cutpoint (y2 ), LGO/heavy gas oil (HGO) cutpoint (y3 ) and
method. measured over flash (y4 ), while the manipulated variables are top

Set−point response for a step in r1 Interactive response for a step in r1

1.2 1.5

1
1
0.8

0.6
y1

y2

0.5
0.4

0.2 Proposed
0
Chen
0 Sarma K L N
−0.2 −0.5
0 200 400 600 800 1000 0 200 400 600 800 1000
Time(s) Time(s)

Set−point response for a step in r2 Interactive response for a step in r2

1.5 0.8

0.6
1

0.4
y2

y1

0.5
0.2

0
0

−0.5 −0.2
0 200 400 600 800 1000 0 200 400 600 800 1000
Time(s) Time(s)

Fig. 11. Servo responses and interactive responses for a unit step in r when model matches.
 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663 657

Manipulated variable responses for a step in d1 Manipulated variable responses for a step in d1 Manipulated variable responses for a step in d1
0 0.5 0.4

−0.2 0.2
0
−0.4 0
u1

u2

u3
−0.5
−0.6 −0.2
−1
−0.8 −0.4

−1 −1.5 −0.6
0 200 400 600 800 0 200 400 600 800 0 200 400 600 800
Time(s) Time(s) Time(s)
Proposed Manipulated variable responses for a step in d2 Manipulated variable responses for a step in d2 Manipulated variable responses for a step in d2
Chen 1.5 2 0.5
Sarma K L N
1 1
0
u1

u2

u3
0.5 0
−0.5
0 −1

−0.5 −2 −1
0 200 400 600 800 0 200 400 600 800 0 200 400 600 800
Time(s) Time(s) Time(s)
Manipulated variable responses for a step in d3 Manipulated variable responses for a step in d3 Manipulated variable responses for a step in d3
0 0.5 0.5

−0.5 0
0
u1

u2

u3
−1 −0.5
−0.5
−1.5 −1

−2 −1.5 −1
0 200 400 600 800 0 200 400 600 800 0 200 400 600 800
Time(s) Time(s) Time(s)

Fig. 12. Manipulated variable responses for a unit step perturbation in d when model matches.

Response for a step in d1 Interactive response for a step in d1

4 6

4
2
2
y1

y2

0
0
−2

−2 −4
0 200 400 600 800 0 200 400 600 800
Time(s) Time(s)
Response for a step in d2 Interactive response for a step in d2
2 6

1 4
y1

y2

0 2

−1 0

−2 −2
0 200 400 600 800 0 200 400 600 800
Time(s) Time(s)
Response for a step in d3 Interactive response for a step in d3
6 10
Proposed
4 Chen
5 Sarma K L N
y1

y2

2
0
0

−2 −5
0 200 400 600 800 0 200 400 600 800
Time(s) Time(s)

Fig. 13. Responses and interactive responses for a unit step perturbation in r when model matches.
658 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663

Response for a step in r1 Interactive response for a step in r1

2 2

1.5
1.5
1
1 0.5
y1

y2
0.5 0

−0.5
0
−1

−0.5 −1.5
0 200 400 600 800 1000 0 200 400 600 800 1000
Time(s) Time(s)

Response for a step in r2 Interactive response for a step in r2

2 1

1.5 0.8

0.6
1
0.4
y2

y1
0.5
0.2
0
Proposed 0
−0.5 Chen
−0.2
Sarma K L N
−1 −0.4
0 200 400 600 800 1000 0 200 400 600 800 1000
Time(s) Time(s)

Fig. 14. Servo responses and interactive responses for a unit step in r when model mismatches.

1 2 3 4 5
0.5 −0.2 0.1 0.5 0

0.4 0 −0.1
u1(1)

u2(1)

u3(1)

u4(1)

u5(1)

−0.25 0
0.3 −0.5 −0.2

0.2 −0.3 −0.1 −1 −0.3

0 100 200 0 100 200 0 100 200 0 100 200 0 100 200
Time(s) Time(s) Time(s) Time(s) Time(s)
6 7 8 9 10
0 0.5 0 1 0.1

0.4 0.5
u1(2)

u2(2)

u3(2)

u4(2)

u5(2)

−0.2 −0.2 0
0.3 0

−0.4 0.2 −0.4 −0.5 −0.1

0 100 200 0 100 200 0 100 200 0 100 200 0 100 200
Time(s) Time(s) Time(s) Time(s) Time(s)
11 12 13 14 15
0.1 0 0.5 0 0.4

0.4 −0.5
u1(3)

u2(3)

u3(3)

u4(3)

u5(3)

0 −0.02 0.2
0.3 −1

−0.1 −0.04 0.2 −1.5 0

0 100 200 0 100 200 0 100 200 0 100 200 0 100 200
Time(s) Time(s) Time(s) Time(s) Time(s)
16 17 18 19 20
0.1 0 0.05 −1 −1

−0.02 −1.5 −1.5

u1(4)

u2(4)

u3(4)

u4(4)

u5(4)

0 0
−0.04 −2 −2

−0.1 −0.06 −0.05 −2.5 −2.5

0 100 200 0 100 200 0 100 200 0 100 200 0 100 200
Time(s) Time(s) Time(s) Time(s) Time(s)
The internal model controller outputs
The sub−PID controller outputs

Fig. 15. The comparison of the internal model controller outputs and the sub-PID controller outputs (1–5, 6–10, 11–15, 16–20 are for a unit step change in r1 , r2 , r3 , r4 ,
respectively).
 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663 659

1 2 3 4 5
4 0 0.2 0.5 0.5

2 0

u1

u2

u3

u4

u5
−0.2 0 0
0 −0.5

−2 −0.4 −0.2 −1 −0.5

0 50 100 0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s) Time(s)
6 7 8 9 10
0.5 1 0 1 0.2
u1

u2

u3

u4

u5
0 0.5 −0.2 0 0

Proposed −0.5 0 −0.4 −1 −0.2

0 50 100 0 50 100 0 50 100 0 50 100 0 50 100
Davison Time(s) Time(s) Time(s) Time(s) Time(s)
Tanttu 11 12 13 14 15
0.1 0.05 1 0 0.2

−0.5
u1

u2

u3

u4

u5
0 0 0.5 0
−1

−0.1 −0.05 0 −1.5 −0.2

0 50 100 0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s) Time(s)
16 17 18 19 20
0.1 0.1 0.1 0 0.5

−1
u1

u2

u3

u4

u5
0 0 0 0
−2

−0.1 −0.1 −0.1 −3 −0.5

0 50 100 0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s) Time(s)

Fig. 16. Manipulated variable responses for a unit step in r when model matches (1–5, 6–10, 11–15, 16–20 are for a unit step change in r1 , r2 , r3 , r4 , respectively).

temperature (u1 ), kerosene yield (u2 ), LGO yield (u3 ), HGO yield
(u4 ) and heater outlet temperature (u5 ). The transfer function is
given as below:
⎡ 3.8(16s + 1) 2.9e−6s −0.73(−16s + 1)e−4s
⎤
0 0
⎢ 140s2 + 14s + 1 10s + 1 150s2 + 20s + 1 ⎥
⎢ ⎥
⎢ 3.9(4.5s + 1) 6.3 16se−2s ⎥
⎢ 0 0 ⎥
⎢ 96s2 + 14s + 1 20s + 1 (5s + 1)(14s + 1) ⎥
Gp (s) = ⎢ ⎥ (32)
⎢ 3.8(0.8s + 1) 6.1(12s + 1)e−s 3.4e−2s 22se−2s ⎥
⎢ 0 ⎥
⎢ 23s2 + 13s + 1 337s2 + 34s + 1 6.9s + 1 (5s + 1)(10s + 1) ⎥
⎣ ⎦
−1.62(5.3s + 1)e−s −1.53(3.1s + 1) −1.3(7.6s + 1) −0.6e−s 0.32(−9.1s + 1)e−s
13s2 + 13s + 1 5.1s2 + 7.1s + 1 4.7s2 + 7.1s + 1 2s + 1 12s2 + 15s + 1
The equivalent model is derived according to (21), and the fol-
lowing low order model is obtained by model reduction method:
e−0.293s e−1.095s
gem1 = , gem2 = ,
1.3717s + 1 7.5297s + 1
In a similar way, the 1 = 0.3494, 2 = 12.1916, 3 = 10.0288,
1.7968e−9.2599s
gem3 = 4 = 23.8345, 5 = 49.8966 are set. From (24), the controller can
3.6779s + 1 be obtained as:

−0.1215(9.1104s + 1)e−3.7554s Table 5

gem4 = ,
(1.1521s + 1)(9.1315s + 1) The ISE values of the three compared methods.

−11.6069(1.0221s + 1)e−4.1259s Method Set-point ISEy1 ISEy2 ISEy3 ISEy4 ISEy–y

gem5 = changes
(0.93s + 1)(167.2569s + 1)
r1 2.4342 0.4910 0.8531 0.4310
r2 0.0099 4.5668 1.3490 0.5597
Table 4 Proposed 22.1926
r3 0 0.0024 6.1762 0.3966
The ISE values of the three compared methods. r4 0.0004 0.0182 0.1147 4.7894
Method Set-point changes ISEy1 ISEy2 ISEy–y r1 3.7014 0.0878 0.0777 0.1009 22.4732
r2 0.5587 6.8820 1.6373 0.6233
r1 20.2901 11.0198 Davison’s
Proposed 37.3748 r3 0.0407 0.0433 4.7687 0.2757
r2 0.7551 5.3098
r4 0.0393 0.0436 0.1057 3.4871
r1 19.4510 14.6980
Chen 41.2808 r1 8.0788 0.2833 0.3081 0.8066
r2 0.0130 7.1188
r2 0.0284 11.0718 0.1849 0.0383
Tanttu’s 41.4419
r1 36.6798 27.8538 r3 0.0443 0.0511 10.0296 0.0712
Sarma K.L.N. 76.4839
r2 4.8201 7.1302 r4 0.0944 0.1931 0.2460 9.9121
660 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663

1 2 3 4
1.5 0.5 1 0.5

1 0.5 0

y1

y2

y3

y4
0
0.5 0 −0.5

0 −0.5 −0.5 −1
0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s)
5 6 7 8
0.5 1.5 0.5 0.5

1 0
y1

y2

y3

y4
0 0
0.5 −0.5

Proposed −0.5 0 −1 −0.5

0 50 100 0 50 100 0 50 100 0 50 100
Davison Time(s) Time(s) Time(s) Time(s)
Tanttu 9 10 11 12
0.1 0.1 1.5 0.5

1
y1

y2

y3

y4
0 0 0
0.5

−0.1 −0.1 0 −0.5

0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s)
13 14 15 16
0.1 0.2 0.2 1.5

1
y1

y2

y3

y4
0 0 0
0.5

−0.1 −0.2 −0.2 0

0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s)

Fig. 17. Servo responses and interactive responses for a unit step in r when model matches (1–4, 5–8, 9–12, 13–16 are for a unit step change in r1 , r2 , r3 , r4 , respectively).

1 2 3 4 5
0 0.5 0.5 2 1

−0.5 0.5
u1

u2

u3

u4

u5
0 0 0
−1 0

−1.5 −0.5 −0.5 −2 −0.5

0 50 100 0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s) Time(s)
6 7 8 9 10
0.5 1 0.5 2 0.5

0 0 0
u1

u2

u3

u4

u5
0 0
−0.5 −1 −0.5

−1 −2 −0.5 −2 −1
Proposed 0 50 100 0 50 100 0 50 100 0 50 100 0 50 100
Davison Time(s) Time(s) Time(s) Time(s) Time(s)
Tanttu 11 12 13 14 15
0.2 0.1 1 2 0.5

0 0
u1

u2

u3

u4

u5

0 0 0
−1 −2

−0.2 −0.1 −2 −4 −0.5

0 50 100 0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s) Time(s)
16 17 18 19 20
0.05 0.05 0.05 0 0.2

−0.5
u1

u2

u3

u4

u5

0 0 0 0
−1

−0.05 −0.05 −0.05 −1.5 −0.2

0 50 100 0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s) Time(s)

Fig. 18. Manipulated variable responses for a unit step perturbation in d when model matches (1–5, 6–10, 11–15, 16–20 are for a unit step change in d1 , d2 , d3 , d4 , respectively).

⎡ 1.3717s + 1 ⎤
0 0 0
⎢ 0.3494s + 1 ⎥
⎢ 7.5297s + 1 ⎥
⎢ 0
12.1916s + 1
0 ⎥ 0
⎢ 3.6779s + 1
⎥
⎢ ⎥
Gemc = ⎢ 0 0 0 ⎥
⎢ 1.7968(10.0288s + 1) ⎥
⎢ (1.1521s + 1)(9.1315s + 1) ⎥
⎢ 0 0 0 − ⎥
⎣ 0.1215(9.1104s + 1)(23.8345s + 1) ⎦
(0.93s + 1)(167.2569s + 1)
0 0 0 −
11.6069(1.0221s + 1)(49.8966s + 1)
 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663 661

1 2 3 4
4 4 4 1

2 2 2 0

y1

y2

y3

y4
0 0 0 −1

−2 −2 −2 −2
0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s)
5 6 7 8
2 4 4 1

1 2 2 0
y1

y2

y3

y4
0 0 0 −1

Proposed −1 −2 −2 −2
0 50 100 0 50 100 0 50 100 0 50 100
Davison Time(s) Time(s) Time(s) Time(s)
Tanttu 9 10 11 12
0.2 0.2 4 1

2 0
y1

y2

y3

y4
0 0
0 −1

−0.2 −0.2 −2 −2
0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s)
13 14 15 16
0.1 0.1 0.1 0.5

0
y1

y2

y3

y4
0 0 0
−0.5

−0.1 −0.1 −0.1 −1

0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s)

Fig. 19. Responses and interactive responses for a unit step perturbation in d when model matches (1–4, 5–8, 9–12, 13–16 are for a unit step change in d1 , d2 , d3 , d4 ,
respectively).

1 2 3 4
1.5 0.5 1 0.5

1 0.5
y1

y2

y3

y4
0 0
0.5 0

0 −0.5 −0.5 −0.5

0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s)
5 6 7 8
1.5 1.5 1 1

1 1
y1

y2

y3

y4

0 0
0.5 0.5

Proposed 0 0 −1 −1
0 50 100 0 50 100 0 50 100 0 50 100
Davison
Time(s) Time(s) Time(s) Time(s)
Tanttu
9 10 11 12
1.5 0.5 1.5 0.5

1 1 0
y1

y2

y3

y4

0
0.5 0.5 −0.5

0 −0.5 0 −1
0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s)
13 14 15 16
1.5 0.5 1 2

1 0.5 1
y1

y2

y3

y4

0
0.5 0 0

0 −0.5 −0.5 −1
0 50 100 0 50 100 0 50 100 0 50 100
Time(s) Time(s) Time(s) Time(s)

Fig. 20. Servo responses and interactive responses for a unit step in r when model mismatches (1–4, 5–8, 9–12, 13–16 are for a unit step change in r1 , r2 , r3 , r4 , respectively).
662 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663

The comparison of the internal model controller outputs and the features, good anti-interference ability, easy to understand and
sub-PID controller outputs is shown in Fig. 15. For the simula- implement, strong decoupling ability and robustness. The proposed
tion example, the sampling time is chosen as T = 1 s and the PID method is also an attempt to accelerate the industry application for
controller can be calculated as:
⎡ ⎤
1.55 2.125s
2.2189 + + 0 0 0
⎢ s 1 + 100.01s ⎥
⎢ 0.075 3.2176s ⎥
⎢ 0 0.5173 + + 0 0 ⎥
⎢ s 1 + 60s ⎥
⎢ 0.025 1.9653s ⎥
CPID = ⎢ 0 0 0.218 + + 0 ⎥
⎢ s 1 + 99.92s ⎥
⎢ 0.305 20.581s ⎥
⎢ 0 0 0 −0.535 − − ⎥
⎣ s 1 + 54.866s ⎦
0.001 11.5772s
0 0 0 −0.1548 − −
s 1 + 100.016s
The Davison’s method and Tanttu’s method [2] are employed for non-square systems using control theory. The simulation results
a fair comparison. The PID controller transfer function matrix of show that the method can offer satisfactory dynamic performance
Davison’s method is:
⎡ 0.133 0.060 0.007 0.018
⎤
0.443 + + 0.399s −0.201 − − 0.181s 0.023 + + 0.021s 0.061 + + 0.055s
⎢ s s s s ⎥
⎢ −0.274 − 0.082 − 0.247s 0.283 +
0.085
+ 0.255s −0.015 −
0.005
− 0.014s −0.037 −
0.011
− 0.033s ⎥
⎢ ⎥
⎢ s s s s ⎥
GC = ⎢ − 0.0003s ⎥
0.0009 0.085 0.088 0.00012
⎢ −0.003 − − 0.003s −0.284 − − 0.256s 0.294 + + 0.265s −0.0004 − ⎥
⎢ s s s s ⎥
⎢ −0.586 − 0.176 0.147 0.088 0.494
− 0.527s −0.491 − − 0.442s −0.294 − − 0.265s −1.647 − − 1.482s ⎥
⎣ s s s s ⎦
0.046 0.023 0.019 0.049
−0.153 − − 0.138s 0.078 + + 0.007s 0.064 + + 0.058s 0.166 + + 0.149s
s s s s
While the PID controller transfer function matrix of Tanttu’s
method is:
⎡ 0.011 0.004 0.002 0.004
⎤
0.143 + + 0.689s −0.071 − + 0.36s 0.021 + 0.027 +
⎢ s s s s ⎥
⎢ −0.072 − 0.005 − 0.453s 0.129 +
0.007
− 0.519s −0.019 −
0.002
− 0.202s −0.025 −
0.004 ⎥
⎢ ⎥
⎢ s s s s ⎥
GC = ⎢ − 0.121s ⎥
0.00002 0.009 0.009 0.00007
⎢ −0.037 − s
− 0.075s −0.032 −
s
− 0.556s 0.068 +
s
− 0.215s −0.013 −
s ⎥
⎢ 0.021 0.017 0.021 0.055 ⎥
⎢ −0.005 − −0.032 + −0.043 − −0.118 − ⎥
⎣ s s s s ⎦
0.008 0.004 0.002 0.005
−0.191 − − 1.959s 0.171 + + 1.309s 0.024 + 0.032 +
s s s s
Fig. 16 shows the manipulated variable responses when model
matches, the corresponding servo responses are shown in Fig. 17.
and robustness, and is evidently superior in industrial practical
The ISE values of the comparison methods are also tabulated in
applicability.
Table 5. The manipulated variable behaviors for a unit step per-
turbation in d when r1 = r2 = r3 = r4 = 0 are shown in Fig. 18 and
the corresponding responses are shown in Fig. 19. It can be seen Acknowledgements
that the proposed method can greatly reduce the number of con-
trollers while still offer good control performance compared with The authors would like to acknowledge the financial support of
the other two methods, thus leading to large economic benefits in the National Grand Fundamental Research 973 Program of China
real process industries. (Grant 2007CB714300), the National Natural Science Foundation
A perturbation uncertainty of +20% in the process gains and time of China (61273132) and the Specialized Research Fund for the
delays is inserted into (32) to ensure the robust of the proposed Doctoral Program of Higher Education (Grant 20110010010). The
method, whereas the parameters of controllers remain the same as authors are grateful to the anonymous reviewers for their valuable
the perfect model. Fig. 20 shows the servo response when model recommendations.
mismatches, where the proposed method can still achieve good
performance even when model mismatches. References

[1] E. Davison, Some properties of minimum phase systems and “Squared-down”

5. Conclusion systems, IEEE Transactions on Automatic Control 2 (1983) 221–222.
[2] K.L.N. Sarma, M. Chidambaram, Centalized PI/PID controllers for non-square
This paper proposes a novel method to design the IMC-PID con- systems with RHP zeros, Journal of Indian Institute Science 85 (2005) 201–214.
[3] E.J. Davison, Multivariable tuning regulators: the feed forward and robust con-
troller based on the NPSO algorithm. The method can contribute
trol of general servo mechanism problem, IEEE Transactions on Automatic
to the control of non-square systems significantly which are quite Control 21 (1976) 21–35.
common in modern process industries. By applying the concept [4] J.T. Tanttu, J. Lieslehto, A comparative study of some multivariable PI controller
tuning methods, Intelligent Tuning and Adaptive Control 39 (1991) 357–362.
of EOTF, the MIMO system can be decomposed to a set of sin-
[5] L. Magni, R. Scattolini, Tracking of non-square nonlinear continuous time sys-
gle loops, and then the IMC-PID controller of each single loop tems with piecewise constant model predictive control, Journal of Process
can be obtained consequently. In this paper, the incremental PID Control 17 (2007) 631–640.
with first-order lag filter structure combining NPSO algorithm to [6] C.E. García, D.M. Prett, M. Morari, Model predictive control: theory and practice,
Automatica 25 (1989) 335–348.
acquire the desired controllers is first employed. The method incor- [7] J. Chen, Z.F. He, X. Qi, A new control method for MIMO first order time delay
porates several advantages over usual methods, i.e. fast tracking non-square systems, Journal of Process Control (2011) 538–546.
 Q.B. Jin et al. / Journal of Process Control 23 (2013) 649–663 663

[8] H.P. Huang, J.C. Jeng, C.H. Chang, W. Pan, A direct method for multi-loop PI/PID [17] C.E. Garcia, M. Morari, Internal model control. 2. Design procedure for
controller design, Journal of Process Control 13 (2003) 769–786. multivariable system, Industrial Engineering Chemistry Process Design and
[9] T.N.L. Vu, M. Lee, Independent design of multi-loop PI/PID controllers for inter- Development 24 (1985) 472–484.
acting multivariable processes, Journal of Process Control 20 (2010) 922–933. [18] C.E. Garcia, M. Morari, Internal model control. 3. Multivariable control law
[10] Q. Xiong, W.J. Cai, Effective transfer function method for decentralized control computation and tuning guidelines, Industrial Engineering Chemistry Process
system design of multi-input multi-output processes, Journal of Process Control Design and Development 24 (1985) 484–494.
16 (2006) 773–784. [19] Y.Y. Li, A.D. Sheng, Y.G. Wang, Synthesis of PID-type controllers without para-
[11] J. Kennedy, R.C. Eberhart, Particle swarm optimization, in: International Con- metric models: A graphical approach, Energy Conversion and Management 49
ference on Neural Networks, Perth, 1995. (2008) 2392–2402.
[12] P.K. Triapthi, S. Bandyopadhyay, S.K. Pal, Multi-objective particle swarm opti- [20] Y.H. Tao, Y.X. Yin, L.S. Ge, New Type of PID Controller and its Application,
mization with time variant inertia and acceleration coefficients, Information Mechanical Industry Press, Beijing, 1998.
Sciences 177 (2007) 5033–5049. [21] H. Liu, W.Y. Yuan, D.Q. Jiang, Matrix Theory and Its Application, Chemical Indus-
[13] J.J. Ma, Y.H. Luo, L.D. Pan, Improved NLJ algorithm in close-loop parameter iden- try Press, Beijing, 2003.
tification and filter design, The Journal Of The Beijing University Of Chemical [22] R.A. Seshagiri, M. Chidambaram, Smith delay compensator for multivariable
Technology 30 (2003) 95–98. non-square systems with multiple time delays, Computers and Chemical Engi-
[14] R. Luus, T.H.I. Jaakola, Optimization by direct search and systematic reduction neering 30 (2006) 1243–1255.
of the size of search region, AIChE Journal 19 (1973) 760–766. [23] C. Valchos, D. Williams, J.B. Gomm, Solution to the shell control problem
[15] Q.B. Jin, Z.J. Cheng, J. Dou, L.T. Cao, K.W. Wang, A novel closed loop identification using genetically tuned PID controllers, Control Engineering Practice (2002)
method and its application of multivaria e system, Journal of Process Control 151–163.
22 (2012) 132–144. [24] P.Y. Chen, L.L. Ou, J. Sun, W.D. Zhang, Modified internal model control and its
[16] M. Morari, E. Zafirious, Robust Process Control, Prentice-Hall, New Jersey, 1989. application in non-square processes, Control and Decision 23 (2008) 581–584.