ISA Transactions 60 (2016) 244–253

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Research Article

An optimal PID controller via LQR for standard second order plus time
delay systems
Saurabh Srivastava, Anuraag Misra, S.K. Thakur, V.S. Pandit n
Variable Energy Cyclotron Center, 1/AF, Bidhan Nagar, Kolkata 700064, India

art ic l e i nf o a b s t r a c t

Article history: An improved tuning methodology of PID controller for standard second order plus time delay systems
Received 17 July 2013 (SOPTD) is developed using the approach of Linear Quadratic Regulator (LQR) and pole placement
Received in revised form technique to obtain the desired performance measures. The pole placement method together with LQR is
21 September 2015
ingeniously used for SOPTD systems where the time delay part is handled in the controller output
Accepted 19 November 2015
Available online 4 December 2015
equation instead of characteristic equation. The effectiveness of the proposed methodology has been
This paper was recommended for publica- demonstrated via simulation of stable open loop oscillatory, over damped, critical damped and unstable
tion by Dr. Ahmad B. Rad. open loop systems. Results show improved closed loop time response over the existing LQR based PI/PID
tuning methods with less control effort. The effect of non-dominant pole on the stability and robustness
Keywords: of the controller has also been discussed.
Linear system
& 2015 ISA. Published by Elsevier Ltd. All rights reserved.
PID controller
System matrix
Linear Quadratic Regulator (LQR)
Time delay
Closed-loop

1. Introduction The design techniques based on linear Quadratic Regulator

(LQR) are well known in modern control theory and have been
Proportional-Integral-Derivative (PID) controller, though very widely used in many applications [10–12]. In a recent article Saha
old design, is still one of the favorite and most widely used con- et al. [11] have obtained the PID parameters for second order
troller for many industrial process control applications. This is due systems via LQR using the dominant pole placement technique.
to its simple structure, satisfactory control effect and acceptable However, their approach is applicable only for systems having no
robustness [1–3]. The PID controller is easier to understand due to time delay. Most of the real industrial plants have time delay in
intuitive simplicity of the algorithm and simple meaning of its their transfer function. Since the presence of time delay in a
tuning parameters proportional (K p ), integral (K i ) and derivative control loop is a source of instability and performance degrada-
(K d ). In order to provide good and robust performance these PID tion, it is, therefore, necessary to design the PID controller opti-
parameters are required to be tuned individually to match the mally to achieve good stability. Many researchers have worked on
process dynamics. An improper PID setting results in sluggish, the tuning of controller for the systems having time-delay [13–20]
oscillatory time response and poor robustness. The PID controller with pole placement and mentioned the challenges due to the
presence of exponential term in the characteristics equation which
tuning first proposed by Ziegler–Nichols [4] has been improved by
leads to the infinite roots. They have used different approaches to
several researchers [5]. As the high performance is always desired
design the controller with some limitations. He et al. [10] have
from the controller and due to the availability of fast computa-
proposed an analytical method to tune the PI/PID parameters in an
tional power, tuning methods based on optimization approach
optimal way using LQR techniques with user specified closed loop
have received more attention in the recent years [6,7]. Many
damping ratio and natural frequency for the first order plus time
techniques have been developed and still research is going on for
delay (FOPTD) model. His method is based on the decomposition
better tuning of the PID controller using complex numerical of state equation in two parts one for toL and another for tZ L in
optimization procedures [8,9]. such a way that the state equation for tZ L become independent of
L and then applied the usual LQR approach for obtaining the PID
n parameters for FOPTD. They have compared simulation results of
Corresponding author.
E-mail addresses: [email protected] (S. Srivastava), their method with the gain-phase margin method [21] and pre-
[email protected] (V.S. Pandit). sented much improved results.

http://dx.doi.org/10.1016/j.isatra.2015.11.020
0019-0578/& 2015 ISA. Published by Elsevier Ltd. All rights reserved.
 S. Srivastava et al. / ISA Transactions 60 (2016) 244–253 245

Most of the real plants can be more closely approximated using where P is the symmetric positive definite Riccati coefficient
second order plus time delay (SOPTD) model compared to FOPTD matrix which can be obtained by solving continuous algebraic
model. The SOPTD processes are very rich in dynamics as they Riccati equation
include under damped, critically damped and over damped sys-
AT P þ PA þ Q  PBR  1 BT P ¼ 0 : ð6Þ
tems. Very few tuning rules are available for such processes. He
et al. [10] have also extended their approach for SOPTD systems by From (5) we can write
equating the larger process pole with the derivative term of the
uðtÞ ¼ um ðt þ LÞ ¼ R  1 BT PXðt þ LÞ : ð7Þ
PID controller and then applied the PI tuning approach using LQR
to obtain other two parameters. This approach works satisfactory Here we see that uðtÞ gives the control signal in the whole time
for SOPTD model if the system poles have real roots, but does not horizon of t Z 0, however X(tþL) is not directly known at time t.
provide the optimum parameters of the PID controller as one of With the use of Eqs. (2), (3) and (5), X(tþ L) can be expressed in
the PID parameters is prefixed. This technique cannot be applied terms of X(t) [10]. The optimal control vector uðtÞ for the present
for SOPTD systems with complex poles (such as highly oscillatory case, thus can be written as
processes) of the system as they are always in pairs and cannot be
eliminated with single complex zero of the controller. uðtÞ ¼  R  1 BT PeðAc Þt eAðL  tÞ XðtÞ ; 0 rt o L ; ð8Þ
In the present work we have combined the concept of LQR
based PI/PID tuning method together with the dominant pole uðtÞ ¼  R  1 BT PeðAc ÞL XðtÞ ; t Z L ; ð9Þ
placement approach to derive the PID parameters analytically for where
SOPTD systems. It is shown that the present technique gives a
good closed loop time response for various processes as compared Ac ¼ A  BR  1 BT P : ð10Þ
with the existing PI/PID tuning methods using LQR. In order to The beauty of above mathematical formulation lies in the fact
illustrate the utility of the present technique, simulations per- that the optimal control vector u(t) handles the delay part as given
formed in MATLAB [22] have been presented for different types of by Eqs. (8) and (9). As the system matrix Ac given by Eq. (10) does
SOPTD models such as critically damped and over-damped pro- not contain any time delay for t Z L, one can easily apply the
cesses as well as processes having complex poles. The effect of approach of direct pole placement to get `the desired closed loop
non-dominant pole on the control signal and on the stability of the time performance measures. In order to obtain the optimal feed-
closed loop system has also been discussed. back gain uðtÞ we need to calculate eðAc Þt and eAðL  tÞ . By sub-
stituting um ðtÞ from Eq.(5) into Eq. (3) we have for t Z L,
_
XðtÞ ¼ Ac XðtÞ : ð11Þ
2. LQR based PID controller design for SOPTD processes
The matrix Ac can be determined by setting  the characteristic
In this section we briefly outline the LQR solution for time delay equation of the closed loop system ΔðsÞ ¼ sI  Ac j equal to the
systems formulated by He et al. [10] for the FOPDT models where desired closed loop equation. For example, in the case of FOPTD
the motion equation is reformulated into a first order differential process, where the matrix Ac is a 2  2 matrix, we have
equation that contains no time delay and then the optimal con-  
troller is designed according to the classical control theory. We ΔðsÞ ¼ sI  Ac  ¼ ðs þ p1 Þðs þ p2 Þ ¼ ðs2 þ 2ςcl ωcl s þ ωcl 2 Þ ; ð12Þ
then extend this approach utilizing the dominant pole placement where
technique to find the optimal PID parameters for SOPTD systems. qffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffi
p1 ¼ ζ cl ωcl þ iωcl 1  ζ cl ; p2 ¼ ζ cl ωcl  iωcl 1  ζ cl ;
2 2

2.1. LQR solution for SOPTD systems

with ζcl and ωcl as the desired closed loop damping ratio and
A linear plant with time delay can be represented as natural frequency.
For the SOPTD process the dimension of matrix Ac will be 3  3.
_
XðtÞ ¼ AXðtÞ þ Buðt LÞ t Z 0 ; ð1Þ Utilizing the help of dominant pole placement technique matrix Ac
where A, B, X and L are the state transition matrix, control matrix, can be evaluated in terms of known parameters ζcl and ωcl from
state matrix and the time delay term respectively. For t o L, no the equation
 
control signal will be effective and thus we have Eq. (1) as control sI  Ac  ¼ ðs þ p Þðs þp Þðs þ p Þ
1 2 3
free equation. Control signal will be effective only for t Z L. So by
¼ ðs þ mζ cl ωcl Þðs2 þ 2sζ cl ωcl þ ω2cl Þ ; ð13Þ
decomposing Eq. (1) into two components, one for t o L and other
for t Z L, we have where the location of non-dominant pole p3 ¼ mζ cl ωcl is placed m
_ times away from the real part of the dominant closed loop poles.
XðtÞ ¼ AXðtÞ ; 0 r t o L ; ð2Þ
We call this m as the relative dominance and as per the literature
_ its value should be chosen around 3 or more [2].
XðtÞ ¼ AXðtÞ þ Bum ðtÞ ; t Z L ; ð3Þ
where u ðtÞ ¼ uðt LÞ. Since Eqs. (2) and (3) are now delay free,
m
2.2. Determination of state weighting and Riccati coefficient
one can easily apply the standard LQR approach [12] for delay free matrices
processes to find the optimum control vector um ðtÞ subjected to
the minimization of the cost-function defined by In the case of second order process matrices Q, R and P are
Z 1  generally taken as
J¼ XT ðtÞ Q XðtÞ þ umT ðtÞ R um ðtÞ dt ; ð4Þ 2 3 2 3
0 q1 0 0 p11 p12 p13
6 7 6 7
where Q is the semi positive definite state weighting matrix and R Q ¼ 4 0 q2 0 5; R ¼ ½r ; P ¼ 4 p12 p22 p23 5 : ð14Þ
is the positive definite control weighting matrix. The LQR solution 0 0 q3 p13 p23 p33
gives the optimal control vector um ðtÞ as
In the optimal control it is a standard practice to design reg-
um ðtÞ ¼  R  1 BT PXðtÞ ; ð5Þ ulator by varying Q and keeping R fixed [10,11]. A schematic of
246 S. Srivastava et al. / ISA Transactions 60 (2016) 244–253

PID ControllerC(s) s3 þ ðð2 þ mÞζ cl ωcl Þs2 þ ðωcl 2 þ 2mζ cl ωcl 2 Þs þ mζ cl ωcl 3 :
2
ð23Þ
x2(t) Kp By comparing the coefficients of powers of s from both sides of
Plant G(s) Eq. (23), the elements p13 , p23 and p33 can be obtained as
+
1 (t) K mζ cl ω3cl
(t) (t) x1(t) K + e − sL y(t) p13 ¼ ;
s i 2
s + as + b η
+
+
- ω2cl þ 2mζ 2cl ω2cl  b
s x3(t) p23 ¼ ;
Kd η
ð2 þ mÞζ cl ωcl  a
p33 ¼ : ð24Þ
Fig. 1. Schematic of closed loop system with PID controller. η
The remaining three elements of the matrix P and three ele-
closed loop system with PID controller for SOPTD process is shown
ments of the matrix Q can be obtain by solving Riccati equation Eq.
in Fig. 1.
(6), which gives six equations for six variables in terms of known
The state variables for the present case are
parameters. With some algebraic manipulations we obtain,
XðtÞ ¼ ½x1 ðtÞ x2 ðtÞ x3 ðtÞT ; ð15Þ
mζ cl ω5cl ð1 þ 2m ζ cl Þ
2
p11 ¼ ;
where η
Z
ð2 þ mÞ m ζ cl ω4cl
2
deðtÞ
x1 ðtÞ ¼ eðtÞ dt; x2 ðtÞ ¼ eðtÞ; x3 ðtÞ ¼ ; ð16Þ p12 ¼ ;
dt η
2 ω3cl ðζ cl þ 2m ζ cl þ m2 ζ cl Þ  ab
3 3
with error eðtÞ ¼ rðtÞ  yðtÞ. Here r(t) and y(t) are the reference and
p22 ¼ ;
output signals respectively. η
From Fig. 1, the control signal can be expressed in terms of the
m2 ζ cl ω6cl
2
state variable as q1 ¼ ;
η
uðtÞ ¼ K p x2 ðtÞ þ K i x1 ðtÞ þ K d x3 ðtÞ : ð17Þ
ω4 ð1 þ 4m2 ζ 4cl  2m2 ζ 2cl Þ b2
q2 ¼ cl ;
The transfer function of the PID controller can be express in s η
domain as ω2 ð4ζ 2cl þ m2 ζ 2cl  2Þ þ 2b a2
q3 ¼ cl : ð25Þ
CðsÞ ¼ K p þ
Ki
þ K ds : ð18Þ
η
s
In the case of unity output feedback system such as shown in 2.3. Evaluation of eAðL  tÞ
Fig. 1, if we put the reference signal rðtÞ ¼ 0 ; we have eðtÞ ¼  yðtÞ :
With this condition, the second order transfer function with time In order to obtain the value of eAðL  tÞ we proceed as follows.
delay can be written as h i
eAðL  tÞ ¼ ℓ  1 ðsI AÞ  1
yðsÞ K e  sL  eðsÞ 02 t ¼ Lt
31
GðsÞ ¼ ¼ ¼ ; ð19Þ 1 ðs þ aÞ 1
uðsÞ s2 þ as þ b uðsÞ s sðs þ p Þðs þ p Þ sðs þ p01 Þðs þ p02 Þ
B6 01 02
7C
in which a ¼ 2ζ ol ωol and b ¼ ω2ol , where ζol and ωol are the B6 sþa 1 7C
¼ ℓ  1 B6 0 ðs þ p01 Þðs þ p02 Þ ðs þ p01 Þðs þ p02 Þ 7C
@4 5A
damping ratio and natural frequency of the open loop plant 0 ðs þ p Þðsb þ p Þ s
01 02 ðs þ p01 Þðs þ p02 Þ
respectively. Using Eq. (16) we can express Eq. (19) in terms of 2 0 3
0 0
state variables as f 11 ðtÞ f 12 ðtÞ f 13 ðtÞ
6 f 0 ðtÞ f 0 ðtÞ f 0 ðtÞ 7
x_ 3 ðtÞ ¼  ax3 ðtÞ  bx2 ðtÞ  Kuðt  LÞ : ¼ 4 21 22 23 5 : ð26Þ
0 0 0
f 31 ðtÞ f 32 ðtÞ f 33 ðtÞ
In terms of state-space formulation the derivative of the state t ¼ Lt

variables can be written as Here p01 and p02 are the poles of the open loop system (see Eq.
2 3 2 32 3 2 3 (19)) given by
x_ 1 ðtÞ 0 1 0 x1 ðtÞ 0
6 x_ ðtÞ 7 6 76 x ðtÞ 7 6 7 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
4 2 5¼40 0 1 54 2 5 þ 4 0 5 uðt  LÞ : ð20Þ a  a2  4b a þ a2  4b
p01 ¼ ; p02 ¼ ; ð27Þ
x_ 3 ðtÞ 0 b a x3 ðtÞ K 2 2
0 0 0 0 0 0 0 0
Using partial fraction approach f 11 ,f 12 ,f 13 , f 21 ,f 22 ,f 23 ,f 31 ,f 32 and
Comparing Eq. (20) with Eq. (1), it is straightforward to 0
f 33 can be evaluated as
α ¼ r  1 K 2 obtain matrices A and B as
2 3 2 3 0
f 11 ðL  tÞ ¼ 1 ;
0 1 0 0
6 7 6 7 0 p02 e  p01 ðL  tÞ p01 e  p02 ðL  tÞ a
A¼40 0 1 5; B ¼ 4 0 5 : ð21Þ f 12 ðL  tÞ ¼  þ ;
p01 ðp01  p02 Þ p02 ðp01  p02 Þ b
0 b a K
0 e  p01 ðL  tÞ e  p02 ðL  tÞ 1
Using Eqs. (10), (14) and (21) we have f 13 ðL  tÞ ¼  þ ;
p01 ðp01  p02 Þ p02 ðp01  p02 Þ b
  0
 s 1 0  f 21 ðL  tÞ ¼ 0 ;
   

sI Ac  ¼  0 s 1 : ð22Þ 0  p02 e  p01 ðL  tÞ p01 e  p02 ðL  tÞ
  f 22 ðL  tÞ ¼ þ ;
 η p13 b þ η p23 s þ a þ η p33  ðp01  p02 Þ ðp01 p02 Þ
0 e  p01 ðL  tÞ e  p02 ðL  tÞ
where η ¼ r  1 K 2 : Now from Eqs. (22) and (13) we have f 23 ðL  tÞ ¼  þ ;
ðp01  p02 Þ ðp01  p02 Þ
s3 þ ða þ η p33 Þs2 þ ðb þ η p23 Þs þ η p13 ¼ 0
f 31 ðL  tÞ ¼ 0 ;
 S. Srivastava et al. / ISA Transactions 60 (2016) 244–253 247

2 3 2 3
0 be  p01 ðL  tÞ be  p02 ðL  tÞ p13 T f 11 ðtÞ f 12 ðtÞ f 13 ðtÞ
f 32 ðL tÞ ¼  ;
ðp01  p02 Þ ðp01  p02 Þ 6 7 6 7
¼ r  1 K 4 p23 5 4 f 21 ðtÞ f 22 ðtÞ f 23 ðtÞ 5
0 p01 e  p01 ðL  tÞ p02 e  p02 ðL  tÞ p33 f 31 ðtÞ f 32 ðtÞ f 33 ðtÞ
f 33 ðL tÞ ¼  : ð28Þ
ðp01  p02 Þ ðp01  p02 Þ 2 0 0 0 32 3
f 11 ðL  tÞ f 12 ðL  tÞ f 13 ðL  tÞ x1 ðtÞ
6 0 0 0 76 7
 4 f 21 ðL  tÞ f 22 ðL  tÞ f 23 ðL  tÞ 54 x2 ðtÞ 5 : ð31Þ
2.4. Evaluation of eðAc Þt 0 0 0
f 31 ðL  tÞ f 32 ðL  tÞ f 33 ðL  tÞ x3 ðtÞ

Using Eqs. (10), (14) and (21) we have By comparing the coefficients of x1 ðtÞ, x2 ðtÞ and x3 ðtÞ in Eqs. (17)
2 3 and (31), one can easily obtain the PID parameters for 0 r t oL ;as
0 1 0
6 1 7
Ac ¼ ðA  BR  1 BT PÞ ¼ 4 0 0 5: X
3
0
X
3
0
K i ðtÞ ¼ r  1 K p13 f 1i ðtÞf i1 ðL  tÞ þ p23 f 2i ðtÞf i1 ðL  tÞ
γ β α
i¼1 i¼1
!
where X
3
0
þ p33 f 3i ðtÞf i1 ðL  tÞ ;
γ ¼ η p13 , α ¼ a þ η p33 and β ¼ b þ η p23 . i¼1
Now
h i X
3
0
X
3
0
K p ðtÞ ¼ r  1 K p13 f 1i ðtÞf i2 ðL  tÞ þp23 f 2i ðtÞf i2 ðL  tÞ
eðAc Þt ¼ ℓ  1 ðsI  Ac Þ  1
i¼1 i¼1
0 2 2 31 !
s þ αs þ β sþα 1 X
3
0
B 6 þ p33 f 3i ðtÞf i2 ðL  tÞ ;
s 7 C
1
¼ ℓ  1 @ 4 γ s2 þ α s 5A
sI  Ac  i¼1
 sγ  sβ  γ s 2
X
3 X
3
2 3 K d ðtÞ ¼ r  1 K p13
0
f 1i ðtÞf i3 ðL  tÞ þp23
0
f 2i ðtÞf i3 ðL  tÞ
f 11 ðtÞ f 12 ðtÞ f 13 ðtÞ
6 7 i¼1
!
i¼1
¼ 4 f 21 ðtÞ f 22 ðtÞ f 23 ðtÞ 5 : ð29Þ
X
3
0
f 31 ðtÞ f 32 ðtÞ f 33 ðtÞ þ p33 f 3i ðtÞf i3 ðL  tÞ ; ð32Þ
i¼1
Using Eq. (13) in Eq. (29) and with some algebraic manipula-
tions, it is straightforward to get f11, f12, f13, f21, f22, f23, f31, f32 and f33
as 2.6. Evaluation of PID parameters for t ZL

X
3
p2  α pi þ β Similarly using Eqs. (9) and (30) the optimal control u(t) for
f 11 ðtÞ ¼ i
e  pi t
;
i¼1
Di t Z L can be evaluated as
X3
 pi þ α  p i uðtÞ ¼  R  1 BT PeAc L XðtÞ ;
t
f 12 ðtÞ ¼ e ; 2 3 2 32 3
i¼1
Di p13 T f 11 ðLÞ f 12 ðLÞ f 13 ðLÞ x1 ðtÞ
X 6 7 6 f 23 ðLÞ 7 6 7
3
1  pi t ¼ r  1 K 4 p23 5 4 f 21 ðLÞ f 22 ðLÞ 5 4 x2 ðtÞ 5 : ð33Þ
f 13 ðtÞ ¼ e ;
D p33 f 31 ðLÞ f 32 ðLÞ f 33 ðLÞ x3 ðtÞ
i¼1 i
X3
 γ  pi A comparison of coefficients of x1 ðtÞ, x2 ðtÞ and x3 ðtÞ in Eqs. (33)
t
f 21 ðtÞ ¼ e ;
i¼1
Di and (17) gives the PID parameters for t Z L as
 
X
3
p 2 α p K i ¼ r  1 K p13 f 11 ðLÞ þ p23 f 21 ðLÞ þ p33 f 31 ðLÞ ;
i i  pi t
f 22 ðtÞ ¼ e ;  
i¼1
Di K p ¼ r  1 K p13 f 12 ðLÞ þp23 f 22 ðLÞ þp33 f 32 ðLÞ ;
 
X3
 pi  pi K d ¼ r  1 K p13 f 13 ðLÞ þp23 f 23 ðLÞ þp33 f 33 ðLÞ : ð34Þ
t
f 23 ðtÞ ¼ e ;
Di
i¼1 Note that for L ¼0, the matrix elements fij ¼1 when i¼ j and
X3
γ pi  pi fij ¼0 for ia j and Eq. (34) leads to the optimal PID parameters for
t
f 31 ðtÞ ¼ e ;
Di systems having no time delay.
i¼1
X3
β pi  γ  pi t
f 32 ðtÞ ¼ e ;
i¼1
Di 3. Results and discussion
X
3
p2 i  pi t
f 33 ðtÞ ¼ e ; ð30Þ In order to demonstrate the application of the PID tuning
i¼1
Di
methodology proposed in this paper, we now present simulation
where results for different processes performed using MATLAB. We have
D1 ¼  ðp1 p2 Þðp3  p1 Þ ; considered the examples of under damped, critically damped and
over damped SOPTD processes. The present day control challenge
D2 ¼  ðp1 p2 Þðp2  p3 Þ ; is to design a controller to tune unstable and highly oscillatory
processes [3]. Therefore, in Examples 4 and 5 we have discussed
D3 ¼  ðp3 p1 Þðp2  p3 Þ : two plants; one with unstable open loop time response and other
with highly oscillatory behavior.
2.5. Evaluation of PID parameters for 0 r t o L ;
3.1. Example 1: non-minimum phase process
Using Eqs. (28) and (30) in Eq. (8) the optimal value of control u
(t) for 0 rt o L ; can be expressed as In this example we will consider an over damped SOPTD model
of a non-minimum phase process and evaluate the PID parameters
1 T Ac t AðL  tÞ
uðtÞ ¼  R B Pe e XðtÞ ; for 0 r t o L and t ZL. The closed loop time response is compared
248 S. Srivastava et al. / ISA Transactions 60 (2016) 244–253

with the previously developed LQR-based PI/PID tuning method

1.2
[10], where the derivative term of the PID controller for SOPTD
process was set equal to one of the process pole and thus is not
1.0
obtained in an optimum way. For fair comparison, the desired
closed loop damping ratio ζcl ¼0.8 and natural frequency
0.8
ωcl ¼0.793 rad/s are taken same in the simulation. The non- Time varying PID with m=6
dominant pole is placed 6 times away from the desired domi- 0.6 Constant PID with m=6

y(t)
nant real poles i.e. m ¼6. Constant PID with m=5
The transfer function of the non-minimum phase process is 0.4 Constant PID with m=4
Constant PID with m=3
1s
P1 ¼ ð35Þ 0.2 Constant PID with m=10
ð1 þ sÞ2 ð2 þ sÞ Constant PID with m=50
and its corresponding over damped SOPTD model is 0.0 Constant PID with m=200
1
P1 ¼ e  1:64s ð36Þ -0.2
s2 þ 3s þ 2 0 10 20 30 40 50 60 70
Matrices A and B can be obtained from Eqs. (36) and (21) as Time(s)
2 3 2 3
0 1 0 0 Fig. 3. Comparison of time responses of time varying PID with constant PID
6 7 6 7 parameters for process P1 at ζ cl ¼ 0:8 and ωcl ¼ 0:793 rad=s at different m.
A ¼40 0 1 5 and B ¼ 4 0 5 : ð37Þ
0  2 3 1 PID parameters are very high at the beginning (t ¼ 0s), followed by
Using Eqs. (24) and (25), matrices P and Q with R ¼[1] can be a decrease with t up to t ¼ 1:64 s and then remain constant
evaluated as thereafter. Note that design of PID controller at higher value of m
2 3 leads to lower values for all the PID parameters for t Z 1:64 s
5:7158 0 0
6 7 whereas the situation is completely reverse in the case of
Q ¼4 0 1:4889 0 5; t o1:64 s.
0 0 9:8290 The time response of the step input for process P1 with 20%
2 3
13:0394 12:1288 2:3908 disturbance at t ¼ 40 s is shown in Fig. 3 by solid black line. The
6 7
P ¼ 4 12:1288 19:2785 3:4540 5 : ð38Þ observed behavior of the closed loop time response during the
2:3908 3:4540 2:0732 initial period is due to the high values of initial PID parameters,
which are responsible for the decrease in the system rise time and
The eigen values of matrices P and Q are hence enhancement in the overshoot. Note that PID parameters
2 3 2 3
1:4050 5:7158 between 0 rt o L are time varying and large initially. This leads to
6 7 6 7 a comparatively larger control efforts and may cause the actuator
eig ½P ¼ 4 3:6579 5 and eig ½Q  ¼ 4 1:4889 5 : ð39Þ
saturation in some cases. It is also difficult to implement them
29:3282 9:8290
practically, particularly in analog domain. It is obvious that a
The positive eigen values of matrices P and Q indicate that the choice of constant PID parameters throughout eases the practical
positive definite condition of LQR is satisfied. Finally, the PID implementation, needs low control effort and maintains the state
parameters for 0 r t o 1:64 s can be obtain using Eq. (32). The optimality for all values of t Z L. The plots of time response using
time varying PID parameters are plotted in Fig. 2. The PID para- only constant PID parameters throughout (i.e. for t Z 0) obtained
meters for t Z1:64 s can be calculated using Eq. (34) as for t Z 1:64 s using Eq. (34) for various values of relative dom-
½K p Ki K d  ¼ ½0:6984 0:4602 0:1543 : ð40Þ inance m are also shown in Fig. 3 for comparison. In the simulation
all other parameters are kept constant.
Fig. 2 shows the variation of PID parameters used in the It can be seen from Fig. 3 that as the value of m is decreased
simulation at m ¼ 6. For comparison, PID parameters evaluated at from 6 to 3, the PID controller based on constant parameters tries
m ¼ 3 and m ¼ 10 are also shown. It is clear that values of all the to cope with the actual time varying PID controller and produces
overshoot with improved rise time. An increase in the value of m
6.0 reduces the overshoot but at the same time increases the rise time.
t = 1.64 s At higher values of m, say around m ¼50, this effect saturates and
m = 10
further increase in m has no significant effect on the time
Kp

3.0 m=6
m=3 response. Thus the choice of m depends upon a particular
requirement whether one needs fast rise time or less overshoot. In
0.0 our experience a good choice for m is between 3 and 10.
0.511.522.533.544.5
3.0 It is interesting to point out here that increase in the rise time
with m in the delayed processes is just opposite to the LQR based
Ki

2.0 PID tuning with no delay [11], where an increase in m decreases

1.0 the rise time of the closed loop time response. For a given process,
0.0 our simulation results indicate that an increase in the value of m
0.5
1.0
1.5
2.0
2.5
3.0
3.54.5 results in the lower values of PID parameters as shown in Fig. 2
1.5
and thus a reduction in the control effort. This fact can also be
Kd

1.0 explained using Eqs. (9), (24) and (25) where an increase in m
0.5 increases the value of matrix elements of P. This finally causes
0.0 reduction in the control effort u(t) due to the presence of the term
0.0 1.0 2.0 3.0 4.0 5.0 PeðAc ÞL which decreases with increase in the value of elements of
Time (s) matrix P. Note that in the case of delay free process eðAc ÞL ¼ 1 and u
Fig. 2. PID parameters Kp, Ki and Kd as a function of time. (t) is proportional to matrix P.
 S. Srivastava et al. / ISA Transactions 60 (2016) 244–253 249

1.0
a 1.0
y(t)

y (t)
0.5
Present 0.5 Present
He et. al. He et. al.
0.0

2.5 0.0

2.0 1.0
b
u(t)

u (t)
1.5
0.5
1.0

0.5
0 10 20 30 40 50 60 70 80 0.0
0 20 40 60 80 100 120 140
Time (s)
Time (s)
Fig. 4. Time response and controller response for process P1 with 20% disturbance
at t ¼ 40 s. Fig. 5. Time response and controller response for higher order process P2 with 20%
disturbance at t ¼ 70 s.

Table 1
Closed loop performance measures. Fig. 4(b) compares the control energy required to achieve good
closed loop time response. Since the cost-function is optimized
Processes Kp Ki Kd ζcl ωclL m %OS Tr(s) Ts(s)
properly in the present method, the required control energy is also
P1 (He et al.) P1 0.6138 0.5561 1 0.8 1.3 14 4.5 15 less. PID parameters and closed loop performance measures such
(Present) 0.6984 0.4602 0.1543 0.8 1.3 6 4 4.5 8 as percentage overshoot (%OS), settling time (Ts) and rise time (Tr)
P2 (He et al.) P2 0.2873 0.0851 1.0753 0.9 1.3 4 16 23
are presented in Table 1 for comparison.
(Present) 0.3919 0.0912 0.2834 0.9 1.3 4 0 12 20
P3 (He et al.) P3 1.7342 2.1759 1 0.98 0.4 35 1.2 8
(Present) 3.7238 1.9858 1.6867 0.98 0.4 4 15 1.1 5
3.2. Example 2: higher order process

Observing the simulation results shown in Fig. 3, it appears that Now we consider a higher order process [1] given by
the time varying part of PID parameters though, improves the rise
time of the closed loop time response, but at the same time it 1
P2 ¼ : ð41Þ
produces substantial overshoot as compared to the cases where ð1 þsÞ8
only constant PID parameters are used. As it will be easy to
implement practically, we therefore, in subsequent examples The corresponding over damped SOPTD model of this process is
consider only constant value of PID parameters evaluated for t Z L
using Eq. (34). 0:3360
P2 ¼ e  4:3s : ð42Þ
To show the effectiveness of the present method, we now s2 þ 1:3878s þ 0:3360
compare our results with the previously developed LQR based
PI/PID tuning method at same values of closed loop damping ratio The controller parameters for this model calculated using the
and natural frequency. The optimal PID controller for process P1 method presented by [10] where one of the pole is taken equal to
with ζ cl ¼ 0:8 ; ωcl ¼ 0:793 rad=s and m ¼ 6 obtained for t Z 1:64 s the Kd are given in Table 1. The optimal PID controller for the
is above process using present method for m ¼ 4 and t Z 4:3 s is
0:4602
C 1 ½present ¼ 0:6984 þ þ 0:1543 s ; 0:0912
s C 2 ¼ 0:3919 þ þ 0:2834s : ð43Þ
s
and the PID controller used in Ref.[10] is
In both the cases same values for desired closed loop para-
0:5561
C 1 ½He et al: ¼ 0:6138 þ þ 1:0 s : meters ζ cl ¼ 0:9 ; ωcl ¼ 0:3 rad=s are used. The eigen values of
s
matrices P and Q for this model are also positive and therefore
Fig. 4(a) compares the step responses with 20% disturbance at
satisfying the condition of LQR. Fig. 5(a) compares of the step
t ¼ 40 s. It is easy to observe that the present method gives very
response with 20% disturbance at t ¼ 70 s. Due to optimal design
less overshoot, only 4% as compared to the 14% of the earlier
of all the three parameters, overshoot is almost negligible with an
method. Note that the value of derivative gain in controller C 1 ½He
improvement in rise time and disturbance rejection. The control
et al: is larger than the controller C 1 ½present. From the simula-
tion it is clear that a choice of larger derivative gain not necessarily effort required for desired time response, plotted in Fig. 5(b), is
reduced the overshoot. The main reason for the reduction in also slightly less in the present optimization method.
overshoot is the optimal tuning of derivative parameter Kd in the In order to test the present method with large time delay, we
present case, which was taken as one of the real pole of the open have varied the time delay of process P2 from 4.3 s to 44.3 s in
loop system in earlier case and thus, was not optimum one. Due to steps of 10 s and performed simulations. In all the cases fixed
the optimal design of all the three parameters in the present value of ωcl L ¼ 1:3 is used. Simulation results indicate that the
method, there is almost 46% reduction in the settling time toge- satisfactory closed loop time response. As usual, we found the
ther with a substantial reduction in the overshoot. response time to become slow as the time delay L increases.
250 S. Srivastava et al. / ISA Transactions 60 (2016) 244–253

1.5 1.2

0.8
1.0

y(t)
y (t)

0.4
0.5 Present
He et. al. 0.0
0.0
5.0
4.0

u(t)
4.0
3.0
u (t)

3.0
2.0
2.0
0 10 20 30 40 50
1.0 Time (s)

0 5 10 15 20 25 30 35 40 Fig. 8. Time response and controller response for process P5 with 20% disturbance
Time (s) at t ¼30 s.

Fig. 6. Time response and controller response for process P3 with 20% disturbance
at t ¼ 20 s.
3.4. Example 4: unstable SOPTD process

Now we consider an unstable plant [23] given by

4.0
1:5
P4 ¼ e  0:3s : ð46Þ
ð0:5s þ 1Þðs  1Þ
y(t)

0.0
With some algebraic manipulation we can easily write Eq. (46)
Present
in standard second order TF as given in Fig. 1 and get the value of
He et. al. a¼ 1, b ¼  2 and K ¼3. The optimal LQR based PID controller
-4.0
obtained with ζ cl ¼ 0:9 ; ωcl ¼ 0:8rad=s and m ¼ 4 is
4.0
0:1688
C4½present ¼ 1:2153 þ þ 0:5682 s : ð47Þ
s
u(t)

0.0
The controller designed by adopting the method of He et al.
[10] taking K d ¼ 2, the larger real system pole with same ζcl and
-4.0
ωcl is given by
0 10 20 30 40 50 0:0824
C 4 ½He et al: ¼ 0:5619 þ þ 2:0 s : ð48Þ
Time (s) s

Fig. 7. Time response and controller response for process P4 with 20% disturbance The time response plotted in Fig. 7 clearly shows the advantage
at t ¼ 30 s. of the proposed method for control of unstable plant dynamics.
Expect for slightly higher percentage overshoot all other closed
3.3. Example 3: critically damped SOPTD process loop performance measures are quite reasonable (T r ¼ 1:8 s, %OS
¼ 150 and T s ¼ 5 s). Simulation results indicate that the range of
Consider a critically damped SOPTD process [5] given by ωclL is limited. In the case of stable system the appropriate range is
ωcl L A ð1:0; 1:5Þ and for the case of unstable systems it is
e  0:2s ωcl L A ð0:1; 0:4Þ.
P3 ¼ : ð44Þ
ð1 þ sÞ2
3.5. Example 5: highly oscillatory SOPTD process
The optimal PID controller designed for ζ cl ¼ 0:98 ; ωcl ¼ 2 rad=s
and m ¼ 4 is Here we consider a SOPTD process with highly oscillatory open
loop response [24] with transfer function given by
1:9858 1
C 3 ¼ 3:7238 þ þ 1:6867s : ð45Þ P5 ¼ e  0:1s : ð49Þ
s s2 þ s þ5

Fig. 6(a) shows the comparison of step responses of the criti- Our aim is to design a controller with very small percentage
cally damped SOPTD process with 20% disturbance at t ¼ 20 s. overshoot and settling time. Since the roots of the process are
Clearly, the present method gives an improved performance. Both complex, the method used by He et al. for SOPTD process cannot
the overshoot and settling time are improved by considerable be applied here. The controller with ζ cl ¼ 0:9 ; ωcl ¼ 1:5 rad=s and
m ¼ 4 is
amount (see Table 1) with slight improvement in the rise time and
disturbance rejection time. Although the present tuning method 5:8325
C 5 ¼ 3:9434 þ þ 3:6339s : ð50Þ
takes slightly more control signal initially (Fig. 6(b)), but one can s
easily verify that the total control cost is almost identical in both Simulation result presented in Fig. 8 shows a remarkable time
the cases. response (T r ¼ 1:5 s, %OS ¼ 2 and T s ¼ 3:3 s) for process P5.
 S. Srivastava et al. / ISA Transactions 60 (2016) 244–253 251

a a
1.2 1.2

0.8 IAE 0.8 IAE

y(t)

y(t)
ISE ISE
0.4 ITAE 0.4 ITAE
ITSE ITSE
Present Present
0.0 0.0

3.2
P1
X Data
b 1.6 P2 X Data b

u(t)
2.4 1.2
u(t)

1.6 0.8

0.8 0.4

0.0 0.0
0 10 20 30 40 50 60 70 80 0 20 40 60 80 100 120 140
Time (s) Time (s)

a 1.2 a
1.2
0.8 IAE
IAE
y(t)

y(t)

0.8 ISE
ISE ITAE
ITAE 0.4 ITSE
0.4 ITSE Present
Present
0.0 0.0
7.2 P3 X Data b 10 P5 X Data
b
5.4 8
u(t)

u(t)

3.6 6
1.8 4

0.0 2
0
0 10 20 30 40 0 10 20 30 40 50
Time (s) Time (s)
Fig. 9. Time response and controller response for processes P1, P2, P3 and P5 obtained using methods based on different time domain performance measures.

3.6. Comparison with other time domain tuning methods cases the optimization started with same initial value of PID
parameters equal to 0.3. i.e. K p ¼0.3, K i ¼0.3, K d ¼0.3. Results
In order to check the relative merits and demerits of the pre- are compared in Fig. 9. It is easy to observe that present
sent method, simulations have been performed for processes P1, method gives overall satisfactory closed loop time response.
P2, P3 and P5 using other time domain tuning methods [1,3] such In other cases the response time is fast but with substantial
as Integral of Square Error (ISE), Integral of Time Square overshoot and oscillations. We have also calculated the
Error (ITSE), Integral of Absolute Error (IAE), Integral of Time control energy using the square of the MATLAB function norm
Absolute Error (ITAE). We have used the fmincon() function (u(t),2). Except for process P3, where ITAE and ITSE require
of the MATLAB [22] optimization toolbox for finding the sets slightly less control energy, controllers designed with present
of optimized PID controller parameters subject to a given method need comparatively less control energy for all
time domain performance index based cost-function. In all the other cases.
252 S. Srivastava et al. / ISA Transactions 60 (2016) 244–253

shown by black solid line. It can be readily seen that an increase in

the value of time Lm from the designed value of L ¼ 2 s, causes an
overshoot in the time response and finally leads to the oscillation
if the value of Lm becomes larger. In contrast, the mismatched
value of Lm less than L is responsible for the increase in the rise
time as well as in the settling time and thus making the response
sluggish.
Since in the present LQR based PID method, we have an extra
tuning factor that is the value of relative dominance m which one
can utilized to improve the robustness of the time response in the
case of a mismatch in the delay time. To explain the effect of m on
robustness of the controller we have designed another optimal
LQR-PID controller keeping all the parameter same except the
value of m. The optimal PID controller for m ¼ 10 is

0:1586
C m10 ¼ 0:0658 þ þ 0:0029s : ð53Þ
s

A comparison of time response curves for cases Lm ¼ 4 s, m ¼ 3

and Lm ¼ 4 s, m ¼ 10 clearly indicates that a controller designed at
higher m shows less overshoot and thus will be more robust in the
case of mismatch between the process delay time and the delay
time at which the controller is designed. However, the penalty one
has to pay is the increase in the rise time.

4. Conclusion

In this paper an improved design methodology of PID con-

troller for standard SOPTD system has been developed by com-
Fig. 10. The plots of (a) time response and (b) controller response for under bining the optimal approach of LQR and the dominant pole pla-
damped SOPTD process P6 with 20% disturbance at t ¼ 30 s. (c) Time response of
cement technique. The proposed tuning method allows more
process P6 at various mismatched delay time.
flexible pole placement, which results in better time response. The
3.7. Example 6: robustness test PID parameters have been calculated analytically using user
defined closed loop damping ratio and natural frequency. It is
Most of the real plants operate in a wide range of operating demonstrated by simulation that present tuning methodology
conditions and it is required that the controller must be able to gives improved closed loop time response with less control effort
stabilize the system with slight change in the operating conditions. as compared to the earlier developed LQR based PI/PID tuning
In such situation, the robustness of the closed loop system is an method. Simulation results indicate that present method works
important feature. The purpose of studying this process is to check well for most of the SOPTD models such as under-damped, criti-
the robustness property of the optimal LQR-PID controller when cally-damped, over-damped, unstable and highly oscillatory pro-
there is a mismatch between the delay time of the process and the cesses. It is observed from the simulation that most appropriate
delay time for which the PID controller is designed. We consider range of ωcl L for stable SOPTD processes is ωcl L A ð1:0; 1:5Þ and for
an under damped SOPTD process given by the unstable SOPTD process is ωcl L A ð0:1; 0:4Þ. A comparison of
9 simulations results with other time domain performance indices
P6 ¼ e  2s : ð51Þ indicates that the present methods gives an overall better closed
s2 þ 1:2s þ 9
It is easy to observe that the open loop system poles are loop time response with comparatively less control effort.
complex. We have designed the PID controller with closed loop It is observed that the location of non-dominant pole (value of
parametric demand of ζ cl ¼ 0:98 ; ωcl ¼ 2 rad=s. The optimal LQR m) affects the closed loop time response provided all others
based PID controller obtained with m ¼ 3 is parameters are kept constant. An increase in the value of m,
increases the rise time with a substantial control on the overshoot.
0:1913
C 6 ¼ 0:0979 þ þ 0:0111s : ð52Þ This observed behavior of the closed loop time response with m in
s
the case of processes with time delay is completely opposite to the
Fig. 10(a) shows the closed loop step response of the under
cases of delay free processes. A slightly higher value of m adds an
damped SOPTD process P4 with 20% disturbance at t ¼ 30 s and
extra robustness to the closed loop time response in the case of
the corresponding control effort is plotted in Fig. 10(b). It can be
mismatch between the process delay time and the delay time at
readily seen that the stabilization of load disturbance by present
controller is quite satisfactory. which the controller is designed. The proposed analytical tuning
We have also studied the robustness of the present controller method to obtain optimum PID parameters for SOPTD process will
by varying the mismatched delay time Lm from 0.5 s to 4 s covering be helpful for the on-line applications. We like to point out here
both sides of the actual delay L ¼ 2 s for which the controller is that the present approach cannot be applied to integrating pro-
designed. Results of simulation are presented in Fig. 10(c) for cesses because they cannot be represented in the form of standard
comparison. The time response with designed parameters is second order transfer function.
 S. Srivastava et al. / ISA Transactions 60 (2016) 244–253 253

References [12] Naidu DS. Optimal control system. CRC Press, USA; 2003.
[13] Cai. GP, Huang JZ, Yang SX. An optimal control method for linear systems with
time delay. Comput Struct 2003;81:1539–46.
[1] Astrom KJ, Hagglund T. PID controller, theory design and tuning. NC: Instru- [14] Panda RC, Yu CC, Huang HP. PID tuning rules for SOPDT systems: review and
ment Society of America Research Triangle Park; 1995. some new results. ISA Trans 2004;43(2):283–95.
[2] Norman SN. Control systems engineering. USA: John, Wiley & Sons, Inc.; 2011. [15] Michiels W, Engelborghs K, Vansevenant P, Roose D. Continuous pole place-
[3] Vilanova R, Visioli A. PID control in the third millennium. London: Springer; ment for delay equation. Automatica 2002;38:747–61.
2012. [16] Yeung KS, Wong WT. Root-locus plot of systems with time delay. Electron Lett
[4] Ziegler JG, Nichols NB. Optimum settings for automatic controllers. Trans 1982;18(11):480–1.
ASME 1942;64:759–68. [17] Engelborghs K, Dambrine M, Roose D. Limitations of a class of stabilization
[5] Dey C, Mudi KR. An improved auto-tuning scheme for PID controllers. ISA methods for delay systems. IEEE Trans Autom Control 2001;46(2):336–9.
Trans 2009;48(4):396–409. [18] Suh Il H, Bien Z. A root-locus technique for linear systems with delay. IEEE
[6] Pan I, Das S, Gupta A. Tuning of an optimal fuzzy PID controller with stochastic Trans Autom Control 1982;27(1):205–8.
algorithms for networked control systems with random time delay. ISA Trans [19] Michiels W, Vyhlídal T, Zítek P. Control design for time-delay systems based
2011;50:28–36. on quasi-direct pole placement. J Process Control 2010;20(3):337–43.
[7] Shabani H, Vahidi B, Ebrahimpour M. A robust PID controller based on [20] Sen M, De La. On pole-placement controllers for linear time-delay systems
imperialist competitive algorithm for load-frequency control of power sys- with commensurate point delays. Math Probl Eng 2005;1:123–40.
tems. ISA Trans 2013;52:88–95. [21] Ho WK, Hang CC, Cao LS. Tuning of PID controllers based on gain and phase
[8] Das S, Pan I, Halder K, Das S, Gupta A. LQR based improved discrete PID margin specications. Automatica 1995;31(3):497–502.
controller design via optimum selection of weighting matrices using fractional [22] MATLAB toolbox: user's guide. Mathworks, Inc.; 2009.
order integral performance index. Appl Math Model 2013;37:4253–68. [23] Xiang C, Wang QG, Lu X, Nguyen LA, Lee TH. Stabilization of second-order
[9] Zhao YM, Xie WF, Tu XW. Performance-based parameter tuning method of unstable delay processes by simple controllers. J Process Control 2007;17:675–
model-driven PID control systems. ISA Trans 2012;51:393–9. 82.
[10] He JB, Wang QG, Lee TH. PI/PID controller tuning via LQR approach. Chem Eng [24] Wang QG, Zhang Z, Astrom KJ, Chek LS. Guaranteed dominant pole placement
Sci 2000;55(13):2429–39. with PID controllers. J Process Control 2009;19(2):349–52.
[11] Saha S, Das S, Das S, Gupta A. A conformal mapping based fractional order
approach for sub-optimal tuning of PID controllers with guaranteed dominant
pole placement. Commun Nonlinear Sci Numer Simul 2012;17:3628–42.