Академический Документы
Профессиональный Документы
Культура Документы
Modeling
Volume 4, Issue 1 2009 Article 29
∗
CLRI (CSIR), panda@clri.res.in
†
Easwari Engg College, sobana subramani@yahoo.co.in
Copyright
2009
c The Berkeley Electronic Press. All rights reserved.
Design of a PID Controller from a Predictive
Control Algorithm∗
Rames Chandra Panda and Sobana Subramaniam
Abstract
Over decades, dynamic matrix control (DMC) was used as an advanced control strategy in
processing industries that yields good performance, especially in the presence of uncertainties and
undesirable noises. As the implementation of PID algorithm in digital platform is easier, an equiv-
alent PID controller is synthesized where proportional gain (KC ) of a controller is obtained by
approximating a dynamic matrix control algorithm. Integral time (τ i ) and derivative time (τ d )
are obtained from process model and step response coefficients. The performances of the equiv-
alent PID controller are tested on different process models (dynamics) and closed loop responses
are analyzed. This performance is compared with the DMC-SISO response. The designed PID
controller is robust and stable. Results obtained from a computer simulation are presented.
KEYWORDS: DMC, PID controller, prediction horizon, control horizon, model based control
∗
Please send correspondence to panda@clri.res.in. Scientist, Dept Chem Engg, CLRI, Chennai-
20, India. Fax: +91(44)24911589.
Panda and Subramaniam: PID Equivalent of DMC
1. Introduction
Model based controllers (MBC) provide accurate and precise control around the
operating point and thus, internal model control (IMC) schemes got huge
popularity, and are widely used in industries and in research. Among salient
advantages, IMC offers, (i) easy controller design (inverse of the process
model),(ii) incorporation of process model inside control algorithm, (iii) robust
closed-loop stability etc. Easy implementation of this control strategy (Figure 1)
has been possible by synthesizing equivalent PID structure (namely, IMC-PID).
But, in presence of model uncertainties and undesirable noise, a special case of
IMC, dynamic matrix control (DMC) (Cutler and Ramaker,1979), has proved to
be successful and widely used technique in many continuous process industries
over last few decades. The process model is used to create a prediction horizon
over which an objective function is optimized inside the algorithm. It is highly
efficient and its performance is very satisfactory even in the presence of model
uncertainties. DMC is successfully implemented in many process industries
(Panda1 and Rao, 1994). The main features of DMC algorithm can be accounted
for as mentioned by QinSY and Badgwell (1996) as
- Uses linear step response mode to predict the process behavior.
- A quadratic objective function over a finite prediction horizon is employed
to find performance.
- Future plant outputs are specified to follow the set point as close as
possible.
- Optimal output is calculated to track set point using least square method.
As a model predictive controller (MPC), DMC has shown good
performance in the control of long time-delayed processes (Lundstrom et.al,
1995). MPC controllers are designed under the consideration that the prediction
horizon is larger than the maximum amount of dead time. Since uncertainty is
associated with the estimation of dead-time of a process model which is directly
used in MPC, the performance of MPC depends on the accuracy of the process
model. Moreover, implementation of predictive controls on discrete platform is
costly as it involves computation & solution of an objective function. These
problems can be avoided and performance can be improved by approximating an
MPC and DMC by PID control structure by different methods. Moreover, as PID
controllers are easy to implement and maintain and are robust in nature, a PID
equivalent of predictive controller is sought. Most of the process industries use
PID loops till today. Thus the objective of this paper is to synthesize PID
controller parameters from DMC strategy.
Saeki (2006) synthesized a PID structure for a H∞ controller. PID controller has
found wide application in industries because of their simple structures, ease in
implementation and relatively good robustness against process uncertainties.
There are several methods to tune parameters of this controller. Haeri (2005)
proposed a PID controller approximating a dynamic matrix control in such a way
that both of manipulated and controlled outputs satisfy some criterion. The
accuracy of the approximation depends on derived PID as well as sourced DMC
parameters. The method involves complex computations.
L L
E E y
R R
G C GP y G C GP
‐ U
U ‐
Gm
‐
(a) (b)
Figure 1. Schematic diagrams of (a) model based control scheme and (b)
equivalent PID strategy.
2 Problem Definitions
Dynamic matrix control uses a time domain step response model of the process to
calculate the future changes of the manipulated variables that will minimize some
objective cost function or performance index. In this approach one would like to
have NP future output response (step response coefficients) match some optimum
http://www.bepress.com/cppm/vol4/iss1/29 2
DOI: 10.2202/1934-2659.1362
Panda and Subramaniam: PID Equivalent of DMC
Let’s consider a process having transfer function (GP) and load transfer function
(GL) in the following form
q1s + q2 s 2 + q3 s 3 + ... + qm s m − DP s
GP ( s ) = e (1)
1 + p1s + p2 s 2 + p3 s 3 + ... + pn s n
with DP as process time delay, DL as load transfer function-time delay and n>m.
The openloop step response for set-point change case of DMC can be given by
∞
yk +1,OL = ∑ ai Δuk +1−i + d k +1 (3)
i =1
Where ai s are step response coefficients, u is input and d is response due to load
change. The closed-loop response can be modeled over a prediction horizon,
h=1,2,...,p, as
h
yk + h ,CL = yk + h ,OL + ∑ ai Δuk + h−i (4)
i =1
calculated values of manipulated variable, the value that gives the smallest sum of
squares error between the set point and the predicted value of the controlled
variable is implemented. Thus, future control moves in DMC controller are given
as (Luyben, 1992)
Δu = K DMC E (5)
K DMC = ( AT A + WI ) AT
−1
(6)
E is the prediction error vector (differences between desired & model output and
output due to disturbances) defined as
E = Yd − Ym − ε (7)
Thus
Δu = ( AT A + WI ) −1 AT E (8)
⎡ a1 0 0 ...0 ⎤
⎢a a1 0 ...0 ⎥
⎢ 2 ⎥ ⎡ a11 a12 ...a1M ⎤
⎢. . . ...0 ⎥ ⎢a a22 ... ⎥
A=⎢ ⎥ B = ⎢ 21 ⎥ (9)
⎢a j a j −1 a j −2 ...a j −M +1 ⎥ ⎢. . ... ⎥
⎢. ⎥ ⎢ ⎥
⎢
. . ...
⎥ ⎣ aP1 aP 2 ...aPM ⎦
⎢⎣ aP aP −1 aP −2 ...aP − M +1 ⎥⎦ PXM
j L +1
Yk + j = ∑ ai Δuk −i + j + ∑ a Δu i k −i + j + aiuk − L−2+ j (10)
i =1 i = j +1
with
http://www.bepress.com/cppm/vol4/iss1/29 4
DOI: 10.2202/1934-2659.1362
Panda and Subramaniam: PID Equivalent of DMC
⎡ Δuk −1 ⎤
⎢ Δu ⎥
⎢ k −2 ⎥
Δ u = ⎢. ⎥ (11)
k
⎢ ⎥
⎢ Δuk − L+1 ⎥
⎢ Δuk − L ⎥
⎣ ⎦ LXN
In Eqn.(7), the term ε is disturbance and the control moves are penalized
by coefficient W. Here M is control horizon, P is prediction horizon and N is
model horizon. A small value should be chosen for W if faster response is
required (Sridhar and Cooper, 1997). Eqn.(8) will be helpful to calculate control
moves.
In case of load disturbances, the step response model can be given by:
n
Or d k = ∑a
i =1
Li ΔLk −i + d 0 (13)
Where aLi s are load-step response coefficients and ΔLk-i are load changes at ith
instant. At instant h, of prediction horizon, p, the output prediction will be
p
d k + h = ∑ aLi + h ΔLk −i + d 0 where h=1,2,…,p (14)
i =1
These load-step response coefficients are related through an, average factor,
generally called, μL such that
where μL is a tuning factor. Based on the cases of setpoint change or load change,
Eqn. (4) & (14) are used to calculate predicted outputs using Eqn. (10) that
requires Eqn. (3) & (15) to find out step response coefficients (ah or aL). These
coefficients are needed to calculate control moves using Eqn(8).
Guidelines for choosing design parameters P(prediction horizon),
M(control horizon), N (model horizon)could be found in (Luyben, 1992). W
(control move weighting coefficient) and α (pole of the reference input filter)
could be found in (Shridhar and Cooper,1997).
Using equation (2), the proportional gain value of PID controller is given by
From which the first element is implemented. The step response coefficients for
2nd order overdamped system
K P e− Ds
GP ( s ) = (16.1)
(τ P1s + 1)(τ P 2 s + 1)
⎡ τ P1 τ P2 ⎤
y (t ) = K P ⎢1 − e − (t − D )/τ P1 + e − ( t − D )/τ P 2 ⎥
⎣ τ P1 − τ P 2 τ P1 − τ P 2 ⎦
The values of the DMC parameters (as per guidelines of Luyben, 1992) used in
simulation to find the linear and nonlinear process response are given in Table-1.
The step response of the discrete form of process transfer function (Eqn. 1) is
given by
http://www.bepress.com/cppm/vol4/iss1/29 6
DOI: 10.2202/1934-2659.1362
Panda and Subramaniam: PID Equivalent of DMC
q0 q q p p p
yk = uk + 1 uk −1 + ... + n uk −n − 1 yk −1 − 2 yk −2 − ... − m yk −m (17)
p0 p0 p0 p0 p0 p0
Assuming that the input and outputs are scaled equally, the sum of coefficients
(ci) for unit step change will be:
Using ramp change, these coefficients (ci) can be found out using following
equations (Eqs. 17.3 & 17.4) in Eqn (17.1)
uk = 0 , uk −1 = 1 , uk −2 = 2 , ..., uk −n = n (17.3)
yk −1 = yk −1 , yk −2 = yk −1 + 1 , ..., yk −m = yk −1 + (m − 1) (17.4)
As the rate of change of ramp is one unit per iterations, the lag for a second order
system can be found out as
2 − c2 − c1' − 2c2'
τP = .T (17.5)
1 − c1 − c2
b2 s 2 + b1s + b0
Thus for a system of GP = (17.5.1)
p2 s 2 + p1s + p0
it is possible to find the numerator and denominator coefficients using recurrence
formula as
T2 −T ln(−c1 ) c2 ± c22 − 4c1
p2 = , p1 = , σ1, 2 = (17.6)
ln(σ1 ) ln(σ 2 ) ln(σ1 ) ln(σ 2 ) 2
b2 = p2 − α T 2 − p1β T , b1 = p1 − β T (17.7)
d2y dy
for a typical second order system, p2 2
+ p1 + y = u or
dt dt
1 1 p1
GP = = we write τ P = p2 and ξ =
2
p2 s + p1s + 1 τ 2p s 2 + 2ξτ P s + 1 2τ P
From which we see that coefficients of discrete inputs in RHS are functions of pi
and qi which are the denominator and numerator coefficients of process transfer
function. Comparing step responses given by Eq. (4 or 14) and Eq. (17.1), one
can easily develop the relation between step response coefficients (ai) and ( ci/ ).
Thus
b1 −T
KP = , τP = , DP = n * T (17.10)
1 + p1 log( p1 )
Hence, for a given transfer function, one can find coefficients ( ci/ ) for step inputs
using Eqns. (17.6) to (17.8). Eqn (17.9) can be used to find step response
coefficients of Eqn. (4) to find open-loop (predicted) response.
GPGC e − Ds
= (18)
1 + GPGC λ s + 1
e − Ds
Or GC ( s ) = = f ( s) (19)
GP [λ s + 1 − e − Ds ]
http://www.bepress.com/cppm/vol4/iss1/29 8
DOI: 10.2202/1934-2659.1362
Panda and Subramaniam: PID Equivalent of DMC
1⎡ s 2 // ⎤ 1
GC ( s ) = ⎢ f (0) + sf (0) + f (0) ⎥ = K C [1 +
/
+ τ D s] (20)
s⎣ 2! ⎦ τ I s
Where from controller parameters are found (Panda, 2004 & Lee et. al., 1998) as,
f / (0) f // (0)
τI = and τ D = (21)
f (0) 2 f / (0)
Thus, with different structures of GP(s) the analytical expressions of integral and
derivative times are obtained. It is found that the expressions are functions of step
response coefficients (pi and bi) of process transfer functions [Eqn.(17.5.1) or
Eqn.(16.1)]
D2
τ i = p1 − b1 + (22)
2(λ + D)
b1D 2 D3
p2 − b2 + −
2(λ + D) 6(λ + D)
τd = (22.1)
τi
It can be noted that integral and derivative time constants, given by eqns.(22) and
(22.1) have similar forms that obtained from IMC-Maclaurin (Panda et. al, 2004)
as
FOPDT K P e − Ds τI
G= KC = ,
τ Ps +1 K P (λ + D )
D2 D2 ⎡ D ⎤
τI =τ + ,τ D = ⎢1 − ⎥ …(23)
2(λ + D ) 2(λ + D) ⎣ 3τ I ⎦
SOPDT K e − Ds τI 2λ 2 − D 2
G= 2 2 P KC = , τ I = 2ξτ − ,
τ P s + 2ξτ s + 1 K P (2λ + D) 2(2λ + D )
D3
τ2 −
6(2λ + D)
τ D = τ I − 2ξτ + …(23.1)
τI
To evaluate the closed loop performance of the new PID controller parameters
using presently developed DMC algorithm, seven different process model (Table-
2) responses are considered from practical examples from literature and are
analyzed. Equations 16, 21 and 22 are used to calculate PID parameters. It can be
noted from equation (15) that with KL=1, the model becomes a step response
model whereas with KL=0, the model reduces to set-point DMC. In case of PID-
DMC, the closed-loop characteristic equation becomes I + GC GF (GP − Gm ) = 0
where Gm is the process model. In case of process model mismatch, this algorithm
is supposed to work satisfactorily. Table 2 gives the transfer function of the
process and their prediction horizon (P), control horizon(M), control weight (W)
and the λ values.
http://www.bepress.com/cppm/vol4/iss1/29 10
DOI: 10.2202/1934-2659.1362
Panda and Subramaniam: PID Equivalent of DMC
Example1: A first order plus dead time system (Chen & Seborg, 2002) is
considered. An equivalent PID control is synthesized for the dynamic matrix
control using parameters (prediction horizon, P=6, control horizon, M=3, and
weighing factor, W=0.01) as mentioned in Table 2. With the parameters given in
Table 3, the designed controller parameters obtained are: KC=1.672, τi=1.0026
4 e −5 s 50 26 40 1.25
UNDER DAMPED:
9s 2 + 2.4s + 1
5 Higher Order System: 26 13 1 0.001
−14 −9 −5
2.035*10 s + 3.628*10 s + 9.979 *10 s + 4.8
3 2
and τd =0.0025. The closed-loop simulation was done for 100 sampling time
from which the performance was noted to be IAE (Setpoint)=0.52,
IAE(Load)=9.6 for unity step changes in setpoint and load-disturbances
respectively. The response (Fig. 2, Ex-1) settled after time of 2 and 60 samples
respectively for set-point and load-disturbance cases. An overshoot of 10% was
observed in set-point tracking.
Example2: An overdamped, 2nd order plus dead time system (Chen & Seborg,
2002) is used for study. The prediction and control horizons are taken to be 76
and 38 sampling units. The calculated PID controller parameters are found to be:
KC=0.287, τi=19.2 and τd =2.16. After simulation the performance of the present
scheme is noted to be IAE (setpoint) = 60.2, IAE (Load) =73.2. The load response
(Fig. 2, Ex-2) takes little more time to settle than set-point case where a 60%
overshoot is observed.
http://www.bepress.com/cppm/vol4/iss1/29 12
DOI: 10.2202/1934-2659.1362
Panda and Subramaniam: PID Equivalent of DMC
Example4: An second order underdamped system with dead time from Lee
et.al(1998) is considered. The designed controller parameters are found to be:
KC=0.0363/KP, τi=4.4 and τd =50.57. The closed-loop performance is noted to be
IAE (setpoint)=80.53 and IAE(Load)=85.36. The response settled at 150
sampling time and at 400 samples respectively.
Example6: This example of Jyh-cheng Jeng et.al (2005) considers a Non linear
process. The calculated controller parameters based on step response (when
approximated as FOPDT) model are: KC=0.06/KP, τi=16.1 and τD=9.09. The
closed-loop performance showed an IAE (setpoint)=119.6, IAE(Load)=262.6.
The response (Fig. 2, Ex-6) showed longer settling time.
http://www.bepress.com/cppm/vol4/iss1/29 14
DOI: 10.2202/1934-2659.1362
Panda and Subramaniam: PID Equivalent of DMC
1
.
(a) (b)
Ex-1 Ex-2
http://www.bepress.com/cppm/vol4/iss1/29 16
DOI: 10.2202/1934-2659.1362
Panda and Subramaniam: PID Equivalent of DMC
Ex-3 Ex-4
Ex-5 Ex-7
The stability of the designed controller can be tested by finding out gain and
phase margins. Generally, PI settings that give larger phase margins are more
stable. With this knowledge, present PID controller is tested for closed-loop
control of a FOPDT process. The forward loop transfer function becomes
K C (τ Dτ I s 2 + τ I s + 1) K P e− DP s
GC ( s )GP ( s ) = . . (25)
τIs τ Ps +1
Using Nyquist stability frame, the point (-1+ j.0) can be solved for gain
margin as
0.5π + tan −1 (ωcpτ D ) + tan −1 (ωcpτ I ) − tan −1 (ωcpτ P ) − tan −1 (αωcpτ D ) − ωcp DP = 0 (27)
If Nyquist curve produces a phase of -π, then the phase margin becomes
φm = 0.5π + tan −1 (ωcgτ I ) + tan −1 (ωcgτ D ) − tan −1 (ωcgτ P ) − tan −1 (αωcgτ D ) − ωcg DP (28)
1
S ( jω ) = (30)
1 + GC ( jω )GP ( jω )
6. Conclusion
Most of the process industries still use conventional PID loops as their tuning,
implementation and maintenance are easier. In this paper a new method of tuning
http://www.bepress.com/cppm/vol4/iss1/29 18
DOI: 10.2202/1934-2659.1362
Panda and Subramaniam: PID Equivalent of DMC
PID controller parameters has been proposed. PID control law is equated to DMC
control algorithm and then parameters of PID controller are related to that of
DMC. After obtaining the openloop step response of the system, closed-loop
feedback gain matrix is calculated from which PID controller gain is synthesized.
Integral and derivative times are extracted from desired closed-loop response
which is a function of step response coefficients and tuning parameter (λ), often
called as closed-loop time constant. Due to the presence of tuning parameter, in
the present control algorithm, user can make the response faster or sluggish. The
robust controller designed in this way is used to simulate setpoint and load
changes for several practical examples.
The simulation study shows that the new method has got good
performance in term of time constant, settling time and IAE for various process
model. This idea could be used for auto-tuning scheme for the controller.
Implementation of the present PID scheme is easy and its performance is robust.
Notation
http://www.bepress.com/cppm/vol4/iss1/29 20
DOI: 10.2202/1934-2659.1362
Panda and Subramaniam: PID Equivalent of DMC
References
Chen, D., Seborg, D.E., “PI/PID controller design based on direct synthesis and
disturbance rejection,” Ind. Eng. Chem. Res, (2002), 41, 19, pp.4807-22.
Haeri Mohammad,“Tuning rules for the PID controller using a DMC strategy”
Asian journal of control, (2002), 4 (4), pp. 410-417.
Lavanya, K., “Estimator based multirate control of dynamic systems,” PhD thesis,
submitted to Anna University, Chennai – 25., 2007.
Lee, Y., Park, S., Lee, M., and Brosilow, C.,“PID controller tuning for desired
closed-loop responses for SI/SO systems,” AIChE J., (1998), 44, pp.106–
115.
Lundstrom, P., Lee, J.H., Morari, M. and Skogestad, S., “Limitations of dynamic
matrix control,” Comp. Chem. Eng., (1995), 19, pp. 409-21.
Luyben,W. L., Process modeling simulation and control for chemical engineers,
2nd edition, McGraw Hill, NY, p. 656, 1992.
Panda, R.C. and C.C. Yu and H.P. Huang, “PID tuning rules for SOPDT systems:
Review and some new results,” ISA Transaction, (2004), 43, pp.283-295.
Panda, R.C. and Ramachandra Rao, V.S., “Model based control of a continuous
fluidized bed dryer”, 3rd IEEE conference on Control Applications,
Glassgow, U.K., 1994.
Qin S.J. & Badgwell. T., “An overview of model predictive control technology”,
Proceeding of fifth International conference in chemical process control,
Tohoe, California, 1996, pp.232-256.
Shridhar and Cooper D. J., “A tuning strategy for unconstrained SISO model
predictivecontrol,” Ind. Eng. Chem. Res., (1997), 36, pp. 729-746.
Saeki Masami, “Fixed structure PID controller design for standard H∞ control
problem”, Automatica, (2006), 42, pp. 93-100.
http://www.bepress.com/cppm/vol4/iss1/29 22
DOI: 10.2202/1934-2659.1362