TM

DYNAMIC MATRIX CONTROL

A COMPUTER CONTROL ALGORITHM
PRESENTED AT:
THE NATIONAL MEETING OF
THE AMERICAN INSTITUTE OF CHEMICAL ENGINEERS

HOUSTON, TEXAS
APRIL 1979

BY:

DR. CHARLES R. CUTLER

[DYNAMIC MATRIX CONTROLCORPORATION)

&
DR. B.L. RAMAKER
[SHELL OIL COMPANY]

[DMC]TM

DYNAMIC HA1lUX a:MROL - A a:HVIm cnnROL ALOORTIlI1

presented by

C_ R. C11l1D and B. L. RAHAKER

at the 86th NATIONAL

AMERICAN INS1T1UJE OF OIEHIoo. mGINEDS

JaSTRACT

The Dynami e 'iatri x Control (011C) A19orfthm

is control technology that has been used
successfu11y in process c~uter appH cat ions
I. Shell for the last six years. The general
deweloplllent of the Dr-te Algoritha to incorporate
'"dforward and multivariable control is covered
il this paper. The Dr1t Algorithn evolved frOlll
technique of represent i ng process dynaoi es

.Ith a set of numerical coeffic:tents. The

_meric.l technique, in conjunction with a

lellSt square fon:ulation to f.'Iinlmize the. inte

V.l of the error/tin! curve, make it possible

to solve co~lex control probler.lS on a digital
t~ter which are not solvable with traditional
'ID control concepts. The incorporation of the
NlCess dynamics into the synthesis of the
desfgn of the ONe, make it possfble to naintain
1ft awareness of deadt ime and unusual dynar:tlc
Mbavior.

COLUHH
..J
.....

>

..J

STEP CKM'GE IN
FEED TO COLUMN

0"'----'---------
FURNACE

CHANCE DUE TO

SOOT BLOWER

3TllAMlC MATRIX CONTROL OEVELOPIHT

The Dynamic Hatrix Control (Drle) AlgorHhn

's a control technology that has been used
sutcessfully In process cerevt er applications
III Shell for the last six years. The general
~yelopment of the ONC Algoritho to Incorperate
.ftedforward and multivariable control wi 11 be
toyered In this paper. A subsequent paper wi 11
e~and the scope of the algorlth" to address
~e problem of constrained multivariable con
trol. The need for the expanded algorithn
~'o1ved from the application to a catalytic
~~acklng unit of a non-linear steady state opti
~latlon which characteristically drives the
~ess to a number of constraints.
. The D/1C Algorithr.l evolved from a techniquc
.f representing process dynaoi cs with a set
of nUllll!rical coefficients. The nUr.!Crical
technique, in conjunct ion with a least square
lonulation to minioize the integral of the
error/tfme curve, r.l4ke it possible to solve
:~lex control probler.lS on a digital conputer
-"'ch are not solvable with traditional PlD
amtrol concepts. The incorporation of tne
:rocess dynami cs into the synthesis of the
~slgn of the oro\(. make it possi ble to r.I4i ntai n
p awareness of deadtime and unusual dyna~ic
~havior. An awareness of deadtime alone
:revents the controller frOr.l overco~ensating
.tllch can only be obta i ned in the PI0 contro11er
~ suppress1ng the integral action with the
.
;orrespondlng degradation of the control.
~Xill91es of unusual dynamic response to a step
:h.nge are illustrated fn Figure 1. The first
curve is characteristic of a system which is out

TIME
FlC. 1

UNUSUAL DYNAMIC BEHAVIOR

of naterial balance. The second curve is the

response of a furnace tranSfer temperature to
a soot blowing operation on a large preheater.
:Iote in both illustrations, the response curve
cannot be adequately described by a first or
~econd order differential equation typically
used in control analysis.
Any system which can be described or
approximated by a systeo of linear differential
equations can utilize the Dynamic Matrix Control
technique which is based upon the nu~rical
representation of the system dynamics. Two
properties of linear systems makcs the numeri-.
cal representation possible. The first of
these principles is the preservation of the
scale factor. It is illustrated in Figure 2
where the response of .the output variable 0 is
shown for a changc in the input varfable I.
The solid line represents the response of the
output variable to a unit change in the input
variable and the dashed line illustrates the
response for a two-unit change in the input
variable. Note the response of the two-unit
change has twice the amplitude of the Dne-unit
change. For a linear system, the response of
the output variable for any size change in the
fnput variable may be obtained by multiplying
the scalar value of the input variable times

the input variables. Hathemitically these

concepts are given by:

/ '

"

, '"

--

, ,

6 01

a1 6 11

b1 6 12

az

11

b2

6 12 +

6 ~3 a3

11

b3

6 02

TWO UNIT CH.eJ:GE

HI UfPUT VARIABLE

(1)

12 +

. /

, ""

I
I

"

TIME

rtc.

PRESERYATlOIl OF SCALE FACTOR

the unit response curve for the output variable.

Further note on Figure 2 that the unit response
curve can be approxi~ted by a set of numbers
if the curve is broken into discrete intervals
of tiMe. The two-unit response curve in Figure
2 can be obtained by ~ltip1ying the set of
numbers for the unlt response by 2. The second
characteristic of a linear system is the prin
ciple of superposition. This principle is
illustrated in Figure 3 where the response Of
the output variable is shown for a unit change
.in two input variables. Also, the response
of the output variable to a si~ltaneous unit
change in both input variables is shown. The
response for this curve was obtained by s~irig
the responses for the unit response curves for

0i ai 6 11 + bi A 12

where the 6 0i are the changes in the output

variable frGQ its initial value to its value
at ti-e interval t and the' I j are the
changes in the input variables from their
initial value at ti-e equal to zero. The ai
and bi are the numerical coefficients referred
to in the preceding paragraphs. Fi9lre 4
illustrates the response of the outlet te~er
ature of a preheat furnace to i step change in
the fuel to the furnace and to a step change
in the inlet te~erature of the feed to the
furnace.
The fuel coefficients shown ln figure 4
are an illustrative exa~le of the at and the
inlet te~erature coefficients are i lustrat1ve
of the bi' The input variables can be ~nipu
lated control variables or ~asured distur
bances. For exa~le. in the furnace control
problem to be described in this paper. the
fuel is a ~nipulated input variable, the
inlet feed tecperature is a !leISured distur
bance. and the controlled variable is the
furnace transfer teqerature.
b1

!.O
Ifl _
INPUT VARIABLE I.

UNIT INCREASE IN
INPUT VARIABLE tb

~~. ~-----

.5

1--0

FUEL
COEF.

INlET
TEMP.
COEF.

.014

0.0

at

usrr mCREASE

W
I - -J

;:) a::

0<
>

.086

.240

.214

.340

.414

.465

.600

.540

.736

.590

.836

.6ZZ

.904

.640

.949

.653

.986

.658

UNIT INCREASE IH

INPUT VARIABLES

la & Ib

TlHE

TIME
FIGURE 3

FIC. 4

PRINCIPLE OF SUPERPOSITION

WPS-B

FURNACE RESP.ONSE, FUEL, TEHPERAllJRE

The change in the controlled variable fro"

time equal to zero to s~~ future tine th!t will
result from changes in the ~nipulated in~ut
can be represented ~ the following equations:
1

01 al A 1

03 a3

(2)

1+

A 11

0i ai

2
a~ A 1

a,

11

where I:lOvement of manipulated variable I, is

considered for three intervals of ti~ into
the future with the superscripts o~ 11 repre
senting the ti~ interval. The A II is the
step change in the ~nipulated input variable
from its value at the end of the first .tine
interval and the beginning of the second.
Similarly A 311 is the etlange in the ~nipulated
input variable from its value at the end of the
second ti~ interval and the beginning cf the
third. Further. note the sa~ set of coeffi
cients is used for each colunn. The coeffi
cients in each column are shifted down one
time interval for each successive column to
correspond-with .the first time interval which
the future inputs can impact on the outrut var
iable.

FEED FORWARD DMC

Feedforward control is accomplished by
moving the measured disturbance input variables
in equation set (1) to the left hand side.
For example. if input 12 is a disturbance input.
equation set (1) becoDes:
01

bl

12

al

11

02

bZ

IZ

aZ

11

03

b3

IZ

04

b4

12

Os

bS

06 + b6

11

11

12

a3
a4
as

11

12 a6

11

61 + bi A 12

02

b2 A 12 aZ A I ~ +

123
I + ai-l A I + ai-2 A I +
1
1
1

01 + b1 A 12 a1 6 I'

(3)

ai A 11

Combining equation set (Z) with set (3)

yields the general form of the equations used

to do feedforward predictive control of a

variable 0.

The desired response of the systeGI is deter

mined by subtracting the predicted response of
the system from the setpoint. which is deter
mined from the past histo~ of inputs to the
system. With the desired output response 6 0i
known. the measured input disturbances A 12
known. and the numerical coeff~ients ai and
bi known. all the infonnation is available
to solve equation set (4) for the set of time
dependent moves in the manipulated input
variable.

DHC SOLUTION TECHNIQUE

The set of equations is over-detengined
which prevents direct solution. but can be
solved using a least square criterion. Such a
solution produces a projected set of moves in
the manipulated variable that minimizes the
error in the output variable from its setpoint.
The obvious difficulty with the use of the least
square method to calculate the movement of the
manipulated variable is the unconstrained
nature of the solution. The method without
constraint will yield very large changes in
the manipulated variable that would not be
physically realizable.
One technique for sappressing the change
in the manipulated variable is to multiplY by
a number greater than one. the main diagonal
elementsof the square matrix that evolves frOm
the least square formulation. The effectiveness
of such a multiplier is illustrated in Figure 5
where a square wave change in setpoint was made.
The unconstrained least square reduction in
the error resulted in a total change in the
absolute value of the manipulated input variable
of 37.57 taken over 10 intervals of time. With
a multiplier of 1.005 the total change was
4.91 and with 1.010 the change was 3.42. The
multiplier effectively adds another row to
the original data for the input variable for
each interval in which it is allowed to move.
All elements in the row are zero except for the
. specific input variable which has a coefficient
related to the size of the multiplier. As can
be seen from figure 5. the suppression of the
manipulated input variable moved by an order
of magnitude did not significantly impair the
reduction in the projected error.
The matrix of coefficients which describe.
the clYnu1cs of the systl!lll 1s the basis for
WP5-B

SET POI~:T CHArlGE

- - - HAIN DI~CimIAL
f'!UL TI PLI ER OF

1.005
------- HAIII IH~~OtIAL

TlPl,lER OF
1.000

~UL

'"'
2 .....-;-1-"

a::

<
=:

calculations are summed with newly calculated

fuel moves and the current fuel ~ve is imple
mented.' The pattern is repeated at each suc
cessive interval of time.
Feedforward control is iitplemented by
measuring the change in the feed inlet teapera
ture between time intervals. ~ltiplying by
the nuaerical coefficient for the effect of
the inler t~erature on the'outlet, and
s~ing this response of the outlet te~erature
into the vector of predicted ,values for the
output variable.

w.I

e,

...'"'::c::

FIG. 5
the m1C

TU-I

HOVE SUPPRESSIon FUEL TO

FURr~CE

Algorit~.
For the furnace control
the response of the output variable was
considered for 30 intervals of ti~ and the
movel:\ent of the fuel gas was considered for 10
intervals. Thirty intervals of ti~ represents
about 4 lIZ time constants for the response of
the outlet temperature to a change in the fuel.
At the tenth time interval. the outlet te~er
ture has three ti~ constants to settle fro~
the last change in the fuel. This choice of
time intervals results in a Rltrix with 10
,colunns and 30 rows. To initialize the algor
it~, the measured outlet te~erature is
stored into the 30 el~nt vector that repre
sents predicted values of the output variable.
This assumes the system is at steady state,
but is not a necessary criterion. An error is
then calculated from the projected value of the
output variable and the setpoint for the 30
intervals of ti~e. This vector of errors
becomes the right hand side for the 10 by 30
~trix.
The least square solution of this set
.of equations yields the best set of fuel ~Yes
to eliminate the projected errors for 30 time
intervals. The projected set of fuel moves is
used to calculate the outlet temperature change
for the forthcoming 30 intervals of time, and
the temperature changes are then added into
the 30 element vector for the predicted value
of'the dependent variable. The first fuel
move is implemented and the entire vector of
predicted output variable values is shifted
forward one interval of ti-.. At the start of
the next interval of time the predicted value
of the output variable is compared with the
measured value. The error in the projection
is used to adjust all 30 values in the predicted
output variable vector. This adjus~nt in
the prediction provides the feedback to cocpen
sate for u~asured disturbances and errors in
the dynaQic prediction. At the next interval
the ,set of errors between the setpoint and the
predicted values of the output variable is used
to solve for another set of 10 fuel aoves.
The ntne remaining fuel ~ves fram the prevtous
probl~

The description of the technique for the

DHC was given in some detail to foster under
standing. The actual calculations involved
in the technique are at least two orders of
IIIIgnitude less than would be required by the
outlined procedure. The first siaplification is
to recognize that the matrix of coefficients
representing the dynamics are fixed and only
the right hand side changes from one time
interval to the next. Furthermore, the square
aatrix that results frOQ the formulation of
the least square procedure does net change.
which also means the inverse matrix does not
change. The errors between the projected
vector of the output variable and the setpoint
appear in the caluclation of the right hand
side for the least square formulation. The
right hand side of the formulation 15 the
transpose matrix of the matrix given fn-equa-'
tion set (2) times the vector of projected
errors in the output variable frOQ its setpoint
Consequently, the inverse matrix tices the
trcnspose of the original matrix times the
error vector gives the projected set of moves
in the manipulated input variable. Since -the

inverse matrix and the transpose are also con

stant. the calculation is reduced to the RUlti

plication of a constant ~atrix tiaes the pro

jected error vector. The preceding description

of the DHC calculations can be expressed con

veniently in the following matrix ,notation:

- (ATA}-l

AT

E.

(5)

where the A gatrix is the Qatrix of coefficients

describing the dynamics [the ai of equation set
(2), E is the vector ,of projected errors, and
the I is the vector of Drojected ~ves in the
input variable. Yhe (ATA}-lAT is the constant
matrix referred to above. As a further'siQPli
fication it can be shown that only the first
row of the constant matrix and the error vector
are needed to solve the control problem. This
calculation yieldS the control move to make at
the present interval of time. The other control
moves in the input variable are iabedded in
the first row of the constant matrix. in the
successive updating of the error vector, and
by the coupling of time intervals from adding
in the projected response from the. fuel moves
in each tntrrval of time. Since.the calculation
of (ATA)-lAT can be done offline in the fbrnace
control discussed earlier. the actual real
time calculations for the PiC Algorithm are
reduced to the .ultiplication of the 30 eleaents

in the first row of the constant .atrlx times

the projected error vector.

WP5-B

A comparison can be made between Figures 6,

7. and 8 to illustrate the merit of the OKC

a:
~
<
a:

II:LET

TEfo!PEP.ATI~P.E

i:

...'"'
TH!?ERA TURE
FIG. 6 COIlVEllTl O::Al AIlAlGG COl/TROl OF FliRflACE

IrllET TE:-1PERATURE

control relative to conventional control tech

niques. Displayed in each of these figures is
the response of the outlet teaperature of the
preheat furnace to a disturbance in the inlet
feed te~erature. Each figure represents real
data taken off the operating unit. In Figure 6
. the perforaance of the units standard analog
teaperature controller is illustrated. The
teaperature controller is cascaded to flow
controller on the fuel and the tenperature con
troller's settings were established after years
of experience by the operating personnel. In
Figure 7, the temperature control is being
accOQPlished via the computer using a PID
algoritha with lead/lag compensation on the
feed inlet teaperature. The conputer output is
directed to the flow controller on the fuel.
In Figure 8. the DHC is on the tenperature
control with feedforward control of the inlet
temperature. The fuel is on flow control
which is reset by the computer. As can be seen
froQ this c~arison. the OMC is substantially
better than either of the conventional aethods
of control. A further example which displays
the robustness of the DHC is shown in Figure 9.
The set point for the outlet temperature of
the furnace in this case was being reset by
another controller within the cO"Puter. The
DHC algorithm effectively handled a large
temperature disturbance while having a load
changeil:lpOsed on it by the outer control loop
of the temperature controller. Fuel QOves to
handle the disturbance are also shown.
_______ J

- ---

-_.'
DonED: FURNACE OUTLET TEMPERATURE
SOLID: SET POINT FOR FURNACE OUTLET
TEMPERATURE

CUiLET
TEr-l?~RATURE

--

Tlr-:E

FIG. 7 Crn-IPUTER PID/FEED FORWARD CONTROL

OF FURtlACE

INLET TEMPERATURE.

'"
>

OUTLET TEliPERATURE

TIME

FIG. 8 DYNAMIC MATRIX COIlTROl OF FURNACE

FIG. 9 DYNAMIC

WPS-B

TlHE
CONTROL OF A FURNACE

HAT~IX

HUL TtVARIABLE Iw.C

al
a2 al
a3 a2

In equation set (I), both input variables

may be manipulated. In fact, any number of
~nipulated input variables that influence the
output v,riable may be included, e.g.
6

01 al

6 02
6

1
1

al

6 11

bl A 1
2

(6)

bZ 6 I + bl

.
ci

2
I +

al

A ai ai -1 'i-Z
cl
c2 cl
c3 C2
CJ

6 1
2
2
2
1
Z
aZ 6 I + b3 A 1 + bZ 6 1
2
1
2

aZ 6 I "+

03 '3

6 11

bl
b2 bl
b3 bZ
bi bi-l
dl
dZ dl
d3 dZ

(8)

bl
bi-Z
dl

di di-l di_2

ci-l ci-2

The vector I of moves in the manipulated v,ri

abIes for the .ultivariable probleQ is solved
the same way as vith the single variable problem
and is given by:

_ (ATA)-l AT E

(9)

where E is the vector of project errors in the

outputs.
where the subscript on I represents a different
variable, the superscript on 1 indicates the
time period in which the input is changed, and
and bi ,re the numerical coefficients
the
assoc ated with II, IZ, respectively.

The first application in Shell of the"

multivariable control vas on a catalytic crack
ing unit. The algorithm has worted well except

under conditions when one or more of the mani

pulated variables beclQt constr!ined as a

result of being integrated vith a real time

optimization of the process. The next paper

will describe the optimization of the catalytic

cracking unit and the enhancements to the

algorithm which allow the projected moves in

the input variables to be constrained in tilllt.

'1'

To incorporate the control of more than

one output variable in the OMC Algorithm, the
matrix of coefficients is expanded below to
illustrate the case of two manipulated inputs
and two outputs:
6 011

al

aZ

A I

(7)

1
11 + al

6 0.1 1

ai A 11

6 012

cl

Ol

+ bZ A

ai_l

1
1

2
A 11

c3 6

I~

+ b i
+

6 11

6 OZ2 Cz 6 11 +

SlM4ARY .

6 I

"I
dl 6 1
2

cl A 11

+ dZ

b 1

I~

+ d)

6 1

+ c2 6

1
2
2 + b1- 1 A 1 2

+ dl

6 1

+ dZ

The superscript on the output variables indi

cates a different variable. The nuaber of
input and output var1ables is unlimited in this
formulation. The only constraint is that the
number of manipulated input variables be equal
to or greater than the number of output vari

ables and that the system of equations be

overspecified. The multivariable problem is

solved in the same manner as illustrated for

the furnace example. The matrix of coefficient

in the last equation set is defined as:

2
2
1
2

The nHC algorithm descr4bed. in this paper

has permitted the on-line cOQPuter solution of
lIIultivariable control problems. Dynamic data
taken from"the process can be directly inserted
to configure the controller. The controller
calculates moves for the present time and a
nudber of future time periods. This feature
always for the correct compensation for pro
cesses with dead time. feedback has been
incorporated into the control algorithm to
ccnpensate for unmeasured disturbances and
dynamic lllOdel errors. feedforvard is silllPly
configured into the controller by the addition
of the dynamic data for measured disturbances.
This technique has been successfully applied to
control problems vithin Shell for the last six
years.
HlJ1JlCLATURE
ai.bi,ci,di Numerical Coefficients Describing
Process Dynamics
A Matrix of Coefficients Describing
Process Dynamics
"
E Vector of Projected Errors "in Tillie
I Vector of Projected Moves in the
Input Variable
6 Ij Changes in the Input Variables
6 0i Changes in the Output Variables

WP$-B
"