You are on page 1of 9

T I M E OPTIMUM CONTROL OF SECOND-ORDER

OVERDAMPED SYSTEMS WITH


TRANSPORTATION LAG
L. B. KOPPEL AND P.R. LATOUR
Purdue Uniuersity, Lafayette, Ind.

Time optimum control of a second-order overdamped system with transportation lag was studied for given
maximum and minimum constraints on the manipulated variable. Pontryagin's maximum principle and a
phase plane analysis were used to derive the two-position control law which would drive the system
output to its dlesired value in the shortest time. A two-position programmed controller, based on switching
time, was synthesized on an analog computer. This controller verified the optimum control law and pro-
vided significant response time improvement over a well tuned proportional-integral-derivative industrial
controller. The degree of improvement is best for large set point changes on processes with large t h e
constants, large transportation lags, or narrow limits of asymmetric saturation. The results of this study
are useful for startups and for transitions between steady-state operating conditions, for the class of processes
whose dynamics are adequately represented as second-order with delay.

HEMICAL process dynamics, as distinguished from those of necessity for knowledge of the theoretical process dynamics,
mechanical or electrical systems, are characterized by which are usually impossible to obtain. Since Equation 1
large time constants, distributed parameters, and sluggish has been found to be broadly applicable and experimental
response. Although thi: criterion for optimum control in the procedures for fitting the parameters have been established,
chemical industry is generally maximum profit, a criterion of the results of the present study should be applicable on a prac-
minimum time would seem to be almost equivalent and is tical basis to a reasonably broad class of process control prob-
clearly more convenient for analysis. The time optimum lems (5, 7). However, it is clear that processes which cannot be
programmed control function specifies the manipulated vari- approximated by Equation 1: such as a highly exothermic
able, M ( t ) , that drives the process output, C ( t ) , to the desired chemical reactor, will require specific formulation and solution
value as rapidly as poscible. The optimum feedback control for the time optimum control law; this topic receives no atten-
law is the manipulated variable specified as a function of the tion in the present work, but has been investigated in a t least
output, rather than of time. two recent papers (4, 9 ) .
I n this work, control of single manipulated input-single T h e manipulated variable in a real process is always con-
controlled output processes, where dynamics may be repre- strained between some maximum limits; hence we assume
sented by a transfer function of the form k 5 M 5 K , where k and K are known. We neglect constraints
on dM/dt. If saturation occurs elsewhere in the loop, k and K
are determined by the most narrow limits.
The block diagram is shown in Figure 1 . R is the set point
is considered. We assume a unity process gain with no loss or desired value of the output C. The error is defined to be
of generality. C(s) and M ( s ) are the normalized, transformed e R - C. The problem of controller design is to specify
process output and input variables, T is the predominant M such that e is reduced to zero in minimum time.
process time constant, b T is the smaller process time constant Although the process itself is linear, the constraints on M
(0 < b < I ) , and U T if3 the transportation lag or dead time. introduce nonlinearity to the control loop. The investigation
Since the effects of distributed parameters and higher order of this problem without transportation lag was pioneered by
dynamic characteristics may often be lumped with the delay Bushaw in 1 3 5 2 (3) and, independently, by Feldbaum (8).
constant, U T ,this transfer function has been found to represent In 1 9 5 6 , a principle leading to the solution of the general
adequately the behavior of many chemical process operations, problem of finding a n optimum control function for this type
including a liquid-liquid extraction column (2), a cracking of constraint was formulated by Pontryagin (74). Desoer (6)
furnace (729, heat exchangers, pneumatic transmission lines, has shown from Pontryagin's maximum principle that, for
agitated mixing vessels (70), and large distillation columns (75).
T h e three dynamic parameters, T , 6 , and a, are assumed to
be known. Methods for fitting these parameters by practical
dynamic testing, to processes whose true (unknown) dynamics
K IL
are more complex than the transfer function of Equation 1 ,
have received extensive recent attention as reported in the

-1
k
monograph of Hougen (70). In contrast with recent studies
on synthesis of time optimum controllers (4, 9), a major objec-
tive of the present study is a practically implementable con-
troller, applicable to s~ broad class of processes, without Figure 1 . Block diagram of the system

VOL. 4 NO. 4 N O V E M B E R 1 9 6 5 463


linear processes, the minimum time solution always results in a
control law \\-ith two-position action-that is, on a time opti-
mum trajectory, M is always a t the boundary of its limits
(either R or k ) with a possible finite number of switches be-
tween them. except when the system returns to the origin.
For nth-order linear systems with real roots, Feldbaum (8)
has sho\vn that the optimum number of switches between K
and k (discounting the final switch to maintain equilibrium
a t the origin) is a t most n - 1, and the optimum switching
points depend upon the initial error and its first (n - 1) initial
M=R ' M=k M =K
(a) (b) (C)
time derivatives.
The solution to the time optimum control problem for
second-order overdamped systems is well known, but little
work has been done for the case with dead time. Kurzweil
( 7 7) has investigated the control of multivariable systems with
time delays for fairly general process transfer functions. His
algorithm for minimum settling time control is complex and
requires a digital computer for implementation and storage of
past process output history. Further, it assumes accurate
knolyledge of process transfer functions and all state variables.
I n the present study, these requirements are relaxed to provide I
a simple implementation of minimal time control of processes (d)
which are adequately represented by the transfer function given Figure 2. Phase plane analysis of second-order over-
in Equation 1. This study provides the analytical solution for damped systems
this case, a comparison between the optimum control and well
1. Trajectory from Figure 2b, M = k
tuned P I D control, and a simple suggestion for practical 2. Trajectory from Figure 2c, M = K
realization of the optimum control. 3. Origin of Figure 20, M = R

Application of Maximum Principle


Pontryagin's maximum principle states (74) that a necessary
The equation relating e and M from Equation 1 is
condition for the time optimum M is that H ( M , x, X) be a
bT2 a ( t ) + (6 + 1) T i ( t ) + e ( t ) = -M(t - at) +R (2)
maximum, with respect to M , for all t along the trajectory.
I t can be seen that H i s maximum if
if R is piecelvise constant. The initial conditions are e(Oi) = K , Xz < 0
e l , i ( O + ) = 0, and an initial forcing function M ( t - U T )= M(t - U T )= (9)
k , Xz > 0
R,, 0 _< t < U T . Now define X I = e, x2 = d, and form the state
equations It is clear that X 2 ( t ) cannot change sign more than once.
Since H is maximum when M is maximum in magnitude, the
dl = x2 = F1 (x,f,t ) (3) maximum principle establishes the form of the control function
as two-position with no more than one switch. To determine
the time for the single switch from full on, M = K , to full off,
M = k, or vice versa, we next study the phase plane trajec-
tories.
T h e forcing function, f, corresponds to --M(t - UT) + R,
cector x to the pair (XI,xg).
Phase Plane Analysis
We form a conjugate system defined by
Equation 2 will be studied for three forcing functions
(4) set point equilibrium

For the present system


--M 7 R =
["-k
-K
+R
+R
minimum
maximum

For the present, we neglect transportation lag and consider


(10)

(5)
M = M ( t ) , to = 0. The phase plane trajectories are shown in
Figure 2 for the three forcing functions. Time is a parameter
along the trajectories. In each case, a stable node occurs on
the e axis at the value of the forcing function.
At time t = 0, a step change is made in the set point from R,
Next formulate the Hamiltonian of the conjugate system
to R. For time t < 0, e = e = 0 = R, - M . The conditions
HE Cib(t)Fi(x,f,t ) (7) immediately after the set point change are e (0) = el = R R,, -
and i ( 0 ) = 0. The desired final conditions are e ( t J = & ( t 5 )=
to get 0, where t j is unspecified,but is to be minimized.
The heavy trajectories of Figure 2, b and c, which pass
H=
- M ( t - UT) + R -~
(6 + 1)
bT
x2 1'
- b T2 (8)
exactly through the origin, constitute the desired final paths,
and hence are switching lines. When e l is finite and negative,
M should be k (because the system output must be reduced)

464 I&EC FUNDAMENTALS


and the trajectory of Figure 26 beginning a t (el, 0) will be
followed. This trajectory has been redrawn on Figure 2d.
If, a t the instant this tr2ijectox-y crosses the heavy switching line
of Figure 2c, M is switched to K , the switching line will be
followed to the origin, and the trajectory of Figure 2d results,
A final change to M == R is made a t the origin to keep the
system a t the origin; this is not counted as a required two-
position switch, since the system is a t equilibrium. For positive
el, the values of k and K are interchanged, so that M = K
for any state above the complete switching line, shown in
Figure 4,and M = k below this line.
When the transportxtion lag is included, the effect of a
switch in M ( t ) is not observed until a n elapse of time a T .
As shown in Appendix I, the optimum programmed control
function for a system wi t h dead time is the same as for the case
with no dead time. Lsing this M ( t ) , the first trajectory will Figure 3. Phase plane trajectory for second-
order overdamped system with transportation
not move from the point ( e l , 0) in Figure 3 until a time aT
after application of M ( t ) = k. Further, M ( t ) must be switched lag
to K while on the first trajectory, a t a time aT before reaching Time Interval M(f) M(f - aT)
to-11 k Ro
the final trajectory. This switching time is shown as t 2 , and f1-t2 k k
t 3 = tz + aT. The locus of the points tz for various e l con- f2-f3
h-f4
K
K
k
K
stitute the phase plane switching line in the presence of a delay. f4-k R K
If the final switch of M ( t ) to M = R is made a t t l , a time a T tr- R R
before reaching the origin, the system will stop and remain a t
the origin. This leads to synthesis of the control law for initial
conditions of the form given with Equation 2. As shown by
el) exp (aT - t d / T and (k -R f el) exp (aT - t3)/bT and
substituted into Equation 13 to give
Balakirev (7), the control law for arbitrary initial conditions
is more difficult to realize.

Derivation of Switching Equations


For negative el, the parametric equations for the first part of
the trajectory from ( e l , 0) to the point a t t 3 , determined from
solution of Equation 2 with M = k , are:

e = which is the equation for the switching line in the left-hand


e - i plane. Subscript s is used to designate that e and are
sk1 - b
O [ e x p ( F > - bexp(s)] - k +R (11) on the switching line. K and k should be interchanged to give
the right-hand plane equation for positive e l . [Balakirev
(7) has derived a less general form of this equation.] If the
process of Equation 1 has a steady-state gain K,, K and k in
Equation 14 are replaced by K K , and kK,, respectively. If
and for the second part of the trajectory from t3 to the origin continuous measurement of e and i is available, the first switch
are : should occur when e and e' solve Equation 14. This equation
e = should be used when lei1 is somewhat greater than the inter-
section of the switching line and the e axis (given by Equation
K
A
R1 - b [exPC+) - bexpcG)] -K +R (12) 14 when i s = 0), to ensure that tz > tl. This may not be
satisfied when errors are small and dead time is large. For the
processes we have studied, however, switching time was greater
e'= than dead time. An example switching line is shown in Figure
T(l - b) 4.
We have assumed a set point change from Ro to R, so that e l = An expression for t ~ the
, time of the first switch, can be ob-
R - R,. Simultaneous solution at t 3 , the phase plane inter- tained from Equation 13 by direct substitution of t z = t S - aT
to give
section, gives two equations in t 3 , t j , and e l . These in turn can
be solved simultaneous1.y to eliminate t 5 and give a n implicit
equation for t 3 and e l :
[E:+( k -KR- +
Rel)exp(s)]=

K - R
- + ( k -K - R + "'> (9)
exp (15)

____
K - k If e' is not readily available, the controller design can be based
K-I?+ -
K - R+ e x p t T ) (13) on time. Equation 15 gives tz implicitly for initial conditions
$1 = 0, e l < 0. When e l > 0, K and k should be interchanged.
Equations 1 1 of the first part of the curve, evaluated a t t = The switching time, t 2 , is independent of the dead time. This
tz = t 3 - aT, give equations for the switching error and error important result allows synthesis of a programmed timing
derivative. These can 'be solved for the quantities ( k - R + controller for process startups and transitions, without the need

VOL. 4 NO. 4 NOVEMBER 1 9 6 5 465


of specifying a basic process parameter, the transportation lag. A repetitive operation computer was used to find settings which
T h e error a t the final switch can be determined from Equa- minimized the integral of absolute value of error for step load
tion 12 when t d > ta and is given by changes, with an oscillation constraint of no more than thre:
visible changes in the sign of the error derivative. This was
K - R
e(tJ = ~

1 - b
[exp (u) - b exp(a/b)] -K +R done in the linear range of operation-Le., no saturation-
to ensure some theoretical significance of the results. Mini-
for e < 0. mum response time or settling time for load changes was not a
useful criterion for optimum P I D controller settings, because
Performance of Two-Position Controller
the “optimum” transients showed undesirable characteristics.
The performance of the time optimum two-position con-
As an example, if one chooses to minimize the 5% settling
troller was compared with that of a well tuned proportional-
time, t j % , settings can be found for some processes (small u )
integral-derivative controller on a n analog computer. Equa-
for which the error never exceeds 0.05L, so that ts% = 0.
tion 1 5 was the basis for logic circuits to provide the switching
However, such a transient had excessive oscillation and was not
time controller, A second-order Pad6 approximation was
acceptable for practical applications. Latour (73) gives more
used to simulate the process transportation lag. Latour (13)
details of this topic. T h e optimum P I D settings based on
gives the details of the circuits.
absolute area under the error curve gave transients, such as
To assure fair comparison for each set of process constants,
that shown in Figure 5, which were more likely to be con-
u and 6, “optimum” P I D controller settings were determined.
sidered as acceptable for industrial practice. For convenience,
the responses generated by the two-position control are labeled
BB on the diagrams (for bang-bang).
A series of comparisons between the well tuned P I D con-
troller and the two-position controller was made. The
output of the PID controller was now limited a t K and k for
fair comparisons. Figure 6 is an example phase plane resulting
from one of these tests. Each trajectory begins a t e l = -30,
e‘ = 0, and remains a t this position for an elapse of time U T
after application of M = k . T h e first switch on the two-
position trajectory, a t t = tz, is indicated. At t = t 4 , the final
switch from two-position back to PID control occurred. This
was done to approximate the action that would occur in prac-
tice if a control room operator forced this two-position control
by manual set point adjustment. (Also, the theoretical final
I I switch to M = R is difficult to achieve in practice, because of
-60 -30 60
inaccuracies in knowledge of transduction constants, such as
change in flow per unit change i n pneumatic pressure.) This
imperfect final switch caused slight undershoot past the origin,
T = 1.0 -60-
K = 40 since a small nonzero error is applied to the P I D controller a t
k E -100 t 4 . The reported theoretical minimum response time, t s , was
R - -20
b = 0.5 not affected by this final switch approximation. I t is im-
portant to emphasize that the switch a t t z is from full forward
to full reverse (k to K ) to give maximum deceleration and
bring the system rapidly to the origin. For simple on-off
control, the switch occurs at e = 0 rather than a t the indicated
switching line, and a limit cycle results.
I n comparing the trajectories of Figures 5 and 6, it may be
Figure 4. Example phase plane switching line seen that the P I D trajectory initially follows the two-position

2or

I I t
3 4

-10

/
-v
b= Q5
R = -30
K = 100

-20

Figure 5. Error transient comparison between two-position and PID control for test 20

466 l&EC FUNDAdENTALS


60

f \

K =Kx)
..
k =-I00
$:

Figure 6. Phase plane comparison between two-position and PID control for test 20

trajectory, since each is saturated a t k . Maximum decelera-


tion is not available i n the P I D case, and hence considerable
overshoot results. T h e point at which the 5% settling time
occurs is indicated on each trajectory.
T h e error transients, e ( t ) us. t , corresponding to Figure 6 ,
are shown in Figure 5. T h e time optimum two-position
transient is clearly superior. T h e manipulated variable
transients in Figure 7 indicate the initial saturation of both
controllers.
These tests were repeated over a range of possible conditions.
Table I shows the results of comparisons for various levels of
set point changes, process parameter values, saturation range,
and controller output asymmetry levels. Some of these
results are presented graphically in Figures 8 to 12.
The two-position co:ntroller gives significant improvement Figure 7. Manipulated variable transient comparison
even for small set point changes, as indicated in Figure 8 . between two-position and PID control for test 20
Greater improvement is realized for large set point changes.
(The P I D t57o is a funciion of step size because of manipulated
variable saturation, as shown in Figure 7 . ) On this graph
t j % for P I D is compared with what is effectively to% for two-
position. at = PI0 tSol0- BE ttj, DEGREE OF
IMPRWEMENT
As shown in Figure ‘9, response is slower and the improve- a= 0.1
b= 05
ment of two-position over P I D increases when the minor time K= 40
constant, 6, is large. T h e minimum response time appears to k =-100
be approximately linear in b (ts/T = 0.54 + 0.87b) in this
case.
Transportation lag appears as a n additive factor i n t 5 , as
shown in Figure 10. For this case t j / T = 0.91 + Q. The
switching time t 2 is independent of the dead time (see Figure
3). For the conditions listed in Figure 10, the final switch
point, t 4 , is 0.91 T for all Q , and the delay is additive to give t 5 .
T h e degree of improvement appears to be better when a is
large.
T h e total available range of manipulated variable, K - k ,
affected each controller performance as shown in Figure 11.
As expected, response was faster when the spread between the
1 I
-20
I
- I

40
I I
-60

maximum and minimum M was wide. However, the two- SET POINT CHANGE, A -&
position improvement over P I D is best when the saturation Figure 8. Effect of initial error size on response time for
limits are narrow. two-position and PID controllers

VOL. 4 NO. 4 NOVEMBER 1 9 6 5 467


Table 1. Results of Comparison between Two-Position and PID Control
PIDta,
Variable Test R b a K k K -k BBtj BBtj, PIDtj, Atss, - ts
Set point 1 -90 0.5 0.1 + 40 -100 140 3.10 4.56 15.8 11.2 12.7
2 -60 0.5 0.1 + 40 -100 140 1.60 2.35 4.8 2.5 3.2
3 -30 0.5 0.1 + 40 -100 140 1.00 1.72 2.85 1.10 1.8
4 -10 0.5 0.1 + 40 -100 140 0.60 1.30 1.89 0.59 1.3
5 - 3 0.5 0.1 + 40 -100 140 0.40 1.14 1.50 0.36 1.1
Minor time 6 -30 0.12 0.1 + 40 -100 140 0.62 1.18 1.70 0.52 1.08
constant 7 -30 0.3 0.1 + 40 -100 140 0.83 1.57 2.35 0.78 1.52
3 -30 0.5 0.1 + 40 -100 140 1.00 1.72 2.85 1.10 1.80
8 -30 0.7 0.1 + 40 -100 140 1.15 1.81 3.10 1.38 1.95
9 -30 0.9 0.1 + 40 -100 140 1.30 1.90 3.32 1.42 2.02
Dead time 10 -30 0.5 0 + 40 -100 140 0.91 0.91 Limit cvcle
+ 40
~~

11 -30 0.5 0.05 -100 140 0.97 0.97 2.62 1. ~ 6 5 - - 1 .65


3
12
-30
-30
0.5
0.5
0.1 ++ 40 -100 140
140
1.00
1.13
1.72 2.85 1.10 1.80
0.2 40 -100 2.41 3.31 0.90 2.19
13 -30 0.5 0.5 + 40 -100 140 1.44 4.25 2.87 -1.38 1.43
Saturation level 14 -30 0.5 0.1 + 10 - 70 80 1.30 2.02 3.40 1.38 2.10
3 -30 0.5 0.1 + 40 -100 140 1.00 1.72 2.85 1.10 1.80
15 -30 0.5 0.1 + 70 -130 200 0.92 1.57 2.49 0.92 1.57
16 -30 0.5 0.1 $100 -160 260 0.77 1.45 2.29 0.84 1.52
Asymmetry 17
3
-30
-30
0.5
0.5
0.1 ++ 4010 -130
-100
140
140
0.96
1.00
1.64
1.72
1.98
2.85
0.34
1.10
1.02
1.80
0.1
18 -30 0.5 0.1 + 70 - 70 140 1.21 1.90 3.58 1.68 2.37
19 -30 0.5 0.1 +lo0 - 40 140 2.08 3.17 7.03 3.86 4.95
20 -30 0.5 0.1 +lo0 -100 200 0.98 1.65 2.83 1.18 1.85
T = l

T h e effect of asymmetry on response time is shown in Figure tions studied. For test 3, Table I, if T = 1 hour, the two-
12. Asymmetry occurs when R is not midway between k and position response would be 1.8 hours faster than the PID
K . Asymmetry in the K direction (refer to Figure 7 once tb%. The degree of improvement is best for large set point
again) means shifting K and k upward while maintaining K - k changes on processes with large a and b or narrow limits of
constant a t 140, with fixed R, = 0 and R = -30. The re- asymmetric saturation.
sponse time becomes very large near k - R = 0, since this set
point requires the saturation minimum, k , a t steady state. Applications
T h e two-position improvement is better for greater asym- An immediate application of this work is the use of the time
metry. switching equation for manual set point control on a typical
I n summary, the time optimum controller is significantly industrial controller. An experienced operator often applies
superior to a well tuned PID controller for nearly all condi- full on or off saturating action for some period of time before

l I , I , I I , I ~
0 0.I 0.2
0.5 I.o
MINOR TIME CONSTANT, b TRANSPORTATION LAG, a

Figure 9. Effect of minor time constant on re- Figure 10. Effect of transportation lag on re-
sponse time for two-position and PID controllers sponse time for two-position and PID controllers

468 I&EC FUNDAMENTALS


3.5r 7
At ’ PID ~ ~ O / ~ -t5B B i
a a 0.1 At = PID ~ ~ O / ~ -t5B B
b= 0.5 6 a 0.1
3.0 - O\ R :-30
a
b = 0.5
R =-30
K-k m I40 I
2.5 -
PID t5el0
\ 5

4
c
\
c I-
z
+
w‘ 2.0-
z
c
\ O
g
I-
3

W
v)
z
2
v)
1.5- ‘“Y O U
w
i2
a
2 0

1.0-
L? ‘0
t5 I

-- -
I l l 1 I I I I I I I 0
80 140 200 260
SATURATION RANGE, K -k 100
40
-70
k- R
100
40 -10
130
70 K-R
Figure 1 1 . Effect of saturation range on response ASYMMETRY IN K DIRECTION
time for two-position and PID controllers
Figure 12. Effect of asymmetry on response
time for two-position and PID controllers
putting the control set point to a new desired value. If
parameters T , b, K , and k can be adequately approximated,
Equation 15 can be used to make a table of values for t z for r
various el = R - R,. Thus, to make a set point change, the
Ro= 0
pointer can be moved to cause full force in one direction u p to .
15 -
@I=R
t 2 , then full force in the opposite direction until t 4 . At t 4 , K = I,O in.
k =-IO in.
the pointer is returned to the desired set point value, R. This b = 0.150
procedure was found to function well if the integral action was a = 0.154
turned off during the transient, and subsequently reapplied - T = 6.9min.

when the process reached the vicinity of its new control point. v)
W
This avoids the undesired integration of the errors during the e
z
3
two-position transient. If use of this simple two-position 5 10-
action with full reversal is made for process startups and a
transitions, significant time savings will be made in bringing r
the process to a new steady state. e5 -
Lapse (72) found the transfer function of a cracking furnace v)

to be c
exp (-1.06s)
5
v)
5-
a
(6.9s f 1) (1.03s f 1) :
where the constants are in minutes. For this case, T = 6.9 Y -
c
minutes, a = 0.154, and b = 0.15. Figure 13 indicates the
switching time, t 2 , for various set point changes from R, = 0
to R. Saturation was assumed to occur a t K = 1 inch and
k = - 1 inch from the center of the chart, R,. If the actual
0 -a5 -ID
SET POINT CHANGE, R INCHES OF CHART
process b deviated from the assumed b = 0.15 by loyo,t 4 was
i n error by approximately 0.1 minute. Figure 13. Time for first switch for set point
If the process is subject to loads during programmed two- changes on Lapse’s cracking furnace
position control, a different trajectory will be followed. Since
the operator will switch to a linear feedback controller near the
origin, deviations due to loads will not greatly affect this is also valid for a system with transportation lag in the feedback
procedure. Also, calculations indicate the position of the measuring element.
switching line to be somewhat insensitive to loads when K k - T h e results of this work provide an upper bound on the
is large. possible response time improvement which can be made over
If the switching time is consistently in error, a trial and error “good” P I D control of processes with transportation lag. I n
adjustment of parameter b or T may give improvement. addition, they indicate the class of control situations for which
Equation 1 5 might even be used to determine or confirm these possible response time improvement is greatest. This informa-
model parameters. tion forms the basis for further research on the development of
We show in Appendix I1 that the optimum control function simple, suboptimal control methods, such as the suggested

VOL. 4 NO. 4 NOVEMBER 1 9 6 5 469


manual set point manipulation, for process control. In par- Suppose also that y ( t ) is related to x ( t ) by a fixed dead time
ticular, Equations 14 and 16 provide the basis for a dual mode element,
feedback controller, providing two-position action when the
y ( t ) = x(t - U T ) , for all t (-43)
error exceeds some predetermined value, and standard control
otherwise. T h e results are restricted to the class of processes where U T> 0. We must also specify an initial function M ( t )
whose transfer functions may be approximated by Equation 1. = M,(t), on the interval - a T < t <
0, because dead time
makes G(s) exp (-aTs) history-dependent. With x ( - U T )
Conclusions and M,(t) given we can solve Equation A1 for x ( t ) , - a T <
t 6 0. This determines x(0). Equation A3 gives y ( t ) ,
The time optimum programmed control function, giving 0 <t 6 a T, so the initial part of y ( t ) is predetermined from the
switching time as a function of e l , R, R,, k , K , b, and T, may be arbitrary M,.
directly applied to drive systems with dynamics adequately MI*(t), t > 0 brings the system from x(0) to equilibrium,
represented by Equation 1 from one condition to another in x ( t J = R, in minimum time, t4*. I n theory, Ml*(t) can be
minimum time. determined uniquely from Pontryagin's maximum principle
I n contrast to this programmed control function, which does or variational techniques. Desoer (6) indicates the nature of
not require measurement of state variables or knowledge of the Ml*(t) as two-position and describes the concept of switching
dead time, the time optimum feedback control law requires surfaces in x space. From Equation A2 we see that M l * ( t )
this additional information, and is more difficult to implement = u,R, t > t4*.
(7). We must specify M 2 * ( t ) ,t > 0 such that y ( t J = R, the
Additional research on time optimum control of chemical desired equilibrium position, in minimum time t5*-that is,
processes may be profitably directed toward establishing
procedures for fitting Equation 1 to experimental process
t j * < t j , for k < <
M ( t ) K . y ( t ) ,0 < t a T i s predetermined
6
by M , and is not influenced by Mz*(t).
dynamics in such a manner that the control function of Equa-
tion 15 gives rapid control.
Now, if M ( t ) = M 1 * ( t ) ,t > 0, y(t4* +
U T )= R from Equa-
tion A3 because x(t4*) = R. Furthermore, y ( t ) cannot reach
R in a time t 5 < t 4 * +
aT, since this would imply that x(t)
Appendix I reached R in a time less than t l * . Hence, Mz*(t) = Ml*(t),
Proposition. If M l * ( t ) , t > 0 is the time optimum control
t > 0 and t j * = t 4 * +
a T . The time optimum control func-
tion is independent of dead time, since M1*(t)is determined
function for a stable, completely controllable system, G(s), with
solely by G(s).
no dead time, and if Mz*(t),t > 0 is the time optimum control
The equivalence of M I * and M2* does not carry over to the
function for the same system with dead time, G(s) exp(-aTs),
feedback case where M * is to be written as a function of the
then Mz*(t) = M1*(t), t > 0 .
state variables, M [ y ( t ) ]rather
, than of time. This is the point
a t which Balakirev ( 7 ) left the problem, in an unsuccessful
M(t) Y(t)
G(s 1 exp(- aTs) c attempt to construct a feedback control law for arbitrary x,.

Appendix II
Suppose a dynamical system is governed by the differential
equation Proposition. The time optimum control function for sys-
;(t) = Ax@) + bM(t) (AI)
tems with dead time in the feedback measuring element is the
same as for systems of Figure 1.
k < M ( t ) 6 K , all t Given the system

x (-UT)= x, M(t) exp (-a,Ts) C,(t)


where x = n vector of state variables (Ts + I ) (bTs + I ) c
1 1
b = a constant m vector
A = a real constant matrix with distinct eigenvalues I
with nonpositive real parts. T h e eigenvalues
may be complex
M ( t ) = manipulated variable L I
k, K = preassigned constants
x, = arbitrary initial condition n vector for x at t = U T- M1*(t), t > 0, is the optimum forcing function for Cl(t),
and M z * ( t ) ,t > 0, is the optimum forcing function for C,(t).
If the vectors (b, Ab, A A b , A3b, . . . . ., An-'b) are linearly Since M1*(t),t > 0 is independent of dead time, alT; the
independent, the system is said to be completely controllable. system can be rewritten as
This ensures that it is possible to bring any initial state xo to
the origin in finite time, t 4 ; and the system equation can be
written in the form of a single nth-order differential equation:

so that Mz*(t),t > 0, is independent of dead time, (a1 a2) T; +


and the two systems differ only in dead time. I t follows that
M**(t) = MI*@),t > 0.
where Cl(t) is the system output. Higher derivatives of Cl(t)
correspond to other elements of the x vector. We can there-
Acknowledgment
fore take Equation AI to be in phase form, so that x 1 = C1.
The nth-order transfer function, G(s) 3 Cl(s)/M(s), can be We gratefully acknowledge the support of Purdue University
found. T h e desired rest position is x = R = (R,0, 0, . . . . . , and a National Science Foundation Fellowship to P. R. Latour
0). during this work.

470 l&EC FUNDAMENTALS


Nomenclature t 6% = time for error to reach and remain within
=t5% el of origin for set point changes
a = normalized dead time (within i570L for step load changes)
aT = transportation lag or dead time x = state variable
b = normalized minor time constant X = state vector
bT = minor process time constant, 0 < b < 1 = conjugate system variable defined by Equation
A&)
= process output 4
= normalized, transformed process output h = conjugate system vector
= constants of integration
= process error, e = R - C literature Cited
(1) Balakirev, V. S., Automation Remote Control 24, 948 (1963).
(2) Biery, S. C., Boylan, D. R., IND.ENG.CHEM.FUNDAMENTALS
Pe 2, 44 (1963).
“‘dtz (3) Bushaw, D. W.? “Differential Equations with a Discontinuous
Forcing Term,” in “Contributions to the Theory of Nonlinear
= initial crror after set point change Oscillations,” Vol. 4, Princeton University Press, Princeton,
= error ai: first switch, e ( t 2 ) N. J., 1958.
= error dlerivative a t first switch, e ( t 2 ) (4) Cotter, J. E., Takahashi, Y., Chem. Eng. Progr. Symp. Ser.
= functions defined by Equation 3 46, 59, 119 (1963).
= forcing function (5) Coughanowr, D. R., Koppel, L. B., “Process Systems Analysis
= process transfer function and Control,” McGraw-Hill, New York, 1965.
= Hamiltonian function (6 Desoer, C. A., Information Control 2, 333 (1959).
= upper limit on M, a constant (71 Eckman, D. P., “Automatic Process Control,” Wiley, New
York, 1958.
= lower limit on M, a constant (8) Feldbaum, A. A., Automatika Telemekhanika 14, 712 (1953).
K, = process steady-state gain (9) Grethlein, H. E., Lapidus, L., A.Z.Ch.E. J . 9, 230 (1963).
L =
=
=
process load variable
manipulated variable
normaliized transformed manipulated variable
I 10) Hougen, J. O., Chem. Eng. Progr. Monograph Ser. 4, 60 (1964).
11) Kurzweil, F., ZEEE Trans. Auto. Control AC-8, 27 (1963).
(12) Lapse, C. G . , I S A J . 3, 134 (1956).
aT) = manipulated variable a delay time in the past (13) .Latour, P. R., “Application of Pontryagin’s Maximum
= desired value of C; controller set point Principle to Time Optimum Control of a Second Order Over-
= initial set point damped System with Transportation Lag,” thesis for M.S.
R* degree, Purdue University, W. Lafayette, Ind., June 1964.
T = predominant process time constant (14) Pontryagin, L. S., BoltyanskiI, V. G., Gamkrelidze, R. V.,
= time, measured a t controller Mishchenko, E. F., “The Mathematical Theory of Optimal
= time zero of set point change Processes,” Wiley, New York, 1962.
= time for first switch (15) Williams, T. J., et al., “Approximation Models for the
=
=
tz+ a?”
-
time for final switch, t 5 aT
Dynamic Response of Large Distillation Columns,” Fourth
Joint Automatic Control Conference, Minneapolis, 1963.
= minimum time to reach origin from e(0) = el, RECEIVED for review July 20, 1964
e(0) = 0 ACCEPTED May 3, 1965

MONTE CARLO SIMULATION


OF REACTION KINETICS
G E R A L D D. C O H E N l
Operations Research Department, Allied Chemical Corp., New York, N . Y.

Complex set!; of equations describing the concentration-time profiles of a chemical reaction can be solved
by random number (Monte Carlo) simulation techniques. The basis for a valid solution is a stochastic model of
a chemical reaction in which the reaction path is not fixed but allowed to vary from trial to trial. An example
is presented to illustrate the convergence of Monte Carlo simulation to results predicted by the laws of mass
action. A relationship between Monte Carlo sampling and numerical integration is demonstrated which
should prove useful in the numerical solution of complex reactions (involving a series of equations) difficult to
solve by analytic procedures.

differential equations describing concentration-time


HE teristics. T h e simulation of reaction kinetics under these
T changes in a chemical reaction can often be formulated, conditions can provide information that may reduce the
but not solved by direct analytic means. Various methods amount of laboratory or plant experimentation. I t could also
have proved useful in this situation. T h e two most common be used for experiments which cannot be physically performed,
are analog computer simulation and numerical integration o n a and a fruitful area of application of this idea is in the field of
digital computer. A third method, Monte Carlo simulation polymer chemistry. I t is possible to consider reactions con-
o n a digital computer, can also be used, but has only recently taining several hundred species (representing different molec-
received attention for this application ( 5 ) . I t has the im- ular weights and types, etc.), and arrive a t results either
portant ability to include the realities of a n industrial situation, difficult to obtain or undeterminable by direct measurement,
such as changing temperatures and pressures, the addition or such as molecular weight distribution.
removal of catalysts and inhibitors, and general process charac- T h e basic idea in a Monte Carlo simulation is to devise a
set of sampling rules which, when followed sequentially,
1 Present address, Mat:hematica, Princeton, N. J. provide a realization of the reaction mechanisms. The

VOL. 4 NO. 4 NOVEMBER 1 9 6 5 471

You might also like