Академический Документы
Профессиональный Документы
Культура Документы
ABSTRACT
This paper outlines a linear model-based control design procedure, suitable for processes involving twophase phenomena, where the transient response is often subject to nonminimum phase characteristics:
inverse response and time delays. We show that in order to capture these features in the process models
implicit in the control scheme, it may be necessary to model the transients in each phase separately,
thus precluding simple homogeneous approaches. Furthermore, as shown by two examples, combinations of nonminimum phase characteristics can be accurately modelled by the Parallel Linear Structure,
introduced here, which can be the basis for improved linear control design.
KEYWORDS:
Process Dynamics, Two-Phase Systems, Inverse Response, Evaporators, Boiling, Feedback Control
1. INTRODUCTION
Inverse response phenomena are commonly found in chemical processing equipment. They may occur
when a single input perturbation affects the process output in two opposing directions, with one
direction being the dominant steady state effect, and the other being the dominant transient effect.
Consider the simple example of a liquid flowing in a heated pipe. Since the thermal expansion coefficient is usually positive, a sudden temperature increase at the inlet will soon cause the first part of the
pipe contents to heat up and also to expand. Therefore, the residence time of the fluid will
progressively decrease, so that the outlet temperature will also decrease before the expected (positive)
thermal wave reaches the outlet (Fuchs et al, 1991).
Since the mean density of a two-phase mixture is usually a very sensitive function of intensive
variables, one would expect that inverse response phenomena should be common in two-phase systems.
Selected examples amongst the many reported cases in the literature are:
1) Distillation Column Tray Dynamics: Luyben (1969), analyzed the response of a binary distillation
column to a step increase in the internal vapor flow rate. This change affects the transient of the
light component concentration on a given column tray through two mechanisms. The first,
dominated by the tray hydrodynamic time constant, will cause the concentration to rise because the
additional vapor, rising through the tray and rich in the lighter component, will displace fluid held
on the tray, previously richer in the heavier component. The long-term effect is controlled by
thermodynamics and will therefore reflect the overall tray steady state gain: as the temperature
increases, the concentration of the lighter component will decrease. The overall transient therefore
exhibits inverse response.
2) Azeotropic Distillation Column Dynamics: Andersen et al (1991) analyzed the dynamics of
homogeneous azeotropic distillation, and found that the effect of internal flows in the azeotropic
column affect the overall separation in two directions, each dominated by different time constants.
Thus, depending on the chosen operating point, both overshoot and inverse response phenomena are
possible.
3) Two-phase Reactor Dynamics: Tsai and Tsao (1991) describe the dynamics of a solid-catalyzed,
liquid-phase reaction in a CSTR. The existence of inverse response in the bulk fluid concentration
transient for such a system is shown to be caused by an appropriate temporal balance between
reactant supply rate to the bulk fluid and the mass transfer rate to the solid catalyst.
SI49
5150
Bearing in mind that inverse response is usually not a desired property, it is of interest to be able to
predict under which conditions such a characteristic will occur. This is the justification for detailed
modelling of dynamical systems which potentially exhibit this phenomenon. Given a dynamic model
which captures the process' true nonminimum phase nature, the unified approach offered by Internal
Model Control design principles (Morari and Zafiriou, 1989) can be utilized to perform adequate
feedback control design. In this paper, we shall consider two common chemical processing units, both
featuring two-phase flow: an evaporator and a crude oil preheating furnace. We shall show that careful
modelling of these systems generates uncertain, linear approximations of the process dynamics which
includes nonminimum phase components. Control designed on the basis of these uncertain models can
be guaranteed to satisfy robust performance specifications.
2. NONMINIMUM PHASE BEHAVIOR IN LINEAR PROCESS DYNAMICS
Processes which exhibit inverse response and delay time are potentially difficult to control. These two
phenomena are termed nonminimum phase because they impart additional phase lag to the open loop
response. Consider for example, the net first order processes:
p
(s) -
k(T 3S+1)
d
k( - T3s+1)
an p (s) - ~-'--:-:-"":;---':-",
(T1S+1)(T2s+1)
2
- (T1s+1)(T2 s+1)
(1)
The two processes have the same frequency-dependent magnitude. The expected high-frequency phase
which is the case for Pl' However, the presence of the positive
lag is therefore ;'moo<p(pj(iw)) = zero (s= I/T 3) in P2 adds an additional phase lag of -1r radians at high frequency.
I'
1.5
---
1.0
0.5
0.0
-0.5
<a>
-1.0
10
-1.0
15
<b>
-t-----.---r--~--y---.__1
I
Time
10
15
Time
Fig 1: The open loop response of pds) and P2(s) to unit step changes.
Parameters are: k= 1, T}=l , T2=5 and (a)T3=O.5; (b) T3=5.
The presence of the delay time in the process transfer function also has serious consequences regarding
achievable closed loop performance. The relationship between the two nonminimum phase components
can be gaged by the use of a Pade approximation:
e-
(J
s -
-~s
2
+1
(2)
- 8 + 1
Thus, as shown by the approximation (2) and by the simulated responses in figure I, the inverse
response caused by the positive zero depends on its magnitude relative to that of the dominant process
pole. The larger the values of T 3 in P2(s), equivalent to large (J in (2), the more problematic will be the
control problem . Mathematically, the transfer function form given as P2(s) can result from two
parallel first order responses to the same input signal affecting the same output in different direction s:
k}
k2
P2(s) = T}s+1 - T2s+1
where k = k} - k2 and
T3
k( - T3s+1)
k
.
= 21"}-k}1"2
k
k
. For Inverse response,
1
(3)
= (T}s+1)(T2s+1)
2
T3
..
must be positive,
3.1 Overview.
In order to successfully design model-based feedback control, the following items of information must
be available: (a) a model describing the response of the output to control variable perturbations; (b) a
quantitative estimate of the reliability of the model; (c) a definition of the disturbances to be rejected
SI51
and/or the types of setpoint changes to be tracked; (d) specifications concerning the closed loop
performance.
r-r
t=:;r'
(b)
(a)
Fig 2: (a) The fMC feedback structure; (b) The classical feedback structure.
The concept of Internal Model Control (Morari and Zafiriou, 1989), enables systematic control synthesis for processes with uncertain linear models. IMC design utilizes the structure given in figure 2(a),
which explicitly includes the nominal process model, p(s), and a model-based controller, q(s). Following
the IMC design procedure, the controller is generated in two steps. Firstly, an ISE (Integral Square
Error) optimal controller, q(s), is computed based on the nominal process model, disregarding both
model uncertainty and any constraints. In its simplest form, the optimal controller is the inverse of the
invertible part of the nominal model. This is then used to form the IMC controller, by augmenting it
with a low-pass filter, q(s) = q(s)f(s). This ensures that the controller is realizable (the number of poles
must at least equal the number of zeros), and that it can be detuned to compensate for the
destabilizing effect of process uncertainty. The IMC and classical feedback structures are related,
making it possible to use the IMC procedure to carry out the design and revert to the classical structure for implementation. The classical controller, c(s), can be expressed as: c(s) = q(s)(l- p(s)q(s))-I.
A prerequisite to the design of linear model-based controllers is the generation of linear approximations
for the process dynamics. These can be obtained by either linearizing a non-linear process model or by
empirically fitting linear transfer functions to plant data or simulation results. As a result, the linear
model parameter values may have significant associated uncertainty, which can be expressed as a
bounded range. This defines the set p(s) E 7I"(s), where 7I"(s) defines the set of processes computed from
all possible combinations of extreme parameter values. A multiplicative uncertainty bound, lm(s), can
be defined, on the basis of a nominal process model, p(s):
lm(w) =
(4)
Dependent on the efficacy of this uncertainty description, a necessary and sufficient condition can be
derived for robust stability. Assuming that the controller stabilizes the nominal process model, then
closed loop stability is guaranteed on the uncertain process p(s), if and only if
lii(iw)llm(w)<l 'tIw, 'tip E 71"
(5)
where ii = pc(l+pcrl is the nominal complementary sensitivity function, which can also be related the
IMC controller, ii = pqf. The form of the complementary sensitivity function is a design decision. Two
important factors playa role in its correct definition:
1) Nonrninimum phase process characteristics: If the process dynamics includes noninvertible nonminimum phase components, the effect of these characteristics will also appear in the closed loop
response, and also form part of the complementary sensitivity function.
2) Input type: The nominal loop transfer function, pc, must have as many integrators as the number of
integrators in the input signal (disturbances to be rejected or setpoints to be tracked). Thus, for
example, if the input is effectively a ramp, jic must contain two integrators. This has consequences
regarding the form of ii. In the IMC design procedure, the requirements of input type are met by
choosing a filter of appropriate form. Thus, for step inputs, the simplest form of the filter is:
f(s) =(As+lr n while for ramp inputs, the simplest form is: f(s) = (nAs+l)(As+lr n. In both cases,
the number of filter poles, n, must be equal to or greater than the net order of the open loop process.
3.2 Design Methodology.
A. Modelling Process Dynamics.
Often heterogeneous models will be required in order to capture inverse response characteristics
observed in two-phase systems. In addition, the thermal feedback associated with heat integrated
systems will result in delayed shock wave effects, modelled by delay times. A generalized Parallel
Linear Structure (PLS) which will allow both inverse response and delayed shock waves to be modelled
is shown in figure 3. In this structure, PI' P2 and P3 are all minimum phase transfer functions. The
CACE 16 Suppl.-L
SI52
presence of the delay terms, permit delayed shock waves to be modelled. The parameters for the PLS
model can be derived by an analytical approach involving linearizing a lumped parameter non-linear
model or by empirically fitting of the parameters in order that the PLS prediction matches the
simulated response of the non-linear model to disturbances in the input variables. This is illustrated in
figure 4, where the simulated response of the effluent temperature of a crude preheating furnace to a
step change in the fuel feed rate, as predicted by a detailed , distributed parameter, non-linear model is
compared with the response using a PLS model with a significant delay on the third term (Fuchs ei ai,
1991). In general, the analytical approach is preferred since this will yield condit ions for the existence of
inverse response phenomena.
1'_.
I:
u
100
200
300
400
500
600
700
800
Time (secollds)
In order to design adequate control, it is imperative that th e numerical values of the linear model
parameters used in the design be bounded within a predefined range. Only then can closed loop
stability and performance be guaranteed. The analytical linearization approach can be used to define
the extreme parameter limits given the process' operating range. If analytically defined coefficients are
not available, the range of parameter limits can be estimated by performing a series of random changes
on each process input in the vicinity of a given operating point , and fitting the parameters of the PLS
to each simulated response. The result will be a set of possible numerical values for each coefficient of
the model, which will delineate extreme values.
B. Control Design.
The modelling step will have produced a PLS approximation for the process dynamics. The best
possible feedback performance will be obtained if the same structure is conserved and integrated into an
IMC controller structure. Conventional, classical feedback can also be designed, but this will require
model reduction to generate a single rational transfer function approximation for the PLS. In general,
the price paid for this simpler control implementation is reduced performance, although if there is
significant uncertainty in the PLS coefficients, the loss of performance relative to the full-order PLSbased controller may be negligible.
If required, model reduction should be carried out while taking closed loop performance specifications
into account, thus simultaneously defining the controller's tunable parameters. For example, for the
control of furnace effluent temperature, the design of a PI controller required that the thre e block PLS
model be approximated by a first order lag. This approximation should be carried out by matching the
closed loop transfer functions obtained with the reduced order model, with that obtained with the PLS
model. This allows the simultaneous tuning of reset tim e and controller gain , in a control-relevant
manner, and will maximize the closed loop bandwidth . In general, the IMC parametr ization simplifies
the implementation of this approach because model-based control synthesis reduces the number of
tunable parameters.
4.1 Modelling.
Figure 5 shows a schematic representation of the single stage pilot-scale vertical evaporator used in this
study. The subcooled liquid feed enters the riser tubes from below, while steam enters the shell around
the tubes and condenses on their outer surfaces. The two phase mixture flows upward into the overhead
drum, which is a large separation vessel, where the vapor produced separates out of the fluid phase
and exits to a condenser. The concentrated solution is withdrawn from the bottom of the separation
vessel as product.
SI53
-+vapor
~zQ~'if:.-L '
: : 0-
bo
"': .f:'"
00
/0
ovorb.ad drum
- + concentrate
steam chest
SI54
marked in the figure. As explained, the system topology requires that pressure changes be modelled
with a delayed effect on x, Then, given the current value of the vapor fraction entering the overhead
drum, as well as all process inputs, mass balances on the overhead drum vapor and liquid phases will
constitute the state equations. With current volumes of gas and liquid in the overhead drum fluid, the
output of interest, the evaporator level, can be easily computed.
The accumulation of liquid in the drum depends on a balance between the unevaporated feed entering
via the evaporator tubes and the exiting concentrate flow:
dV
dtI = L t(l - x) -
(6)
The accumulation of vapor in the fluid depends on a balance between the gas-phase feed entering from
the riser tubes, and the vapor escaping from the fluid, whose rate we assume to be proportional to the
vapor fraction in the fluid, to the superficial vapor velocity, u g, and to the cross-sectional area of the
drum, At: dV
V
g
PI
g
<It = Ltx pg -ugA t VI + V
(7)
g
A value for the superficial vapor velocity will be determined by empirical fit with plant data. The
instantaneous level in the reflux drum depends of the total volume of fluid:
A
VI + Vg
A
(8)
........
....
: I::
Minimum Value
Maximum Value
Units
K1 = J8x/8L tJ
Tin,Pst
- 4.67 x 103
- 2.99 X 103
(~c)
K2 = 18x / 8T jn l
:::0
4.2 x 10-4
(~r
6.47 x 10-4
26.4 X 10-4
(:Sl
Pst' Lt
K3 = 18x / 8P st l
Tjn,Lt
3 -1
Having fixed the model parameters in such a way that the experimental vapor fraction data is
predicted for the range of operating conditions used, our dynamic modelling approach requires that the
sensitivities of vapor fraction with respect to changes in process inputs be determined. This was
accomplished by solving the model describing conditions in the evaporator tubes at steady state, as
described by Faigon (1991). The numerical results are summarized in Table I. Bearing in mind that the
typical disturbance perturbations in the three inputs will be: ilL t 1= 2 X 10-5 m 3/sec, ilTin =20c and
ilP = 100 kN/m 2, this means that the sensitivity to the average perturbation in the inlet temperature will be 2% of the sensitivity to the average pressure perturbation. This is the justification for
ignoring the effect of inlet temperature on the evaporator dynamics.
I atl
SI55
Linearization and Laplace transformation of the model given by equations (6) - (8) yields transfer
functions relating the evaporator level to perturbation in the control variable, B, and in the three
disturbances (Pst. Tin and Lt). The response of the evaporator level to perturbations in the concentrate
flow rate, B, is given by:
lL)s+ 1)
h) ((
g
H(s)
(
u ss
)
B(s) = p(s) = - AtVl 5S ((A h 2)
s
s+1
lUg
ss
(9)
The transfer function relating perturbations in the total feed rate to the evaporator to perturbations in
the evaporator level is:
(()
h(l- b+ab)
s+
55
)
u g(l- b)
((A h 2)
H(s) _
_ (h(l- b))
Lt(s) - Pdl(s) VI
5S
[ox]
oL
t
-V
lUg
)
1
+1
(10)
S5
Tin'
Pst
Noting that a ~ 1, and that Ibl=o(I), the expression (10) can be simplified to:
(I(
H(s)
Lt(s)
hab )S+I)
h(l - b)) \: u g (1 - b) ss
(
((AVl~gh
2
VI
5S s
(11)
+ 1)
ss
The sign of the process zero depends of the sign of b. For the operating range of the evaporator, the
limits for x and Lt (8X/ 8Lt) are: 0.1 :S x :S 0.6 and - 0.375 :S Lt (8X/ 8Lt) :S - 0.075. As a consequence,
the parameter b is limited by the values: - 0.275 :S b :S 0.525. Note that the steady state gain is always
positive. This result is interesting because it indicates that inverse response is possible as a result of a
perturbation in the flow rate. Analysis of the parameter values which give a positive zero, we find that
they occur when x is minimum (i.e. Pst is minimized) and the term Lt l8x/ 8Lt Iis maximized (i.e. Lt is
maximized), which makes sense physically.The transfer function relating perturbations in pressure to
perturbations in the evaporator level is:
H(s)
P
= Pd3(s) =
5t(s)
(12)
Here, the steady state gain is always negative, while the process zero is always positive. From analysis
of the relative sensitivity of the evaporator hold-up to the two disturbances considered, it is apparent
that the pressure disturbance is the more significant of the two. We shall now consider the design of
control to reject this disturbance.
4.2 Control Design.
From the linear approximations describing the response of the evaporator level to changes in the
concentrate (eq.9) and steam pressure (eq. 12), it is apparent that a step change in Pst is seen as a
ramp (l/s 2) in the process output. Thus, the IMe controller that will guarantee set point tracking for
such a disturbance will be:
q(s)
s(rp2s+1)(2.\s+1)
= p(_ s)-12.\s+1
- - =-'"''-------:--:----,-;;(.\s+ 1)2
(13)
where kp=(A
Ah
Jy
),
Tpl=(~) and TpZ=(Vtu ). This is equivalent to the classical feedback controller:
tlss
gss
Igss
q(s)
(Tp2s+1)(2.\s+l)
c(s) = I-p(s)p(s) = k .\2s(T
p
p1S+1)
(14)
Sl56
..
1..
n.
....
n.
.....
system is illustrated in figure 8. We note t h a t .
Time(.ec)..
the level can be kept almost constant, at the
FIg 8: Closed loop response to a dIsturbance In Pst
cost of a rather violent manipulation of B. The large control action is a result of the relatively low
sensitivity of the level to concentrate flow.
.....
.....
.....
'_1
5. CONCLUSIONS
This paper has outlined a control design methodology for processes featuring two-phase flow. The
evaporator example has shown the need for accounting for mass accumulation in each phase separately,
which is the cause of potential inverse response effects which would be absent in a homogeneous model.
The Parallel Linear Structure (PLS) has been introduced as a means of modelling complex dynamic
response. Such a structure can be used to model both inverse response (as brought by positive zeros),
and transportation time delay, which can enter in a complex fashion. In the furnace model, the effect of
all inputs considered (feed rate, inlet temperature and fuel feed rate) can accurately modelled by three
first order transfer functions in parallel, appropriately delayed to account for shock-wave effects. In the
evaporator, two at most are required. Finally, the concept of Internal Model Control has been shown to
provide a suitable framework for robust control design.
ACKNOWLEDGEMENT
We are thankful to Prof. Samuel J. Wajc for his helpful comments. This research was partially
supported by the Henri Gutwirth Fund for Advancement of Research.
REFERENCES
Andersen, H. W., Laroche, 1. and Morari, M. (1991): "Dynamics of Azeotropic Distillation Columns",
Ind. Eng. Chern. Res., Jill, 1846-1855
Andre, H. and Ritter, R. A. (1968): "Dynamic Response of a Double Effect Evaporator, Can. Jnl. of
Chern. Eng.,1., 1259-1264
Chen, J. C. (1963): "A Correlation for Boiling Heat Transfer to Saturated Fluids in Convective Flow",
ASME Paper 63-HT-34
Faigon, M. (1991): "Simulation and Robust Control of a Two Stage Evaporator", M.Sc. Thesis (in
Hebrew), Technion
Fuchs, A., Lewin, D. R. and Wajc, S. J. (1991): "Modelling, Simulation and Control of a Crude Preheating Furnace: I - Approximate Crude Description", submitted to Chern. Eng. Sci.
Luyben, W. 1. (1969): "Feedback and Feedforward Control of Distillation Columns with Inverse Response", I. Chern. E. Symp. Series.al, 6:39-48
Morari, M. and Zafiriou, E. (1989): Robust Process Control, Prentice-Hall Inc., Englewood Cliffs, NJ
Newell, R. B. and Fisher, G. D. (1972): "Model Development, Reduction and Experimental Evaluation
of an Evaporator", Ind. Eng. Chern. Proc. Des. Dev., 11(2), 213-221
Serizawa, A. (1979): "A Study of Forced Convective Subcooled Flow Boiling", in Two Phase
Momentum, Heat and Mass Transfer, ed. Durst et ai, Hemisphere Publishing Corporation.
Tsai, J-G. and Tsao, G. T. (1991): "Dynamic Response of a Two-phase Continuous Flow Reactor",
Chern. Eng. Sci., 12(3), 881-888