Anda di halaman 1dari 8

Published at Proceedings of the 3rd IEEE Conference on Control Applications, Glasgow, 281-288 (1994)

μ-Optimal Advanced PID Control of an Industrial


High Purity Distillation Column
H.E. Musch, M. Steiner
Measurement and Control Laboratory
Swiss Federal Institute of Technology (ETH)
CH-8092 Zurich, Switzerland

Abstract ties and consequently a high controller performance. Usually,


distillation columns are operated in a wide range of feed flow
State-space controllers are difficult to implement in a distrib- rates and feed compositions. Therefore any linear time-invariant
uted control system. Therefore a control design based on the controller must exhibit a satisfactory performance not only at a
common PID control structure is desirable. This paper presents single operating point, but for the entire operating range of the
the μ-optimal tuning of advanced PID control structures as well distillation column. In earlier publications ([6], [7]) it has been
as a simple feedforward control design for an industrial binary shown that this design task can be handled within the frame-
distillation column. The design is based on an optimization work of the structured singular value μ. However, the common
approach. The resulting controllers guarantee stability and a synthesis methods (DK-iteration [2], μK-iteration [4]) lead to
high performance for the entire operating range of the distilla- state-space controllers of a high order. As mentioned before, the
tion column. A summary of the practical experiences comple- acceptance of state-space controllers is still low and the imple-
ments the theoretical results. mentation in a DCS is troublesome. Therefore, it is very desir-
able to obtain robust controllers using the common PID or
1 Introduction advanced PID control structures.
Modern control theories provide the process control engineer This paper presents the design of μ-optimal PID control struc-
with increasingly sophisticated tools for a robust, model-based tures including a simple feedforward controller for an industrial
controller design. The advantages of these controllers over on- binary distillation column. The theoretical results are supported
line tuned PID control structures have been shown in numerous by the results of the controller implementation at the real plant.
publications. Nevertheless, only a few applications of the
modern control theories have been reported. Therefore the 2 The distillation process
mismatch between theory and practice is still evident. Some of
the reasons for this situation are listed below. The industrial distillation column is equipped with 50 sieve
trays, a total condenser, and a steam-heated reboiler. The feed of
Distributed Control Systems: For a control engineer in the this column consists of the bottom product streams from two
process industry, process control in the first place is a hardware parallel operating columns, buffered by a tank. In the main it is
problem. His perspective is the installation and configuration of a mixture of two components. The feed is split into the two
a Distributed Control System (DCS) [1]. Even a modern DCS is product streams D and B. While the level of the reflux accumu-
often limited to PID and advanced PID control. For the DCS, lator is controlled by overflow, the top pressure is controlled by
the implementation of modern state space controllers requires a vacuum pump. The most important operating data including
either the coupling with an external computer or the program- the operating range are summarized in Table 1. Due to high
ming of software modules. Both ways are troublesome and investment and maintenance costs of composition analysers,
expensive. pressure-compensated temperatures on trays 10 and 44 are
Economic benefits: The economic benefits of improved control controlled instead of the product compositions. The manipu-
tend to be significantly underestimated. A benchmark study by lated variables chosen to control these two tray temperatures are
ICI indicated that “the effective use of improved process control the reflux L and the boilup V.
technology could add more than one third to the worldwide ICI
Group’s profits” [1]. Another study shows smaller, but still 2.1 The control objectives
massive benefits [5].
The main objective of the control system is to keep the
Of course it is not necessary to replace all PID controllers by controlled tray temperatures close to their setpoints in presence
model based control structures. Most control problems in the of disturbances. The most important disturbances are changes of
process industry are handled well with simple on-line tuned PID the feed flow rate. In order to keep the level of the buffer tank
control. However, strongly nonlinear or ill-conditioned plants within the specified limits, the feed flow rate typically is set by
require advanced control techniques for a high controller hand every couple of hours. This means that the feed flow rate
performance. is changed stepwise. Variations of the feed composition are the
High-purity distillation columns are such ill-conditioned and second important disturbance input. Being more smooth, these
strongly nonlinear plants. The economical and ecological feed composition disturbances affect a lower frequency range
optimal operation requires a tight control of both product puri- than the feed flow disturbances. All other disturbances, such as
Table 1: Steady-state data The performance specification: A controller design or a robust
performance analysis using this uncertainty model requires the
Column data framework of the structured singular value μ (SSV), which
No. of trays 50 expects the disturbance inputs d (feed composition, flow rate),
Column diameter (m) 0.8 the reference inputs r, and the control error e to be in an H ∞ -
Feed tray 20 norm bounded set. The performance objective is defined as
making the weighted control error to be in the set
Murphree tray efficiency ≈0.4
Relative volatility α 1.61 sup W e ( jω )e ( jω ) ≤1 ∀ω ∈ R +
2
Operating data d
≤1 (1)
r
Top composition xD (mol/mol) 0.99 2

Bottom composition xB (mol/mol) 0.015 The following H ∞ bound is equivalent to this specification:
Feed composition xF (mol/mol) 0.7-0.9
Feed flow rate F (mol/min) 20-46 T ( jω ) ≤1 (2)
d
→p
Top pressure (mbar) 60 r ∞
Nominal operating point
W e ( s ) is a matrix of transfer functions (usually diagonal)
Feed composition (mol/mol) 0.8
which shapes the maximum allowed amplitude of the transfer
Feed flow rate (mol/min) 33 function from [ d, r ] T to e. If W e is large in a certain frequency
Reflux L (mol/min) 65 range, only a small control error is allowed there.
Boilup V (mol/min) 104
The matrix W d ( s ) shapes the frequency content of the distur-
bances and setpoint changes. In the case of our distillation
variations of the cooling water temperature, the feed tempera- column, variations of the feed composition and feed flow rate
ture, or the ambient temperature, are less significant. The affect the medium and lower frequency range. First-order lags
second design objective is a fast enough setpoint tracking which shape the frequency content for these two disturbances quite
exhibits small overshoots. These two objectives — the best well. Because measurement noise enters the control loop at the
possible disturbance compensation and a reasonable setpoint same position as the reference inputs, the corresponding
tracking — must be guaranteed within the entire operating weights are chosen as constants to model measurement noise as
range of the distillation column. well. The block Δ P within the performance specification is an
arbitrary uncertainty block which is used only for controller
3 Feedback Control Design design by μ-synthesis.
Input uncertainty: Ill-conditioned plants are known to be very
3.1 The structured uncertainty model sensitive to input uncertainty [10]. Therefore, a multiplicative
The distillation column is not operated at a single operating input uncertainty according to
point, but it covers a wide range of feed flow rates and composi-
tions. A highly promising method for a control design which ũ ( jω ) = [ I + Δ u ( jω )W u ( jω ) ]u ( jω )
(3)
guarantees stability and performance for the entire operating with Δ u ( jω ) ≤1

range is based on a structured uncertainty model. Such an
uncertainty model, which describes the column dynamics quite is an essential part of the uncertainty model.
well for all operating points, is illustrated in Fig. 2. The basic
Column nonlinearity and output uncertainty: The column
ideas and the mathematical background concerning this struc-
nonlinearity, on the one hand, is modelled by an uncertain linear
tured uncertainty have already been discussed in [7]. Therefore,
combination of the column models for maximum column load
merely the three parts of the uncertainty structure are briefly
(GI) and minimum column load (GR) which models the varia-
described here.
tions of the steady-state operating point

Performance Input uncertainty Nonlinearity and


ηo ηi
specification output uncertainty δy 0
1
p Wy
0 δy
2
GI 1/2
ΔP We
ρο ρi +
ξo ξi
+ +
+
Wu Δu δGI2 +
d e + - +
Wd +
K GR 1/2
+ ~u
r -
T10, T44

Fig. 2: The structured uncertainty model


G I ( jω ) + G R ( jω ) G I ( jω ) – G R ( jω ) 3.3 Control structures
G̃ ( jω ) = ------------------------------------------ + δ G ( jω ) ------------------------------------------
2 2 (4) Before we start to optimize controller parameters, we have to
with δ G ( jω ) ∞ ≤ 1 define potential control structures. The diagonal PI control
structure (Fig. 3) is probably the most frequently used composi-
and, on the other hand, by a multiplicative uncertainty at plant
output which represents the uncertainty due to movements of
the composition profile in the mid- and high-frequency range

r10 L Distillation T10
δ y ( jω ) 0 PI 1
+ Column
ỹ ( jω ) = I + 1
W y ( jω ) y ( jω )
0 δ y ( jω ) (5) r44
with
T44
2 V inventory
PI 2
+ control
with δ y ( jω ) ≤1 –
i ∞

This output uncertainty may include the uncertainties of the


measurements as well. The selection of the input weights Wd, Fig. 3: Diagonal PI control structure
the uncertainty weights Wu, Wy, and the performance weights
We is described in [7] and [8].
tion control structure in industry [11]. The corresponding matrix
3.2 Controller synthesis transfer function is given by

In the process industry PID or advanced PID control structures


1 + TI 1 s
are very common. As mentioned above, the implementation of a KR 1 -------------------- 0
controller design and its acceptance are significantly improved L(s) = TI 1 s e 10 ( s )
(10)
if the design is based on PID control structures. The corre- V(s) 1 + TI 2 s e 44 ( s )
sponding design objective is the μ-optimal tuning of simple 0 KR 2 --------------------
TI 2 s
control elements, such as PID controllers or first-order lags,
within a fixed control structure: The major disadvantage of this control structure is its neglect of
the large interactions between the two control loops. These
K = arg inf μ (M) (6)
K stabilizing Δ̃ ∞ interactions lead to a very sluggish disturbance compensation.
K with fixed structure
They can be partially cancelled by use of decoupling tech-
The solution of this design objective is difficult. Because no niques. A simple controller structure with decoupling is shown
synthesis methods exist, the design objective (6) must be solved in Figure 4. If we assume a real PID control law (i.e., a first-
by a parameter optimization approach. During this optimization
the structured singular value μ (SSV) has to be calculated
repeatedly for a number of frequency points. However, for a r10 – L T10
sharp maximum of the SSV, the calculated maximum may be PID 1
+ Distillation
very sensitive to the number of frequency points calculated. In C2 Column
order to simplify the numerical treatment, the design objective
with
can be approximated by a summation of the cube of the SSV for C1 inventory
all k frequency points: r44 V control T44
PID 2
k +

θ = arg inf ∑ μ 3 { F l [ P,
Δ̃
K(θ)]} (7)
θ Fig. 4: PID control structure with decoupling
i=1
Summing the cube, large values of the SSV carry much more
weight such that the design objective approaches Eq. (6).
The calculation of the SSV presumes nominally stable control order lag in series with an ideal PID controller), this control
loops. Within μ-synthesis, the controllers are calculated by structure is mathematically described by the matrix transfer
solving an H ∞ problem, which always guarantees nominal function
stability. However, during a parameter optimization, nominally
unstable control loops may be generated. Therefore, the design L(s) = 1 C 1 ( s ) PID 1 ( s ) 0 e 10 ( s )
objective (7) must be supplemented with the boundary condi- V(s) C2 ( s ) 1 0 PID 2 ( s ) e 44 ( s )
tion for nominal stability: (11)
2
1 + TI i s + TI i TD i s
Re [ λ max ( F l { G i, K } ) ] < 0 (8) with PID i ( s ) = KR i ------------------------------------------------
TI i s ( 1 + TL i s )
A second boundary condition is the robust stability criterion,
The decoupling elements C i can be static (static decoupling) or
which should be fulfilled for the final parametrized controller:
dynamic (dynamic decoupling). Often decouplers are based on
μ Δ { F l [ P, K ( θ ) ] } <1 (9) an inversion of the plant’s transfer function G ( s ) . The

resulting closed-loop behaviour is very sensitive to input uncer-
This constrained parameter optimization problem is solved by tainty and decoupler errors. Summarizing the research results,
sequential quadratic programming [3]. Skogestad [11] concludes that two-way decouplers should never
be used for high-purity distillation columns with the LV-config- overshoots. The singular value plots for the disturbance
uration. On the other hand, one-way decoupling ( C 1 = 0 or compensation show a maximum in the mid-frequency range
C 2 = 0 ) seems to be less sensitive to input uncertainty and (Fig. 6 b), which has two reasons. First, the small sensitivity to
should be preferred [9]. disturbance in the high-frequency range is caused by the low-
pass characteristics of the plant itself. Second, the controller
4 Design results gains are large in the low-frequency range. Using a real PID
control law, the improvements of the controller properties are
4.1 Diagonal PI control small. Due to measurement noise, the high frequency gains of
the controller are bounded. Therefore the additional degrees of
Table 2 summarizes the results of the parameter optimization. freedom in the design are insignificant.

Table 2: Results for the diagonal PI control structure 4.2 Results for static two-way decoupling
KR1 TI1 KR2 TI2 The simplest decoupling structure is static decoupling. Here the
(mol/min/°C) (min) (mol/min/°C) (min) two decoupling elements C 1 and C 2 are constant factors. The
–14.09 137 2.49 34 results for this structure are obtained with the same weighting
functions and with the same uncertainty model as were used for
the diagonal PI control structure.
The upper bounds for robust stability and performance are
shown in Figure 5. While stability is guaranteed for the speci- Table 3 summarizes the μ-optimal parameters for real PID
fied uncertainties and for the entire frequency range, the
Table 3: μ-optimal parameters for PID control with
performance specification is not met in the lower frequency
range. However, robust performance is achieved within the static decoupling
upper frequency range. Controller or KR TI TD TL C
decoupler No. (mol/min/°C) (min) (min) (min) (–)
Structured singular value

2 1 –13.1 51.6 7.83 8.43 –0.217


RP
2 4.56 62.1 5.11 3.07 1.03
1.5

1
control with static decoupling. The high-frequency gains of the
0.5 RS PID controller are small enough. The results for the decouplers
are somewhat surprising. They indicate that the optimal decou-
0 -5 -3 -1 1
pling is very close to a one-way decoupling! Let us examine this
10 10 10 10
Frequency (rad/min) control structure in detail:
The decoupler parameter C 2 is close to one. Therefore any
Fig. 5: Robust performance and stability for diagonal PI control variation of the output of the top composition controller causes
a simultaneous increase or decrease of reflux and boilup by
almost the same magnitude. Thus this controller shapes the
The singular value plots of the transfer function from the refer- composition profile within the column by an adaptation of the
ence signals to the output signals using a linear model for the separation.
nominal operating point (Fig. 6 a) show high condition numbers
in the medium and upper frequency range and a relatively high The other decoupler parameter C 1 is small. Consequently the
maximum in the upper frequency range. Consequently, we have output of the bottom composition controller has little effect on
to expect both a strong dependence of the setpoint tracking on the reflux. This controller moves the composition profile within
the spatial direction of the setpoint changes and significant the column.

Tr → y The experiences with a control design including an additional


Td → y
10
1 1
10
temperature measurement in the middle of the column have
shown that the maximum controller performance is increased
0 0
only by a small amount. Based on the interpretation above, the
Magnitude

Magnitude

10 10 limited advantage of this additional temperature measurement is


easily explained. Since no setpoint is available for such a
10
-1 -1
10 temperature measurement, an improved feedback may be calcu-
lated neither for the composition profile nor for the composition
-2 -2 profile’s position.
10 -5 -3 -1
10 -5 -3 -1 1
10 10 10 10 10 10 10 This special behaviour of the control system has significant
Frequency (rad/min) Frequency (rad/min) advantages for the closed-loop behaviour. The μ-plots in
a) b) Figure 7 demonstrate the superior robust performance of the
Fig. 6: Singular values for the nominal closed-loop with decoupling control structures. While the optimal tuning for PI
diagonal PI-control control shows a peak of the robust performance plot within the
a) Transfer function from reference to output signals lower frequency range, the PID control structure shows nearly
b) Transfer functions from disturbance to output signals flat and significantly smaller structured singular values. Using a
solid line: feed composition disturbance; decoupling control structure, the additional degrees of freedom
dash-dotted line: feed flow disturbance in the controller design allow a significantly improved
Table 4: μ-optimal parameters for PID control with dynamic

Structured singular value


1 decoupling
RP Controller KR KC TI TD TL
(mol/min/ (–) (min) (min) (min)
0.5 RS °C)
PID 1 –22.2 – 80.2 19.6 44.8
PID 2 5.68 – 59.4 12.6 24.7
1
0 -5 -3 -1 1
10 10 10 10 10 C1 – –0.138 – 117 7.42
Frequency (rad/min)
C2 – 1.07 – 53.0 71.43
Fig. 7: Robust performance and stability for easy to implement in a distributed control system and is simple
real PID control with static decoupling to initialize. In order to keep the decoupler as simple as
possible, the discussion is limited to static one-way decoupling.
controller performance, especially in the important mid- Two different decoupler structures are possible if we set either
frequency range. C 1 or C 2 of the control structure shown in Figure 4 to zero.
The singular value plots of the transfer function T r → y ( s ) While the results for the two-way decoupling lead us to expect a
(Fig. 8 a) illustrate the improved controller performance as good performance for the first case ( C 1 = 0 ), no inference is
well. The condition numbers and the maximum are much possible for the second case ( C 2 = 0 ). In fact, the optimization
smaller than those of the diagonal PI control structure. Conse- results show insufficient performance for the second case
quently, the tracking behaviour is significantly improved. ( C 2 = 0 ). Therefore an overturn of the decoupling control
Concerning the disturbance compensation, a similar maximum structure, with shaping of the composition profile by the bottom
of the singular values is achieved (Fig. 8 b). However, the peak composition controller and moving the composition profile
is more narrow and both singular values are significantly position by the top composition controller, does not lead to
smaller in the low-frequency range. results comparable to those obtained with the other decoupling
structure.
Tr → y Td → y
1 1 The tuning parameters for the controller with C 1 = 0 can be
10 10
found in Table 5. The corresponding μ-curves (Fig. 9) let us
expect a performance similar to that of the PID controller with
Magnitude

Magnitude

0 0
10 10 static two-way decoupling.

-1 -1
Table 5: μ-optimal parameters for PID control with static
10 10
one-way decoupling

10 -5
-2
10 -5
-2 Controller or KR TI TD TL C
-3 -1 -3 -1 1
10 10 10 10 10 10 10 decoupler No. (mol/min/°C) (min) (min) (min) (–)
Frequency (rad/min) Frequency (rad/min)
1 –10.5 45.7 2.18 5.01 0
a) b)
2 5.35 67.4 13.4 13.9 1.05
Fig. 8: Singular values for the nominal closed-loop system for
PID control with static decoupling
1
a) Transfer function from reference to output signals
Structured singular value

RP
b) Transfer functions from disturbance to output signals
solid line: feed composition disturbance;
dash-dotted line: feed flow disturbance 0.5 RS

4.3 PID control with dynamic decoupling


Using lead-lag transfer functions for the decoupler elements C 1 0 -5 -3 -1 1
and C 2 10 10 10 10
Frequency (rad/min)
1 + TD i s
C i ( s ) = KC i --------------------- (12) Fig. 9: Robust performance and stability for real
1 + TL i s PID-control with one-way decoupling
a dynamic decoupling structure is realized. The additional
degrees of freedom allow a further improvement of the control
design. The resulting optimal tuning constants are listed in 4.5 Simulation results
Table 4. The frequency domain results presented above are supported by
nonlinear simulation of the closed-loop response to step
4.4 PID control with one-way decoupling
changes of the feed composition and feed flow rate (Fig. 10).
The results for two-way decoupling have shown optimal results The steady-state offsets of the product composition are caused
for decoupling structures which are close to one-way decou- by controlling the pressure compensated temperatures on trays
pling. In this section the optimal tuning results for one-way 10 and 44 rather than the two product compositions. A compar-
decoupling are discussed. This control structure is particularly ison of the performances confirms the frequency domain
results. In contrast to the very sluggish disturbance compensa-
PI PID+static deco. tion of the diagonal PI control structure, the compensation is
0.02 0.02
fast enough using decoupling techniques. The performance of
the PID controller with dynamic decoupling is very close to the
Composition (mol/mol)

Composition (mol/mol)
best results obtained with μ-optimal state-space control. Fixing
0.015 0.015
the decoupler structure to one-way decoupling, the maximum
errors of the product qualities are larger than those of the two-
way decoupling structures. However, the performance of this
0.01 0.01 simple feedback controller is still considerably superior to that
obtained with diagonal PI control.

0.005
0 20 40
0.005
0 20 40
5 Feedforward control design
Time (h) Time (h) It is a drawback of feedback control that any corrective action
PI PID+static deco. necessitates a deviation of the controlled variables from their
0.02 0.02 setpoints. This disadvantage can be overcome by the use of
feedforward control. The most important disturbance of this
distillation column is a change in the feed flow rate. Because
Composition (mol/mol)

Composition (mol/mol)

0.015 0.015 the feed flow rate is measured, it can be used as a controller
input. An appropriately designed feedforward controller takes
most of the necessary corrective action before the product
compositions and the controlled tray temperatures change.
0.01 0.01
However, because of model errors and other unmeasured distur-
bances, a feedforward controller alone will never be able to
yield perfect control. Feedback control will remain necessary as
0.005 0.005 well.
0 20 40 0 20 40
Time (h) Time (h) A simple and easy-to-implement feedforward control structure
is a first-order lag with different gains for the outputs to the
PID+dynamic deco. PID+one-way deco. reflux L and the boilup V according to
0.02 0.02

KR L 1
KF ( s ) =
Composition (mol/mol)

Composition (mol/mol)

--------------- (13)
KR V 1 + Ts
0.015 0.015
The parameters of this simple control structure are computed by
a constrained parameter optimization [3]. Most of the feed flow
0.01 0.01 disturbances entering this distillation column are step changes.
Consequently, we are able to define an appropriate design
objective in the time domain. It is the minimum absolute control
0.005 0.005 error for a step change in the feed flow rate. However, the distil-
0 20 40 0 20 40
lation column is not operated at a single operating point. The
Time (h) Time (h)
feedforward control should improve the feed flow disturbance
PID+dynamic deco. PID+one-way deco. compensation at all operating points within the defined oper-
0.02 0.02 ating range. Similarly to the feedback control design, this
design task is approximated by a simultaneous design for the
Composition (mol/mol)

Composition (mol/mol)

column models GI(s) and GR(s). The design objective becomes


0.015 0.015
[ T, KR L, KR V ] = arg inf E (14)
[ T, KR L, KR V ]

with
0.01 0.01
Te
E = ∫ { e10 ( t )
R
+ e 44 ( t ) + e 10 ( t ) + e 44 ( t ) }dt
R I I
.(15)
0.005 0.005
0 20 40 0 20 40 0
Time (h) Time (h) The performance measure E is calculated for a step response to
the plant input F, using the plant illustrated by Figure 11.
Fig. 10: Simulation results for the μ-optimal tuned control struc- If we select the μ-optimal PID-controller with one-way decou-
tures for an increase in feed composition (0.8 → 0.9 mol/mol) pling as the feedback controller K and limit the time constant T
at t=0 h and an increase of feed flow rate (+3.6 mol/min) at by a lower bound of 5 minutes, the following simple optimal
t=20 h feedforward controller results:
Solid lines: 1-xD Dashed lines: xB
1
Upper plots: Ft=0 =20 mol/min Lower plots: Ft=0=46 mol/min K F ( s ) = 1.5 ------------------- (16)
2.6 1 + 5.0s
illustrate the improved compensation of feed flow disturbances
F as well. Of course, the compensation of the feed composition
GR ( s ) step change is still the same. However, feed composition distur-
u
e 10 bances are less significant.
– R
+
K (s) e 44
F uF + R
6 Implementation results
KF ( s )
F The PID controller with one-way decoupling and the simple
GI ( s ) feedforward controller have been implemented in the distrib-
u
+ – e 10 uted control system of the real plant. In Figure 14 the excellent
I
performance of the control scheme is demonstrated. The
K (s) e 44
+ I controller error in presence of these feed flow disturbances
remains extraordinarily small. In fact, the controller error is
Fig. 11: Plant structure for the optimization of hardly distinguishable from the measurement noise and from
feedforward controller parameters the effect of all other unknown disturbances. This proves the
high performance of this simple control scheme. Its advantages
The singular values of the feedforward controller are shown in over the former situation (i.e., manual operation of the column)
Fig. 12 a. In Fig. 12 b we find the singular values of the transfer are
functions T d → y for a nominal closed-loop system with this • more uniform product quality
feedforward controller. It demonstrates the low sensitivity of the • better average product quality in column bottom
feedback and feedforward controlled distillation column to vari- • no need to install an additional column
ations of the feed flow rate. The simulation results in Fig. 13 • cost savings potential in the area of $250000 annually
KF Td → y The installation took significantly more time than expected,
1 1
10 10 however. While the controller demonstrated a very robust
behaviour without any particular difficulties, the compensation
of the effect of the large pressure variation (110-180 mbar in the
Magnitude

Magnitude

0 0
10 10

Feed
-1 -1
Deviation of feed

10 10
flow rate (l/h)

40

-2 -2
10 -5 10 -5 20
-3 -1 1 -3 -1 1
10 10 10 10 10 10 10 10
Frequency (rad/min) Frequency (rad/min) 0
a) b)
0 20 40 60 80 100
Time (h)
Fig. 12: a) Singular values of the feedforward controller with
fixed structure

sated temperature (°C)


Tray 10

Deviation of compen-
Deviation of estimated
composition (mol-%)

5
b) Singular values of the transfer functions for the nom- -0.5
inal closed-loop system from the disturbance inputs d
to the controlled output signals y (Feedback and feed- 0
forward control)
solid line: feed composition disturbance;
0.5
dash-dotted line: feed flow disturbance -5
0 20 40 60 80 100
Time (h)
F(t=0)=20 mol/min F(t=0)=46 mol/min

sated temperature (°C)


Tray 44
Deviation of estimated

Deviation of compen-
composition (mol-%)

0.02 0.02 6
-0.5
Composition (mol/mol)

Composition (mol/mol)

0.015 0.015 0

0.5
-6
0 20 40 60 80 100
0.01 0.01 Time (h)

Fig. 14: Recorded operating data with installed PID control


including one-way decoupling and feedforward control.
0.005 0.005
0 20 40 0 20 40 Top: Deviation of feed flow rate from 260 l/h
Time (h) Time (h) (49 mol/min)
Middle: Deviations of estimated tray composition and of
Fig. 13: Simulation results for the PID with one-way decoupling pressure-compensated temperature from setpoint
and simple feedforward control for an increase in feed com- on tray 10
position (0.8 → 0.9 mol/mol) at t=0 h and an increase of Bottom: Deviations of estimated tray composition and of
feed flow rate (+3.6 mol/min) at t=20 h. pressure-compensated temperature from setpoint
Solid lines: 1-xD. Dashed lines: xB on tray 44
bottom of the column) on the tray temperature was more trou- The design results for an industrial distillation column have
blesome. As a first step the pressure compensation was demonstrated that the potential performance of “simple”
designed by regression of Tpx data of the substance mixture. advanced PID control structures is comparable to that of state-
However, as a second step, extensive steady-state simulations space controllers. Moreover the tuning of PID control structures
became necessary in order to adjust the parameters of the including a two-way decoupling have shown that the optimal
compensation for a good correlation with the quality of the control structure is very close to one-way decoupling. However,
products analysed once a day. The resulting overall perform- these results are obtained by unusual settings for the individual
ance is shown by Figure 15. Since a perfect pressure compensa-
control elements. The performance of these feedback control
tion can not be achieved (lacking pressure measurements on
structures is significantly improved by a simple feedforward
trays 10 and 44), a minor adjustment of the controller setpoints
controller, tuned by an optimization in the time domain. The
is necessary depending on the feed flow rate. Since the results
easy-to-initialize PID control with one-way decoupling and the
presented are achieved with almost constant setpoints, these
results will improve even further as the operators gain more simple feedforward control have been implemented in the real
extensive experience choosing the setpoints. It must be empha- plant with good success. The high pay-off fully justifies the
sized here that — despite the relatively high effort — the use of extensive computational effort.
pressure compensated temperatures as controlled outputs is
indispensable for this column. An attempt to install an on-line References
gas chromatograph at a similar column which separates the
same components failed due to plugging of the device.
[1] Brisk, M.L.: “Process Control: Theories and Profits,”
Preprints of the 12th World Congress of the International
7 Conclusions Federation of Automatic Control, Sydney, July 18-23, 7,
The μ-optimal tuning of PID control structures has shown to be 241-250 (1993)
an effective method for the optimal tuning of easy-to-imple- [2] Doyle, J.C.: “Structured Uncertainty in Control System
ment control structures which guarantee stability and a high Design,” Proc. of the 24th Conf. on Dec. and Control, Fort
performance for the entire operating range of a distillation Lauderdale, FL, 260-265 (1985)
column. The application of design methods for unstructured
uncertainty to this control problem [8] has shown that it is [3] Grace, A.: Optimization Toolbox — User’s Guide, The
extraordinarily difficult to obtain performance and robustness MathWorks, Inc., Natick, MA (1990)
properties comparable to those of the μ-optimal controllers. [4] Lin, J.-L., I. Postlethwaite, and D.-W. Gu: “μ-K Iteration:
A New Algorithm for μ-Synthesis,” Automatica, 29, 219-
Manual operation Controlled 224 (1993)
0.3 0.3

[5] Marlin, T.E., J.D. Perkins, G.W. Barton, and M. L. Brisk:


Composition (mol/mol)

Composition (mol/mol)

0.25 0.25
“Benefits from process control: results of a joint industry-
0.2 0.2
university study,” J. Proc. Cont., 1, 68-83 (1991)
0.15 0.15
[6] Musch, H.E. and M. Steiner: “Modeling distillation
0.1 0.1 column nonlinearity for μ-synthesis,” Proc. of the 1993
American Control Conf., June 2-4, San Francisco, CA,
0.05 0.05
1177-1178 (1993)
0 0
0 35 70 0 22 44 [7] Musch, H.E. and M. Steiner: “μ-optimal Control of an
Days Days Industrial Binary Distillation Column,” Preprints of the
12th World Congress of the International Federation of
Manual operation Controlled Automatic Control, Sydney, July 18-23, 1, 49-54 (1993)
0.25 0.25
[8] Musch, H.E.: “Robust Control of an Industrial High-Purity
Composition (mol/mol)
Composition (mol/mol)

0.2 0.2 Distillation Column,” Ph. D. thesis, ETH-Zurich (1994)

0.15 0.15 [9] Skogestad, S., and M. Morari: “Implications of Large RGA
Elements on Control Performance,” Ind. Eng. Chem. Res.,
0.1 0.1 26, 2323-2330 (1987)

0.05 0.05 [10] Skogestad, S., and M. Morari: “Robust Control of Ill-
Conditioned Plants: High-Purity Distillation,” IEEE Trans.
0
0 35 70
0
0 22 44
Automatic Control, 33, 1092-1105 (1988)
Days Days
[11] Skogestad, S.: “Dynamics and Control of Distillation
Fig. 15: Analysis results of top and bottom products Columns – A Critical Survey,” Preprints of the 3rd IFAC
Top: Top composition 1-xD Symposium on Dynamics and Control of Chemical Reac-
Bottom: Bottom composition xB tors, Distillation Column and Batch Processes, April 26-
Dashed line: Average composition 29, 1992, College Park, MD, 1-25 (1992)

Anda mungkin juga menyukai