Professional Documents
Culture Documents
Abstract
This section is an introductory reference to process control. It discusses feedback
control algorithms and controller tuning in depth. The unique requirements of level
controller tuning are covered separately in Section 331. The importance of understanding the various forms of the proportional-integral-derivative (PID) control
algorithm and the impact on various tuning rules is analyzed.
The benefits and application of common multiple-loop control configurations such
as cascade, ratio, and feedforward are described. The control objectives analysis
(COA) process is described. COA is a proven methodology for gathering the necessary information to ensure that a process control system will meet plant objectives
for optimal performance, and provides a sound basis for control loop design.
An introduction to advanced control and optimization is given. Finally, resources
and references are provided to allow the reader to pursue more advanced topics
about process control.
Contents
Page
310
300-3
311
Technology Hierarchy
312
Operational Benefits
313
Economic Benefits
320
Basic Control
321
Control Loops
322
Feedback Controllers
323
324
On/Off Control
325
326
327
328
ChevronTexaco Corporation
300-9
300-1
December 2003
December 2003
330
Controller Tuning
300-27
331
332
333
334
340
Multiple-Loop Control
341
Cascade Control
342
Ratio Control
343
Feedforward Control
350
351
Summary
352
COA Products
353
COA Participants
360
Advanced Control
361
Overview
362
363
364
370
Online Optimization
371
Introduction
372
373
374
380
Resources
381
382
390
References
300-54
300-67
300-70
300-90
300-93
300-96
300-2
ChevronTexaco Corporation
PROCESS
300-1
The Basic Regulatory Controls (BRC) consists of the simple control loops
provided to ensure safe, efficient regulation of the process. Examples include
simple single-loop control of flows, pressures, levels, and temperatures, as well
as simple cascades and ratios.
The Intermediate Regulatory Controls (IRC) are somewhat more complicated than BRC loops and include such control strategies as steam drum level
control, boiler combustion control, fuel gas BTU control, feedforward control,
separation factor control for distillation columns, and furnace pass balancing.
The basic and intermediate loops are typically implemented in a Distributed Control
System (DCS) such as provided by Honeywell or Yokogawa. These loops nominally operate once per second. At this level in the technology hierarchy, PID
(proportional, integral, derivative) controllers are typically used.
ChevronTexaco Corporation
300-3
December 2003
increased throughput,
increased yield,
maximum production of a more valuable product, and
lower energy costs.
This section illustrates how improved control allows the process to run closer to
constraints or setpoints. Figure 300-2 shows typical performance data from a control
loop. The controller attempts to keep the controlled variable at the target. However
due to disturbances and other factors, the controlled variable deviates from the
December 2003
300-4
ChevronTexaco Corporation
target. The target has to be positioned away from the constraint or specification to
achieve an acceptable level of performance.
Controlled Variable
Controlled Variable
Fig. 300-2
Target
Constraintor
Specification
300-2
Normalized Frequency
of Occurance
Time, days
1.5
= 1/4
Constraint/
Specification
1.0
= 1/2
=1
0.5
0.0
3
+1
+2
+3
Reducing the standard deviation brings improved stability to the process, which can
be beneficial in reducing or eliminating upsets (Figure 300-4).
ChevronTexaco Corporation
300-5
December 2003
Fig. 300-4
Shifting Target
= 1/4
Target
(mean)
Constraint/Specification
1.5
1.0
0.5
= 1/2
=1
300-4
0.0
3
+1
+2
+3
Figure 300-5 quantifies several aspects of the previous curves, which are assumed to
be normal distribution curves. As such, there will always be a small percentage of
off-spec data, no matter how far the target is from the constraint/specification.
Fig. 300-5
0.1%
% of Data Exceeding
Constraint / Specification
+2.0
2.5%
5.0%
10.0%
+1.0
0.0
0.0
0.5
1.0
For example, to limit the off-spec data to 2.5%, the setpoint (or target) must be
two standard deviations from the constraint/specification, assuming a one sigma
December 2003
300-6
ChevronTexaco Corporation
variation in the data. But, if we were able to reduce the standard deviation in half
due to improved control, we could move the setpoint one standard deviation closer
to the constraint/specification.
Online
Optimization
Total
Atmospheric Distillation
10
15
Vacuum Distillation
10
14
Coking
20
27
Catalytic Cracking
18
10
28
Hydrocracking
18
10
28
Reforming
15
22
Alkylation
15
22
Isomerization
11
15
22
Gasoline Blending
10
18
Process Unit
The numbers reflect typical incentives for advanced control and optimization above
a base level of performance achieved by regulatory (DCS) controls. For example, an
atmospheric distillation unit with a throughput of 100,000 Bbl/Day would have a
ChevronTexaco Corporation
300-7
December 2003
mid-range incentive of $3,650,000/year for advanced control. Since these are midrange estimates, actual incentives at specific sites could differ substantially.
There is some evidence the Solomon averages are strongly affected by plants that
gain feed max benefits. Typically, only one or two units in a refinery are a bottleneck to production or are required by economics to run at maximum feed rate.
Note Feed maximization benefits are substantially larger than yield and energy
saving benefits.
The relatively high cost for the basic regulatory controls (BRC) reflects the cost
of the infrastructure that is required (e.g., distributed control system, instrumentation and control valves).
Once the infrastructure is there, more advanced applications can be added for a
relatively low cost (relative to the benefits that can be achieved).
Advanced control and online optimization applications offer the possibility of
very large benefits for a relatively small incremental cost.
Fig. 300-6
Relative Cost
300-6
Online
Optimization
Advanced
Control
IRC
BRC
0
Relative Benefits
100
Typically, the biggest bang for the buck comes from advanced control (e.g.,
AspenTechs DMCplus or Honeywells RMPCT).
Depending on the scope of the application and the type of process, costs can range
from $100,000 to $1,000,000, with payout times of from one month to a year.
December 2003
300-8
ChevronTexaco Corporation
A manual control loop requires a human being to observe the value of the
controlled variable. If this variable is not at the setpoint, the human observer
adjusts a manipulated variable.
An automatic control loop employs a controller to keep the controlled variable at the setpoint.
Feedback Control Loops. Figure 300-7 shows a typical feedback control loop. In
the process furnace, a temperature controller monitors the outlet temperature
(controlled variable) of the furnace. If the outlet temperature is not at the desired
value (setpoint), the controller changes the fuel flow (manipulated variable) by
changing the position of the fuel gas control valve (final control element). A typical
disturbance would be the furnace feed rate. This type of control is called a closed
loop feedback control system. Perfect feedback control is impossible in all cases
since the controlled variable must deviate from the setpoint before any control
action takes place.
Feedforward Control Loops. In contrast, feedforward control uses a measured
disturbance to generate a corrective action which minimizes the deviations of the
controlled variable from its setpoint (outside of any feedback action). Perfect feedforward control is (theoretically) possible in some cases. But, practically speaking,
there will always be errors.
ChevronTexaco Corporation
300-9
December 2003
Fig. 300-7
Furnace
Feed
Stream
Furnace Outlet
Temperature
Temperature
Transmitter
Burners
TC
301
Control Valve
Temperature
Setpoint
Temperature
Comtroller
Fuel Gas
Supply
Feed
Disturbance
Variable
FI
TC
Manipulated
Variable
Controlled
Variable
FFC
Feedforward
Feedback
Fuel Gas
302
Note also that because of control valve non-linearity, feedforward control normally
would be used in conjunction with a furnace outlet temperature to fuel gas flow
cascade feedback control configuration.
December 2003
300-10
ChevronTexaco Corporation
Setpoint, %
Error, %
6
-
Control
Algorithm
Controller
Output, %
Measurement, %
303
There are two key elements: the comparator and the control algorithm. The setpoint
(the desired value of the controlled variable) is compared with the actual measured
variable to form an error. As shown in the block diagram, error is usually defined
as follows:
Error(t) = Setpoint(t) - Measurement(t)
(Eq. 300-1)
Note There is inconsistency in the industry on the above definition; error is just as
often defined as measurement minus setpoint.
For single-loop control, the controller output signal is sent to the control valve
(final control element).
For cascade (multiple-loop) control, the controller output becomes the setpoint
of the secondary controller.
On/Off
ChevronTexaco Corporation
300-11
December 2003
These algorithms will now be discussed (along with some less-commonly used variations).
on (100%)
off (0%).
It only responds to the sign of the error, that is, whether it is above or below the
setpoint.
On/Off control is not generally suitable for continuous automatic feedback control
because it results in constant cycling of the controlled variable.
On/Off with differential gap control. This is a refinement of on/off control.
Instead of changing output from on (100%) to off (0%) at a single setpoint, differential gap action changes output at high and low limits called boundaries. As long as
the measurement remains between the boundaries, the controller holds the last
output. A typical application of differential gap control is the operation of a dump
valve or pump to keep a vessel level within an acceptable range.
Integral: Controller output changes by an amount related to the size and duration of the error.
Derivative: Controller output changes by an amount related to the rate-ofchange of the error.
December 2003
300-12
(Eq. 300-2)
ChevronTexaco Corporation
where:
CO(t) = Controller output [=] %
KC = Controller Gain [=] %/% (dimensionless)
E(t) = Error [=] %
t = Time [=] minutes
The controller gain, Kc, is also called the controller sensitivity. It represents the
proportionality constant between the control valve position and controller error.
Figure 300-10 shows the relationship between the controller output (valve position)
and error that is characteristic of proportional control.
Fig. 300-10 Proportional Mode Output is Proportional to Error (Open loop)
Controller
Output
Error
0
304
Time, Minutes
The valve position changes in exact proportion to the amount of error, not to its rate
or duration. The response is almost instantaneous, and the valve returns to its initial
value when the error returns to its original value.
Figure 300-11 shows how controller gain affects valve opening for constant change
in error.
Fig. 300-11 Proportional Mode Plots Step Response (Open loop)
KC=1.5
Controller
Output
KC=1
KC=0.5
Error
305
Time, Minutes
ChevronTexaco Corporation
300-13
December 2003
(Eq. 300-3)
where:
( 100 )
CO ( t ) = K C E ( t ) + B = ------------- E ( t ) + B
PB
(Eq. 300-4)
Controller Output
(Control Valve)
"Throttling Range"
100%
PB=50
PB=100
PB=200
50%
KC=0.5
KC=1
0%
-50%
KC=2
0%
+50%
Error
306
Offset. A controllers error is the difference between its setpoint and measurement.
In a proportional only controller, a change in setpoint or load introduces a permanent error called offset.
December 2003
300-14
ChevronTexaco Corporation
Furnace
Outlet
Temperature
Setpoint
Offset
Furnace
Feed Rate
307
Time, Minutes
If the feed rate to the furnace increases, more fuel will be needed. This disturbance
represents a load change to the furnace. To get more fuel, the fuel valve must be
opened more. As is suggested by the equation for proportional action, the only way
that the valve can be at some value other than its starting point is for an error to
exist. Thus, the proportional controller alone cannot return the outlet temperature to
its setpoint. As mentioned, some controllers allow the operator to adjust the bias
until the value of the error (or offset) is zero.
The proportional only controller is the easiest continuous controller to tune. It
provides rapid response and is relatively stable. If tight control is not required,
proportional only control can be used.
CO ( t ) = K I E ( t ) dt' + CO 0
(Eq. 300-5)
where:
CO(t) = Controller output [=] %
KI = Integral mode gain [=] 1/minutes
E(t) = Error [=] %
t = Time [=] minutes
CO0 = Initial controller output [=] %
ChevronTexaco Corporation
300-15
December 2003
With integral action, controller output is proportional to both the size and duration
of the error. As long as a deviation from setpoint exists, the controller continues to
drive its output in the direction that reduces the deviation.
The rate of change of controller output is proportional to the magnitude of the error.
dCO
(t) = K E(t)
-----------------I
dt
(Eq. 300-6)
Integral
Mode
Output
Error
0
308
Time, Minutes
Integral action responds to the sign, size, and duration of the error:
TIME 0 A constant error appears. The integral action drives the output
higher at a constant rate proportional to the size of the error
TIME A The size of the error doubles. The integral action drives the output
higher twice as fast.
TIME B The sign of the error changes. The integral action drives the output
in the other direction.
TIME C The error goes to zero. The integral action stops, holding the
existing output.
TIME D The error ramps down at a constant rate. The integral action drives
the output down at an ever increasing rate.
TIME E The error returns to zero. The integral action stops, holding the
existing output.
December 2003
300-16
ChevronTexaco Corporation
1
CO ( t ) = K C E ( t ) + ---- E ( t' )dt' + CO 0
I
(Eq. 300-7)
where:
CO(t) = Controller output [=] %
KC = Controller gain [=] %/% (dimensionless)
E(t) = Error [=] %
I = Integral (reset) time [=] minutes
t = Time [=] minutes
CO0 = Initial controller output [=] %
Note that the effective gain for the integral mode in the above (standard) equation
for a PI controller is KC / I. The overall controller gain KC affects both the proportional and integral action.
On some controllers, integral settings are in repeats, meaning repeats per minute; on
others, settings are in minutes, meaning minutes per repeat. One setting is the reciprocal of the other. Decreasing the integral time increases the amount of integral
action and visa versa. Integral time is also called reset time.
Figure 300-15 shows how the PI algorithm responds to a step change on error (open
loop/no feedback from the process):
Fig. 300-15 PI Step Response (Open Loop)
Controller
Output
KCA
CO0
KCA
WI
Integral (Reset) Time, Minutes
Error
0
309
ChevronTexaco Corporation
A
0
300-17
Time, Minutes
December 2003
Integral time is quantified as the time required for the controller output to change by
an amount equal to the change caused by the initial proportional kick. In other
words, it is the time required for the contribution of the integral mode to repeat
the contribution of the proportional mode.
Large Overshoot
Furnace Outlet
Temperature
(DegF)
Setpoint
Offset
Reset Windup
Temperature
Controller
Output
(%)
Controller
Un-winds
Valve Starts
Moving
100%
Control Valve
Wide Open
Feed Rate
Disturbance
(MBD)
Time, minutes
310
The temperature controller responds to the disturbance in feed rate by opening the
control valve. But if the control valve capacity is not large enough, it may saturate
before the furnace outlet temperature (controlled variable) has returned to the
setpoint. A persistent error (offset) will then be present. The integral mode keeps
increasing its output to try to eliminate the offset, but there will be no effect on the
process. This effect is called reset (integral) windup.
If at some later time the feed rate (disturbance) returns to its original value, the
furnace outlet temperature (controlled variable) will drift up to the setpoint due to
the decreased load on the system. The integral action cannot start unwinding until
the error changes sign (when the temperature crosses the setpoint). Then, the
temperature controller output starts un-winding. Since there is no valve movement
until the controller output drops below 100%, furnace outlet temperature overshoots the setpoint significantly.
December 2003
300-18
ChevronTexaco Corporation
All industrial implementations of the PID algorithm have provisions for preventing
reset windup. For standard control loop configurations such as single loop control or
cascade control, anti-windup is generally built in. More complicated, non-standard
control structures may require some custom user configuration.
Lets look at the performance of the same control system with anti-windup included
(Figure 300-17).
Fig. 300-17 Integral Anti-Windup - Furnace TC
Less Overshoot
Furnace Outlet
Temperature
(DegF)
Setpoint
Offset
Temperature
Controller
Output
(%)
No Reset Windup
100%
Controller
Starts Closing
Valve Immediately
Control Valve
Wide Open
Feed Rate
Disturbance
(MBD)
Time, minutes
311
There is no difference in the first part of the plot. But with no reset wind-up, the
temperature controller can start closing the control valve immediately when the
disturbance returns to its initial value. As a result, there is substantially less overshoot in the furnace outlet temperature.
(Eq. 300-8)
where:
CO(t) = Controller output [=] %
KD = Derivative mode gain [=] minutes
ChevronTexaco Corporation
300-19
December 2003
Derivative
Mode
0
Output
312
Error
Time, Minutes
(Eq. 300-9)
where:
CO(t) = Controller output [=] %
KC = Controller gain [=] %/% (dimensionless)
E(t) = Error [=] %
t = Time [=] minutes
D = Derivative time [=] minutes
CO0 = Initial controller output [=] %
Note that the effective gain for the derivative mode in the above (standard) equation for a PI controller is KC D. The overall controller gain KC affects both modes.
December 2003
300-20
ChevronTexaco Corporation
Figure 300-19 shows how the PD algorithm responds to a ramp change on error
(open loop/no feedback from the process).
Fig. 300-19 PD Ramp Response (Open Loop)
Derivative
Mode
0
Output
312
Error
Time, Minutes
In this case, the derivative time is the time for the proportional contribution to
repeat the initial derivative kick. Notice that derivative action introduces a lead
(or anticipatory) element into the controller.
Derivative Filters. Note that derivative action would produce a spike if the error
were to undergo a step change. However, in all real implementations of the derivative function, the derivative is filtered. The filter time constant is D, with alpha
typically ranging from 1/6 to 1/10. Use of a derivative filter limits the size of the
derivative spike on sudden changes (Figure 300-20).
Fig. 300-20 Derivative Filter
Input
Step
M
Gain
Filter
Devivative
Practical
KC M
f
M KC / D WD
Theoretical
314
ChevronTexaco Corporation
300-21
December 2003
(Eq. 300-10)
When the setpoint is not changing, its derivative is zero, and we can use the
following expression for derivative.
dM ( t )
K C D --------------dt
(Eq. 300-11)
1
dE ( t )
CO ( t ) = K C E ( t ) + ---- E ( t' ) dt' + D ------------- + CO 0
I
dt
(Eq. 300-12)
where:
CO(t) = Controller output [=] %
KC = Controller gain [=] %/% (dimensionless)
E(t) = Error [=] %
t = Time [=] minutes
I = Integral (reset) time [=] minutes
December 2003
300-22
ChevronTexaco Corporation
f
D
Controller
Output
Theoretical Derivative
Filtered Derivative
I
KCA
CO0
KCA
WI
Integral (Reset) Time, Min.
Error
A
315
0
0
Time, Minutes
Note how the individual control modes (P, I, and D) combine to form the complete
controller output. The real controller response includes the derivative filter
discussed earlier.
Figure 300-22 shows the open loop response of the PID controller to a ramp change
in error
Fig. 300-22 PID Ramp Response (Open Loop)
KCB
t2
2WI
Controller
Output
KCBt
P
D
CO0
KCWDB
WD
Derivative Time, Min.
316
Error
B
1
0
0
ChevronTexaco Corporation
300-23
Time, Minutes
December 2003
1
dM ( t )
CO ( t ) = K C E ( t ) + ---- Et' dt' D --------------- + CO 0
I
dt
(Eq. 300-13)
1
dE ( t )
CO ( t ) = K C E ( t ) + ---- E ( t' ) dt' + D ------------- + CO 0
I
dt
(Eq. 300-14)
However, with microprocessor-based implementations of the algorithm in distributed control systems (DCS), programmable logic controllers (PLC), and supervisory control and data acquisition systems (SCADA), discrete approximations are
used. For example, here is the discrete (incremental) equivalent of the above equation.
t s
D
CO n = K C E n + ------- E n + ------- ( E n )
I
t s
(Eq. 300-15)
Or
t s
D
CO n CO n 1 = K C ( E n E n 1 ) + ------- E n + ------- ( E n 2E n 1 E n 2 )
I
t s
(Eq. 300-16)
December 2003
300-24
(Eq. 300-17)
ChevronTexaco Corporation
where:
CVS, PVPS, SPPS [=]%
K [=]%/% (Controller Gain)
T1 [=] minutes (Reset Time)
T2 [=] minutes (Derivative Time)
a = 0.1 (Derivative Limit Factor)
Honeywell also has interactive versions of the PID equation.
Below is how Yokogawa documents their Centum CS3000 PID Equation (Noninteractive):
100
T
TD
MV n = K S --------- E n + ------- E n + -------- ( E n )
PB
TI
T
(Eq. 300-18)
where:
MVn, En [=] Eng Units
KS = Scale Conversion Factor
PB [=] % (Proportional Band)
TI [=] seconds (Reset Time)
TD [=] seconds (Derivative Time)
T [=] seconds (Control Period)
(Effective Derivative Limit Factor = 0.125)
Yokogawa does not have an interactive PID alternative.
ChevronTexaco Corporation
300-25
December 2003
5 4 3
Offset
2
1
317
Time, Minutes
Controlled
Variable
No Control
2
3
5
Offset
4
318
Setpoint
Time, Minutes
December 2003
300-26
ChevronTexaco Corporation
Controller
CVSP
KM
(EU)
Control Valve
KC GC
CO
KV GV
(%)
MV
K P GP
CV
(EU)
(EU)
CVSP%
CVM
Controlled Variable
Transmitter
K M GM
318a
(%)
where:
CVSP Controlled variable (CV) setpoint [=] EUCV
EU Engineering units
KM Controlled variable transmitter gain [=] %/EUCV
CVSP% Controlled variable %-setpoint [=] %
KC Controller gain [=] dimensionless (%/%)
GC Controller dynamics (integral, derivative)
KV Control valve gain [=] EUMV/%
GV Control valve dynamics
MV Manipulated variable, [=] EUMV
KP Process gain [=] EUCV/EUMV
GP Process dynamics
D Disturbance [=] EUD
KD Disturbance gain [=] EUCV/EUD
GD Disturbance dynamics
ChevronTexaco Corporation
300-27
December 2003
A high performance control loop would have rapid, smooth responses to setpoint
changes and disturbances with minimal control valve movement. A robust control
loop would have good performance for a wide range of process conditions.
However, it is not possible to achieve all of these goals simultaneously. There are
inherent conflicts and tradeoffs that must be considered:
December 2003
First, the controller is changed to proportional-only by turning off the integral and derivative modes.
300-28
ChevronTexaco Corporation
2.
Then the controller gain is increased in small steps, each time changing the
setpoint if required to induce cycling (Figure 300-26).
Time, Min.
Controlled
Variable
319
Time, Min.
Controlled
Variable
3.
Increase
Controller
Gain
This is repeated until the controller measurement cycles with constant amplitude (Figure 300-27).
(KC KCU)
Controlled
Variable
Time, Min.
PU
(Minutes)
320
The final controller gain setting is called the ultimate gain, denoted KCU. The
period of oscillation at the ultimate gain is called the ultimate period, measured
in minutes and denoted PU.
4.
The ultimate controller gain and the ultimate period are then used to calculate
tuning constants per the following table:
The ultimate controller gain and the ultimate period are then used to calculate tuning
constants per the following table:
ChevronTexaco Corporation
300-29
December 2003
Integral Time,
Min.
Derivative Time,
Min.
0.50 KCU
PI
0.45 KCU
PU / 1.2
PID
0.60 KCU
PU / 2.0
PU / 8.0
This method was the first systematic method developed for tuning industrial
controllers.
Shortcomings. Note that the Z-N tuning objective was quarter amplitude
damping (the response oscillates with each peak being one quarter that of the
previous peak).
Advantages. On the other hand, the Z-N procedure is simple and the tuning rules
are easy to remember.
Advanced tuning methods address most of these shortcomings. They are generally
model-based and address robustness (directly or indirectly). Model-based tuning
will be described in Section 333.
2.
Change the controller valve position by a small amount and record the
controlled variable.
The controlled variable response curve is called the process reaction curve.
Refer to Figure 300-28.
December 2003
3.
4.
The point that this line crosses the initial value of the controlled variable
measurement is used to determine P.
5.
The quantity X is the size of the controller output step and Y is the final
steady-state response of the controlled variable.
300-30
ChevronTexaco Corporation
1st-Order Lag
+ Dead Time
Approximation
TP
WP
'Y
KP
'Y 'X
322
Controller
Output (%)
'X
Time, Minutes
These values will be used to fit the response curve to a first-order lag plus dead time
model.
dCV (t )
+ CV (t ) = K P CO(t P )
dt
(Eq. 300-19)
The model parameters are determined as follows. The quantity P is the dead time
(minutes) and is determined graphically as explained above. The dead time is the
delay between a change in valve position and the resulting change in the controlled
variable. The process time constant is the time required for the controlled variable to
reach 63% of its final value. It can be determined graphically as sketched on the
response plot or calculated from the following equation:
P = Y/S [=] minutes
Finally, the process steady-state gain is calculated from the following equation:
KP = Y/X [=] % / %
An alternative approach to fitting the model, which is more accurate for noisy
processes, is illustrated below (Figure 300-29).
The process steady-state gain is found as before. The dead time and time constant
are calculated from the following equations:
P = 1.5 (t63% - t28%) [=] minutes
P = t63% - P [=] minutes
ChevronTexaco Corporation
300-31
December 2003
Controlled
Variable (%)
"Process
Reaction
Curve"
0.63' Y
'Y
0.28' Y
'Y 'X
KP
323
Controller
Output (%)
'X
t 28% t 63%
Time, Minutes
Having estimated a process model, we then apply the Ziegler-Nichols reaction curve
tuning rules:
Integral Time,
Min.
Derivative Time,
Min.
(1.0/KP)(P/P)
PI
(0.9/KP)(P/P)
3.3P
PID
(1.2/KP)(P/P)
2.0P
0.5P
As with the ultimate sensitivity tuning method, the controller objective function is
quarter amplitude damping. To reduce the oscillatory behavior, simply reduce the
recommended controller gain by 50 to 100%.
Note that the controller gain is proportional to the ratio of the time constant to the
dead time, so be cautious about applying this method when the dead time is small!
December 2003
300-32
ChevronTexaco Corporation
Note that the response is oscillatory for both common forms of the PID algorithm.
Refer to Forms of the PID Equation on page 300-44 for more information.
QIn
A
LI
QOut
324
Pump
ChevronTexaco Corporation
300-33
December 2003
Level Control Objectives. Ideally, we would maintain a constant level, and minimize the effect of inflow disturbances on downstream units. However, these are
conflicting objectives. To maintain constant level, outflow would have to mimic
every inflow change. To smooth the outflow, the level would have to change to
absorb the inflow fluctuations.
As a result, two distinct types of level control have evolved:
1.
2.
In most cases, averaging level control is more appropriate. As long as the level stays
within a defined range, we can take advantage of a vessels surge capacity to
smooth out the flow. Averaging level control takes advantage of whatever surge
volume is provided in the vessel. The degree of effectiveness in smoothing the flow
depends on the size of the surge volume relative to the magnitude of the flow disturbances.
We will investigate the level process and develop recommendations for proportional-integral (PI) controller tuning.
The Level Process. The dynamic response characteristics of the level process can
be determined by writing a dynamic material balance (inflow-outflow = rate of
accumulation):
dV ( t )
Q In ( t ) Q Out ( t ) = ------------dt
(Eq. 300-20)
where:
QIn(t) = Inflow [=] GPM
QOut(t) = Outflow [=] GPM
V(t) = Volume [=] Gallons
t = Time [=] Minutes
The volume can be calculated from the measured level as follows (assuming the
cross-sectional area is constant):
V(t) = k A L( t)
(Eq. 300-21)
where:
k = 7.481 Gal / Ft3
A = Cross-sectional area [=] Ft2
L(t) = Level [=] Ft
December 2003
300-34
ChevronTexaco Corporation
then,
dkAL ( t )
Q In ( t ) Q Out ( t ) = -------------------dt
or,
(t)
----------------Q In ( t ) Q Out ( t ) = dCL
dt
Gal
where C k A [ = ] -------Ft
The quantity C is called the capacitance of the vessel. It is effectively the
volume per foot of level. Since C is a constant, it can be moved outside of the
derivative term.
(t)
Q In ( t ) Q Out ( t ) = CdL
----------------(Eq. 300-22)
dt
Typically, the pump head is large compared to the static head provided by the level,
and thus, changes in level have very little effect on outflow (The process is non self
regulating). That is,
QOut f(L)
(Eq. 300-23)
L(t ) =
[Q
In
(t ) QOut (t )] dt + L0
(Eq. 300-24)
Level, L(t)
QNet /C [=] Ft/Min
1
L0
325
QIn(t)-QOut (t)
QNet
0
0
ChevronTexaco Corporation
300-35
Time, Minutes
December 2003
Unlike most processes, the level process is non self-regulating; it does not come to
steady state. For the level process to be at steady state, the net inflow must be zero.
Notice that the slope of the ramp response is QNet/C. Thus, the capacitance of the
vessel can be determined by introducing a known imbalance between inflow and
outflow and measuring the slope of the level response.
Solving for C
Slope = QNet/C
(Eq. 300-25)
Q Net
Gal Min
Gal
C = -------------- [ = ] ---------------------- = -------Ft Min
Slope
Ft
(Eq. 300-26)
Level Control Configurations. There are two common level control configurations:
1.
QIn
LC
FI
QOut
and
2.
QIn
LC
FC
QOut
326
December 2003
300-36
ChevronTexaco Corporation
Level Control Response Equations. The closed-loop response equations for both
single-loop and cascade configurations are second-order (and identical) when we
assume the following:
Both configurations have the same maximum flow (valve max or flow
controller setpoint max).
d Q Out
dQ Out
dQ In
H I ------------------ + I --------------- + Q Out = I ------------ + Q In
2
dt
dt
dt
where:
(Eq. 300-27)
C H T
H -------------------------- [=] minutes
K C F MAX
HT = Level transmitter span [=] Ft
The degree of flow smoothing between the inflow and outflow depends on the
values of the parameters in this equation.
Note that the measurable volume (within the level transmitter range) is given by
VolMeas = C HT
(Eq. 300-28)
Then
1
1 Vol Meas
H = ------- -------------------= ------- H [ = ]Minutes
KC
K C F Max
where H = VolMeas/FMax [=] minutes
The quantity H is the vessel residence time based on the maximum outflow FMax.
In other words, it is the time to fill the measurable volume (that is, within the level
transmitter range) with an inflow of FMax and with the outflow valve closed.
The following equation describes the level setpoint-to-level response:
2
dL SP
dL
d L
- + L SP
H I --------- + I ------ + L = I ----------2
dt
dt
dt
ChevronTexaco Corporation
300-37
(Eq. 300-29)
December 2003
This equation has exactly the same form and parameters as for the inflow to outflow
response. The following equation describes the inflow-to-level response:
2
H I dQ In
dL
d L
H I --------- + I ------ + L = ----------- ----------- C dt
2
dt
dt
(Eq. 300-30)
This equation tells us how much the level will vary as the inflow changes. Note that
the left-hand side of this equation (known as the characteristic equation) has
exactly the same form and parameters as for the previous two response equations.
Level Control Period and Damping. We will now compare the equations derived
for the level control system with the standard equation for a second-order system.
2 d2 Y ( t )
dY ( t )
n ---------------- + 2 n ------------- + Y ( t ) = K X ( t )
2
dt
dt
(Eq. 300-31)
where:
Y(t) = Dependent variable
X(t) = Independent variable
n = Natural time constant
= Damping coefficient
K = Steady-State Gain
t = Time
The response of a second order system to a step change in the independent variable
is shown in Figure 300-35. The shape of the response varies from a smooth,
S-shaped curve to a highly oscillatory one depending on the value of the damping
coefficient .
Comparing the level control systems equations with the standard form for the
second-order equation we can find the closed-loop period of oscillation, T
(minutes/cycle), and the damping factor, (dimensionless) for the level control
system:
T=
2
1 2
December 2003
300-38
I Vol Meas
KC
FMax
(Eq. 300-32)
1 I K C FMAX
2
Vol Meas
ChevronTexaco Corporation
K*'X
] = 1 (Critically Damped)
Y0
] > 1 (Overdamped)
328
X
'X
X0
0
Time
These equations show how the level controller tuning parameters affect the period
and degree of damping of the closed-loop response. A close examination reveals
several important (and surprising) facts about level control systems.
Note that increasing the level controller integral time, I, increases level control
stability (i.e., ) and increases control loop period, T. Both of those results are
expected. However, note that increasing level controller gain, KC, decreases control
loop period, but also increases stability (i.e., ). The latter result is exactly opposite of what one would typically expect.
In real-world level control systems, increases in KC eventually will result in an
unstable system because other lags in the system (that we didnt model) will become
significant. (The fact that increasing controller gain initially increases stability, but
ultimately destabilizes the system makes level controllers conditionally stable
systems.)
These observations show that tuning level controllers is non-intuitive.
Averaging Level Control Tuning. Page Buckley of Dupont (1964) developed a
tuning approach for averaging level control that has been applied throughout
ChevronTexaco. First, he proposed that the closed-loop response be critically
damped ( = 1). This will produce a smooth, non-oscillatory response.
Recall that
=
1 I K C FMax
2
Vol Meas
(Eq. 300-33)
ChevronTexaco Corporation
300-39
December 2003
4 Vol
4
H
Meas
I =
F
=
K
C Max K C
(Eq. 300-34)
1.5
(] = 1)
1.14
1.0
Outflow
1.0
Inflow
0.74
KC
0.5
0.5
Level
0.0
0.0
0
Dimensionless Time, t /W H
10
329
Note that, as expected, there is no oscillation in the response. But the output will
always temporarily exceed (overshoot) the inflow. (With the level process there
always needs to be an imbalance between inflow and outflow to change the level).
The plot shows that at peak level we have the following:
L peak
H
T
FMax
Q
In
K C =0.74
(Eq. 300-35)
December 2003
300-40
(QIn FMax )
(LPeak H T )
(Eq. 300-36)
ChevronTexaco Corporation
LPeak 1
=
H T
2
(Eq. 300-37)
Substitution into the previous equation allows us to solve for controller gain.
K C = 0.74
(QIn
(LPeak
FMax )
(1 2) = 0.74
= 0.74
(1 2)
H T )
(Eq. 300-38)
We can now use this value for controller gain to find the controller integral time.
I=
4
KC
Vol M
FMax
4
=
K H
C
(Eq. 300-39)
FMax
I = 5.4
(Eq. 300-40)
In summary, for averaging (flow smoothing) level control (Buckley tuning), use a
controller gain of 0.74 and a controller integral time of 5.4 times the vessel residence time. For example, for a vessel with a six minute residence time, controller
gain would be 0.74 and controller integral time would be 32.4 minutes.
The following plot (Figure 300-37) shows the level and outflow response to an
inflow change equal to half the maximum outflow with Buckley tuning. (The vessel
has a residence time of H = 6 minutes).
Notice how the vessel surge volume is used to smooth out the inflow change.
Tight Level Control Tuning. Buckley has also solved the response equations for
the general case (that is, for all values of the damping coefficient, ). See
Figure 300-38.
Note that outflow overshoots inflow for any (any controller settings). We will use
these curves to develop tuning guidelines for tight level control
For tight level control, we choose = 0.707 = 1 2 as this will provide the fastest
possible non-oscillatory response. The plot shows that the level peak for = 0.707 is
LPeak
H T
FMax
QIn
K C = 0.64
(Eq. 300-41)
Solving for KC
K C = 0 .64
ChevronTexaco Corporation
300-41
( Q In /FMax )
( L Peak / H T )
(Eq. 300-42)
December 2003
100
100
75
75
50
50
25
25
0
0.0
12.8
25.6
38.4
Time, Minutes
Inflow, GPM
51.2
64.0
Outflow, GPM
330
200.0
200.0
150.0
150.0
100.0
100.0
50.0
50.0
0.0
0.0
0.0
12.8
25.6
38.4
Time, Minutes
51.2
64.0
1.4
1.0
0.74
1.3
0.75
0.64
1.22
1.2
0.5
' HT
' QIn
KC
1.14
0.25
1.1
1.0
0.0
0.0
0.5
0.707
Underdamped
1.0
1.5
2.0
331
Overdamped
Then we specify a tight level range, e.g. 40% to 60% (starting from 50% level, the
level peak would be one tenth of the level transmitter range) for an inflow disturbance of half the maximum outflow. In mathematical terms, we have:
QIn 1
=
FMax 2
December 2003
300-42
LPeak
1
=
10
H T
(Eq. 300-43)
ChevronTexaco Corporation
(QIn /FMax )
( 1/ 2 )
= 0.64
= 3.2
(LPeak /H T )
( 1/ 10 )
(Eq. 300-44)
We can now use this value for controller gain to find the integral time. Recall that
=
1 I K C FMax 1 I K C
=
2
Vol Meas
2
H
Substituting = 0.707 = 1
(Eq. 300-45)
I=
2
KC
Vol M
FMax
2
=
K H
C
(Eq. 300-46)
FMax
I = 0.625
(Eq. 300-47)
In summary, for tight level control, use a controller gain of 3.2 and a controller integral time of 0.625 times the vessel residence time. For example, a six minute
vessel would have a controller gain of 3.2 and controller integral time of 3.75
minutes.
The following plot (Figure 300-39) shows the level and outflow response to an
inflow change equal to half the maximum outflow with tight tuning (vessel residence time of H = 6 minutes).
ChevronTexaco Corporation
300-43
December 2003
100
100
75
75
50
50
25
25
0
0.0
12.8
25.6
38.4
51.2
64.0
Time, Minutes
Inflow, GPM
Outflow, GPM
332
200.0
200.0
150.0
150.0
100.0
100.0
50.0
50.0
0.0
0.0
0.0
12.8
25.6
38.4
Time, Minutes
51.2
64.0
Notice how the level controller quickly moves the outflow to keep the level near the
setpoint.
1
CO(t ) = K C E (t ) +
E (t )dt +
0
dE (t )
+ CO0
dt
(Eq. 300-48)
This is the ISA standard form, and is sometimes called the parallel or ideal form.
The interacting form of the PID algorithm is given below.
1
CO(t ) = K C E (t ) +
dE (t )
E (t )dt 1 + D
+ CO0
dt
0
(Eq. 300-49)
December 2003
300-44
ChevronTexaco Corporation
In the Honeywell DCS, for example, both the interacting and non-interacting forms
of the PID equation are offered. Yokogawa offers only the non-interacting form.
It is important to note that the tuning parameters are different in the two forms.
Using the same tuning parameters in the two versions will not produce the same
results!
1
CO(t ) = K C E (t ) +
E (t )dt + D
dE (t )
+ CO0
dt
K C = K C 1 + D
I
I = ( I + D ) = I 1 +
D =
D I
=
( I + D ) D
(Eq. 300-50)
D
1 +
I
1
CO(t ) = K C E (t ) +
K C =
dE (t )
E (t )dt 1 + D
+ CO0
dt
0
KC
2 D / I
1 + 1 4 D / I = K C
2
1 1 4 D / I
I =
D =
(1+
2
(1+
) (1
1 4 D / I
(1
1 4 D / I
1 4 D / I =
2 D
1 4 D / I
I
2
2 D
Note that, because of the square root term, the equivalent factored version is valid
only for D/I 1/4.
ChevronTexaco Corporation
300-45
December 2003
Note also that if D/I 1/4 (in the non-interacting/ideal form), then D = I (in the
interacting/factored form).
No conversion necessary for P-only or PI control; there is only one equation form.
I = PU/2 D = PU/8
(Eq. 300-51)
But, what form of the PID equation did they assume? The controllers of the day
were closer to the interacting form than the non-interacting/ideal form. If we assume
that the Z-N tuning rules apply to the interacting form, then the following would be
a complete statement of their rules:
1
CO(t ) = K C E (t ) +
K C = 0.6 K CU
E (t )dt 1 +
0
I = PU 2
dE (t )
+ CO0
dt
D = PU 8
But, suppose the PID equation that was available in our control equipment had the
non-interacting form.
1
CO(t ) = K C E (t ) +
E (t )dt + D
dE (t )
+ CO0
dt
(Eq. 300-52)
We could simply use the conversion equations shown earlier to convert to the
Equivalent Values for the non-interacting form.
KC = (0.6 KCU) (1.25)
(Eq. 300-53)
We would then get the same results as if we had used the original values in the interacting PID equation.
However, if the interacting to non-interacting conversions were not made, the effective proportional gain would be 25% too low (less aggressive), effective integral
time would be 25% shorter (more aggressive), and the effective derivative time
would be 25% longer (more aggressive)
But what if Z-N assumed the non-interacting/ideal formulation? Most textbooks and
many journal articles apply Z-N to the ideal form!
December 2003
300-46
ChevronTexaco Corporation
CO(t ) = K C E (t ) +
KC = 0.6 KCU
E (t )dt + D
dE (t )
+ CO0
dt
I = PU/2 D = PU/8
(Eq. 300-54)
But, suppose the PID Equation we were using had the interacting form.
1
CO(t ) = K C E (t ) +
E (t )dt 1 +
0
dE (t )
+ CO0
dt
We could simply convert to the equivalent values for the interacting form.
K C = (0.6 K CU ) 2
I = (PU 2) 2
D = (PU 8) 2
If we did so, we would then get the same results as if we had used the original
values in the non-interacting PID equation.
However, If the non-interacting to interacting conversions were not made, the effective proportional gain would be 100% too high (more aggressive), the effective integral time would be 100% longer (less aggressive), and the effective derivative time
would be 100% shorter (less aggressive).
The following shows Z-N tuning with and without PID form conversion
(Figure 300-40).
The results for the parallel PID and series PID (converted) are very similar but not
precisely the same because the conversion equations used didnt consider the derivative filter term found in real controllers.
The main points to take away from all this when applying any tuning rules are to
1.
Know the form of the PID equation the tuning rules assumed, and
2.
Either apply the rules as intended or convert the tuning parameters to the
equivalent values for the other PID form.
ChevronTexaco Corporation
300-47
December 2003
Fig. 300-40 Z-N Tuning With and Without PID Form Conversion
December 2003
300-48
ChevronTexaco Corporation
The DS/IMC tuning rules have been widely accepted as the successor to ZieglerNichols tuning rules.
Ideal PID
KCK
KCK
--
--
-----------+
-----------+
------------------+2
/2
+2
------------------+2
+
--2
-------------2 +
1
---------------------- + 3 +
1 + 2 +
------------------------- + 3 +
Model
K ------------s + 1
Ke ------------s + 1
Ke
-------------s + 1
K ( 3 s + 1 )e
------------------------------------------( 1 s + 1 ) ( 2 s + 1 )
1 + 2 +
1 2
+ -------------------------1 + 2 +
where:
3
= ---------------------- + 3 +
(Eq. 300-55)
The parameter lambda, , in the table is the desired closed-loop time constant, and
is the only tuning parameter the control engineer has to adjust! All other parameters
in the tuning formulas come from the estimate of the process model.
ChevronTexaco Corporation
300-49
December 2003
Controlled
Variable
361
Controller
Output
Time
We can approximate the response with a second-order plus dead time model.
ps
KP e
= -------------------------------------------------G
P
( 1P s + 1 ) ( 2P s + 1 )
(Eq. 300-56)
Lambda = 2 min.
Lambda = 3 min.
Controller Gain
0.2907
0.1938
0.1453
Integral Time,
min.
2.5000
2.5000
2.5000
Derivative Time,
min.
0.6000
0.6000
0.6000
Figure 300-43 shows IMC/DS (lambda) tuning for the case when the process model
is an exact representation of the actual process (perfect model). Both a setpoint
change and disturbance are simulated.
December 2003
300-50
ChevronTexaco Corporation
Notice that the lambda based controller responses are smooth, both for the setpoint
change and the disturbance. A value of lambda of 2 or 3 is a good choice. The
Ziegler Nichols tuning is also shown, for comparison. Notice that the Z-N tuning is
quite oscillatory for the setpoint response.
Figure 300-44 explores the robustness quality of the model-based controller. Here
the actual process gain has changed to 150% of the nominal value.
Fig. 300-44 Model-based Controller (Process Gain 150% Nominal Value)
ChevronTexaco Corporation
300-51
December 2003
The lambda controllers (with the original tuning) perform very well even for this
large process change. The Ziegler Nichols tuned controller is on the verge of instability.
Figure 300-45 explores the robustness to dead time. Here we have increased the
process dead time by 50 percent over the nominal value.
Fig. 300-45 Model-based Controller (Process Dead Time 150% of Nominal Value)
Again the lambda controllers (with the original tuning) are performing very well
even for this large process change. The Ziegler Nichols controller is quite oscillatory. The robustness benefits of IMC/DS (lambda) based tuning are evident in these
plots.
Tuning Tools
The Chien and Fruehauf tuning rules (along with tools to facilitate model identification) are built into two internally-developed software packages available to ChevronTexaco personnel:
Both packages are available from the Process Control Team (Process Automation
Unit) of ChevronTexaco Energy Research & Technology Company in Richmond,
CA.
The ICM Tuning Tool is a Honeywell TDC3000 applications module (AM) application. This easy to use package helps users apply the IMC-PID tuning rules on
Honeywell TDC3000 distributed control systems. The user can select a loop to tune,
December 2003
300-52
ChevronTexaco Corporation
perform step testing, plot the data, model the data, implement tuning coefficients,
and examine the closed loop performance all from within the two screens of the tool
while sitting at the TDC console. After entering the desired closed-loop time
constant, tuning coefficients are calculated automatically for the user on the basis of
IMC-PID theory. The tuning parameters automatically account for the Honeywell
PID equation form that is in use (interacting or non-interacting). The package also
includes provision for averaging level control (based on Buckleys approach
described earlier). There are numerous help pages to guide the user through the
tuning process.
For those sites without a TDC3000 system, there is an excel spreadsheet called
TUNE, for model identification and tuning based on the DS/IMC-PID tuning rules.
This spreadsheet imports input/output data (for example, from a controller step test),
facilitates model identification, and then recommends tuning parameters for various
implementations of PID controllers.
There are also a number of excellent commercial tuning packages. For example:
ExperTune (www.expertune.com)
These commercial packages often go well beyond loop tuning. For example, the
ExperTune advanced package includes support for the numerous versions of the
PID control algorithm available in different DCS systems. It also facilitates the
process of linearizing control loops, calculates optimal filtering and tuning to reduce
valve wear, provides statistics on control variability, provides robustness plots,
etc.
Gain
%PB
I, Min.
D, Min.
Flow (Fast)
0.5
200
0.1-0.5
None
Flow (Slow)
Noisy/Sticky Valve
None
None
0.1-0.5
None
Pressure (Fast)
Liquid Packed
0.5
200
0.5-1.5
None
Pressure
1.0
100
1.0-2.0
0-0.5
Level (Tight)
3.2
30
0.6 H(1)
None
(1)
None
Level (Averaging)
ChevronTexaco Corporation
0.74
300-53
135
5.4 H
December 2003
Type
Gain
%PB
I, Min.
D, Min.
Temperature (Fast)
Bypass-Type
0.5-1.0
200-100
1.0-2.0
0-0.5
Temperature
0.5-1.0
200-100
4-20
1-5
These values are approximate. One of the tuning methods discussed previously
should always be used. For future reference, always record the control loop ID
number (e.g., FRC 123), the date, and the tuning constants when you have finished
tuning the loop.
Process
Furnace
334
TC
TC
Primary
Burners
Primary
Burners
PC
Secondary
FC
Secondary
Fuel Gas
December 2003
Fuel Gas
300-54
ChevronTexaco Corporation
In the first example, the furnace temperature controller cascades to the fuel gas flow
controller. In the second, it cascades to a burner pressure controller. In both cases,
the secondary controllers (PC or FC) are much faster than the primary controller
(TC) and would be tuned with much shorter reset times. In fact, for cascade control
to be significantly better than single-loop control, the secondary controller should be
at least 5 to 10 times faster than the primary controller.
Cascade control improves performance in three ways when properly applied:
1.
2.
Disturbances measured by the primary controller also die away more quickly
because the primary controller can be tuned to be faster
3.
The secondary loop linearizes that part of the process. For example, in the TCFC cascade, the temperature controller output effectively represents fuel gas
flow. The flow controller will position the valve wherever necessary to get the
flow asked for by the temperature controller. If the temperature controller set
the control valve position directly, the nonlinear flow-versus-valve position
relationship would adversely affect the temperature controller performance.
Cascade control is most effective for disturbances that are measured by the
secondary controller.
Lets look at how a furnace outlet temperature to fuel gas pressure control cascade
deals with some typical furnace disturbances: The key to how effective the cascade
will be in rejecting these disturbances is where the disturbance enters the loop
(Figure 300-47).
Fig. 300-47 Furnace TC-PC Disturbance Location Block Diagram
Fuel Gas
Supply
Pressure
335
Temp
SP
Temperature
Controller
Pressure
Controller
Control
Valve
Burner
Dynamics
Feed
Inlet Rate
Temp Ambient
Conditions
Furnace
Dynamics
Outlet
Temp
Secondary Loop
Pressure
Transmitter
Primary Loop
Temperature
Transmitter
ChevronTexaco Corporation
300-55
December 2003
The sketch clearly shows that only the fuel gas supply pressure upset will be picked
up by the (faster) pressure controller, and so we can expect that the biggest improvement over single-loop control would be for that upset.
Figure 300-48 contrasts the effectiveness of the cascade loop in rejecting the feed
flow and fuel gas supply pressure disturbances:
Fig. 300-48 Furnace TC-PC Disturbance Responses
Outlet
Temp
Outlet
Temp
Fuel Gas
Supply
Presure
Process
Flow
336
Time
Time
Response to the process flow upset is not much better than for single-loop control
and the response is slow because the (slower) temperature controller must respond
to the upset.
On the other hand, for the supply pressure upset, burner pressure is affected almost
immediately, and the (faster) pressure controller compensate quickly for the change
in fuel gas flow. The fuel flow remains relatively steady while the pressure is
changing so the furnace temperature will be more constant. Fuel flow changes
almost immediately when the control valve is moved. Therefore, the pressure
controller can be tuned to eliminate most of the disturbances in fuel flow.
The temperature-to-flow cascade control option is the normal choice for fired
heaters. With cascade temperature-to-flow control, the flow controller linearizes the
valve response which helps with temperature controller tuning. However, in some
cases the fuel system is so dirty the flow measurement cannot be made reliable.
When fuel gas heating value (and gravity) fluctuates, the temperature-to-flow
cascade arrangement can be easily modified to fuel gas BTU control by
combining the flow measurement with an online fuel gas heating value measurement (or inferential).
Temperature-to-pressure cascades are sometimes used. Some fuel gas burners have
a very narrow operating range. The cascade temperature-to-pressure control option
allows high and low limits to be easily set (via pressure controller setpoint limits).
Supply pressure upsets are easily handled by the temperature-to-pressure cascade,
but not as directly for the temperature-to-flow cascade. The temperature-to-flow
cascade responds to the pressure upset partially, to the extent the flow measurement is affected by the pressure.
The two cascades respond very differently when adding or removing burners. When
adding (or removing) burners with the temperature-to-flow cascade, the initial
December 2003
300-56
ChevronTexaco Corporation
response of the flow controller is to redistribute the flow keeping the total constant,
whereas with the temperature-to-pressure cascade, the pressure controller will
initially increases (or decreases) heat flow.
FFC
201
FC
101
Ratio
Controller
337
Controller FFC-201 is a flow fraction controller (a.k.a. ratio controller). It is basically a flow controller with a setpoint equal to
FFC_SP = R * Flow_101
where:
R = ratio
The operator enters the desired ratio, and the FFC keeps Flow-201 in the proper
ratio to Flow-101. The signal to the FFC can be either the measurement of FC-101
(as shown) or the setpoint. Using the setpoint of FC-101 has the advantage that the
signal will be noise free. On the other hand, if FC-101 becomes saturated, its
measurement and setpoint will differ, so the ratio is normally based on the measurement.
ChevronTexaco Corporation
300-57
December 2003
DM
Disturbance
Transmitter
D
(EU)
KMD GMD
(%)
FF Controller
Output, %
Process
CVSP
KM
(EU)
CO +
K C GC
KSL GSL
(%)
K D GD
Secondary
Control Loop
Primary Controller
MV
K P GP
CV
(EU)
(EU)
CVSP%
Controlled Variable
Transmitter
CVM
KM GM
338
(%)
where:
CVSP Controlled variable (CV) setpoint [=] EUCV
EU Engineering units
KM Controlled variable transmitter gain [=] %/EUCV
CVSP% Controlled variable %-setpoint [=] %
KC Primary controller gain [=] dimensionless (%/%)
GC Primary controller dynamics (integral, derivative)
KSL Secondary control loop closed-loop gain [=] EUMV/%
GSL Secondary control loop closed-loop dynamics
December 2003
300-58
ChevronTexaco Corporation
Response of
Controlled Variable
to Manipulated
Variable
Feedforward
Compensation
(Adjust Gain
& Dynamics)
Controlled
Variable
Mirror
Images
Net Result:
No Change in Controlled Variable
Response of
Controlled Variable to
Disturbance Variable
339
ChevronTexaco Corporation
Time, Minutes
300-59
December 2003
CV(t)
63% of Final Response
KP 'CO
339a
CO(t)
'CO
TP
TPWP
dCV ( t )
P ------------------ + CV ( t ) = K P CO ( t P )
dt
(Eq. 300-57)
where:
KP Effective process gain [=] %/% (alternatively, EUCV/%)
P Effective process time constant [=] minutes
P Effective process dead time [=] minutes
Note that the word effective is used in the above parameter definitions to indicate
that the process dynamics include the closed-loop dynamics of the secondary
control loop.
December 2003
300-60
ChevronTexaco Corporation
Then, the response of the controlled variable to the disturbance is also modeled as a
first order lag plus dead time.
dCV ( t )
D ------------------ + CV ( t ) = K D D ( t D )
dt
where:
(Eq. 300-58)
Dead Time
Output
343
Dead Time
Input
0
TDTP
Time, Minutes
The quantity (D - P) represents the dead time (in minutes) that the feedforward
controller must delay the manipulated variable to line up the manipulated-variableto-controlled-variable response with the disturbance-to-controlled-variable
response.
The quantity (D - P) must be positive for it to represent a delay; if it were negative it would be represent a prediction. Thus, for the feedforward controller dead
time to be realizable, the disturbance response dead time must be greater than the
process response dead time (D P).
Feedforward Controller Lead-Lag Compensation. The remainder of the feedforward controller is a lead lag function. The following equation describes the dynamic
response of the lead-lag function to a unit step in its input.
Lead Lag
LLOut (t ) = K LL 1 +
Lag
ChevronTexaco Corporation
300-61
t Lag
e
(Eq. 300-59)
December 2003
where:
KLL Lead-lag gain [=] % / %
Lead Lead time constant [=] minutes
Lag Lag time constant [=] minutes
The lead-lag function gain, KLL, is used to set the overall feedforward controller
gain. The value of this parameter determines the ultimate size of the feedforward
controllers response to a disturbance.
The values of the lead and lag time constants affect the dynamic response (or transient response) of the feedforward controller.
For the case when the lead time constant is greater than the lag time constant, the
response to a unit step at time zero looks somewhat like the (filtered) derivative
mode of the PID algorithm (Figure 300-54).
Fig. 300-54 Feedforward Lead-Lag Step Response (Lead > Lag)
KLL
WLead
WLag
Lead Lag
Output
63%
344
KLL
0
0
WLag
Time
For the case when the lead time constant is less than the lag time constant, the
response to a unit step at time zero looks as follows (Figure 300-55).
Fig. 300-55 Feedforward Lead-Lag Step Response (Lead < Lag)
Lead Lag
Output
KLL
KLL
345
63%
WLead
WLag
0
0
December 2003
WLag
300-62
Time
ChevronTexaco Corporation
The lead-lag parameters must be set to the following values for proper feedforward
compensation:
KLL = -(KD/KP) (- disturbance-to-process gain ratio) [=] % / %
Lead = P (process time constant) [=] minutes
Lag = D (disturbance time constant) [=] minutes
Both the disturbance and process gains must be on a percent of span basis to put
the feedforward controller gain on a dimensionless basis (%/%).
Furnace
FI
Furnace
Outlet
Temperature
Feed Rate
Transmitter
Feed
Stream
Temperature
Transmitter
Burners
FY
Feedforward
"Controller"
Temperature
Setpoint
TC
F(t)
Temperature
Comtroller
Control Valve
6
FC
FY
Summer
Fuel Gas
Flow Controller
PI
340
Fuel Gas
Supply
ChevronTexaco Corporation
300-63
December 2003
Temperature
Measurement
(%)
341
Temperature
Controller
Output (%)
0
Time, minutes
We can use either of the two methods described earlier to fit this response to a firstorder plus dead time model.
For the above manipulated variable response data, the model parameters are found
to be as follows:
KP = 6.0 %/% (effective process gain)
P = 1.2 minutes (effective process time constant)
P = 1.0 minutes (effective process dead time)
Disturbance Variable Response. Furnace feed rate is the disturbance we want to
compensate for with feedforward control. We need to either wait for a change to
occur naturally or deliberately make a change (Figure 300-58).
Fig. 300-58 Disturbance Response
Temperature
Measurement
(%)
342
Feed Rate
Measurement
(%)
0
Time, minutes
December 2003
300-64
ChevronTexaco Corporation
Again, we can fit the response to a first-order lag plus dead time model. For the
above response disturbance response data, the model parameters are found to be as
follows:
KD = -12.3 %/% (disturbance gain)
D = 0.52 minutes (disturbance time constant)
D = 2.0 minutes (disturbance dead time)
Feedforward Controller Design. The required feedforward controller (compensator) is a lead-lag and dead time function in series. The feedforward controller
dead time parameter is as follows.
(D - P) = 2.0 - 1.0 = 1.0 minutes.
The feedforward controller lead-lag function parameters are as follows:
KLL = -(KD/KP) = -(-12.3/6.0) = 2.05 (%/%)
Lead = P = 1.2 minutes
Lag = D = 0.52 minutes
Feedforward Controller Performance. Figure 300-59 shows the response of the
furnace temperature feedback plus feedforward control system to a step disturbance
in feed rate and then to a setpoint change.
Fig. 300-59 Furnace Feedforward Control Performance
Furnace Outlet Temperature, DegF
525.0
525.0
512.5
512.5
500.0
500.0
487.5
487.5
475.0
475.0
0.0
7.0
14.0
21.0
Time, Minutes
346
28.0
35.0
105.5
65.0
87.5
55.0
70.0
45.0
52.5
35.0
35.0
25.0
0.0
ChevronTexaco Corporation
7.0
14.0
21.0
Time, Minutes
300-65
28.0
35.0
December 2003
The feedforward controllers dead time plus lead lag response to the feed rate disturbance is clearly evident in the plot. A derivative spike does not occur with the
setpoint change because of the PID option chosen (derivative on measurement).
Figure 300-60 shows how well the fuel gas flow controller is able to track the feedforward controllers lead-lag command signal.
Fig. 300-60 Fuel Gas FC Response to Lead Lag Command
Fuel Gas Flow Controller SP, MSCFH
347
105.5
105.0
95.0
95.0
85.0
85.0
75.0
75.0
65.0
65.0
0.0
2.0
4.0
6.0
8.0
10.0
Time, Minutes
The lags in the flow control loop do not allow perfect tracking. But recall that the
closed-loop dynamics of the secondary controller (the fuel gas flow controller) were
included in the overall process dynamics.
Controlled
Variable
348
Disturbance
0
Time
Then adjust the dead time and lead-lag parameters to minimize deviations from the
setpoint (Figure 300-62).
December 2003
300-66
ChevronTexaco Corporation
Controlled
Variable
349
Disturbance
0
Time
Less re-work
COA sessions are typically held in the early stages of a new plant construction
project or before a re-instrumentation or advanced control project to develop a
sound basis for the control design. For new projects, the Control Objective Analysis
(COA) is best done when Process Flow Diagrams are finalized and the P&IDs are in
early development. Often the COA will impact the P&IDs so its good to perform
the COA early in the P&ID development phase. COAs are often performed during
late CPDEP Phase 2 or early Phase 3. CPDEP (ChevronTexaco Project Development and Execution Process) is ChevronTexaco's process for conducting projects.
The Control Objectives Analysis (COA) technique serves as a Quality Improvement method for process operations. The COA meetings provide a forum for
collecting all relevant information on opportunities to improve plant performance. A
snapshot of the current process control related opportunities are documented in the
form of prioritized lists. Each of these lists is then assigned to a different group
(process engineering/designs engineering/maintenance/operations) for follow up.
ChevronTexaco Corporation
300-67
December 2003
The COA participants form a Quality Improvement team. The COA team is
comprised of operators the people who know the process and are directly
involved in the work. Engineers and plant supervisors supplement the team on an
equal ranking basis.
The COA technique, through the defining of control objectives, analyzes the operators work process, provides a better understanding of why things are done, and
empowers the operators to get further involved in improving their work.
The COA process can be applied to any continuous process. Within
ChevronTexaco, it has been used in the refining, chemicals production, and
upstream processing facilities.
December 2003
Individual Process Control Objectives There are as many individual objectives as there are control valves. Note: A valve is a discussion focal point to
talk about the process. A control objective does not have to be identified with a
specific valve. The COA discussion should attempt to develop objectives that
are clear, concise and true all the time. (Although in practicality, it is realized
that there may be times that this is not possible.) An example control objective might be: Control the flow to R-100 at the operator Setpoint. Detailed
control strategy design is beyond the scope of the COA process. ChevronTexaco uses a process called Control Design Analysis or CDA for developing detailed control strategies.
List of Identified Process Upsets Typical upsets are: feed rate/composition/temperature changes, feed contamination, utility changes (fuel BTUs, fuel
availability, steam pressure, cooling water temperature), and downstream
restrictions/requirements. The frequency and process impact of these disturbances are identified.
List of Identified Process Constraints Typical constraints are capacity limitations on the following equipment: overhead condensers, reboilers/furnaces,
compressors/blowers, hydraulics (pumps, lines, valves), and tray loading. The
300-68
ChevronTexaco Corporation
List of Advanced Control Opportunities A generalized list includes: maximize plant throughput, maximize upgrade/conversion/severity, and balance
performance in parallel equipment, control product specifications, smooth plant
operations, minimize column pressure, and minimize energy requirements.
ChevronTexaco Corporation
300-69
December 2003
Two or three operators from different crews Its important to have several
operators so that its evident that its an operator focused meeting and so that
the operators can bounce their thoughts off each other. The selected operators
should have strong process knowledge and good communications skills.
A shift supervisor from a crew not represented above that has had operating
experience on the process.
For new plant construction projects, the COA participants would be modified to
include the operations representative, I&E project manager, as well as engineering
contractor representatives.
The control engineer and moderator can alternate between facilitating and taking
notes. It is important that the words used during the COA process be captured as
stated so that the operators recognize the work as their own.
Other optional participants may include: process unit operating assistant or section
supervisor, corporate process expert, designs engineer, planner. Note: The optional
participants need to understand their limited role and not hinder full participation by
the operating people. For effective group dynamics, the maximum COA size is
about eight people.
December 2003
300-70
ChevronTexaco Corporation
The controller contains specific dynamic response information about the plant
obtained through plant testing.
The controller is predictive and accounts for the impact of past moves. The
predictions are updated after each processing cycle with actual measured
values.
Constraints are defined for all manipulated and controlled variables. The
controller is not allowed to make moves that exceed the limits on the manipulated variables. The controller adjusts the process to ensure that all controlled
variables stay within their limits both now and in the future.
ChevronTexaco Corporation
300-71
December 2003
LP Targets or Steady-State Targets (SS Target) are the values that the MPC wants to
move the plant to. A linear program, or LP, calculates the targets each time the
controller does a calculation cycle.
Past
Future
Targets (LP)
with Future MV Moves
Predicted CVs
without Future MV Moves
Planned Future
MV Moves
k k+1 k+2
k+p
Prediction Horizon (Time)
A process model, which is determined from step testing, is used to project the future
value of the controlled variables. Given the past history of the manipulated and
controlled/constraint variables, the controller can use the model to predict the future
CV values.
An embedded linear program (LP) is part of the MPC controller and is used to
determine optimal targets for the controlled and constraint variables. At any given
time, k, the controller has knowledge of all the past manipulated variable moves and
the past values of the controlled variables. It then can calculate an optimal series of
future manipulated variable moves which will bring the controlled/constraint variables up to the LP targets
After one time step has passed the controller takes the actual control variable
measurements and compares them with the predicted values. The difference is
added as bias to future predictions. This feedback action brings robustness to the
control algorithm.
December 2003
300-72
ChevronTexaco Corporation
2.
The process runs at a better operating point. For example, if the MPC objective
is to maximize feed, the controller will increase feed to the plant up to multiple
simultaneous limits.
ChevronTexaco Corporation
300-73
December 2003
353
DMC OFF
DMC ON
Upper
Limit
Compressor
Interstage
Pressure
01:32
04:37
07:41
10:46
13:50
16:55
19:59
354
DMC OFF
DMC ON
Lower
Limit
Regenerator
Excess O2
21:52
05:53
13:53
21:53
05:54
13:54
21:54
December 2003
300-74
ChevronTexaco Corporation
DMC ON
355
Day-to-Night
Temperature Variation
Heavy
Conversion
08/30
08/31
09/01
09/02
09/03
09/04
09/05
09/06
1 week of 1 hr Averages
Project
Scope &
Objectives
Select / Install
Analyzers &
Inferentials
Develop
LP
Economics
Tune Controllers
Repair Valves
Add Instruments
Install
Additional
Inferentials
Plant
Pre-Test
Plant
Test
Build/Test
Controller
Offline
356
Commission
Controller
LongTerm
Support
ChevronTexaco Corporation
300-75
December 2003
Since some sites do not have experienced MPC engineers, MPC projects are often
led by either the MPC vendor or a third party contractor who performs MPC project
work. Within US Refining, MPC projects are being increasingly done by experience ChevronTexaco engineers without the assistance of outside project consultants.
Typical MPC project activities include pre-testing, testing, model identification,
development of inferential measurements, controller building and off-line tuning,
and commissioning. Project duration would typically be 6 to 9 months. If an outside
MPC contractor is used, he or she would be on site for pre-testing, testing, commissioning. Data analysis, model identification, inferential measurement development,
controller building and off-line tuning can be done at the contractors offices. The
control engineer/technician who will be responsible for maintaining the MPC
system should participate in all of these activities to the extent practical. This participation by the control engineer will enhance his/her ability to support the MPC
application in the future.
A ChevronTexaco experienced Control Engineer should be assigned to the MPC
project. Preferably this person would come from the plant site. However, if a person
with the right skills is not available, an experienced engineer from the central
process control group in Richmond could serve in that role.
Operator proprietorship in the finished product is also crucial to the success of the
project, so an operations representative should be involved in the project. The Operations representative will help to define the design objectives for the MPC
controller. Also, the Operations representative will assist with operator training and
help to promote understanding of the MPC application.
The list of manipulated variables (MVs) should include all the significant
handles that the operator uses.
The controlled/constraint variable (CV) list must include all the product specifications as well as all of the constraints that may limit the operation at one time
or another during the course of the run.
The disturbance variables (DVs) should include all those disturbances that can
be measured and that have a significant impact on the process.
A final design cannot be completed without some plant testing to determine actual
plant constraints. It should be expected that the list of manipulated, disturbance, and
controlled/constraint variables will be revised/updated after plant testing has been
completed.
December 2003
300-76
ChevronTexaco Corporation
constraints are available to the controller and that all the interactions can be effectively dealt with.
However, large MPC controllers are harder for the operator to understand and
harder for the control engineer to maintain. One must balance the benefits gained
from the larger scope against the added complexity.
Controllers with less than 10 MVs are considered small. Controllers with 10 to 20
MVs are considered medium size. Any controller with more than 20 MVs is considered large.
Plant Pre-Test
The objectives of the MPC pre-test are to:
Ensure the data collection system is gathering all the necessary data and storing
it properly
Ensure that the regulatory controllers are tuned well and configured properly
ChevronTexaco Corporation
300-77
December 2003
Below are some examples of the things one must look for in the pre-test.
Each of the controllers in the MPC scope must be examined for tuning and robustness. In addition, because of the interaction between controllers, one should
examine controllers outside of the MPC scope as well. The dynamics of these loops
will also impact the MPC models.
Much of the work of the pre-test involves improving the tuning of controllers, as
required. But it is also sometimes necessary to re-configure some of the loops for
better performance. For example:
Use a variable gain option on the control algorithm to account for gain differences in split range valves (vent and natural gas makeup).
Add fuel gas flow pressure, temperature, and specific gravity compensation for
a fuel gas flow.
Replace a flow controller with a duty controller to break the interaction with
another section of the plant.
New instruments may need to be added to support 1) new constraints that have been
identified, or 2) improved control configurations that are required:
The MPC design may indicate need for a number of calculated variables and hence,
tags will need to be created and added to the data collection.
Note that by simply plotting controller output versus flow rate for all key control
valves as they respond to set point changes and disturbances during the testing, will
give a quick indication of the need for linearizing valve transformations. More
focused testing can then be scheduled for suspect valves.
December 2003
300-78
ChevronTexaco Corporation
It is best to linearize the valve in the DCS loop itself rather than in the MPC. This
improves DCS loop performance and eliminates the need to re-tune the controller as
the operation changes.
Plant Test
General. The key to developing a good MPC model is a good plant test. There are
many things that can go wrong in a plant test that can invalidate some or all of the
data. For example, control valves can saturate, large unmeasured disturbances can
enter the system, atypical events can occur, etc. Therefore, it is crucial to do everything possible to ensure the quality of the data.
A good practice is to take data for all variables (tags) in the plant even if theyre not
in the original controller scope. New plant constraints may reveal themselves during
the testing that need to be included in the controllers list of CVs. It is less of a
disruption to operations if the new model can be identified from existing data, rather
than having to go back and repeat part of the plant test.
Typically, 10-15 moves in each MV are targeted for the plant test. It is helpful to
keep a running record of the number good moves for each MV to aid in monitoring the progress of the test. Occasionally, moves have to be rejected because of
control valve saturation, controllers in the wrong mode, major plant upsets, etc.
Various vendor packages are available to partially automate/facilitate the step
testing process. These automated step testing packages can reduce project costs by
reducing the time spent during the plant test.
For, example, Honeywell has a product called the Step Test Builder (unofficially
known as the Robo-Tester). It has two parts:
A display schematic shows the sequence graphically, and allows the amplitude to be
changed, the execution started/stopped, etc.
AspenTech has a product called SmartStep which works with DMC. Unlike other
automated step test packages, SmartStep is best suited for project revamps as it
conducts step testing with a DMC controller in closed loop. Special logic is
employed to enforce manipulated variable limits while keeping the controlled variables within their prescribed limits. SmartStep ensures the dataset contains sufficient MV moves at all frequencies (low, medium and high) to ensure good test data
for the model identification.
SmartStep also has some special features that can be used during the pretest to identify sticky valves and to develop data for model-based PID loop tuning.
Analyze Data and Build Dynamic Model. Analyzing step test data and building
the MPC model requires skill and experience. The step test data is analyzed with a
model identification software package such as AspenTechs DMCplus Model.
ChevronTexaco Corporation
300-79
December 2003
The first step in the process is to carefully analyze the step response data (vector
files) for problems such as control valve saturation, data dropout, unchanging regulatory controller outputs (indicating the controller was put in manual), etc. That
portion of the data must be manually marked bad (sliced out). For major problems affecting all the data, global slices can be used. Sliced out data is not used
for the analysis. The software is then used to generate model curves similar to that
in Figure 300-68 below:
Fig. 300-68 Sample Model Curves
Selecting Model Curves. The engineer will select a curve from this figure to use in
the final MPC model. Typically, the engineer will pick the curve with the shortest
time-to-steady-state where the steady state has been reached. Each curve used in the
final model can have a different TSS. When building the overall controller model,
the shorter curves will automatically be extended.
For Disturbance Variables (DVs), choose curve with lowest gain (other factors
being equal). Otherwise the MPC controller would predict too much CV movement
and move the MVs excessively for control.
For Manipulated Variables (MVs), choose curve with highest gain (other factors
being equal). The MPC controller will plan MV moves based on the models high
gain and so the controller will be less likely to overshoot.
In the response above, the CV comes to steady-state about 50 minutes after the MV
step is introduced.
Figure 300-69 shows another example of the MV-CV step response model. In this
case the CV does not come to steady-state, and is known as a ramp or integrating variable. Levels often respond in this manner.
December 2003
300-80
ChevronTexaco Corporation
The Dynamic Matrix. Step response models are developed for each MV - CV pair.
In addition the step response models for each measured disturbances are identified
(DV - CV). All the significant models are assembled into the dynamic matrix. A
typical example of a complete controller model is shown in Figure 300-70.
Fig. 300-70 Example of a Complete Controller Model
The above plot (an AspenTech DMC model) is arranged with the
controlled/constraint variables (CV) listed across the top and the manipulated vari-
ChevronTexaco Corporation
300-81
December 2003
ables (MV) and disturbance variables (DV) listed vertically on the left. Honeywell
RMPCT models are arranged with the MVs across the top and the CVs vertically.
When deciding whether or not to include particular response curves, be sure to scale
the response plots with typical moves. Note that if the typical move value for a
particular MV or DV is set too low, the response curves for that variable will seem
insignificant.
Singularities in Matrices. The gain matrix must be checked for a condition called
near-singularity using singular value decomposition (SVD). If one observes an LP
target switching back and forth between and high and low limit, this may indicate a
problem with a near-singularity. Matrices should be analyzed for near singularity
before commissioning the MPC on line.
MVs with similar effects on a CV can cause problems from a matrix inversion point
of view, as in pass balancing, for example. In that case it is necessary to manually
make the gains exactly the same so the LP will not trade off small differences.
Singular value decomposition is a mathematical technique that decomposes a matrix
G into three others, typically referred to in the literature as U S and V matrixes
G = UV
(Eq. 300-60)
The diagonal matrix contains the singular values where the condition number of
the matrix is the ratio of the highest singular value to the lowest. A condition
number of infinity means that the matrix is exactly co-linear.
In control engineering MPC terms, this means that there could be a substantial
difference in MV movement from one LP optimum to the next, depending on what
happens to move the LP solution. Events that can cause this to occur are model gain
changes, LP cost changes and variables switched in and out of the controller.
A plot of MV movement from one solution point to another vs. the condition
number for several 2 by 2 matrixes is shown in Figure 300-71.
Fig. 300-71 MV Movement
200
R = 0.9779
MV Movement
150
100
50
Condition 50
100
December 2003
300-82
ChevronTexaco Corporation
in parallel processes (e.g. two feed trains) or similar manipulated variables (e.g.
furnace fuel gas and fuel oil).
The whole area of the effect of ill-conditioned matrices on the performance of MPC
has only recently been fully recognized. Dealing with this issue is crucial to the
success of the controller.
Sometimes it is advisable not to include certain curves in the dynamic matrix model.
For example, to prevent the LP from finding a solution that is undesirable from a
process standpoint, or to prevent the controller from using a particular MV for
controlling a particular CV.
Commission Controller
Initially, MPC is turned on in the prediction-only mode no control moves are
sent to the field, but trends of the planned future moves are available. After evaluation by the project team and given the go-ahead by operations, commissioning can
begin.
ChevronTexaco Corporation
300-83
December 2003
Formal advanced control training must be given to each MPC user prior to commissioning. It must be tailored to their needs. It should not be given too early or key
information might be forgotten. The training manuals and documentation should be
reviewed and accepted by Operations before the formal training commences.
December 2003
300-84
ChevronTexaco Corporation
Maximum potential ($) a measure of benefits achievable based on the difference between a baseline condition without advanced control and a stretch goal.
ChevronTexaco Corporation
300-85
December 2003
December 2003
300-86
ChevronTexaco Corporation
increased significantly, and the first stage had clearly become the limit to second
stage throughput.
All of the relevant loops from the first stage could have been added to the original
scope, but it became evident that first-stage feed control valve capacity was the
constraint. Thus, the increased DMC scope was limited to the feed valve. This ultimately worked, but two show-stopping problems had to be surmounted:
1.
Valve non-linearity
2.
The first problem was easily handled with a standard valve transformation. See
Figure 300-72. A linearized first-stage flow control valve position was used as a
constraint in DMC. The valve had to be linearized because it normally operated in
the range of 90 -99% open and the flow/valve characteristics were nonlinear. An
abnormally high pressure drop in the first stage reactor due to plugging contributed
to the valve operating in the 90 -99% region.
Fig. 300-72 Trend Valve Transformation
'Flowrate
Linearized VP
800
100
(Valve Gain is Linear with Flow)
357
95
2 Hrs
By linearizing the valve response, it was possible to develop a better model between
the flow controller setpoint and valve position. As a result, DMC performed better.
Linearizing this valve position allowed DMC to push more flow through the unit.
The second problem required some creative re-structuring of the regulatory loops
controlling the levels in the plant. The original regulatory control scheme is shown
in Figure 300-73.
ChevronTexaco Corporation
300-87
December 2003
LC
1
To & From
Furnaces and
Reactor
Feed
Decoupler
358
LC
2
FC
Feed from
Storage
To 2nd-Stage
Flow Control
Note that the level controllers were configured to manipulate in the opposite direction as the flow. Note also that the catalytic reformer first stage fed the second-stage
directly, with no intermediate tankage normally used. Thus, second-stage feed rate
set the throughput requirement for the first stage.
In the original scheme, DMC manipulated second-stage feed, which drew down the
levels in the first stage. The first-stage levels then manipulated first-stage feed to
match the second-stage feed demand. The problem was that the two first stage
levels were often out of phase, which caused the first-stage feed valve to go through
unnecessarily large and unpredictable swings.
In the revised control scheme, a calculated variable, 1st-stage total volume, was
included as a DMC controlled variable (Figure 300-74).
Fig. 300-74 Post-DMC First Stage Level Control Configuration
PC
1
From Furnaces
and Reactor
C-1
359
LC
1
Post-DMC
Loop "Open"
Feed
Decoupler
To Furnaces
and Reactor
f(VP) CV
Post-DMC
Loop Remains
"Closed"
CV
1STVOL
V-2
DMC
FC
Feed from
Storage
December 2003
LC
2
To 2nd-Stage
Flow Control
300-88
ChevronTexaco Corporation
This variable was representative of first-stage inventory. The two main inputs to the
calculation were the volumes in C-1 and V-2. Because V-2 was a horizontal cylindrical vessel, the non-linear relationship between level and volume had to be considered. We also found it necessary to correct for the variable mass in the vapor space
upstream of C-1. We didnt anticipate the need for this variable and only came to it
after several iterations.
With the revised configuration, the upstream level controller was opened up and
DMC adjusted first-stage feed rate to keep 1st-stage total volume within range.
Figure 300-75 displays the two individual levels and 1st-stage total volume, and
clearly shows how this variable stabilized the plant response.
Fig. 300-75 Trend First Stage Volume Balance
20 MIN
HM HM HM
100
LC1.PV
50
25
0
360
75
LC2.PV
20
18
16
14
12
10
In this plot, the two levels are swinging and out of phase, but 1st-stage total
volume is essentially constant, so DMC does not have to make any unnecessary
moves to first-stage feed rate. DMC could not have been commissioned without this
variable!
Plant Test Case Study: FCC DMC Project. The importance of the underlying
regulatory control loops on DMC performance was dramatically illustrated in one of
ChevronTexacos FCC DMC projects. The plant was a Model IV (pressure
balance) FCC, in which a reactor-regenerator differential pressure controller (PDC)
manipulated the regenerator vent gas slide valves to adjust catalyst circulation.
Performance of the PDC was marginal, tuning it was difficult, and the PID
controller settings were slow.
DMC was initially tested and commissioned with the differential pressure controller
in AUTO. However, performance was not acceptable. DMC can only perform as
well as the underlying control system.
One of the keys to the ultimate success of this application was the decision to
configure DMC to manipulate the differential pressure controller output (PDC in
MAN) rather than its setpoint (PDC in AUTO). The plant was re-tested with the
new configuration and the controller commissioned. Two significant benefits were
achieved:
1.
The process responded much more quickly to changes in the manipulated variables (the slowest responses settled out in 60, rather than 90 minutes)
ChevronTexaco Corporation
300-89
December 2003
2.
The process responses were simpler (inverse response behavior was eliminated
in most cases).
The dynamic response curves for the two cases are compared below
(Figure 300-76).
Fig. 300-76 Dynamic Response Curves
The time base for the plots is 90 minutes. The shorter curves are for the case when
DMC manipulates the PDCs output; the longer curves, the PDCs setpoint. Each
curve represents the dynamic response of a given control or constraint variable to a
unit change in the manipulated variable, with all other manipulated variables held
constant.
The availability of detailed process response data like this gives us a unique look at
the behavior of a plant and can lead to greater process understanding.
December 2003
300-90
ChevronTexaco Corporation
changing feedstock quality and quantity, changes in product slate and/or demand,
variations in utility costs, etc.
When interfacing an online optimizer to DMC, for example, the optimizer replaces
the steady-state targets that the embedded linear program would otherwise calculate. Specifying the target on the CVs gives DMC the freedom to move the manipulated variables when a significant disturbance enters the system and shifts the level
of operation. Constraining the manipulated variables in the controller would take
away its degrees of freedom, which would reduce the probability the controller
could handle a significant disturbance.
2.
ROMeo Invensys/SimSci
3.
4.
The optimizer applications are built from fundamental first principles models of
individual unit operations. Standard off-line modeling packages (e.g., Aspen Plus
and Simulation Sciences Pro-II) are used for model building.
Each vendor offers a similar approach/functionality. However, ease of use and maintainability features vary from package to package.
To get an idea of what is involved in setting up and running a real-time optimizer,
the Invensys/Simulation Sciences package ROMeo will now be described. A
graphical user interface (GUI) is used to construct the model, configure data links,
preprocess data, generate reports and diagnose problem. ChevronTexaco refineries
have used the Simulation Sciences ProII/ProVision steady-state simulation package
ChevronTexaco Corporation
300-91
December 2003
for process design and off-line simulation work, and the ROMeo GUI is very
similar. ROMeo is an NT based application.
The model is constructed by dropping objects (typical unit operations) onto a flowsheet. When constructing the model, degrees of freedom are automatically
checked to ensure the problem is always kept square.
Measurements are easily added to the model. The connection to the real time historian is configured once, and from then on, all the variables are available to ROMeo.
A gross error detection methodology is used for data reconciliation. Measurements
are given weighting factors for the reconciliation objective function based on the
typical standard deviation for a particular measurement device.
Custom equations can be easily added to the flowsheet using a macro language
called TCL. Links to databases also can be made using TCL. Third-party models (e.
g., reactor kinetic models) can be incorporated via an object-oriented language
called Milano. Custom physical properties also can be added to ROMeo. Physical
properties such as crude assay information form the CAL database can be imported.
The optimization objective function is constructed using a GUI. Any variable can be
added to the objective function. Optimization variables are selected in the same
GUI. Pricing information is imported from Microsoft Access before each optimization run. Tier pricing can be easily incorporated into the objective function.
The optimization problem is solved using successive quadratic programming (SQP)
algorithm. Typical process models are described by thousands of equations and variables. A small number of these are optimization variables. The successive quadratic
programming (SQP) technique is used for manipulating and solving these large
sparse problems.
The sequence of operation for the online optimization problem is also setup using a
GUI. The engineer can graphically construct the online sequence, which can include
such things as solution failure routines; e-mail messages, report generation, etc.
Offline case studies are run with the same model in the same GUI environment.
December 2003
300-92
ChevronTexaco Corporation
380 Resources
Centralized support for process control technology throughout ChevronTexacos
operating companies is provided by the Process Control Team (Process Automation
Unit) of the ChevronTexaco Energy Research and Technology Company in Richmond, CA.
ChevronTexaco Corporation
300-93
December 2003
Perform Control Objectives Analysis (COA) gives a sound basis for control
design and generates a list of opportunities to improve process control
Develop Specifications
Manage Projects
Inferential Calculations
Some process properties are difficult or costly to measure on-line at the frequency
required, especially for advanced control. The solution is model-based algorithms,
or soft-sensors, which infer process properties from other measurements that are
readily available. These soft sensors use a variety of techniques, including Partial
Least Square and Neural Networks. Examples of Modeled Properties:
December 2003
300-94
ChevronTexaco Corporation
Training
CTU Process Control: One and one half day course at ChevronTexaco Technical University designed for new engineers.
Basic Control: a five-day course for the process control professional; it can be
given at customer site.
ChevronTexaco Corporation
300-95
December 2003
Implement full-scope APC on the simulator prior to start-up of the plant. Then,
after startup, implement full-scope APC on the actual plant.
2.
The first approach is more costly, as much (approximately 60-70%) of the work of
implementing APC on the simulator would have to be repeated on the actual plant.
On the other hand, the full-scope APC would be available for operator training prior
to startup. Also, you develop more confidence in the final controller design, and
though practice, APC implementation in the actual plant should be smoother. The
second approach would be less costly overall and would be adequate to support the
control engineer mentoring/training objective.
390 References
Chevrons Control Engineering Manual, 2nd edition, 1994 (Available from ChevronTexaco ERTC Process Automation Unit)
McMillan, G.K., Tuning and Control Loop Performance, 2nd ed., Instrument
Society of America, 1990 (3rd ed., 1994)
Marlin, T.E., Process Control: Designing Processes and Control Systems for
Dynamic Performance, 2nd ed., McGraw-Hill, 2000
December 2003
300-96
ChevronTexaco Corporation
Smith, C.A., and A.B. Corripio, Principles and Practice of Automatic Control, 2nd
ed., John Wiley, 1997
Luyben, W.L. and M.L. Luyben, Essentials of Process Control, McGraw Hill, 1997
Astrom, K.J., and T. Hagglund, PID Controllers: Theory, Design, and Tuning, 2nd
ed., Instrument Society of America, 1995
Shinskey, F.G., Process Control Systems - Application, Design and Tuning, 4th ed.,
McGraw-Hill, 1996
Shinskey, F.G., Feedback Controllers for the Process Industries, McGraw-Hill, 1994
Ogunnaike, B.A. and W.H. Ray, Process Dynamics, Modeling, and Control, Oxford
University Press, 1994
Seborg, D.E., T.F. Edgar and D.A. Mellichamp, Process Dynamics and Control,
John Wiley, 1989 (second edition to be published in 2003)
Chin, T.G., Guide to Distillation Pressure Control Methods, Hydrocarbon
Processing, October 1979. pp. 145-153.
Buckley, P.S., Techniques of Process Control, John Wiley, 1964, Chapter 18.
Chen, D. and D.E. Seborg, PI/PID Controller Design Based on Direct Synthesis and
Disturbance Rejection, Ind. Eng. Chem. Res., Vol. 41, No. 19, 2002, pp 4807-4822.
Chien, I.-L. and P.S. Fruehauf, Consider IMC Tuning to Improve Controller Performance, Chem. Eng. Progress, Vol. 86 (10), Oct. 1990, pp 33-41
Rivera, D.E., M. Morari, and S. Skogestad, Internal Model Control. 4. PID
Controller Design, Ind. Eng. Chem. Process Des. Dev., Vol. 25, No. 1, 1986, pp
252-265
Cutler, C. R and B. L. Ramaker, Dynamic Matrix Control - A Computer Control
Algorithm, AIChE 86th National Meeting, Houston, TX, April 1979
Cutler, C. R., A. M. Morshedi, and J.J. Haydel, An Industrial Perspective on
Advanced Control, AIChE National Meeting, Washington, DC, October, 1983
Park, S. (Pembroke Cracking Company), An Application of an Optimized DMC
Multivariable Controller to the PCC Catalytic Cracking Unit, International Symposium, Advanced Process Supervision and Real-Time Knowledge Based Control,
Univ. of Newcastle Upon Tyne, UK, November, 1988
Cutler, C. R. and R. B. Hawkins, Constrained Mulitvariable Control of a Hydrocracker Reactor, American Control Conference, Minneapolis, MN, June 1987
Marchetti, J. L., D. A. Mellichamp and D. E. Seborg, Predictive Control Based on
Discrete Convolution Models, I&EC Process Design and Development, Vol. 22,
1983
Richalet, J., A. Rault, J. L. Testud and J. Papon, Model Predictive Hueristic Control:
Applications to Industrial Processes, Automatica, Vol 14, 1978, pp. 413-428
ChevronTexaco Corporation
300-97
December 2003
December 2003
300-98
ChevronTexaco Corporation