Professional Documents
Culture Documents
PredictBack White Paper
PredictBack White Paper
PredictBack White Paper
Contents:
1. Introduction
2. Concepts
3. Implementation Guidelines
4. Model Identification
5. Model Building
6. Examples
INTRODUCTION
This purpose of this document is to provide a detailed description on how to use two of
the more advanced features of RMPCT, Predict-back and CV Measurement
Compensation. The discussion is mainly focused on practical issues relating to model
identification, model building and implementation. General recommendations are
shown in Italics wherever applicable.
CONCEPTS
Predict-back: is an optional RMPCT feature that provides additional feed forward
information and disturbance rejection to Controlled Variables (CV) when the process
value (PV) of a Manipulated Variable (MV) or a Disturbance Variable (DV) deviates
from its setpoint (SP). Predict-back is particular useful when the disturbance rejection
of the MV or the DV is slower than the effects of the unrejected (leak-through)
disturbances on the CVs.
When an MV is the SP of DCS-based PID controllers, it is usually good practice to keep
the PID cascade closed if the control loop performs well, and use their SPs as MVs
within a RMPCT controller design. Keeping higher level cascades closed can help
linearize the process, improve disturbance rejection and thus improve RMPCT
controller performance across different operating conditions.
Examples include heater outlet temperature, tower overhead temperature, accumulator
pressure, etc.
The dynamic responses of higher level PID cascades are generally slower than the
basic regulatory controllers (i.e., flow controllers), thus disturbances leak through these
Manipulated Variables and can take longer to reject. In a traditional RMPCT controller
design where only the SP of the MV is considered, process disturbances that result in a
deviation of the MV.PV from its SP are not detected until the CVs are affected. When
the CVs do respond and violate their limits, RMPCT will move the MVs accordingly,
even though the MV.PV may have already returned to its SP.
Example:
Consider the heater outlet temperature illustrated in Figure 1 as a MV and the reactor
outlet temperature as a CV.
DV
FIC
Feed
TIC
MV
FIC
TI
CV
Fuel Gas
Assume the heater is often subject to load disturbances, such as a sudden reduction in
feed flow rate, an increase in fuel gas BTU content or a change in heater inlet
temperature caused by changes in the preheat exchanger train. When these
disturbances occur, the heater outlet temperature PV starts to deviate from its SP then
returns to the SP due to the PID feedback control action as illustrated in Figure 2.
Some time later, the reactor outlet temperature will respond to these changes and
follow a similar path due to the leak through effect of the disturbances. With the
RMPCT controller off, the CV disturbance is effectively rejected in a single open loop
settling time because the MV PV eventually returns to its SP.
CV response to MV deviation
SP open loop
CV.PV
0 1 2 3 4 5 6 7 8 9 10 11 12
MV
disturbance MV return
to SP
MV.PV
MV.SP 0 1 2 3 4 5 6 7 8 9 10 11 12
Figure 3 illustrates the RMPCT controller response to the heater disturbance using a
traditional RMPCT design without Predict-back. In closed loop the heater disturbance
isn’t noticed until the reactor outlet temperature starts to respond, because without
Predict-back the controller does not utilize the MV PV to detect disturbances. By the
time the CV responds to the disturbance, the heater outlet temperature may already
have recovered from the disturbance, so any additional control moves to restore the CV
to within its funnel will be detrimental to the control of the reactor outlet temperature.
Notice how the MV cycles due the lack of information in the controller that accounts for
the MV returning to its SP.
CV response to MV deviation
SP with RMPCT control
No Predict back
CV.PV
0 1 2 3 4 5 6 7 8 9 10 11 12
MV.PV
return to SP
0 1 2 3 4 5 6 7 8 9 10 11 12
Figure 4 illustrates the RMPCT controller response to the disturbance when predict-
back is used. Control of the CV can be improved if we utilize the intermediate feedback
information of the MV.PV. When Predict-back is configured for the heater outlet
temperature MV, a prediction of the MV.PV is maintained based on the MV.SP and
possibly other inputs. When the MV.PV deviates from its predicted value due to a
disturbance, the future predictions of the MV are corrected to account for the effect of
the disturbance and then “predicted back” to the MV.SP, accounting for the PID
1
controllers PID nominal response . In addition CV predictions are augmented with the
MVs deviation and future response using a model of the CV from the MV.PV. The
RMPCT controller response is a set of small, offsetting control moves of the MV.SP to
minimize the predicted deviation of the reactor outlet temperature. The end result is a
better performing controller with less MV movement and less violation of the CV limits.
Note; Predict-back does not include state estimation of disturbances that affect the MV
PV, for causality reasons.
1
When MVs are wound up, the MV.PV is not predicted back to SP and the entire MV prediction error is applied to
the CV predictions.
New CV
CV response to MV deviation from
predictions
SP with RMPCT control action,
with Predict Back
Old CV
predictions
0 1 2 3 4 5 6 7 8 9 10 11 12
MV.PV
disturbance Prediction of MV.PV
MV.PV deviation
return to SP
to be mapped to
CV predictions
Old prediction
of MV.PV
0 1 2 3 4 5 6 7 8 9 10 11 12
Future MV.SP
control moves
(CV LIMITS)
+ +
(MV SP) + (RAW CV)
PID Process
RMPCT Process
Controller G(s) +
+
-
Compensated CV prediction
8
(MV PV)
Figure 5
(MV PV)
Main RMPCT Model CV Compensation (CC) Model
White Paper
+
+ (Model error bias update) -
+
RMPCT with Predict Back and CV Measurement Compensation White Paper
Figure 6 illustrates the closed loop control performance for a type I scenario with a
controller configured with CV Measurement Compensation. A CV target change was
introduced to illustrate the control performance.
Compensated CV Prediction
CV Process Predictions
MV.SP
MV.PV
Figure-6
Hydrogen
Quench
FIC
TIC
TIC MV
DV
TI
FIC
CV
Fuel Gas
• MV response to SP changes
Figure-7
TIC PV
IMPLEMENTATION GUIDELINES
Predict-back: The decision to use Predict-back is typically made at the design stage.
Predict-back is normally recommended when:
• The MVs or DVs experience significant disturbances from measurable or non-
measurable sources and the disturbance rejection is slower than the effects of the
unrejected (leak-through) disturbances on the RMPCT controlled variables.
• The MV PID settling time is large relative (cascade controllers) to the RMPCT
execution period.
• The CV settling times (including deadtime) are large (coupled with 1 and/or 2)
• If tight control of CVs is required (reactor temperatures).
When the dynamics of the MV’s PID controller are fast relative to the controller
execution period, then disturbances to the MV are usually attenuated fast enough to
have little impact on the RMPCT CVs. Flow controllers are an example of “fast” MVs.
For example, it is usually not recommended to break a flow controller cascade and
adjust the valve position as an RMPCT MV.
The following is a partial list of MVs where using Predict-back can or may improve
controller performance:
1. Feed heater coil outlet temperature
2. Distillation Column top temperature
3. Reboiler return temperature
4. FCC feed preheat temperature
5. FCC riser outlet temperature
6. FCC regenerator air flow
7. FCC Gas Con strip-off flow
8. Hydrocracker reactor bed inlet temperature
9. Hydrocracker reactor WABT (for conversion control)
MODEL IDENTIFICATION
Models used for Predict-back are broken down into three categories:
• Predict-back Models
• Estimated Disturbance Models
• Main RMPCT Models
For simplicity, first order transfer functions have been used in the following illustrations.
Higher order models including lead terms are possible provided that consistency among
the model types is preserved.
Predict-back Models
These models are used to predict the MV PVs or DV PVs. These are simply the
models of the MV.PV response to its’ own SP change, as shown in equation 1. Notice
that the process gain is exactly 1 by nature, because the PV always returns to its SP,
unless the controller is wound up.
−td 1
s
dMV.PV 1e
=
dMV.SP T1s + 1 [1]
The Predict-back models may also include other inputs that affect the MV.PV. Normally
these additional disturbance models have a zero gain in steady state and are
sometimes referred to as “hump” models or just zero-gain models. Equation 2 is an
example of a “hump” model used in this context. The purpose of these models is to
predict the effect of external disturbances to the MV PV so that the controller can
anticipate the effect and take action even before the disturbances hit the MV.
−t d 2
dMV.PV K p 2 s ⋅ e
s
=
dX T2 s +1 [2]
= ⋅ = ⋅ = p2 p3 p2 p3
dX dX dMV.PV T s +1 T s +1 (T s + 1)(T s + 1)
2 3 2 3
[5]
Where X = other MV or DV disturbance input
Ks + 1
G(S ) = k
τs + 1 [6]
Ks
G ( s) = k
τs + 1 [7]
The main difference between the two representations is that only [7] is a true zero-gain
model while [6] has a non zero steady state gain, where the value of the offset depends
on the values of the parameters. Consider the following transfer function [8]:
100 s + 1
G ( S ) = 0 .1
(10 s + 1)(5 s + 1) [8]
The limit value of [8] using the final value theorem is 0.1.
100 s
G ( S ) = 0 .1
(10 s + 1)(5 s + 1) [9]
The limit value of [9] using the final value theorem is exactly 0.
This example is illustrated below, with transient response curves to a step input. In this
case both models are adequate representations of the response of a manipulated
variable process value to a step input in a DV. The general recommendation is to use a
true zero gain model to represent the transient response in the MV PV to avoid conflicts
with the steady state optimizer. The model parameters can manually be adjusted via
the model summary page to force a zero steady state offset. Note, that if the results
from the FIR model identification indicate a steady state gain that is significantly
different than zero, forcing the gain to zero may cause a major dynamic mismatch
between the original FIR model and the corrected model. In this case the user should
revisit the results from the model identification and if necessary go back collect
additional process data for further analysis. Note that if a MV is in a wind-up state then
the predictions of MV PV will not return to its SP. Any “hump” model between a DV and
a CV will not reflect this and a model mismatch is introduced. In addition, the anti-
windup feature for the MVs is permanently disabled when using the Predict-back
feature for the MV.
Non zero gain model GNZ(s) with 0.1 steady state offset:
100 s 1
0.1 invlaplace , s 1.8. exp( .1. t ) 1.9. exp( .2.t ) .1
( 10. s 1 ) ( 5 s 1 ) s
GNZ( t ) 1.8.exp( .1. t ) 1.9. exp( .2. t ) .1
Zero gain model GZ(s) with no steady state offset:
100 s
0.1 invlaplace , s 2.. exp( .2.t ) 2.. exp( .1.t )
( 10. s 1 ) ( 5 s 1 ) s
GZ( t ) 2.. exp( .2. t ) 2..exp( .1. t )
1
GNZ( t )
0.5
GZ( t )
0
0 20 40 60 80 100
t
The two curves represents the transient response to a step input.
The solid represents the response of the non zero gain model.
The dotted curve represent the response of the zero gain model.
When “hump” models for the relevant MVs have been identified this way, the
corresponding “hump” models for the CVs can be calculated using a convoluted model.
If we select [9] as the model between the DV and the MV then assume the following
estimated disturbance model between the MV PV and the CV:
1
G(S ) = 2 [10]
5s + 1
The convoluted disturbance model for the CV is:
1 100 s
G(S ) = 2 * 0 .1 [11]
5s + 1 (10 s + 1)(5 s + 1)
It is assumed that the disturbance model [11] has been identified prior to this. It is
recommended to validate that the convoluted model predicts the CV response well.
Models used for CV Measurement Compensation are broken down into two categories:
MODEL BUILDING
Predict-back: APCDE is used to configure a controller with Predict-back. The Regular
RMPCT model file (i.e., *.mdl) is still the main file used to develop the *.cfg, *.xm, *.xs,
etc. files to be used in the on-line controller. Figure 8 illustrates the “Build Controller”
Options that need to be specified for Predict-back.
Figure 9: RMPCT “MV Control” Display with Predict-back Option for MVs 3 and 4
The three model files discussed in the previous section are mandatory in order to
configure Predict -back. Specific dimensions are required for the different model files
based on the number of MVs, DVs and CVs. Table 1 illustrates the dimension and
input/output requirements for each model file.
2
nMV = # of MVs; nDV = # of DVs; nCV = # of CVs;
Identification of Models:
When Predict-back is used, it can be applied to any subset of MVs and CVs within the
control problem. For instance, in the 3 MV by 5 CV and 1 DV controller illustrated
below, you can choose to only apply Predict-back to MV #2 and CVs # 1,3 and 4, even
though there are 5 CV models for that MV in the Main model file.
Main Models3
3
“√” indicates that a model exists for this CV/MV pair
4
The final models need to be saved in each case as the “Build Controller” step reads the file versions from disk.
Predict-back Tuning:
Predict-back feedforward corrections can be turned off, by setting the Predict-back ratio
parameter for the MV to zero. Normally a Predict-back ratio = 1 is recommended for full
Predict-back feedforward corrections. A fractional value less than 1 is recommended to
de-tune the feedforward corrections to the CV. Note; a value different than 0 or 1 will
introduce mismatch in the CV predictions.
Essentially the disturbance model gain is multiplied with the Predict-back ratio to
attenuate the feedforward prediction corrections to the CV when the MV PV deviates
from its SP.
Note; the Predict-back ratio is not a gain-scheduling tool, instead the online gain delay
factors should be used to make corrections to the gains, or alternatively recreation of
the XM and XS files.
It is recommended to lower the Predict-back ratio if MV control moves are deemed to
vigorous in the presence of disturbances, and larger transient errors in the CVs can be
tolerated. However the tuning of this factor is mostly ad hoc based on the observed MV
and CV response and normally a value of 1 is recommended.
5
nMV = # of MVs; nDV = # of DVs; nCV = # of CVs;
EXAMPLES
Predict-back:
Example 1: Simple 1 x 1 controller
Figures E1-1 and E1-2 are APCDE screen dumps that illustrate the main (main.mdl),
predict-back (pb.mdl) and estimated disturbance (ed.mdl) models and variable
configurations.
Hydrogen
Quench
FIC
TIC
TIC MV
DV
TI
FIC
CV
Fuel Gas
The objective of the Profit Controller application presented in this example is to provide
control of bed 2 outlet temperature. As in the previous example, any deviation of a
bed’s inlet temperature PV from its SP will propagate through to its respective outlet
temperature. Since the beds are in series, changes in bed 1 outlet temperature PV
result in a load disturbance on bed 2 inlet temperature and thus also affect bed 2 outlet
temperature. Figure E2-2 illustrates the dynamic response of the bed 1 inlet and 2
inlet/outlet temperatures when bed 1 inlet temperature setpoint is changed:
back to its SP on PID control. Tout2PV then follows Tin2PV subject to the process
dead-time and lag.
As stated earlier, hydrocracker reactions can be highly exothermic. If the inlet
o
temperature control cannot quickly reject disturbances, a 1 F disturbance in the bed
inlet temperature can propagate through the reactor beds resulting in a +10oF
disturbance in the reactor outlet temperature. Coupling this disturbance with
feed/effluent heat exchange (most designs have this) returns a disturbance back to the
bed 1 inlet temperature which can lead to unstable reactor control in the best case or a
run away reaction and catastrophic metallurgical failure in the worst case. Anything that
can be done to attenuate temperature disturbances within a reactor system improves
the control stability.
From Figure E2-2 we can see that there is feed forward potential in knowing that
ThtrPV has changed and that it will affect Tin2PV and subsequently Tout2PV. With this
feed forward information, Tin2SP could be changed to counteract the disturbance and
minimize the effect on Tout2PV. This feed forward effect is configured into RMPCT by:
• Adding the effect of ThtrPV on Tout2PV to the main model file (hcr_main.mdl). This
is a “hump” model. This provides feed forward information to the CV predictions.
Provided that the performance ratio of Tout2PV is sufficiently aggressive, control
moves will be made to minimize this disturbance.
• Adding the effect of ThtrPV on Tin2PV to the Predict-back model file (pb.mdl). This
is a “hump” model. Since the direct effect of the ThtrPV disturbance is already
mapped to Tout2PV as described in item 1, this disturbance needs to be mapped
onto the Tin2PV predictions to eliminate any prediction error in the MV (i.e., to avoid
any double dipping). Should Tin2PV not respond as predicted, its prediction errors
will be mapped onto the Tout2PV predictions using the typical Predict-back
mechanism.
• Notice, the predict back model between the heater SP (ThtrSP) and its PV (ThtrPV)
is not present. The reason for this is the fact that only the heater PV is used for
feed-forward predictions, which makes the need to use predicts back for the heater
DV obsolete since the heater SP is not used for predictions. In most cases of this
nature, the heater temperature would be a MV and predict back would have to be
included for this variable.
The following figures illustrate the Predict-back model configurations used for this
example:
CV Measurement Compensation:
Example 3: Simple 2 x 1 controller
Figures E3-1 and E3-2 are APCDE screen dumps that illustrate the main (main.mdl),
and CV Measurement Compensation (cc.mdl) models and variable configurations.