Model Predictive Control Demonstration

Model Identification Testing and Simulation Results

1015711

Comparison of Actual and Predicted Steam Temperature

15

10

5
Temperature

-5

-10

-15

-20
0 500 1000 1500 2000 2500 3000 3500 4000 4500
Time (seconds)

0
0
 Model Predictive Control Demonstration
Model Identification and Simulation Results

1015711

Technical Update, March 2009

EPRI Project Manager

M. DeCoster

ELECTRIC POWER RESEARCH INSTITUTE

3420 Hillview Avenue, Palo Alto, California 94304-1338 ▪ PO Box 10412, Palo Alto, California 94303-0813 ▪ USA
0 800.313.3774 ▪ 650.855.2121 ▪ [email protected] ▪ www.epri.com
DISCLAIMER OF WARRANTIES AND LIMITATION OF LIABILITIES
THIS DOCUMENT WAS PREPARED BY THE ORGANIZATION(S) NAMED BELOW AS AN ACCOUNT OF
WORK SPONSORED OR COSPONSORED BY THE ELECTRIC POWER RESEARCH INSTITUTE, INC. (EPRI).
NEITHER EPRI, ANY MEMBER OF EPRI, ANY COSPONSOR, THE ORGANIZATION(S) BELOW, NOR ANY
PERSON ACTING ON BEHALF OF ANY OF THEM:

(A) MAKES ANY WARRANTY OR REPRESENTATION WHATSOEVER, EXPRESS OR IMPLIED, (I) WITH
RESPECT TO THE USE OF ANY INFORMATION, APPARATUS, METHOD, PROCESS, OR SIMILAR ITEM
DISCLOSED IN THIS DOCUMENT, INCLUDING MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE, OR (II) THAT SUCH USE DOES NOT INFRINGE ON OR INTERFERE WITH PRIVATELY OWNED
RIGHTS, INCLUDING ANY PARTY'S INTELLECTUAL PROPERTY, OR (III) THAT THIS DOCUMENT IS
SUITABLE TO ANY PARTICULAR USER'S CIRCUMSTANCE; OR

(B) ASSUMES RESPONSIBILITY FOR ANY DAMAGES OR OTHER LIABILITY WHATSOEVER (INCLUDING
ANY CONSEQUENTIAL DAMAGES, EVEN IF EPRI OR ANY EPRI REPRESENTATIVE HAS BEEN ADVISED
OF THE POSSIBILITY OF SUCH DAMAGES) RESULTING FROM YOUR SELECTION OR USE OF THIS
DOCUMENT OR ANY INFORMATION, APPARATUS, METHOD, PROCESS, OR SIMILAR ITEM DISCLOSED IN
THIS DOCUMENT.

ORGANIZATION(S) THAT PREPARED THIS DOCUMENT

Kurz Technical Services, Inc.

This is an EPRI Technical Update report. A Technical Update report is intended as an informal report of
continuing research, a meeting, or a topical study. It is not a final EPRI technical report.

NOTE
For further information about EPRI, call the EPRI Customer Assistance Center at 800.313.3774 or
e-mail [email protected].

Electric Power Research Institute, EPRI, and TOGETHER…SHAPING THE FUTURE OF ELECTRICITY
are registered service marks of the Electric Power Research Institute, Inc.

Copyright © 2009 Electric Power Research Institute, Inc. All rights reserved.

0
 CITATIONS
This report was prepared by
Kurz Technical Services, Inc.
540 Rock View Drive
Spring City, TN 37381
Principal Investigator
C. Taft
This report describes research sponsored by the Electric Power Research Institute (EPRI).
This report is a corporate document that should be cited in the literature in the following manner:
Model Predictive Control Demonstration: Model Identification and Simulation Results. EPRI,
Palo Alto, CA: 2009. 1015711.

0 iii
0
 PRODUCT DESCRIPTION
This report describes the first phase of a two-phase project to demonstrate an advanced control
technique on a fossil power plant. The advanced control technique chosen for the project is
model predictive control (MPC). Although MPC has been a commercially successful technique
in many process industries, it has been used sparingly in the power industry even though
industry-specific products are available. Although the Electric Power Research Institute (EPRI)
has sponsored several advanced control demonstration projects in the past, this was the first
project to use the MPC algorithm. Also new in this project is the fact that the model predictive
controller is implemented in the distributed control system (DCS) control processor rather than
in a separate application processor.

Results and Findings

A simple steam temperature control application has been designed for the model predictive
controller and tested on a checkout simulator. The controller performance in the simulated
environment was satisfactory based on preliminary test results. The model predictive controller
also exhibited adequate robustness to process model changes in gain and time lag. One model
predictive controller function that has not been verified yet is the ability to assume control from
another controller with a smooth transition (bumpless transfer). Bumpless transfer is necessary
when transferring from proportional integral differential (PID) control to MPC.
Tests have been conducted at the host plant to identify a model of the superheat steam
temperature control process. Desuperheater spray control is the steam temperature control
method used at the host plant. Due to seasonal demand, lack of availability of the host unit has
delayed additional tests needed to identify disturbance models that are important to obtaining
optimal performance with the MPC technique. These additional tests will be completed in
subsequent project tasks, and the results will be documented in a final technical report.

Challenges and Objectives

The primary technical challenge in this project was to identify adequate models of the steam
temperature process. There was considerable process noise identified in the steam temperature
data, making determination of actual plant characteristics difficult. Once a good process model
was obtained, the control design process was relatively straightforward, as was the addition of
the new MPC algorithm to the existing temperature control logic.
The primary objective of this demonstration project was to determine whether advanced control
techniques such as MPC can improve control system performance compared with conventional
PID-based control. It is also important to determine whether the implementation of such
advanced systems is too complex or expensive for everyday power plant use.

Application, Value, and Use

Advanced control techniques have shown the ability to improve control system performance in
difficult power plant control loops. Their widespread use has been hampered by their cost and
complexity. MPC is a more user-friendly type of advanced control than other methods and as
such may be more readily accepted in the power industry. If this demonstration project is
successful, it will encourage more use of advanced control, which will improve plant dynamic

0 v
performance. Many power plants are still struggling with quantifying the value of improved
dynamic response, but in competitive markets, the value is quantifiable. Also, as plants become
more complex with extensive back-end emission control systems, it is important that the basic
power island be well controlled at all times.

EPRI Perspective
EPRI’s Generation Sector Instrumentation, Controls, and Automation for Improved Plant
Operations Program has the objective to identify and demonstrate new methods and technologies
to improve control system performance in fossil power plants. Although conventional PID-based
control strategies are adequate for most boiler control applications, there are some difficult loops
where PID does not perform well. The use of advanced control techniques has been studied by
EPRI before, but the MPC method has not. Because MPC products are commercially successful,
they are more fully developed than other advanced control systems. The particular system being
demonstrated in this project is attractive because it operates in the DCS control processor along
with the conventional PID control logic. This reduces cost and simplifies implementation, which
should lead to more widespread use of the technology.

Approach
A host site was selected for the demonstration by the host utility. One reason that the host site
was selected was related to current difficulty with the steam temperature control system. The
main tasks of the first phase of this project were to design and implement the model predictive
controller on a simulator for testing. Plant tests were performed to identify the process model
needed by MPC technology. The simulator was a small DCS similar to that of the plant so that
the test environment was as realistic as possible. In subsequent project tasks, after testing is
completed on the simulator, MPC technology will be implemented in the real plant and tested.
The existing PID control will also be available in the plant for comparison purposes. All results
will be documented in a final technical report.

Keywords
Control system
Instrumentation and control
Model predictive control (MPC)
Proportional integral differential (PID)
Steam temperature control

0 vi
 ACKNOWLEDGMENTS
This project could not have been accomplished without the help of John Sorge, a research
engineer from Southern Company. Additionally, EPRI recognizes the staff at the Gaston plant of
Alabama Power Company (Southern Company) for allowing this project to be demonstrated on-
site.

0 vii
0
 CONTENTS
1 INTRODUCTION ....................................................................................................................1-1
Demonstration Project Plan .................................................................................................1-1
Phase I: Controller Design, Implementation, and Testing on the Simulator...................1-1
Phase II: Implementation and Testing in Gaston Unit 3 .................................................1-3

2 MODEL IDENTIFICATION TESTING.....................................................................................2-1

Plant and Process Description .............................................................................................2-1
Model Identification Tests ....................................................................................................2-3
Model Identification Results .................................................................................................2-7

3 MPC IMPLEMENTATION IN THE SIMULATOR....................................................................3-1

Simulator Description ...........................................................................................................3-1
Installation of the Ovation Advance Process Control Toolkit ...............................................3-1
Configuration of the Control System and Simplified Process Model....................................3-1
Moving the Existing Steam Temperature Logic .............................................................3-1
MPC Configuration.........................................................................................................3-2
Simplified Process Model...............................................................................................3-3

4 SIMULATOR TEST RESULTS...............................................................................................4-1

Test Procedure.....................................................................................................................4-1
Simulator Test Results .........................................................................................................4-2
MPC Tuning Tests .........................................................................................................4-2
Disturbance Rejection Comparison................................................................................4-3
Robustness Tests ..........................................................................................................4-4
Summary of Simulator Tests................................................................................................4-5

5 CONCLUSION........................................................................................................................5-1

6 REFERENCES .......................................................................................................................6-1

A MODEL IDENTIFICATION TEST PLAN............................................................................... A-1

Gaston Unit 3 Dynamic Response Test Plan...................................................................... A-1
Background ................................................................................................................... A-1
Test Description ............................................................................................................ A-1
Test Prerequisites ......................................................................................................... A-1
Test Descriptions .......................................................................................................... A-2

0 ix
0
1
INTRODUCTION
The Electric Power Research Institute (EPRI) has been researching advanced control
technologies for many years [1,2]. When the research began, there were few, if any,
commercially available advanced control solutions for the power industry. As a result, much of
the early work was done with universities and involved evaluation of different control algorithms
and their suitability for power plant applications. In the past few years, several major control
system vendors have begun offering their own advanced control products usually based on the
model predictive control (MPC) methodology. In 2007, EPRI arranged web casts by several
commercial vendors of MPC technology to better understand the technology and its possible
application to the power industry. The web casts were very informative and showed that MPC
technology is very widely used in other industries but is still somewhat rare in the power
industry. A brief summary of the web casts is provided in EPRI TR-1014237 [3].
After the MPC technology review was completed, Southern Company agreed to host a
demonstration of MPC technology at one of their fossil power plants. The unit chosen for the
demonstration was Alabama Power Company’s Gaston plant unit 3, and the application was
steam temperature control. The host unit has an Emerson Ovation control system, and Southern
made arrangements with Emerson to use their Ovation Advanced Process Control Toolkit, which
includes a model predictive controller.
This report describes the overall plan for the demonstration project and discusses progress to date
with the model identification testing, model development, and simulator setup.

Demonstration Project Plan

The objective of the MPC demonstration project was to design and install an MPC system in a
power plant application and compare its performance and usability with those of a conventional
proportional integral differential (PID)-based control system. The project will be completed in
two phases, with phase 1 encompassing the control design, implementation, and testing on a
simulator and phase 2 involving implementation and testing in the actual plant.

Phase I: Controller Design, Implementation, and Testing on the Simulator

Task 1: Setup of the Ovation Simulator in the Southern Company DCS Laboratory
Southern Company set up a small Ovation DCS in their laboratory for use as a simulator on this
project. The setup included assembling the correct hardware, installing the proper Ovation
software, installing the Advanced Control Process Toolkit, and testing it all to ensure proper operation.

0 1-1
Task 2: Configure Conventional and Advanced Steam Temperature Control in the
Simulator
The existing Gaston unit 3 steam temperature logic was configured in the simulator along with
the connections required to interface with the new MPC blocks. Due to the different Ovation
versions at Gaston and the simulator and because only a subset of the Gaston logic will be
included on the simulator, the design was manually transferred to the Gaston system as opposed
to copying the simulator logic sheets.

Task 3: Develop Steam Temperature Process Model and Install in the Simulator
A dynamic model of the steam temperature process was developed and installed in the simulator.
The model was developed using system identification based on open loop and closed loop testing
performed in the plant. Another project on automated control system tuning was also being
conducted on Gaston unit 3 steam temperature control, and the intention was to reuse some of
the test data from that project to identify a model for this project. The best approach for installing
the model in the simulator was not determined ahead of time. One option was to implement it
using the standard Ovation function blocks such as lags and gains. If the model was found to be
relatively simple, this could be straightforward. Other options were also available, using the
Advanced Control Process Toolkit or a special simulation processor.

Task 4: Design Model Predictive Controller

The MPC design was completed using the Advanced Control Process Toolkit. The outputs of the
design process are the matrix coefficients required by the MPC function block. Engineering
judgment was required in the design process to properly weight the loop errors and control
actions to achieve the best overall performance. Real-world constraints were also factored into
the design.

Task 5: Test Model Predictive Controller on the Simulator According to the Test Plan
Once the model predictive controller was implemented in the simulator, it was tested using both
a formal test plan and ad hoc tests. A written test plan was prepared and approved by all parties.
The test plan incorporated normal operating scenarios as much as possible. The purpose of the
testing was to determine whether the model predictive controller provides better performance
than the conventional PID controller. Performance measures include speed of response, stability,
and robustness.

Task 6: Evaluate Test Results

The results of the simulator testing were evaluated by Southern Company to determine whether
to proceed with the implementation of the model predictive controller in Gaston unit 3.

Task 7: Prepare Progress Report

This EPRI progress report was prepared in the first phase of the project.

0 1-2
Phase II: Implementation and Testing in Gaston Unit 3

Task 8: Install Ovation Advanced Control Process Toolkit in Gaston Unit 3

The Ovation Advanced Control Process Toolkit will be installed in the Gaston Unit 3 Ovation
DCS. This may require a unit outage to do.

Task 9: Modify Steam Temperature Logic for Model Predictive Controller

The MPC function block will be installed in the DCS. The existing logic will be modified to
interface it with the new model predictive controller. The configuration of the model predictive
controller and the logic changes for the interface should be essentially identical to those tested on
the simulator in phase 1.

Task 10: Test Model Predictive Controller

A new test plan will be developed for the plant testing. It will include closed loop step tests and
load ramps if possible. Normal operation will also be included in the evaluation process.

Task 11: Prepare Final Report

A final project technical report will be prepared for publication.

0 1-3
0
2
MODEL IDENTIFICATION TESTING
A model predictive controller uses a process model in both its design and operation. There are
several possible methods for developing a model of the process being controlled, and for this
project, the model will be developed based on actual plant test data. The technique is called
model identification or system identification and is widely used in control system design and
analysis.
Identification consists of plant testing, usually in open loop mode (control system in manual
mode) but sometimes in closed loop mode. The input signal can be a simple step change or a
more complex signal such as a series of steps or a sine wave. It is used to excite the system so
that the natural characteristics of the system can be observed. Input and output data are collected
during the test, and the data are analyzed using one or more computerized methods. Commercial
software packages, such as Matlab, provide extensive tools for system identification. The
methods can be parametric or nonparametric. When using a parametric method, the user must
specify a model structure or possibly a family of structures. Each possible structure has one or
more parameters associated with it, and the identification software attempts to find the set of
parameters that allows the model to best fit the test data.

Plant and Process Description

Gaston unit 3 is a 250-megawatt (MW) unit that began operation in 1962. The boiler is a coal-
fired Babcock & Wilcox drum type with front and rear wall burners. The fuel is provided by six
ball and race pulverizers with gravimetric feeders. The turbine is a General Electric cross-
compound type.
Superheat temperature is controlled by left- and right-side desuperheating sprays, and each
desuperheater has two identical spray valves in parallel (Figure 2-1). The two valves are driven
by the same controller, but the second valve does not begin opening until the first valve is 25%
open. The boiler is equipped with two gas recirculation fans to control reheat temperature. The
unit also has reheat spray valves, but they are no longer used.

0 2-1
Figure 2-1
Process overview

The boiler control system is an Emerson Ovation system installed in 2007. The superheat steam
temperature control strategy is a conventional cascaded control system using desuperheater
outlet temperature for inner loop control and final steam temperature for outer loop control. No
feedforward control is used. The reheat temperature control is also conventional with a single
controller for both gas recirculation fans and inlet vanes. Normally the plant operates only one
fan at a time, and the inlet vane position has a high limit of 40%.
One somewhat unusual control strategy in the unit is the fuel control. The fuel is controlled by
changing the primary airflow to the pulverizers. This provides a quick response in the fuel to the
furnace by delivering some of the pulverized coal inventory to the burners. The coal feeders are
used to maintain the differential pressure across the pulverizer at a setpoint, which varies with
primary airflow. This type of fuel control strategy was originally used when two-speed table
feeders were supplied rather than variable-speed gravimetric feeders because, with table feeders,
there is no direct measurement of fuel flow rate available.

0 2-2
Model Identification Tests
Tests were performed in the plant to produce data for identifying process models to be used in
the MPC design. The identification tests consisted of a series of open loop step response tests of
the superheat temperature controller output and the reheat controller output. In addition, other
step tests on disturbance variables such as fuel flow, airflow, and steam flow were planned but
not done during the testing described here. The complete test plan is included in Appendix A.
Some modifications to the test plan were necessary during the tests to accommodate plant needs
and scheduling constraints. Additional tests will be performed at a later date. A brief summary of
the tests is provided here.
Tests were done on the superheat temperature control by placing the spray valve
manual/automatic station in manual mode and making step changes in the spray valve position
demand. There is one control loop for the left-side spray and one for the right side. Each side
has two spray valves in parallel (Figure 2-2). The controller output drives both spray valves
(Figure 2-3), one directly and one through a function generator. The function generator is set to
begin opening the second spray valve at 25% demand and have it fully open at 100% demand.
For the initial tests, the fuel and air control were kept in automatic mode, but the megawatt and
throttle pressure control were placed in manual mode. In later tests, the fuel flow (primary air
dampers) and airflow were also placed in manual mode. Tests were done on both the left- and
right-hand sides with the unit at two different loads, 100% and 68%. It was not possible to do
tests at minimum load because of the operating needs of the plant and time constraints. The step
sequence (Figure 2-4) for the tests at 100% load was two steps up followed by two steps down.
This allowed any hysteresis in the valve positions to be identified. The small spikes in the valve
position data are caused by small disturbances in the valve positioner. The test sequence at 68%
load was one step up and one step down because the hysteresis, if any, was already identified by
the first tests. Similar step tests were done on the reheat control by stepping the gas recirculation
fan damper demand down, up, and down again (Figure 2-5) at 68% load. It was not possible to
do these tests at 100% load because the gas recirculation fan is not needed at full load.

0 2-3
Figure 2-2
Section of superheat control schematic

0 2-4
 Spray Valve Demand Curves

110

100

90

80

70
Valve Demand (%)

60

50

40

30 Valve 1

20 Valve 2

10

-10
-10 0 10 20 30 40 50 60 70 80 90 100 110
Controller Output (%)

Figure 2-3
Superheat spray valve demand curves

Left Hand SH Spray Valve Tests

60 45

SHAT_LHV1_VLV_POS.UNIT3@OVATION
SHAT_LHV1_VLV_DMD.UNIT3@OVATION 40
55 SHAT_LHV2_VLV_POS.UNIT3@OVATION
SHAT_LHV2_VLV_DMD.UNIT3@OVATION

Valve 2 Dmd and Pos (%)

Valve 1 Dmd & Pos (%)

35
50

30

45
25

40
20

35 15
1:15:00 1:30:00 1:45:00 2:00:00 2:15:00 2:30:00 2:45:00 3:00:00
Time

Figure 2-4
Left-hand spray valve test steps

0 2-5
 Gas Recirulation Fan Damper

25

20

A_GRF_INL_DMP_POS.UNIT3@OVATION
Damper Position (%)

15

10

0
9:00:00 9:15:00 9:30:00 9:45:00 10:00:00 10:15:00 10:30:00
Time

Figure 2-5
Gas recirculation fan step tests

Data from the test were collected in three different ways. An Ovation trend was configured with
approximately 50 data points of interest. The trend was configured with a 1-second update rate
and was saved as a tabular trend every 10 minutes. These tabular trend files are just text files in a
space-delimited format and can easily be concatenated and loaded into Excel. The second
collection method was through an overspeed protection controller (OPC) connection between the
Ovation system and a notebook computer. OPC server software was already installed in the
Ovation system and the Matlab OPC toolbox software was installed on the notebook. A small
Matlab program was written to collect the same 50 points as on the trend also at a 1-second
update rate. The last collection method was the plant archive system. The scan rate of most of the
points of interest was changed to 1 second for some earlier tuning work on this loop. However,
some of the points had larger deadbands than ideal for this type of testing, so these data were
primarily used as a backup. Using three data collection methods was more than really necessary
but did ensure that no test data were lost.
In addition to the open loop step response tests, several other closed loop response tests were
also performed for other purposes. These additional test results may also be useful for
identification and validation of process models.
In the second round of testing, several tests are planned to identify disturbance functions for
other process variables such as fuel flow, airflow, gas recirculation fan, and turbine load.

0 2-6
Model Identification Results
Data from one of the spray valve step tests are shown in Figure 2-6. It is apparent that the steam
temperatures are not very steady during the tests. After the first upward step, the final
temperature actually goes up instead of down. The second step up and the first step down
produce reasonable results, although at the end of the first step down, the final temperature is
dropping instead of remaining steady. This unexpected behavior indicates that other disturbances
are influencing the steam temperature. These disturbances persisted throughout the testing, and
although attempts were made to identify the source of the upsets, it was never determined. The
upsets also made the identification process more difficult.

Left Hand Superheat Response

1025 70

975 60
Final and Desuperheater Outlet

Valve Position Demand (%)

925 50
Temperatures (degF)

875 40

LH_SH_OUT_T.UNIT3@OVATION
LH_SHAT_OUT_T.UNIT3@OVATION
825 SHAT_LHV1_VLV_DMD.UNIT3@OVATION 30

775 20

725 10
1:20:00 1:35:00 1:50:00 2:05:00 2:20:00 2:35:00 2:50:00 3:05:00 3:20:00
Time

Figure 2-6
Final steam temperature and desuperheater outlet temperature responses to spray valve step
changes

During the first round of testing, only the spray valve step response tests were performed. The
data were first imported into Matlab and then into the Matlab System Identification Toolbox
graphical user interface for system identification (Ident). The Ident graphical user interface
provides a convenient interface for most of the Matlab system identification functions. The user
may preprocess the data to remove means, designate one subset of the data for identification and
another subset for validation, and select the model structure from several different types of
models and other details. For this initial modeling effort, process model type was selected. The
type selected was a first- or second-order plus deadtime model. An integrating type of model can
also be specified. Both first- and second-order models were developed.

0 2-7
A comparison of the superheat steam temperature test data and the predicted data from one
model is shown in Figure 2-7. The model shown here is a first-order with deadtime. The model
parameters are a gain of −1.35, a time constant of 235 seconds, and a deadtime of 60 seconds.
Comparison of Actual and Predicted Steam Temperature
15

10

5
Temperature

-5

-10

-15

-20
0 500 1000 1500 2000 2500 3000 3500 4000 4500
Time (seconds)

Figure 2-7
Comparison of test data (solid line) with model prediction (dashed line) for superheat temperature
spray valve step test

0 2-8
3
MPC IMPLEMENTATION IN THE SIMULATOR
Simulator Description
The simulator used in this project is a small Windows-based Ovation system consisting of one
nonredundant Ovation controller with no hard input/output system, one operator workstation,
and one software server. The simulator is set up in Southern Company’s DCS Laboratory in
Birmingham, AL. The Ovation software for the simulator is version 5.4.

Installation of the Ovation Advance Process Control Toolkit

Before the Advanced Process Control Toolkit could be installed, the Ovation controller had to be
upgraded from a standard controller to an advanced controller. This is simply a software
upgrade, no new hardware was required. Although the upgrade and software installation process
should have been relatively simple, it ended up taking several weeks to complete. Part of the
problem was that some key personnel were not available at certain times, and there was some
confusion about getting the proper license installed for the system. After the software was
installed and the license was correct, the controller would still not execute a program, although it
would accept downloads. The flash memory on the controller had to be cleared and reloaded to
fix the problem. The simulator has operated without any problems since then.

Configuration of the Control System and Simplified Process Model

To test the integration between the existing PID-based control logic and the new MPC-based
logic, it was necessary to install both sets of logic in the simulator.

Moving the Existing Steam Temperature Logic

The existing PID-based logic sheets were exported from the Gaston unit 3 control system and
imported into the simulator. This worked but turned out to be a mixed blessing because all the
point names associated with inputs and outputs were renamed by the Ovation Control Builder
software when they were imported into the simulator system. Also, much of the existing logic
was related to alarming or interlocks, which were not needed for the simulator testing. As a
result, some of the existing logic was removed to simplify the configuration and the drawings.
Also, new point names were assigned to the key inputs and outputs.
The existing logic had to be modified to accept the new model predictive controller logic and to
interface with the simplified process model blocks. A selection switch was installed just
upstream of the existing manual automatic station with one input to the switch coming from the
PID-based control and the other input coming from the model predictive controller. The operator
can use the switch to select which controller is in service at any time. Refer to Figure 3-1 for the
superheater control logic implemented in the simulator.

0 3-1
Figure 3-1
Superheater control logic implemented in the simulator

MPC Configuration
The model predictive controller is implemented in a single software block. It can have up to five
manipulated variables, six control variables, and five disturbance variables. The block contains
two primary functions. The prediction function computes the future behavior of the process
based on the process model defined in the controller and the trajectory of the manipulated
variable. The optimizer function finds the best trajectory of the manipulated variables to
minimize a quadratic control cost function tuned by the user.
Because of the complexity of the MPC algorithm, a large number of parameters must be
specified to configure the controller. As a result, it is not really practical to manually enter all the
parameter values. The Advanced Control Process Toolkit includes a Model Builder program that
produces a configuration data file that can be read by the MPC algorithm. The Model Builder
program can identify a process model based on test data, or the user can enter a process model
developed by another system such as Matlab.
The MPC algorithm has several parameters that can be adjusted by the user to improve the
control response. The MPC algorithm operates with two different horizons, prediction and
control. These are also referred to as the control variable horizon and the manipulated variable
horizon, respectively. The prediction horizon is how far into the future the controller will
compute the control variable response. The control horizon is the interval over which the

0 3-2
manipulated variable trajectory will be optimized. The optimizer will find the manipulated
variable trajectory, which minimizes the cost function over the prediction horizon. The
prediction horizon should be set to the sum of the control horizon and the steady-state response
time of the process. If the process can be represented by a first-order response plus a deadtime,
then the steady-state response time will be approximately the sum of the deadtime and four times
the time constant of the process. This ensures that the predicted process response will reach
steady state within the prediction horizon. The MPC algorithm has a maximum control horizon
of 10 steps to limit the computational load on the control processor. Also, the MPC algorithm
does not necessarily run at the same cycle time as the regulatory control functions, with the cycle
time selection of the MPC being dependent on the response time of the process. For this testing,
the MPC controller cycle time was set to 10 seconds.
The user can also tune the controller by specifying relative penalties for control variable error
and change in manipulated variable. If the response of one control variable is too sluggish, the
user can assign a larger penalty to that control variable, which will alter the control design to
tighten that variable’s response. For a multivariable application, the weighting parameters are
matrices. For a single-input, single-output system, the weighting parameter is a scalar. The
control variable weighting matrix is called the Q matrix and the manipulated variable weighting
matrix is the R matrix.
For the application in the simulator, the model predictive controller is configured with one
control variable, one manipulated variable, and one disturbance variable. Consideration was
given to using one MPC algorithm to control both the left- and right-side superheat spray valves,
but there did not seem to be any compelling reason to use this approach. Additional disturbance
variables will likely be added when additional plant testing is permitted and more disturbance
data can be collected. Until then, the single disturbance variable will allow some testing of the
MPC’s disturbance rejection capabilities.
Consideration was also given to using the desuperheater outlet temperature in the model
predictive controller, but it was not a control variable and would not work as a disturbance
variable because it would be affected by the manipulated variable.

Simplified Process Model

A simplified process model was configured in the Ovation controller to verify the functionality
of the model predictive controller and to provide an idea of how its performance compares with a
PID controller (Figure 3-2). The process model included four first-order lag plus deadtime
models. The four models are superheater outlet temperature, desuperheater outlet temperature,
superheater disturbance, and desuperheater disturbance.

0 3-3
Figure 3-2
Simplified process model on simulator

The parameters for each of the models in this application are shown in Table 3-1. You may have
noticed that the desuperheater outlet temperature model time constant is 90 seconds in Table 3-1.
This is unusually long for this application, and the cause of this unusual response time is being
investigated. A definitive answer has not yet been found.

Table 3-1
Simplified process model parameters

SH Outlet DSH Outlet SH Disturbance DSH Disturbance

Gain −1.2 −1.1 1.0 0.5
Time constant 120 s 90 s 45 s 20 s
Deadtime 45 s 20 s 15 s 10 s

Notes:
SH means superheater; DSH means desuperheater.

0 3-4
4
SIMULATOR TEST RESULTS
The model predictive controller was implemented and tested on a simulator based on a DCS
platform. The tests were intended to verify the functionality of the algorithm and provide some
insight into the performance capabilities of the technique. The interface between the existing
PID-based logic and the new MPC logic can also be verified. Performance comparisons between
the PID control and the MPC were also included. Because the process model used in the
simulator is quite simple, the precise performance of the model predictive controller cannot be
determined by the simulator testing.

Test Procedure
Testing consisted of tuning tests, set point step response tests, disturbance rejection tests, and
robustness tests. Table 4-1 lists all the tests that were or will be conducted on the simulator.

Table 4-1
Control system tests conducted on the simulator

Test Category Test Description Purpose

MPC tuning Disturbance rejection, R = 100, Q = 1 Quantify effects of Q and R
MPC tuning Disturbance rejection, R = 10, Q = 1 Quantify effects of Q and R
MPC tuning Disturbance rejection, R = 1, Q = 1 Quantify effects of Q and R
MPC tuning Disturbance rejection, R = 1, Q = 10 Quantify effects of Q and R
MPC tuning Disturbance rejection, R = 1, Q = 100 Quantify effects of Q and R
PID disturbance Disturbance rejection Comparison with MPC
rejection
MPC robustness Disturbance rejection with process model Comparison with PID
time constant doubled
MPC robustness Disturbance rejection with process model Comparison with PID
gain doubled
PID robustness Disturbance rejection with process model Comparison with MPC
time constant doubled
PID robustness Disturbance rejection with process model Comparison with MPC
gain doubled
MPC performance Disturbance rejection using MPC Assess benefits of including
controller with disturbance variable input disturbance variables in model
predictive controller
MPC performance Set point step response Performance assessment
PID performance Set point step response Performance assessment
MPC functionality Tracking when PID in control Verify bumpless transfer

0 4-1
Simulator Test Results

MPC Tuning Tests

The MPC algorithm uses a quadratic cost function for the optimization of the control trajectory.
The cost function includes the control variable deviation from set point and the movement of the
manipulated variable. Each of these terms has a weighting parameter associated with it that the
designer can adjust to tailor the control response to the design specifications. The parameter
associated with the control error is called the Q matrix and the parameter for the manipulated
variable is called the R matrix. For this MPC testing, because the control is single input/single
output, each parameter is a single number rather than a matrix. The impact of these matrices on
performance is greatly affected by the scaling of the process variable. To better understand the
impact of changing the weighting parameters, a series of tests were run in which the two
weighting parameters were varied. The values ranged from R = 100 and Q = 1 to R = 1 and Q =
100. The results (Figure 4-1) show that the response gets faster as Q is made larger than R.

MPC Demo Project - MPC Tuning Tests

SH Outlet Temperature Response to Disturbance
1004 17.5

15.5
1002

13.5
1000
11.5
Temperature (degF)

Disturbance
998 9.5

R=100.0, Q=1.0
996 7.5
R=10, Q=1.0
R=1.0, Q=1.0
5.5
R=1.0, Q=10.0
994
R=1.0, Q=100.0
3.5
Disturbance
992
1.5

990 -0.5
0 100 200 300 400 500 600 700 800 900 1000
Time (sec)

Figure 4-1
Disturbance rejection response of the model predictive controller as the weighting parameters
R and Q are varied

0 4-2
When the response is the fastest, the manipulated variable is moving quite aggressively (Figure
4-2) and is probably faster than most plants would allow. For this reason, the tuning where R and
Q were both 1.0 was used for the rest of the tests.

MPC Demo Project - MPC Tuning Tests

Valve Demand Response to Disturbance
55
Vlv - R=100, Q=1
54
Vlv - R=10, Q=1
Vlv - R=1, Q=1
53
Vlv - R=1, Q=10

52 Vlv - R=1, Q=100

Valve Demand (%)

51

50

49

48

47

46

45
0 100 200 300 400 500 600 700 800 900 1000
Time (sec)

Figure 4-2
Manipulated variable movement during disturbance rejection as weighting parameters R and Q
are varied

Disturbance Rejection Comparison

Disturbance rejection tests were run on the model predictive controller and the PID controller to
compare performance (Figure 4-3). The weighting parameters for the model predictive controller
were set at 1.0 for both Q and R. The outer loop PID tuning parameters were set at a proportional
gain of 4.5 and an integral time of 30 seconds. The inner loop PID tuning parameters were a
proportional gain of 30 and integral time of 4 seconds. As shown, at least for this condition, the
model predictive controller was much more effective.

0 4-3
 MPC Demo Project - Disturbance Rejection Comparison

1001
SH Outlet Temp - PID
SH Outlet Temp - MPC

1000
Temperature (degF)

999

998

997

996
0 100 200 300 400 500 600 700 800 900 1000
Time (seconds)

Figure 4-3
Comparison of MPC and PID disturbance rejection

Robustness Tests
Because power plant characteristics change over time, it is important that the plant’s control
system performance not degrade too much as a result of the plant changes. To investigate the
robustness of the MPC algorithm and the PID controller disturbance rejection, tests were run in
which the characteristics of the simplified process model were altered. For the first test, the time
constant of the process model was doubled from 120 seconds to 240 seconds. In the second test,
the process model gain was increased by 50% from –1.2 to –1.8. The results of the time response
change tests (Figures 4-4 and 4-5) show that both controllers can tolerate these changes. The
magnitudes of the disturbances are the same in both tests, only the directions are different. In
fact, the performance of the PID controller is actually better when the time constant is increased.
This is because the ratio of deadtime to time constant in the process decreases as the time
constant is increased. The larger this ratio becomes, the more difficult it is for a PID controller to
control the process. Both controllers can also tolerate the process gain change.

0 4-4
Summary of Simulator Tests
Testing on the simulator is intended to functionally test the MPC algorithm and its interface with
the existing steam temperature control logic. It is not intended to be an exhaustive performance
evaluation, although some insight into the MPC performance and tuning is obtained from the
tests. A few additional simulator tests will be performed before the model predictive controller is
installed in the plant. These will include the tracking ability of the model predictive controller
when the PID controller is active and some more robustness tests.

MPC Demo Project - MPC Robustness Test

Model Time Constant Increased from 120s to 240s
1002

1001

1000
Temperature (degF)

999
SH Outlet Temp - Nominal Model
SH Outlet Temp - Altered Model
998

997

996

995
0 100 200 300 400 500 600 700 800 900 1000
Time (sec)

Figure 4-4
Model predictive controller performance during disturbance rejection with nominal and altered
process models

0 4-5
 MPC Demo Project - PID Robustness Test

1004
SH Outlet Temp w/ nominal model
SH Outlet Temp w/ time constant doubled
Temperature (degF) 1003

1002

1001

1000

999
0 100 200 300 400 500 600 700
Time (seconds)

Figure 4-5
PID controller performance during disturbance rejection with nominal model and model with time
constant doubled

0 4-6
5
CONCLUSION
A project to demonstrate a controller resident model predictive controller is under way at
Southern Company. A series of tests were run on Alabama Power Gaston plant unit 3 to identify
the superheat steam temperature response to spray valve changes. The data from these tests were
used identify process models that will be used in the model predictive controller design process.
A checkout simulator was developed using an Ovation DCS. The checkout simulator allowed the
model predictive controller to be tested in a realistic environment to verify the functionality of
the advanced system. Several tests were done to quantify the selection of weighting parameters
for the MPC. Other tests were done to compare MPC performance with the conventional PID-
based control system.
The controller resident MPC algorithm provides a less complex implementation than model
predictive controllers, which require a special process for the operation. Although documentation
is somewhat limited, the toolkit provided with the controller is adequate for design purposes.
In subsequent project tasks, after testing is completed on the simulator, the MPC will be
implemented in the real plant and tested. The existing PID control will also be available in the
plant for comparison purposes. All results will be documented in a final technical report.

0 5-1
0
6
REFERENCES
1. Demonstration of Advanced Control Techniques on Fossil Power Plants. EPRI, Palo
Alto, CA: 2000. 1001069.
2. Advanced Control Demonstration on a Combined Cycle Plant. EPRI, Palo Alto, CA:
2006. 1010261.
3. Instrumentation & Control Technology Assessment. EPRI, Palo Alto, CA: 2007.
1014237.

0 6-1
0
A
MODEL IDENTIFICATION TEST PLAN
Gaston Unit 3 Dynamic Response Test Plan

Background
Southern Company and EPRI are collaborating on a project to demonstrate the use of an
advanced control method called MPC on a fossil power plant. The host site for the demonstration
will be Alabama Power Gaston plant unit 3. The control application will be superheat and reheat
steam temperature control using spray valves and gas recirculation. As the name implies, the
MPC method uses a process model in its design. For this project, the model will be developed
using system identification from data gathered from the actual plant. This test plan describes the
testing necessary to produce the data required by the identification technique.

Test Description
When testing a process for model identification purposes, the process must be excited so that the
inherent dynamic response of the system is exposed. It is also preferable to have the control
system in manual mode so that the dynamics of the control system do not mask the dynamics of
the process. Several types of excitation signals can be used for the testing, but simple step
changes in the control actuator positions are the easiest to implement. Each actuator is stepped
one at a time; thus, the effect of each is isolated from the others. Such tests are referred to as
open loop step response tests. Because the dynamic response of the unit may change with load, it
is desirable to perform the tests at nearly full load and also at a lower load, depending on the
operating pattern of the unit. In this case, because the process being tested is steam temperature,
the need for lower load tests will depend also on the ability of the boiler to make design steam
temperatures at lower load. If the boiler cannot achieve design steam temperatures at lower load,
then the testing is unnecessary.

Test Prerequisites
For the high-load test, the unit should be between 90% and 95% of full load and in a good
steady-state condition. The unit should be off automatic generation control. The number/
combination of mills in service should be consistent during the same-load test sequence.
The following control loops must be in manual mode:
• Superheat spray valves
• Gas recirculation fans
• Megawatt control
• Fuel flow control
• Airflow control

0 A-1
After the above control loops are placed in manual mode, the unit should be allowed to reach a
good steady-state condition before beginning the tests. This will typically take about 30 minutes.
If after 30 minutes, the unit has not reached a good steady-state condition, an attempt should be
made to determine the cause of the continuing upsets.
The data archive system being used to store the data during the tests should be checked to ensure
that each desired point is in the archive database and its storage parameters have adequate
precision. Storage dead bands may need to be reduced during the tests to provide the desired
precision in the stored data. Process data will be collected through a combination of the
following mechanisms:
• OPC connection
• Manual log sheets
• Plant Aspen historian
• Fleetwide monitoring center Aspen historian
• Ovation tabular trends
Test Descriptions
Data points that should be collected for model identification are listed in Table A-1.
Table A-1
List of data points to be collected

Description Ovation Name

Gross megawatts (MW) TOTAL_MW

Throttle pressure THROTTLE_PRESS

Throttle temperature (temp)

Excess oxygen SELECTED_O2

Steam flow/first-stage pressure STEAM_FLOW

Primary air (PA) coal flow PA_DMP_FUEL

Coal flow OCB005B019-OUT

Airflow (total) AIR_FLW_SUM

LH SH temp setpoint SP OCB0152041-OUT

LH SH temp controller output (CO) OCB0152009-OUT

LH SH temp process value (PV) LH_SH_OUT_T

LH SH DSH temp SP

LH SH DSH temp CO OCB0152008-OUT

LH SH DSH temp PV LH_SHAT_OUT_T
LH SH DSH spray valve 1 position SHAT_LHV1_VLV_POS

LH SH DSH spray valve 2 position SHAT_LHV2_VLV_POS

0 A-2
Table A-1 (continued)
List of data points to be collected

Description Ovation Name

LH SH DSH spray valve 1 demonstration SHAT_LHV1_VLV_DMD

LH SH DSH spray valve 2 demonstration SHAT_LHV2_VLV_DMD

RH SH temp SP OCB0152043-OUT

RH SH temp CO OCB015200A-OUT
RH SH temp PV RH_SH_OUT_T

RH SH DSH temp SP
RH SH DSH temp CO OCB015200F-OUT

RH SH DSH temp PV RH_SHAT_OUT_T

RH SH DSH spray valve 1 position SHAT_RHV1_VLV_POS

RH SH DSH spray valve 2 position SHAT_RHV2_VLV_POS

RH SH DSH spray valve 1 demonstration SHAT_RHV1_VLV_DMD

RH SH DSH spray valve 2 demonstration SHAT_RHV2_VLV_DMD

SH spray flow SH_SPRAYWTR_FLOW

RH temp SP OCBSPREH-OUT
RH temp CO OCB00D4002-OUT
RH temp PV AVG_REHEAT

Gas recirculation damper position A_GRF_INL_DMP_POS

Coal feeder A flow A_FDR_FLOW

Coal feeder B flow B_FDR_FLOW

Coal feeder C flow C_FDR_FLOW

Coal feeder D flow D_FDR_FLOW

Coal feeder E flow E_FDR_FLOW

Coal feeder F flow F_FDR_FLOW

Sootblower sequence (3-BLR-SB-RUN) SB_RUN

Sootblower pressure SOOTBLOWER_P

LH DSH SP bias LHSPRAY_BIAS2

LH SH SP bias LHSPRAY_BIAS1

RH DSH SP bias RHSPRAY_BIAS2

RH SH SP bias RHSPRAY_BIAS1

0 A-3
Test 1: High-load, right-hand (RH) superheat spray valve step
Unit should be in steady-state conditions at a load between 90% and 95%.
1. Step RH superheat spray valve position up approximately 10% and wait 15 to 20 minutes
for the response to complete. The step should be made by entering the new position
demand digitally on the operator’s console. This ensures a consistent rapid step change in
position.
2. Step RH superheat spray valve position up approximately 10%.
3. Step RH superheat spray valve position down the same amount of the step up in 2.
4. Step RH superheat spray valve position demand down to the value before the test began.

Test 2: High-load left-hand (LH) superheat spray valve step

Unit should be in steady-state conditions at a load between 90% and 95%.
1. Step LH superheat spray valve position up approximately 10% and wait 15 to 20 minutes
for the response to complete. The step should be made by entering the new position
demand digitally on the operator’s console. This ensures a consistent rapid step change in
position.
2. Step LH superheat spray valve position up approximately 10%.
3. Step LH superheat spray valve position down the same amount of the step up in 2.
4. Step LH superheat spray valve position demand down to the value before the test began.

Test 3: High-load gas recirculation damper step

Unit should be in steady-state conditions at a load between 90% and 95%.
1. Step gas recirculation damper position up approximately 10% and wait 30 to 40 minutes
for the response to complete. The step should be made by entering the new position
demand digitally on the operator’s console. This ensures a consistent rapid step change in
position.
2. Step LH reheat spray valve position up approximately 10%.
3. Step LH reheat spray valve position down the same amount of the step up in 2.
4. Step LH reheat spray valve position demand down to the value before the test began.

Test 4: Low-load RH superheat spray valve step (optional)

Unit should be in steady-state conditions at a load between 90% and 95%.
1. Step RH superheat spray valve position up approximately 10% and wait 15 to 20 minutes
for the response to complete. The step should be made by entering the new position
demand digitally on the operator’s console. This ensures a consistent rapid step change in
position.
2. Step RH superheat spray valve position up approximately 10%.

0 A-4
Test 5: Low-load LH superheat spray valve step (optional)
Unit should be in steady-state conditions at a load between 90% and 95%.
1. Step LH superheat spray valve position up approximately 10% and wait 15 to 20 minutes
for the response to complete. The step should be made by entering the new position
demand digitally on the operator’s console. This ensures a consistent rapid step change in
position.
2. Step LH superheat spray valve position up approximately 10%.

Test 6: High-load gas recirculation damper step

Unit should be in steady-state conditions at a load between 90% and 95%.
1. Step gas recirculation damper position down approximately 10% and wait 30 to 40
minutes for the response to complete. The step should be made by entering the new
position demand digitally on the operator’s console. This ensures a consistent rapid step
change in position.
2. Step gas recirculation damper position up approximately 10% to return it to its original
position. Wait 30 to 40 minutes for the response to complete.

Test 7: Airflow demand step

Unit should be in steady-state conditions at a load between 90% and 95%.
1. Step airflow demand up approximately 3% and wait 30 to 40 minutes for the response to
complete. The step should be made by entering the new demand digitally on the
operator’s console. This ensures a consistent rapid step change in demand.
2. Step airflow demand down approximately 3% to return it to its original position. Wait 30
to 40 minutes for the response to complete.

Test 8: Turbine valve step

Unit should be in steady-state conditions at a load between 90% and 95%.
1. Step turbine valve position demand up approximately 4% (~10 MW) and wait 30 to 40
minutes for the response to complete. The step should be made by entering the new
demand digitally on the operator’s console. This ensures a consistent rapid step change in
demand.
2. Step turbine valve position demand down approximately 4% to return it to its original
position. Wait 30 to 40 minutes for the response to complete.

Test 9: Closed-loop load ramp

For this test, all the control loops placed in manual before test 1 should be placed back in
automatic mode.
1. Starting from a steady-state condition between 90% and 95% load, the unit should be
ramped down at 2.5 MW/min to approximately 75% load.
2. The unit should be allowed to stabilize at 75% load for 30 minutes, and then the load
should be ramped back up to its original load point at 2.5 MW/min.

0 A-5
0
0
Export Control Restrictions The Electric Power Research Institute (EPRI)
Access to and use of EPRI Intellectual Property is The Electric Power Research Institute (EPRI), with
granted with the specific understanding and major locations in Palo Alto, California; Charlotte,
requirement that responsibility for ensuring full North Carolina; and Knoxville, Tennessee, was
compliance with all applicable U.S. and foreign export established in 1973 as an independent, nonprofit
laws and regulations is being undertaken by you and center for public interest energy and environmental
your company. This includes an obligation to ensure research. EPRI brings together members, participants,
that any individual receiving access hereunder who is the Institute's scientists and engineers, and other
not a U.S. citizen or permanent U.S. resident is leading experts to work collaboratively on solutions to
permitted access under applicable U.S. and foreign the challenges of electric power. These solutions span
export laws and regulations. In the event you are nearly every area of electricity generation, delivery,
uncertain whether you or your company may lawfully and use, including health, safety, and environment.
obtain access to this EPRI Intellectual Property, you EPRI's members represent over 90% of the electricity
acknowledge that it is your obligation to consult with generated in the United States. International
your company’s legal counsel to determine whether participation represents nearly 15% of EPRI's total
this access is lawful. Although EPRI may make research, development, and demonstration program.
available on a case-by-case basis an informal
assessment of the applicable U.S. export classification Together…Shaping the Future of Electricity
for specific EPRI Intellectual Property, you and your
company acknowledge that this assessment is solely
for informational purposes and not for reliance
purposes. You and your company acknowledge that it
is still the obligation of you and your company to make
your own assessment of the applicable U.S. export
classification and ensure compliance accordingly. You
and your company understand and acknowledge your
obligations to make a prompt report to EPRI and the
appropriate authorities regarding any access to or use
of EPRI Intellectual Property hereunder that may be in
violation of applicable U.S. or foreign export laws or
regulations.

© 2009 Electric Power Research Institute (EPRI), Inc. All rights reserved.
Electric Power Research Institute, EPRI, and TOGETHER…SHAPING
THE FUTURE OF ELECTRICITY are registered service marks of the
Electric Power Research Institute, Inc.

Printed on recycled paper in the United States of America 1015711

Electric Power Research Institute

3420 Hillview Avenue, Palo Alto, California 94304-1338 • PO Box 10412, Palo Alto, California 94303-0813 • USA
0 800.313.3774 • 650.855.2121 • [email protected] • www.epri.com