Proceedings of the 17th World Congress

The International Federation of Automatic Control

Seoul, Korea, July 6-11, 2008

Controller Design for a 1000 MW

Ultra Super Critical Once-through Boiler Power Plant
Kwang Y. Lee*, Joel H. Van Sickel**, Jason A. Hoffman**,
Won-Hee Jung+, and Sung-Ho Kim+

* Baylor University, Waco, TX 76798, USA

(e-mail: Kwang_Y_Lee@baylor.edu)
** Pennsylvania State University, State College, PA 16803 USA
USA (e-mail: jhv107@psu.edu)
+ Doosan Heavy Industries and Construction Company, Ltd., Seoul, Korea
(email: sungho.kim@doosanheavy.com, wonhee.jung @doosanheavy.com)

Abstract: A large-scale 1000 MW once-through type ultra super-critical boiler power plant, requires
investigation for the development of an analyzable model for use in the development of an intelligent
control system. Using data from the power plant, a model is realized using dynamically recurrent neural
networks. For proper operation, the plant must be broken into smaller subsystems that are each modeled by
a separate neural network. Modified predictive optimal control is then used to drive the plant to desired
states. Due to the computational intensity of modified predictive optimal control, it was rendered unviable
by the computation time required for each time step of the controller. As an alternative, a reference
governor was implemented along with a PID feedback control system that utilizes intelligent gain tuning.

model. This type of approach is covered in detail in (Lee, et

1. INTRODUCTION
al.,2007a). Only the higher level details will be covered to
Ultra super-critical (USC) boiler power plants are currently show the differences required to deal with a new power plant.
being developed to increase the efficiency of standard fossil
It was desired to use a modified predictive optimal control
fuel power plants. The modelling and control of a large-scale
scheme with this plant to track unit load demand in order to
1000 MW once-through type ultra super-critical boiler power
provide adaptive control that optimized certain functions of
plant is investigated here. Larger more complicated power
the power plant. This scheme was developed successfully,
plants require more sophisticated methods to streamline the
but turned out to be more computationally intensive than
modelling process as well as more sophisticated control
desired for an actual controller. To overcome this difficulty, a
schemes that can be used to further enhance plant efficiency.
reference governor was developed to provide feed forward
The development of large capacity power plants requires new controls in conjunction with a simple PID feedback control
approaches to analyze plant dynamics for control purposes. In system that utilizes intelligent gain tuning. Both approaches
practice, many utility companies utilize simulation programs, are presented with a focus on the reference governor and
such as Modular Modelling Systems (Leavesley, et al., 1996) intelligent gain tuning.
or their own simulation tools for modelling. However, it is a
challenge to extend current models to model larger capacity 2. 1000 MW USC POWER PLANT
plants, and to design new models without component
In this report, the USC boiler power plant consists of four
specifications. To design a control system for a power plant,
processes which are air/flue gas, pulverizer, water/steam, and
a model must be developed in advance. Recently, the study of
turbine/generator. However, for modelling purposes, the
Neural Networks (NN) has become important in designing
number of detailed subsystems will be nineteen. Fig. 1 shows
system identification and control systems in the power
the 1000 MW USC boiler power plant. Most blocks are
systems area. With system data, the NN can be trained to
subsystems, which will be represented by a NN-based
approximate highly nonlinear functions. Since the NN
subsystem model. The proposed scheme will be applicable to
strongly depends on the input/output data but not on the
other types of plants, including nuclear and fuel cell plants.
physical structure of the system, it is flexible and can easily
be adapted to different types of power plants. The power plant under investigation is a 1000 MW coal-
pulverized, once-through type, boiler-turbine-generator unit.
Accurately modelling such a system with a single NN is
There are three economizers used to raise the temperature of
theoretically possible, but it was discovered that in practice,
water entering the boiler from the feedwater system. Two
the training of such a network was not practical. Instead, the
forced draft fans and two primary air fans provide air to the
individual subsystems of the power plant were modelled with
air preheater. The air preheater in turn provides heated air to
separate NN that were combined to form the power plant
the pulverizers, burners, and furnace. The primary air fans

978-1-1234-7890-2/08/$20.00 © 2008 IFAC 13938 10.3182/20080706-5-KR-1001.3464

17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

Figure 1. 1000 MW USC Power Plant.

also provide cold air to the pulverizers. The fuel is provided Table 1. Control actions
to the furnace through the pulverizers and burners. Furnace
pressure is maintained at the desired value by controlling two Control Associated
Control Description
induced draft fans. The waterwall surrounds the furnace Number Subsystem
vertically and spirally. Flue gas exiting the furnace travels primary air
uc1 primary air fan control
through the superheaters and reheaters, economizers, and air subsystem
preheater to raise the temperature of the steam, water or air, secondary air
uc2 forced draft fan control
respectively. There is a separator on top of the furnace which subsystem
supplies high pressure steam to the primary superheater and induced draft fan gas recirculation
uc3
reduces the impurities in the steam. The superheater consists control subsystem
of four parts, primary, division, platen, and finish. The hot primary air damper pulverizer/burner
reheaters reheat the steam after the High Pressure (HP) uc4
control subsystem
turbine using the primary reheater and the reheater finish. cold primary air damper pulverizer/burner
Finally, the turbine generates power from the tandem uc5
control subsystem
compound triple turbines, which consist of three parts: a HP pulverizer/burner
turbine, an Intermediate Pressure (IP) turbine, and Low uc6 coal feeder control
subsystem
Pressure (LP) turbine. feedwater pump control
uc7 feedwater subsystem
The model will be focused on boiler, turbine, and generator
parts. Each subsystem has common inputs and outputs: mass superheater division
uc8 feedwater subsystem
flow rate, temperature, pressure, and enthalpy of fluid. In spray control
addition to these inputs, there are control variables involved superheater platen spray
uc9 feedwater subsystem
in driving each subsystem to the desired state, which are control
listed in Table 1. The proposed model, which is based on the high pressure turbine high pressure turbine
uc10
NN, will use the predefined control action as feedforward valve control subsystem
control. The four process models which are broken up superheater damper gas recirculation
uc11
further into subsystems are shown in Table 2. With the control subsystem
proposed approach, the utility company is able to investigate gas recirculation
the dynamic characteristics of power plants with different uc12 reheater damper control
subsystem
capacities.

13939
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

Table 2. Process models and subsystems Table 3. Control actions

Turbine & Hidden

Water & steam Air & flue Pulverizer Subsystem Inputs Outputs
Generator Neurons
model gas model model
model Pulverizers/
11 3 19
Feedwater Burners
Economizer1 Primary Air Primary Air 2 4 17
Economizer2 Intermediate/Low Secondary Air 2 2 21
Economizer3 pressure turbine Separator 4 4 11
Secondary High Pressure Turbine 5 5 21
Waterwall/Furnace
Air
Separator Intermediate Pressure
4 4 25
Primary Pulverizer/ Turbine
Superheater Burner Platen Superheater 10 4 21
Air
Superheater Primary Superheater 7 4 23
preheater
Division High pressure Primary Reheater 7 4 25
Superheater platen turbine Air Preheater 7 9 17
Superheater Finish
Gas Division Superheater 10 4 23
Primary Reheater
recirculation Economizer1 7 4 23
Reheater Finish
Economizer2 7 4 25
Economizer3 11 4 21
3. NEURAL NETWORK COMBINED MODEL Feedwater 5 11 17
Final Reheater 7 4 25
A neural network (Ku and Lee, 1995) representing each
subsystem is trained many times with different numbers of Final Superheater 7 4 9
hidden neurons. The cost function for each training, which is Furnace 10 7 17
the Mean Squared Error (MSE) between the neural network Gas1 6 16 21
output and the target values, is compared with the others for
different numbers of neurons. The number of hidden neurons Gas2 6 16 15
with the smallest MSE is set as that subsystems hidden
neuron number. The optimal number of neurons depends on
the number of inputs and outputs of each subsystem as well
as the input/output data pattern; therefore, some subsystems
with few inputs and outputs require more hidden neurons to
achieve the best performance. The resulting hidden neurons
for each subsystem are shown in Table 3. The gas
recirculation system was split into two separate networks
because there were six inputs and thirty-two outputs. Each
NN of the two gas recirculation networks uses the six inputs
to generate half of the outputs. Gas 1 delivers outputs to the
division superheater, the platen superheater, the primary
superheater, the final superheater, the primary reheater, and
the final reheater. Gas 2 delivers outputs to the primary
reheater, the economizers 1, 2, and 3, and the air preheater. A
neural network with six inputs and thirty-two outputs will
cause the computer to run out of memory when training. The
final result is referred to as the neural network combined
model (NNCM). Fig. 2. MPOC (Neural Network Output) tracking power
demand (APESS Power Output).
4. MODIFIED PREDICTIVE OPTIMAL CONTROL
next control action to see if it moves the power plant to the
Modified predictive optimal control has already been used desired states. This is different from standard predictive
successfully in (Lee, et al., 2007b), and was the method control which evaluates further than just the next time step.
expected to be used to control this power plant. This This was done to reduce overall calculation time.
particular instance of predictive optimal control uses Unfortunately, the proposed method still did not achieve
recurrent neural networks (RNNs) to implement an online quick enough results to be used real-time for this application.
identifier that models plant behaviour. Particle swarm is used Figs. 2 and 3 show the designed control system successfully
in conjunction with this identifier to test the validity of the tracking the desired performance. The online identifier
(Ghezelayagh & Lee, 2005) is updated so that it can

13940
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

have been easily decreased through the use of parallel

processing, it was decided against adding this level of
complexity to the controller.

5. REFERENCE GOVERNOR AND GAIN TUNING

Since the MPOC did not generate control actions quickly
enough, an older method was modified to work with this
process. Using a two stage system, a reference governor can
provide feedforward control actions as well as setpoints for a
feedback controler, and the feedback controller provides the

Fig. 3. MPOC (Neural Network Output) tracking pressure

demand (APESS Power Output).

Fig. 5. Overall control scheme for reference governor and

feedback control.
Table 4. Set points
Set-Points/Demands
Throttle Pressure Demand
Feedwater Demand
Coal Flow Demand
Final Superheater Temperature Demand
Final Reheater Temperature Demand
Furnace Gas Pressure Demand
Pulverizer Temperature Demand
Air Flow Demand
MW Demand

Table 5. Control actions and coupled set points

Controls Associated Set-points/Demands
Primary Air Fan Coal Flow Demand
Secondary Air Fan Air Flow Demand
Feedwater Pump Feedwater Demand
Fig. 4. Scheme for online identifier with MPOC.
Spray 2 Final Superheater Temperature
accurately model current plant behaviour and can be used by Demand
MPOC to search for the next control action. Fig. 4. details the Spray 3 Throttle Pressure Demand
online identifier because the same scheme is used for an HP Turbine Valve MW Demand
online identifier in the second control approach. Induced Draft Fan Furnace Gas Pressure Demand
Air Flow Demand
4.1 Calculation Time Issue Reheater Damper Final Reheater Temperature
Demand
The trouble with calculating control actions with MPOC, is Superheater Damper Final Reheater Temperature
that the control signal was desired to be updates at least every Demand
0.25 seconds. It was acceptable, while running in Matlab, for Hot Air Damper Pulverizer Temperature Demand
the algorithm to generate an update every second, but the Cold Air Damper Pulverizer Temperature Demand
final speed as closer to 1.5 seconds. While the speed could Coal Feeder Coal Flow Demand

13941
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

Where the variables are as follows:

actual control actions to the plant, or in this case, the NNCM.
This method is visualized in Fig. 5. For this to work, it was α1,α2,α3: Multi-objective weights
required to determine what set points would be used and ULD: unit load demand
which control actions would be coupled to these set points.
The results are shown in Tables 4 and 5. PowerOut: actual power output
CoalFlow: control that determines how much coal is used as
5.1 Calculation Time Issue fuel
The reference governor with gain tuning bypasses the issue of u: feedforward (ff) control actions
how quickly the control signal can be updated, because both
the reference governor and gain tuner work offline, and unom: nominal feed forward control actions (what the power
simply provide updates to the controller as required. While plant would do without optimization)
the reference governor and gain tuner are searching for their There is a disadvantage to using this approach because it
next results, the controller can be set any feasible sample
assumes that the nominal control actions are available. In this
period achievable by the hardware, as only PID control loops case, these nominal control actions were available from
must be calculated. earlier in the power plants design process. If this is not the
case, a simple control system would have to be developed to
5.2 Reference Governor create these control actions, which may be more work than
desired to use this particular approach.
Using a reference governor for providing feedforward control
actions and set points has been shown many times such as in 5.3 Intelligent Gain Tuning
(Garduno-Ramirez & Lee, 2001). As done in previous work,
a steady state model (Heo & Lee, 2005) of the system was Intelligent gain tuning is done using an online identifier and a
trained using a static neural network, and then a heuristic
heuristic search. The online identifier is similar to the one
search method was used to find the feed forward control used for MPOC, but different in that it has outputs that are
actions and corresponding set points that would optimize a
used for feedback control. The heuristic search tries different
cost function made of weighted objectives. gain values and then simulates the system with these gain
For this application, four of the five set points are actually values and the online identifier. It continues to experiment
held constant regardless of unit load demand, and can with different gain values until it finds the set of gains that
therefore be eliminated from the neural network, as their reduce the error between the set points and the plant outputs.
values will never be changing. These set points are Final It is very similar to MPOC except that instead of choosing the
Superheater Temperature Demand, Final Reheater control values, it is choosing the gain values. This change is
Temperature Demand, Furnace Gas Pressure Demand, and
made because the gain values do not have to be updated all of
Pulverizer Temperature Demand. the time, while control values do. For gain tuning, a large
Interestingly, this approach worked poorly at first. With the window size can be chosen for which to tune the gains. This
high order of this system, the search algorithm was able to window could range from the size of a few minutes to
find numerous candidate control actions and set points that multiple hours, depending on how often it is desired the gains
equally satisfied the provided cost function. This was very be tuned. The gain tuning has exactly the time of the window
undesirable as ideally, the cost function should be set up so to search for the next set of gains. Once the window time has
that a single set of control actions and set points provide an passed, the gain tuner reports the best set of gains it has found
optimal solution, or the reference governor will not know to the control system, which is then updated with these new
which set to choose. Using a scheme where different control gain values. Then the gain tuner starts searching again for the
actions have the same fitness is very noisy and inefficient. To best set of gains for the next window. This process is
cope with this problem, the concept of using nominal control repeated indefinitely.
actions was introduced. The nominal control actions are The window size was chosen to be twenty minutes. This is
simply what the conventional control actions would be for a
not the only window size that can be used, but it was the
given unit load demand if a more sophisticated control smallest window size that had smooth operation. Smaller
scheme was not in place. The cost function was then
window sizes can change the gains too often, which causes
modified so that it would optimize specific goals, and then the system to become noisy and if the window size was small
choose the candidate control actions that were closest to the enough, could actually lead to unstable operation. This is not
nominal control actions. This modification served to fix the
the case for a window size of twenty minutes though. The
problem and provided good performance. The result was the power plant is obviously running for longer than one window
following cost function (1). size, so its operation must be split into multiple windows.
f (u ) = α1 ULD − PowerOut + ... With a window size T, and total operational time of Tf, the
(1) operation is split into N = T/Tf windows, with end at T0,
= α 2 CoalFlow + α 3 u − unom T1,.. TN. This is shown in Fig. 6. After an optimization
window ends is when the gain values are updated.

13942
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

Fig. 6. Window operation.

The algorithm works by searching three different gain
matrices, one for the proportional control, integral control,
and derivative control. It takes the possible gain matrices and
simulates the system for the next twenty minutes (the current
window size) with those gain values. It repeats this
simulation for different possible combinations of gains and
then evaluates the gains by choosing which gain has the
smallest total error for set point tracking, using the following
cost function, where setpointn is the nth desired set point, and
outputn is the nth actual plant output of that setpoint:
9 tf

y = ∑∑ setpointn − outputn (2)

n =1 t = to

An online identifier, as with MPOC, is continually updated

and used for the simulation of the different gain values. It
only needs to be updated once every window, so a large
window size means the online identifier has to be updated
less. Shown in Fig. 7 are the results of using the reference
governor to vary the power plant from 1000 MW to 600 MW
to 800 MW. Only the setpoints which actually change are
shown, as the rest simply remain constant. Fig. 7. Set point tracking of reference governor with gain
tuning.
6. CONCLUSIONS
While MPOC was desired for its adaptive capabilities and the
ability to optimize a cost function, it turned out to be too Ku, C. C., and K. Y. Lee, 1995, “Diagonal Recurrent Neural
calculation intensive to implement in a real time controller. Networks for Dynamic System Control,” IEEE Trans.
Instead, a reference governor with intelligent gain tuning was On Neural Networks, vol. 6, no. 1, pp. 144-156.
implemented. To get the reference governor to work, it had to Leavesley, G. H., S. L. Markstrom, M. S. Brewer and R. J.
favour control actions that were closer to the nominal control Viger, 1996, “The modular modeling system (MMS) –
actions over others. If specialized hardware was used for the The physical process modeling component of a database-
OLID, mainly, the neural network calculations, MPOC might centered decision support system for water and power
become a viable option as well. management,” Water, Air, and Soil Pollution, vol. 90, no
1-2, pp. 303-311.
Lee, K. Y., J. S. Heo, J. A.Hoffman, S.-H. Kim and W. H.
REFERENCES
Jung, 2007a, “Neural Network-Based Modeling for A
Garduno-Ramirez, R. and K.Y. Lee, 2001, “Multiobjective Large-Scale Power Plant,” Proc. of the IEEE PES
optimal power plant operation through coordinate control General Meeting, Tampa, FL
with pressure set point scheduling,” IEEE Transactions Lee, K. Y., J. S. Heo, J. A.Hoffman, S.-H. Kim and W. H.
on Energy Conversion, vol. 16, pp 115-122. Jung, 2007b, “Modified Predictive Optimal Control
Ghezelayagh, H. and K. Y. Lee, 2005, “Application of Self- Using Neural Network-based Combined Model for
Organized Neuro-Fuzzy Identifier in Intelligent Large-Scale Power Plants,” Proc. of the IEEE PES
Predictive Control of a Power Plant,” Engineering General Meeting, in Tampa, FL.
Intelligent Systems, vol. 13, no. 2, pp. 113-118.
Heo, J. S. and K. Y. Lee, 2005, “Multi-agent system-based
intelligent steady-state model for a power plant,” Proc.
the 13th International Conference on Intelligent Systems
Application to Power Systems.

13943