Copyright @ IFAC Power Plants and Power Systems Control, Brussels, Belgium, 2000
CLOSED LOOP NO x CONTROL AND OPTIMISATION USING NEURAL NETWORKS.
Jack Gabor 1 Daniel Pakulski 2 Konrad Swirske Pawel D. Domanski4 •
I
Westinghouse Process Control, Inc., A Fisher-Rosemount Company 200 Beta Drive, Pittsburgh, Pennsylvania 15238, e-mail:
[email protected] 2
Transition Technologies Sp.
Z 0.0.,
ul. Dzika 4,00-194 Warsaw, e-mail:
[email protected] 3
Institute of Heat Engineering, Warsaw University of Technology,
ul. Nowowiejska 25, 00-650 Warsaw, e-mail:
[email protected] 4
Institute of Control and Computation Engineering, Warsaw University of Technology,
ul. Nowowiejska 15//9, 00-665 Warsaw, Poland, e-mail:
[email protected].
Abstract: This paper presents real industrial applications of Advanced Control for power plant systems and emission controls. Closed loop NOx control using artificial neural nets and boiler optirnisation was applied. Basis knowledge to build an advanced control tool called IVY is also briefly described. Obtained control performance is very high for all presented aspects of the combustion process. The advanced controller provided higher unit efficiency, while
simultaneously
maintaining environmental and technological constraints. Copyright lfi2000 IFAC
Keywords: neural networks, predictive control, adaptive control, optimisation, boilers.
noises INTRODUCTION
also degrade
system
performance.
The
dynamic properties of a controlled plant may not be very complex, even though its detailed structure and
The main issues to consider for the design of an
parameters are usually unknown.
industrial process control system are the negative
Soft computing methods are considered to be the
effects of non-linearities. Process and measurement
most appropriate in the presence of the imprecise
141
measurements and in the case of non-linear processes
hardware to implement control algorithms has been
with some unknown dynamics (Bartos 1997). This
improved significantly in recent years. Despite the
computing
difficulty in achieving high control quality, the fine-
methodologies that adhere to its guiding principle. At
tuning of the controller's parameters is a tedious task,
this juncture, the principal constituents of soft
requiring experts with knowledge both in control
computing are
theory and process dynamics. All of these call for the
approach
is
the
consortium
fuzzy
logic,
of
neural
computing,
evolutionary techniques, rough sets, probabilistic
development
of new
controllers.
The
modern
reasoning, etc...
advanced control techniques based on soft computing
At the same time, new emissions regulations form the
approach can become such a solution.
challenge for control systems to achieve the goals of
The most important methods
environmental
distinguish the systems are based on fuzzy logic
protection.
These
goals
consider
we can
use to
opacity, NOx, S02, and CO.
(Zadeh, 1965; Czogala, Pedrycz, 1982; Zimmermann
The combustion process and the creation of toxic
1991) and artificial neural networks (Hertz 1992).
particles are strongly connected. Because this process
Both of these methods can successfully deal with
is still the most popular energy source, we need to
non-linear or poorly defined problems.
think about some methods for reduction of these
Fuzzy
negative effects. We can distinguish two approaches:
linguistic or rule-based knowledge into a classical
logic
makes
it possible
to
incorporate
control scheme. This can be achieved in many ways, •
primary methods that use a better low
mainly by: Fuzzy Logic Controllers (FLC), Fuzzy
emission combustion process operation; •
Logic Supervision (FLS), Fuzzy Models, and Fuzzy
secondary methods that try to clean the
Expert Systems.
output
The potential of NNs (Neural Networks) for control
combustion
gases
from
toxic
applications lies in the following properties:
elements.
• In this study the primary methods will be addressed,
they could be used to approximate any continuous mapping;
mainly in connection with measurable emissions. Because these processes are very complex, and
•
they achieve this approximation through learning;
highly non-linear, the suggested way for control is using the soft computing approach. •
The paper consists of the presentation of applied technologies. These methodologies are illustrated with followed industrial applications: •
parallel processing and fault tolerance are easy to be accomplished with NNs;
•
they make it very easy to implement control structures with internal non-linear
NOx reduction at Polaniec Power Plant, unit
modeling, for example in the predictive
#4, Polaniec, Poland,
or adaptive structures. •
Valley Power Plant, unit #4, Milwaukee, Wisconsin 2.1
2
ADVANCEDCONTROLTECHNOLOGffiS
Model Predictive Control - MPC
Model Predictive Control algorithms form the wide range of predictive structures that compute the
Contemporary industrial process control systems
control signal on the basis of the process model (Qin,
dominantly rely on Pill-type controllers, though the
Badgwell, 1996). While the proper model provides
142
support to the controller with the process prediction,
multi-regional
a method is needed for calculating the control signal
1994).
with respect to the process. In this step we
•
incorporate into the controller some non-linear
MODEL
controllers
(Qin,
REFERENCE
Borders
ADAPTIVE
CONTROL (MRAC) algorithm based on a
optimisation methods. The signals must be filtered,
controlled plant model in the form of
conditioned and validated. There also needs be
ARMAX structure.
boundaries placed around the process input and output signals.
•
SELF-TUNING
(STR)
REGULATOR
The control action (in each sample interval) is as
algorithm is quite similar to the MRAC, but
follows:
we do not use ARMAX model explicitly to
•
set control horizon for
compare with the plant.
which the control
scenario sequences will be checked; •
2.3
The process model is the main part of a model
set the prediction horizon (longer than control
predictive controller. This model should be accurate
horizon) for which the system performance will
over a wide range of plant operation, and thus it is
be checked; •
Non-linear Process Modeling
desirable to use the non-linear model for non-linear
run the optimisation algorithm which will be
processes. Non-linear identification using equation-
looking for optimum control actions with respect
based
to the defined performance criteria.
consuming. Many alternate approaches can be found
modeling
can
be
expensive
and
time
This procedure is repeated in each step. Since the
in the literature, such as fuzzy logic, neural networks,
control derivation process is time consuming, an
evolutionary computation, linguistic reasoning, rough
efficient computer system is required.
sets, etc.
Adaptive Control
2.2
3
IMPLEMENTED SOFfWARE: IVY ADVANCED CONTROL TOOL
In most real world control problems, the plant under consideration has its known response characteristics
IVY is an advanced control software tool. It enables
and unknown response characteristics. The primary
designing and tuning of a model predictive controller
purpose of an adaptive control scheme is to
with a receding horizon for base multidimensional
manipulate the controller to cope with the unknown
control and adaptive structures based on the fuzzy
and changing dynamics in the plant environment. In
approach to the gain scheduling algorithm.
addition to that, however, adaptive controllers have
IVY software consists of two parts:
also been advocated as an alternative to prior •
Run-Time Controller
means to relieve the user from modeling uncertainty
•
TT Design - Modeling Design Tool
(black box philosophy). We can enumerate three
MPC
basic
The model is based on fuzzy neural network
analysis: "automatic adaptation" is then seen as a
adaptive
control
structures
(Astrom,
Wittenmark 1994): •
technologies. A fuzzy model can be viewed as a fuzzy
GAIN SCHEDULING - Nowadays this
non-linear
NARMAX
(Non-linear
Auto
Regressive Moving Average with auXiliary input)
approach can be met with a new version, if
model (Horikawa, et al. 1991). This notion comes
we consider fuzzy Takagi-Sugeno based
from piecewise linear systems.
143
Gain Scheduling
wall fired boiler, 650 klbh steam, Foster-Wheeler
To perform the function of Gain Scheduling, an
boiler.
adaptive algorithm or a hybrid PID-IMC (Internal
The main goal for the NOx Reduction System
Model Control) multi-regional control structure can
implemented at the Valley Power Station is to reduce
be used (Zacho, et al. 1993; Domanski, et al. 1997).
NOx emission. The secondary goal is to maintain all
Optimization algorithm
other boiler parameters. It was identified during the
The problem of choosing manipulated variables for a
tests that LOI is strongly connected with NOx
given prediction
concentration, and NOx reduction can make LOI
horizon
is formulated
as
an
optimisation problem. The IVY controller can use
worse. Due to this fact the LOI measurement has
different optimisation algorithms,
been used as a second control goal. Both of these
since the calculated manipulated variables must
signals will be called CVs - Controlled Variables.
satisfy constraints on values and the speed of
The control structure was implemented on a WDPF
changing, a penalty function is added to the
system. The implementation is built around the IVY
performance index.
controller (Model Predictive Controller with receding
Graphical User Interface - GUI
horizon for multidimensional control) with nc2w2
The User can access and tune the controller using
software for the interface to the DCS.
GUI's (Graphical User Interface) written in JAVA.
The base
The interface consists of seven
main graphic
smooth and bumpless co-ordination with the neural
control structures have been updated for
windows that enable the monitoring of the controller,
controller. A specialised graphical screen has been
tuning and setting of its parameters.
prepared, and existing screens have been updated.
System requirements
The system is running on a dedicated WeStation. The
Graphical User Interface was written in Java.
system assures bumpless and safe operation while the
Therefore, an installation of Java on the system is
base control can still work in normal MANUAL or
necessary to execute the interface. The controller can
AUTO mode. To introduce the System, the new
run under Windows NT, Solaris UNIX 2.5 or LINUX
control
operating system.
(Supervisory). When the controls are in the SPV
Applications
mode it means that the NOx Reduction System is
mode
has
been
added,
called
SPV
running. •
NOx emission control (Arabas, et al. 1998)
•
Opacity optimisation (Domanski, et al. 2000)
•
Boiler optimisation
When controls are in AUTO, the WDPF will track the above biases from the NOx Reduction System, when the system is turned ON it starts from actual
(Neelakantan, et al.
boiler
1998)
parameters.
Simultaneously
the
initial
operator's parameters are stored, so that after turning it off, the system goes back to the original parameters.
4
The results of the 48 hour performance test are
APPLICATIONS OF THE DYNAMIC
shown in the table 1. The average steam flow, NOx
EMISSION BOll...ER CONTROL
level, O 2 , and desuperheater spray flow are indicated.
4. J
NOx control
An average of 15 % reduction in NOx emissions was achieved. Figure 1 and 2 presents comparison of the
The project was implemented on boiler #4, Valley
emission NOx without NOx Reduction System
Power Plant, Wisconsin Electric Power Company,
(figure 1) and with it (figure 2) in approximately the
Milwaukee, Wisconsin, USA, a pulverised coal front
same steam flow.
144
Table 1 NOx commissioning test results
•
6 mills system (l mill system for 1 burner row)
Date
NN
Steam
NOx
O2
Despht.
Status
Flow
[#/mmb
[%]
Spray
[kpph]
tu]
Flow
•
Burner angles control ability
•
air flow control: primary air; secondary air,
[kpph] 12/15/99
OFF
416.3
0.468
5.24
27.82
12/16/99
ON
429.7
0.399
4.29
22.37
3.2%
% change
OFA. The aims of the project are: •
NOx emission reduction
•
Increase in boiler efficiency factor
•
Flue gas temperature symmetrization
-14.790 -18.1% -19.6%
Fig. 1. Test with controller not running
The control structure was implemented on the WDPF
I
PPM
system
E
675
using
Aspen
Target-model
predictive
controller with nc2w2 software as the interface to the DCS. To obtain better combustion control, a Burner Management System was added. The process model
600
was implemented using neutral network technology. The model's inputs are:
525 •
manipulated variables: flue and air fan drive positions; O2 setpoint; secondary air dampers positions; mill feeders;
•
disturbances:
O2
concentration,
drum
pressure; total air flow; pressure differential
Fig. 2. Test with controller running
and temperature in combustion chamber; flue
NOx
10 SECOND AVERAGE
285
675
PPM
KLBH
gas fan motor power; unit capacity
The model's outputs are NOx emission level, CO
600
emission level, outlet flue gases temperature, and steam temperatures.
NOX SETPOINT
Obtained results:
2,5 hours
The controller enables the maintenance of the environmental limits as its shown on table 2,
4.2
symmetries, and the flue gas temperature at the same
NOx Control and Boiler Optimisation
time. An average increase of 0.4-0.8 % in the boiler efficiency was obtained. Simultaneously, there was
The project was implemented on unit # 4, Polaniec
no over or under heating of the steam. Nor was an
Power Plant, a pulverised coal fired boiler OP-650.
•
increase of CO emission or more slugging observed.
24 low NOx burners placed on sides in 6 rows
145
Table 2 NOx commissioning test results
stability analysis, European Control Conference '97 CD-ROM Proc., Brussels.
Controler
NOx
status
[mgiNm3]
ON
430
OFF
470
Domanski P.D., Swirski K., Williams J.J (01/2000)
Application of Advanced Control Technologies to the
Emission
Control
and
Optimisation,
Conference on Power Plant Emission Control and Monitoring Technologies, London, UK. Hertz J., Krogh R., Palmer A. (1992) Introduction to
the theory of neural computing, Addison Wesley. 5
Horikawa S. et al. (1991) A study on Fuu:y Modelling
CONCLUSIONS AND FURTHER
using Fuu:y Neural Networks, IFES'91, pp. 562-
IMPLEMENTAnONS
572.
Neelakantan R., Domanski P.D., Swirski K (12/1998)
This paper presents real industrial applications of power plant systems and
Hybrid Neural Network Model Based Control of
emission controls. Obtained control performance is
a Coal Fired Boiler, PowerGen International '98,
very high for all presented aspects of the combustion
Orlando, USA.
Advanced Control for
Qin SJ., Badgwell TJ. (01/1996) An overview of
process. The advanced controller provided higher unit efficiency, while
industrial model predictive control technology,
simultaneously maintaining
CPC-V, Tahoe.
environmental and technological constraints. The
Qin SJ., Borders G. A (02/1994) Multiregion Fuzzy
plant staff, both operators and engineers, have accepted the implementations.
Logic Controller for Non-linear Process Control.
The next step is directed toward the supervision of
IEEE Trans. on Fuzzy Systems, Vol.2, No.I,
the second part of the power energy unit - the
pp.74-81. Zadeh L. (1965) Fuzzy Sets and Systems, Proc.
turbine. The first results are expected by the end of
Symp. Syst. Theory, Polytech. Inst. Brooklyn,
this year.
pp.29-37. Zhao Z.Y., M. Tomizuka, S. Isaka (1993) Fuu:y Gain
Scheduling of PID Controllers. IEEE Trans.
REFERENCES
Systems, Man and Cybernetics, vo1.23, No.5, pp.I392-1398.
Arabas J., Bialobrzeski L., Domanski P.D., Swirski
Zimmermann H.-J (1991) Fuu:y Set Theory and Its
K. (1998) Pulverised Boiler Optimization and
NO x
Reduction
by
Neural
Application.
Networks,
Publishers.
Pragoregula'98, pp.61-69, Praha. Astrom
K.J.,
Wittenmark
B.
(1994)
Adaptive
Control, Addison Wesley, Reading, MA. Bartos FJ., Artificial Intelligence (1997) Smart
Thinking
for
Complex
Control,
Control
Engineering, July 1997. Czogala E, Pedrycz W (1982): Control problems in
fuzzy systems, Fuzzy Set and Systems 7, pp.257274,. Domanski P.D., Brdys M., Tatjewski P (1997) Fuzzy
logic multi-regional controllers - design and
146
Boston,
Kluwer
Academic