IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO.

5, SEPTEMBER 2002

735

Multivariable Robust Controller Design for a Boiler System

Wen Tan, Horacio J. Marquez, and Tongwen Chen

AbstractIn an industrial boiler system, multiloop (decentralized) proportional-integral (PI) control is used because of its
implementational advantages. We show that such control schemes
sacrifice robustness and performance of the overall system. In
particular, under normal boiler operating conditions, we deloop-shaping
sign a robust multivariable controller using
techniques; for consideration in implementation, we then reduce
this controller to a multivariable PI controller. Both the
controller and its PI approximation are tested extensively in
the frequency domain as well as in the time domain, using a
complex nonlinear simulation software; the results show that the
designed controllers are superior in robustness and performance
to the existing multiloop controller.
Index TermsBoiler systems, control applications, controller
reduction,
optimization, loop shaping, robust control.

I. INTRODUCTION

NDUSTRIAL cogeneration systems invariably present

a challenge to control system designers. The Syncrude
Canada, Ltd. (SCL) integrated energy facility located in Mildred Lake, Alberta, utilizes a complex header system for steam
distribution, which includes headers at four different pressure
levels (6.306, 4.24, 1.068, and 0.372 MPa).
The 6.306-MPa header receives steam from three utility-type
(UB) boilers burning refinery gas, three CO-type boilers burning
coker off gas and refinery gas, and two once-through steam
generators (OTSGs). The steam is then distributed through the
header system to several steam turbines to generate electricity.
The overall plant, like many similar available worldwide, is thus
a rather complex nonlinear interconnected system. A simple
diagram of the utility plant is shown in Fig. 1.
The normal plant operation requires tracking the steam demand while maintaining the steam pressure and the steam
temperature of the 6.306-MPa header at their respective setpoints, despite variations of the steam load. Due to the physical
characteristics, utility boilers are used to regulate the steam
pressure. At present, utility boilers are controlled via a multiloop
(decentralized) proportional plus integral (PI) type controller.
These configurations, however, ignore the fact that there exist
interactions among the variables to be regulated. The result is
that in normal operation, the 6.306-MPa header pressure exhibits oscillatory modes that the controller is unable to damped
out as quickly as desired. This motivates us to redesign the
UB boiler controllers to improve overall system performance.
In this paper, we investigate the use of modern multivariable
control techniques as applied to cogeneration systems such as
Manuscript received April 25, 2000. Manuscript received in final form May
10, 2002. Recommended by Associate Editor K. Schlacher.
The authors are with the Department of Electrical and Computer Engineering,
University of Alberta, Edmonton, AB T6G 2G7, Canada (e-mail: wtan@ee.ualberta.ca; marquez@ee.ualberta.ca; tchen@ee.ualberta.ca).
Publisher Item Identifier 10.1109/TCST.2002.801787.

Fig. 1. A simple diagram for the Syncrude utility plant.

the one described here. More explicitly, we will investigate the

following issues.
Multivariable Control Design: We will proceed to design a
multivariable controller for the utility boiler and compare
our new controller to the existing one. The new controller
should be robust, i.e., it should maintain stability as well
as performance despite the existence of modeling errors.
Optimization and Controller Reduction: Our design

will be cast as an
optimization problem. Undoubtedly,
optimization has been the leading design approach in robust control over the last two decades. It is
known, however, that the order of the resulting controller
using this approach is no less than that of the original plant
to be controlled. There is some reluctancy in industry toward the use of high-order controllers, due to complexity
of implementation and difficulty associated with possible
retuning. Thus, we will approximate the optimal controller
by a multivariable PI type controller. It will be shown that
the new PI multivariable controller retains the features of
the optimal control with little performance deterioration.
Simulation and Time-Domain Performance Comparison:
As mentioned, the true system is nonlinear and rather
complex. Under normal operation, however, the plant is
expected to maintain its output variables at a prespecified
value. It follows, then, that only small excursions with
respect to nominal operating conditions are expected and
linear time-invariant (LTI) models constitute a fairly good
approximation of the true nonlinear plant. This argument
serves to justify the approach used to design the controller,
which is entirely based on frequency domain specifications
and manipulations. The final result, however, should be
tested under more rigorous conditions. Namely, it should

1063-6536/02$17.00 2002 IEEE

736

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 5, SEPTEMBER 2002

be tested with a more accurate description which should

incorporate the nonlinearities encountered in the true
plant.
Several boiler models have been proposed in the recent
years, e.g., [1][3], and there are several simulation packages
for steam plants, e.g., [4]. Syncrude Canada, Inc. has also
available a simulation package, known as SYNSIM [5]. The
SYNSIM model was developed with the purpose of simulating
certain upset conditions that have been sporadically detected,
as well as a general tool for stability analysis. The model has
been extensively tested, and correlation between measurements
from the true plant outputs and predictions by SYNSIM are
excellent. The final controller will be simulated in SYNSIM
and compared to the existing design under fairly realistic conditions. We should point out that while SYNSIM constitutes an
invaluable analysis tool, it cannot be used directly for controller
design due to the excessively large complexity of the models
used in the simulation. Thus, the first step of our design will
be to obtain a simple LTI model approximation of the utility
boiler, based on input-output measurements.
The rest of this paper is organized as follows: in Section II
we describe the internal structure of the utility boiler, define
the control problem, and obtain an LTI approximation of this
controller design. In Secsystem. Section III deals with the
tion IV, this controller is reduced to a PI type one, a constraint
imposed by practical implementation issues. The new PI controller will be shown to have similar characteristic to the
controller within the plant bandwidth. Section V contains an
analysis of the results both in the frequency domain and in the
time domain using the simulation package SYNSIM. Section VI
contains the conclusions and final remarks.

this type of boiler is shown in Fig. 2 (where the arrow points out
the direction of the steamwater flow).
As shown in Fig. 2, the principal input variables are
feedwater flow rate (kg/s);
fuel flow rate (kg/s);
attemperator spray flow rate (kg/s);
and the principal output variables are
drum level (m);
drum pressure (MPa);
steam temperature ( C).
The following basic requirements must be satisfied for proper
functioning of the boiler system.
1) Steam pressure of the 6.306-MPa header must be
maintained despite variations in the amount of steam
demanded by users.
2) The amount of water in the steam drum must be maintained at the desired level to prevent overheating of drum
components or flooding of steam lines.
3) The steam temperature must be maintained at the desired
level to prevent overheating of the superheaters and to
prevent wet steam entering turbines.
The nominal operating conditions are specified as follows:
1) Two utility boilers, three CO boilers, and two OTSGs are
on-line. The two utility boilers are operated in parallel at
equal load.
2) The total steam load for the 6.306-MPa header (sum of the
6.306-MPa steam load, the 6.306-MPa4.2 MPa steam
flow, and the unmodeled 6.306-MPa steam load) is 315.08
kg/s.
3) The total steam load for the 0.372 MPa header is 155.69
kg/s.
4) The total steam turbine electrical output is 145.53 MW.
At this operating point, for each of the two UBs, we have

II. UTILITY BOILER

The utility boilers in the plant are watertube drum boilers.
This type of boiler usually comprises two separate systems.
One system is the steamwater system, which is also called
the water side of the boiler. In this system preheated water
from the economizer is fed into the steam drum, then flows
through the downcomers into the mud drum. The mud drum
distributes the water to the risers, where the water is heated
to saturation conditions. The saturated steam-water mixture
then reenters the steam drum in which the steam is separated
from the water and exits the steam drum into the primary and
secondary superheaters. In the two superheaters, the steam
is further heated and then is fed into the 6.306-MPa header.
In between the two superheaters is an attemperator which
regulates the temperature of the steam exiting the secondary
superheater by mixing water at a lower temperature with the
steam from the primary superheater.
The other system is the fuelairflue gas system, which is
also called the fire side of the boiler. In this system, the fuel and
air are thoroughly mixed and ignited in a furnace. The resulting
combustion converts the chemical energy of the fuel to thermal
or heat energy. The gases resulting from the combustion, known
as the flue gases, pass through the superheaters, the risers, and
the downcomers, and leave the boiler. A schematic diagram of

Notice that the normal setpoint of the steam temperature is

499 C and at the operating point considered (466.7 C) the
steam temperature is below the set point value. Therefore,
under these conditions the attemperator valve is closed and
the spray flow is not used. Despite these facts, the spray flow
should not be ignored for controller design since 466.7 C is
not the only operating point used at the plant. Often the boiler
is operated at a higher load, where the steam temperature will
typically exceed the set point. Thus, to operate the controller at
higher operating points where the spray flow will take action,
we considered the full 3 3 system.
To have a comparison with the existing controller, we assume
the steam temperature setpoint as the current temperature. In addition, the following limit constraints exist for the three control
variables:

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 5, SEPTEMBER 2002

737

Fig. 2. Utility boiler.

We note that
at this operating point is small compared
with the magnitude limit, so these limits do not impose hard
constraints for design. However, the rate limit for fuel flow rate
( ) has a significant impact on the system performance.
For this operating point, collecting inputoutput data on
SYNSIM and using the MATLAB Systems Identification
Toolbox, we have identified the following LTI model shown in
(1) at the bottom of the page.
After testing several alternative model structures, it was
found that a second-order model can fit the input output data
fairly well. Pure delays were found to be relatively insignificant
and were ignored. Several tests were conducted to validate this
model using SYNSIM. To show one such a test, we injects a
random input sequence with 20-s sampling interval to each of
the three input variables, then compare the outputs generated
from SYNSIM with the corresponding ones predicted by the
LTI model. The results are shown in Fig. 3. From this experiment we conclude that indeed the LTI model constitutes a fairly
good approximation in the vicinity of the operating conditions.

III. CONTROLLER DESIGN

Various control techniques have been applied to boiler or
boilerturbine controller design, e.g., inverse Nyquist array
[6], linear quadratic Gaussian (LQG) [7], LQG/loop transfer
approach [9], inrecovery (LTR) [8], mixed-sensitivity
ternal-model control (IMC) [10], and predicative control [11].
approach introduced
Here, we will adopt the loop-shaping
by McFarlane and Glover [12]. The essential features of this
approach can be summarized as follows.
It can incorporate both performance and robust stability
requirements.
It is conceptually and computationally simple and retains
many of the features encountered in the design of classical
PI loops using frequency domain techniques. This is important, since experienced engineers working in industry
can incorporate their knowledge of the plant and classical
control design into modern powerful techniques.

(1)

738

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 5, SEPTEMBER 2002

robust stability if
is sufficiently large. The value
is used as a design indicator; usually it should be
-optimizabetween 0.3 and 0.5. (Notice, also, that the
tion problem in (2) can be solved explicitly without iteration, using only two Riccati equations.)
3) The final feedback controller is obtained as

Fig. 3.

Outputs of the utility boiler (solid) and model (dashed).

Given a plant model , the design approach consists of three

steps.
1) Loop Shaping: Using pre- and/or postcompensators
and
, the singular values of the original plant are
a
shaped to give the compensated plant
desired open-loop shape. This step contains all of the ingredients of the classical techniques. The shaping funcand
are controlled by the designer and the
tions
properties of the resulting controller depend upon these
functions in an essential manner. Guidelines for choosing
and
may be found in [13].
2) Robust Stabilization: A feedback controller which robustly stabilizes the shaped plant is found. There is no
need for further design parameters here. There are infinite many controllers which stabilize the shaped plant.
optimization
One such controller is singled out via
to further maximize the stability margin. More explicitly,
for the shaped plant , we solve the following
optimization problem given as

We now apply this method to the utility boiler model. We start

by scaling the model so as to improve the condition number of
the plant, which is not good because the coefficients of the first
row and the first column of the model are too small compared
with other columns and rows. The scaling is very important in
loop shaping design, since for a multivariable control system
each channel has its own unit, the magnitude will probably be
in a large range. One possible way to avoid scaling is to use the
sensor spans and/or actuator spans, which are already properly
scaled. Our model are identified directly from the output measurements, so it needs scaling. In fact, by scaling the drum level
by a factor of 100, the drum pressure by 1000 and the feedwater
flow rate by ten, the condition number of the system frequency
response matrix is improved by about 30 dB at all frequencies.
and set
For the scaled model, we will choose
. Here,
serves as a static decoupler, which
can be found by well-known algorithms such as those given
by [14][16]. We chose to use the ALIGN algorithm proposed
in [15]. The algorithm consists of finding a real matrix which
approximately inverts a complex matrix. Though the plant
has a pole at the origin, we can still compute the approximate-inverse of the frequency response matrix at a certain
rad/s.
is a diagonal
frequency. Here we chose
PI compensator that determines the desired open-loop shapes.
It can also be chosen as a decentralized PI for a decoupled
system, where the functions of the proportional gain and
the integral action are well known. Finally, the following
precompensator was selected:

(3)
. The
The final design indicator for this design is
singular values of the open-loop transfer matrices are shown
in Fig. 4. We observe that the designed loop shapes (solid) do
not change much from the desired (dashed), especially near the
crossover frequency, indicating a good design.

stabilizing
(2)
The importance of this minimization is the following:
is a normalized left coprime
Suppose that
factorization of . It can be shown [12] that the controller
obtained in (4) will guarantee stability of any plant
which belongs to the family of plants
defined as follows:

This also guarantees that the loop shape we selected in

the previous step can be well approximated with good

IV. CONTROLLER REDUCTION

A minimal state-space realization of the designed
controller has order 18. Practical implementation issues dictate the
need to investigate the performance of a reduced order controller. Moreover, all controllers presently used in the plant are
of the PI type. This structure is thus familiar to the operators and
it is easy to implement. Thus, in this section we investigate the
performance of a reduced order multivariable controller of the
PI type. A secondary reason for the preference of PI type
controllers is that antiwindup implementation is relatively easy,

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 5, SEPTEMBER 2002

739

Here we assume
since we are most interested in
the low-frequency band. So we have

It is clear, then, that based on this reduction procedure, the

resulting PID controller achieves good approximation of the
controller at low frequencies.
In the present case, our goal is to get a PID type controller
controller designed earlier.
to approximate the 18th-order
The PID parameters are given as

Fig. 4. Singular values for the open-loop frequency response matrices. (Solid:
controller).
desired. Dashed:

considering practical constraints such as rate limits and saturations of the control inputs.
, given by a state-space realConsider now a controller
ization of the form
Since, as explained earlier, our interest is in a PI structure,
we arbitarily neglect the derivative component. The resulting
controller is given by
Let the rank of the matrix
be . Note that
since by
in (3) into the controller,
always has at
incorporating
least one eigenvalue at the origin. The multiplicity of the zero
is
. We assume that there exist
eigenvalue of
linearly independent eigenvectors for this zero eigenvalue. Then
we find a similarity transformation such that

where
is nonsingular. This transformation can be
computed using the eigenvalue decomposition of . With this
, the new state-space realization is given by

with

and

Fig. 5 compares the singular values of the open-loop frequency

] for both the
controller
response matrices [
and its PI approximation
. It is clear that both controllers
have similar characteristics within the plant bandwidth. Both
of these controllers were extensively simulated in the time and
frequency domain, with the PI structure showing little performance degradation. In the sequel, we will produce some of these
results, with an emphasis on the PI approximation.
Note the controllers above are designed for the scaled model.
The final PI controller for the unscaled model is, thus, given by

A proportional-integral-derivative (PID) approximation of the

form

can now be obtained by truncating the MacLaurin expansion of

the controller with respect to the variable

For comparison purposes, a simplified structure of the present

controller is shown in Fig. 6. It utilizes some variables as feedforward terms and has cascade structures besides the decentralized control structure.

740

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 5, SEPTEMBER 2002

Fig. 5. Singular values for the open-loop frequency response matrices. (Solid:
controller. Dotted: PI controller).

Fig. 6.

Simplified structure of the present controller.

V. SIMULATION RESULTS
In this section, we evaluate the results obtained in the previous sections. We start with a frequency domain comparison
among the three different controllers, namely: 1) the multiloop
controller
PI controllers presently used in the plant; 2) the
of Section III; and 3) the PI controller approximation of Section IV. Note we must use the scaled model and the scaled controllers to compute the frequency domain properties, since for
unscaled model these properties are not in the same magnitude
for comparison.
and
Fig. 7 shows the maximum singular values,
, of the sensitivity function and the complementary
sensitivity function for the three controllers in consideration.
We start with a sensitivity analysis. According to Fig. 7,
the sensitivity function of the original controller has a slope

Fig. 7. Maximum singular values of and

controller, solid: PI controller).
Dashed:

T . (Dotted: original controller.

of 20 dB/decade at low frequencies, consistent with the PI

remains less than 0 dB
structure of the controller and
for frequencies less that 5 10 rad/s, approximately. More
has a peak value of approximately 19 dB
importantly,
rad/s. This means that plant disturbances
when
(in certain direction) at this frequency will be amplified roughly
by a factor of ten. By contrast, the sensitivity function obtained
controller of Section III and the reduced order
using the
PI controller of Section IV show a significant improvement.
First, the 0-dB crossover frequency has been extended from
5 10 rad/s to approximately 2 10 rad/s, thus extending
the maximum frequency for which disturbance attenuation is
is now about
actually attained. Second, the peak value
4 dB when using the PI controller and only about 3 dB when
controller. so disturbance rejection has been
using the
improved at low frequencies.
The analysis of the complementary sensitivity function
shows similar results. Indeed
for the original controller re(or equivalently, 0 dB) for frequencies
veals that
, approximately. This means that sensor noise
attenuation is achieved above this frequency and that robust stability will be assured for multiplicative perturbations whose amplitudes are significant only above this frequency. Once again,
and the reduced order PI controller outperform the
the
original one by a significant margin: The frequency at which
has now being reduced by about one decade and,
is almost 10 dB less than the corresponding
moreover,
values obtained with the existing controller for all frequencies
above 1 rad/s. Finally, notice that the peak of the complementary sensitivity function for original controller has also a peak
of about 19 dB, which has simply disappeared using either the
or the reduced order PI controller.
From the previous analysis, we conclude that a significant
conimprovement has been achieved with either the
troller or the PI controller. Moreover, no significant loss of
controller and the
performance was detected between the
PI approximation of Section IV. Thus, from now on we will

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 5, SEPTEMBER 2002

741

Fig. 8. Boiler time response for the designed PI controller. (Solid: 10%
drum level increase. Dashed: 5% drum pressure increase. Dotted: 10% steam
temperature decrease).

continue with the analysis of

only, which as mentioned
earlier is our preferred choice due to ease in implementation
and antiwindup configuration due to the fuel flow rate limit.
This frequency analysis, however, is based on the LTI approximation obtained in Section II and therefore contains no information regarding the performance of any of these controllers
connected with the true nonlinear plant. Clearly, this frequencydomain analysis is not suitable for this purpose. To this end,
and to test our controller under virtually true plant conditions,
we performed a series of test using SYNSIM, but replacing the
existing controller with the final PI controller.
Fig. 8 shows the step response of the utility boiler with regard to respective setpoint changes using the final PI controller.
Due to the rate limit of the fuel flow rate, the response for
the drum pressure is much slower. However, due to the antiwindup implementation, the response is smoother than that
without the constraints. We can also observe that the drum level
and the steam temperature setpoint changes have little effect on
the drum pressure.
Fig. 9 compares the step responses (the drum pressure and
drum level) for the designed multivariable PI controller and the
existing multiloop controller. We see that the multiloop controller has a slow rising time and a large overshoot for drum
pressure change. However, it has a fast rising time and less overshoot for drum level change. This is due to the fact that the original feedwater controller is a cascade one using the steam flow
as feedforward, which can improve the response of the drum
level. We also observe that the drum pressure drop due to the
change of the drum level for the multiloop controller is much
more than that for the designed PI controller.
The controller outputs are shown in Fig. 10. It is not surprising
to find that the performance improvement is achieved by more
aggressive control action.
To further test the controller we designed, suppose that the
unmodeled 6.306-MPa steam load increases by 20%, Fig. 11
shows the responses of the 6.306-MPa header pressure under
the original controller (dashed) and the designed multivariable
PI controller (solid). We observe that the designed PI controller
damps out the oscillation modes much more quickly.

Fig. 9. Boiler time response. (Solid: multivariable PI controller. Ddashed:

original multiloop controller).

Fig. 10. Controller outputs. (Solid: multivariable PI controller. Dashed:

original multiloop controller).

VI. CONCLUSION
In this paper, we have investigated the virtues and shortcomings of advanced multivariable design as opposed to decentralized PI loops for complex industrial systems. Two designs were
optimal MIMO control and 2)
obtained and tested: 1) an
a PI approximation of this optimal control. Since the optimal
control is expected to (and does) outperform the PI approximation, we only analyzed in detail the results corresponding to the
PI approximation. From the analysis of the results, we conclude
that the designed PI controller outperforms the existing one in
both robustness and performance.

742

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 5, SEPTEMBER 2002

[3] Z. Niu and K. F. V. Wong, Adaptive simulation of boiler unit performance, Energy Conversion Manage., vol. 39, no. 13, pp. 13831394,
1998.
[4] N. Serman and K. Mavracic, Flexible program package for simulation of once-through boiler steady states and dynamics, in Proc. 1st
Int. Forum Math. Modeling Comput. Simulation Processes Energy Syst.,
Sarajevo, Yugoslavia, 1990, pp. 573584.
[5] R. Rink, D. White, A. Chiu, and R. Leung, SYNSIM: A Computer Simulation Model for the Mildred Lake Steam/Electrical System of Syncrude Canada Ltd., Univ. Alberta, Edmonton, AB, Canada, 1996.
[6] L. Johanssson and H. N. Koivo, Inverse Nyquist array technique in the
design of a multivariable controller for a solid-fuel boiler, Int. J. Contr.,
vol. 40, pp. 10771088, 1984.
[7] R. Cori and C. Maffezzoni, Practical optimal control of a drum boiler
power plant, Automatica, vol. 20, pp. 163173, 1984.
[8] W. H. Kwon, S. W. Kim, and P. G. Park, On the multivariable robust
control of a boiler-turbine system, in Proc. IFAC Symp. Power Syst.
Power Plant Contr., Seoul, Korea, 1989, pp. 219223.
[9] G. Pellegrinetti and J. Bentsman,
controller design for boilers,
Int. J. Robust Nonlinear Contr., vol. 4, pp. 645671, 1994.
[10] T. B. Petrovix, D. D. Ivezic, and D. L. J. Debeljkovic, Robust IMC
controllers for a solid-fuel boiler, Eng. Simulation, vol. 17, no. 2, pp.
211224, 2000.
[11] B. M. Hogg and N. M. E. Rabaie, Multivariable generalized predictive
control of a boiler system, IEEE Trans. Energy Conversion, vol. 6, pp.
282288, Jan. 1991.
[12] D. C. McFarlane and K. Glover, Robust Controller Design Using
Normalized Coprime Factorization Description. New York:
Springer-Verlag, 1990.
[13] S. Skogestad and I. Postlethwaite, Multivariable Feedback Control:
Analysis and Design. New York: Wiley, 1996.
[14] M. Morari and E. Zafiriou, Robust Process Control. Englewood Cliffs,
NJ: Prentice-Hall, 1989.
[15] Y. S. Hung and D. C. McFarlane, Multivariable Feedback: A QuasiClassical Approach. Berlin, Germany: Springer-Verlag, 1982.
[16] T. J. McAvoy, Interaction Analysis. Research Triangle Park, NC: ISA,
1984.

Fig. 11. Responses for unmodeled steam load change. (Dotted: original
controller. Solid: multivariable PI controller).

REFERENCES
[1] J. M. Rhine and R. J. Tucker, Modeling of Gas-Fired Furnaces and
Boilers. New York: McGraw-Hill, 1991.
[2] M. E. Flynn and M. J. OMalley, Drum boiler model for long term
power system dynamic simulation, IEEE Trans. Power Syst., vol. 14,
pp. 209217, Jan. 1999.