Copyright 2002 IFAC

15th Triennial World Congress, Barcelona, Spain

BOILER FIRING CONTROL DESIGN USING MODEL

PREDICTIVE TECHNIQUES 1

 2

Wen Tan , Tongwen Chen , Horacio J. Marquez

Department of Automation, North China Electric Power University,

Beijing, China
Department of Electrical & Computer Engineering, University of
Alberta, Edmonton, AB, Canada T6G 2V4

Abstract: PI controllers are typically used in industry for boiler firing rate control for their
simplicity and ease in tuning; this is the case with the utility boilers at Syncrude Canadas
northern Alberta plant. However, instability often occurs in cases of large (load) disturbances,
primarily due to firing rate limit constraints. Using a Syncrude Canadas utility boiler as
an example, we attempt to redesign the firing rate controller. We show that stability and
performance of the closed-loop system can be improved to some extent by properly designed
PID controllers. For further improvement, we adopt a model predictive control (MPC) scheme
which is capable of handling the firing rate constraints directly; a simple MPC algorithm is
implemented on a nonlinear simulation package, and significantly better results are achieved.
Copyright 2002 IFAC.

Keywords: Boiler control, model predictive control, PID design, control applications.

1. INTRODUCTION steam to meet certain steam load demand. Notice

that boilers are usually connected with other compo-
Drum boilers are used a lot in industry in order to nents/systems and steam demands are usually deter-
generate steam; their dynamics and control are widely mined by these systems: For examples, in a unit power
studied, see, for example, (Dukelow, 1991; Maffez- plant, a boiler is typically connected to a turbogenera-
zoni, 1997; Astrom and Bell, 2000; Pellegrinetti and tor, and the steam demand on the boiler should follow
Bentsman, 1994). A drum boiler control system usu- the electrical power demand associated with the tur-
ally consists of the following subsystems: combustion bogenerator; in a co-generation plant, several boilers
control, drum level control, steam temperature control, are connected to a main header; to maintain the main
and furnace draft control. It is well-known that in a header pressure, the steam demand on each boiler is
drum boiler, the drum level can be efficiently con- distributed from the total steam demand on the main
trolled by a three-element level control structure, the header. So boiler combustion control is often designed
steam temperature by a cascade temperature control together with the connected systems. In a unit power
structure, and the furnace draft pressure by a forced plant, the control of both the boiler combustion and
and induced draft control system. These controllers for the electrical power generated by the turbogenerator
the subsystems are typically designed separately. is usually referred to as the coordinated control; while
The objective of a boiler combustion system is to in a co-generation plant, the combustion control has
ensure that the boiler generates sufficient amount of two parts: steam control (to regulate the steam flow)
and pressure control (to regulate the main header pres-
sure). Since boiler combustion is a slow process, while
1 This research was supported by Syncrude Canada Ltd. and the the demand for steam is usually required to be met as
Natural Sciences and Engineering Research Council of Canada. fast as possible, these conflicting facts make the boiler
2 Corresponding author: Tel: (780)4923940; Fax: (780)4921811;

Email: tchen@ee.ualberta.ca
combustion controller design a challenging problem,
especially in cases when there exist physical limit firing rate constraints, we show that a simple model
constraints on the firing rate. predictive control (MPC) strategy is most effective.
Our paper is motivated by a practical situation with
Syncrude Canadas utility plant in Fort McMurray,
Alberta: Because of the firing rate limit constraints, 2. MODELING FOR BOILER FIRING CONTROL
the existing PI type of combustion controllers exhibit
unstable behavior when large disturbances are present To facilitate the control design, we will derive a simple
in the system. Improving stability and performance model to be used in the design of the boiler master
of the firing rate control system is the goal of this firing controller.
investigation.
We note that in the utility plant the CO-type boilers
Some discussion of the overall system at the Syncrude must burn out the coker-off gases from other pro-
utility plant is necessary. The plant currently has three cesses; they are operated with fixed steam load. The
utility boilers, three CO-type boilers, and two once- OTSGs must burn out the waste gases from the gas
through steam generators (OTSG). The total steam turbines; their loads are usually fixed too. So the con-
generated by the boilers is gathered in a 900# header, tributions from CO-type boilers and OTSGs to the
whose pressure should be maintained at 6.306MPa 900# header pressure can be ignored. To regulate the
for normal plant operation. Steam at other levels of 900# header pressure, we need to consider only the
pressure (4.24MPa, 1.0682MPa and 0.3584MPa) is utility boilers. Furthermore, since the three utility boil-
obtained by four letdown stations, namely, from 900# ers in the plant are of the same capacity and are oper-
header to 600# header, from 600# header to 150# ated in parallel with equal load, we need to study just
header, from 600# header to 50# header, and from one of them. Based on the operational conditions, we
150# header to 50# header. A simple diagram of the assume that the air and fuel flow controller functions
steam system in the utility plant is shown in Figure 1. sufficiently well so that we can treat the firing rate as
the input to the boiler.

900# steam
  Near operating conditions of the utility boiler, the
Utiltiy CO
OTSG's
Boilers Boilers transfer function from the firing rate, denoted BUB ,
to the steam flow rate, denoted DUB , can be approx-
 imated as a first-order system with a time delay:
900# Header
 Atmosphere
kUB
 U B s
DUB BUB 
900/600
letdown  e (1)
valve
   G6 Vent
Tumblers TUB s  1
G1 G2 G4

PV25
The dynamics of a header can be simply expressed by
  150/50 50# steam
600/50 50# Header
letdown
valve letdown
valve dPheader  inlet steam  outlet steam
 dt header capacity 

600/150 150# Header 150# steam

letdown
valve where Pheader is the header pressure. Based on this we
 
600# Header
600# steam obtain the equation for the 900# header:

 DBoiler  D900
Fig. 1. A simple diagram of steam system in the P900  (2)
Syncrude utility plant C900 s
Here, C900 is the capacity of the 900# header, D900 is
For this co-generation plant, the utility boilers are used the total steam demand for the 900# header, including
to regulate the 900# header pressure, while the CO- steam demand from electricity generation and other
type boilers are operated in the steam regulation mode. headers, DBoiler is the total steam generated by the
The combustion controller for each boiler is composed boilers. By the argument above, we have DBoiler 
of two parts: a master firing controller which generates DUB .
a firing rate command, and a fuel-air flow controller Replacing DBoiler in (2) by DUB and then eliminat-
which generates desired fuel flow rate and air flow ing DUB by considering (2), we arrive at the follow-
rate according to the firing rate command. The existing ing model for the firing control design:
master firing rate controller for the utility boilers is of

PI type, with the parameters manually tuned. As we P900 GBUB  Gd D900 (3)
commented earlier, the closed-loop system exhibits
with
instability when there are large steam demands its
performance in regulating the 900# header pressure in
kUB U B s
these cases is unacceptable. We show in this paper that G  e
we can improve the firing rate control performance TUB s  1  C900 s 

by introducing PID type firing rate controllers and  1

Gd
fine tuning them. Furthermore, to directly handle the C900 s 
At the operating point where each utility boiler gen- parameters shown in Table 1 are recommended by the
erates 90.6kg/s steam, the parameters involved in the references.
above model are estimated as follows:
Figures 2 shows the responses of the closed-loop
   
kUB 4 TUB 80 UB 12 C900 441  systems with the three PID controllers due to a step
  
disturbance of magnitude 1, where the linear model
As we mentioned earlier, firing rate limits are im- in (3) is used and the controller output (firing rate)
portant constraints to be considered in control de- constraints are ignored. The PID controller tuned by
sign; these reflect physical characteristics of the boiler the W-C method has a very weak integral action so
which prevents the firing rate from responding as fast the output takes a long time to return to its setpoint.
as desired. These limits are given by The T-L-T method has the largest proportional gain
and integral action, and thus has the best disturbance
 0  16  60  BUB  0  16  60 rejection ability; however, the controller response is

0  BUB  1 very aggressive. In fact, if we consider the controller
rate limit constraints, none of the three PID controllers
The above rate limit and saturation constraints are shown in Table 1 will make the closed loop system
critical for controller design and closed-loop perfor- stable.
mance.  ' 

 ' 
3. PID CONTROL DESIGN  ' !

 ' #
First, we will try to introduce a derivative term in the
existing PI controller and re-tune the resultant PID  ' %

controller for firing rate control. There are two reasons  '
for doing so: First, we would like to see how much  ' 
improvement is possible using just PID controllers,
 ' !
which are still quite implementable in practice; sec-
ond, this sets up a performance benchmark for more  ' #     ! " # $ % & 
advanced control scheme such as the MPC technique
which we will study in the next section. Process output vs. time (sec.)
Since the utility boilers are used to maintain the 900# Fig. 2. PID control (solid: P-P; dotted: T-L-T; dashed:
header pressure, the main objective of the firing rate W-C)
control design is to reject the disturbance, namely, the
steam demand at 900# header. Note that the model If we use anti-windup PID configuration (Hanus,
in (3) is an integrating process. Unlike the case of 1980) for the PID controllers shown in Table 1, the
stable processes, there are few methods for tuning PID closed-loop systems become stable but the responses
controllers for integrating processes. Nevertheless, we are too oscillatory to be acceptable. One can detune
have found three methods in the literature which can the PID parameters in order to get better responses;
be directly used here. Table 1 shows the PID parame- however, due to space considerations, we will not in-
ters tuned by these methods. clude the results in this case.
Table 1. PID parameters tuned by three
different methods
4. MODEL PREDICTIVE CONTROL DESIGN
Tuning Parameters Kp Ti Td
P-P Mr  4 3.322 222 51.2
W-C  0  707   0  7 2.591 732.4 13.3 In the previous section we mentioned that anti-windup
T-L-T  0 1 5.86 162 40.5 PID controllers can be detuned to accommodate the
firing rate constraints. However, two major problems
exist:
The P-P method (Poulin and Pomerleau, 1996) uses
(1) Since the PID parameters are detuned because
the maximum peak resonance (Mr ) of the closed-loop
of constraints, the performance may be poor in
system as a specification. A higher Mr indicates that
cases such as small load disturbances, where the
the system is less damped and has larger overshoots.
constraints are not active.
The W-C method (Wang and Cluett, 1997) has two
(2) The anti-windup scheme does not work well for
tuning parameters: , the damping factor, and , the
very large (load) disturbances.
time constant of the desired control system. The T-
L-T method (Tan et al., 1998) has one parameter In order to take full advantage of the boiler capac-
which reflects the trade-off between the systems ity, we need to design a firing rate controller which
time-domain performance and robustness. The tuning achieves optimal performance whether the constraints
are active or not, and no matter how large the load dis- next sampling time t  1, the optimization process is
turbances are. This calls for a method which explicitly repeated a receding horizon control scheme. In the
handles the constraints in controller design; the model special case when there are no constraints, the solution
predictive control framework is a good candidate for to the optimization problem can be obtained analyti-
this purpose. cally, and the controller can be implemented easily in
a feedback form. However, if there are input or output
Model predictive control (MPC) was originally pro-
constraints, the solution requires a QP solver, which
posed in the process control community and has
is available in, e.g., the MPC Toolbox associated with
had wide applications in the process control industry
Matlab.
(Camacho and Bordons, 1999); MPC is, perhaps, the
most general way of posing process control problems In our design, we choose the following parameters:
in the time domain, and the only technique accepted in
p  100 m  10 y 
1 u 
1
the process industry for handling multivariable control   
systems. The relevance of MPC in this project lies For a step (load) disturbance of magnitude 1 (using
in two aspects: First, it is capable of directly incor- the linear model), we see in simulations that closed-
porating input and output constraints, and slew rate loop responses for the MPC/DMC controller with
constraints in the optimal control law design; second, constraints is almost the same as the PID controller
there exists an efficient solution technique quadratic tuned by the P-P method. However, for small and
programming (QP) for the resultant optimization large disturbances, the MPC controller is much better
problem. than the PID controller, see, e.g., Figure 3 for a step
disturbance of magnitude 2. These clearly show that
For our application, we adopt a special MPC tech-
the MPC controller handles the constraints effectively,
nique, the so called DMC (dynamic matrix control)
and good performance is achieved regardless of the
approach proposed by Cutler and Ramaker (1980). In
magnitudes of disturbances.
this method, open-loop step responses are required for
3 96
the algorithm, which can be obtained easily by either
sampling continuous-time models, or by identification 3 95
in discrete time. The step response model to be used
has the following form: 3 94

 
y t 
gi u t  i

(4) 3
i( 1
3 94
where y and u are the output and input in discrete
time, gi are the step response coefficients,

and

is the 3 95
discrete-time difference operator: u t   u t ) u t 
1  . Note that we have selected the sampling period to 3 96

be 2 seconds, which is quite reasonable in view of 3 97 3 433 533 633 733 8333 8433 8533 8633 8733 4333
the DCS system currently in use with the Syncrude
plant. In order to have a fair comparison with the PID Process output vs. time (sec.)
controllers in Section 3, we will use the step response
model obtained by directly sampling the continuous- Fig. 3. MPC (solid) and PID (dash) control
time model discussed in Section 2 in the DMC design.
The objective of DMC is to drive the output as close In the Syncrude boiler system, it is a fact that the steam
to the setpoint as possible in a least-squares sense with demand (load disturbance) is measured. A natural
the possibility of the inclusion of a penalty term on question to ask is whether the performance of the
the control moves, i.e., at sampling time t, we need closed-loop system can be improved if the disturbance
to minimize the following cost function to obtained model is incorporated in the DMC design. The answer

u t  j  1  ( j  1 2 m  1): is no; the reason is due to the firing rate constraints:
 *+*+*, Since the initial response of the controller has already
p reached its maximum capacity, advance prediction
  
J t 
y - y t  j . t / w t  j 10 2  of the disturbance using the model will not help in
j( 1 speeding up the response.
m 
u - u t  j  1 20 2  (5)
j( 1

5. CLOSE-TO-REAL SIMULATION
Here

y t  j . t  is the j-step prediction of the output,
w t  is the reference trajectory, p is the output predic- The simulation results reported in the preceding sec-
tion horizon, m is the input prediction horizon, y is tion were based on the linear model discussed in Sec-
the output weighting, and u is the input weighting. tion 2; while they serve the purpose of illustrating
When the optimal 
control moves are computed, only possible performance improvement by re-turning PID
the first u t + is in fact implemented; and at the controllers and by designing MPC controller in the
 Q RS
boiler firing control system, extensive testing on the
actual Syncrude boiler system is desirable in order
to justify the performance and robustness properties Q RP M
of the designed controllers, the reason being that the
actual system is a complex nonlinear system and the Q RP
linear model is just an approximation of the real plant
at a certain operating point. Unfortunately, such real
tests are not possible at this stage, mainly due to safety Q RO M

reasons in the utility plant. However, Syncrude has

available a complex and nonlinear simulation package Q RO
called SYNSIM (Rink et al., 1996), which is built
in Matlab/Simulink environment and has undertaken
Q RN M L M NL NM OL OM PL
thorough tests to closely reflect the real plant. In this
section, we report the testing results done on SYNSIM
(a) 900# Header Pressure(MPa) vs. time(min.)
for the performance and robustness of the PID and V
DMC controllers designed earlier.
T Y] U
= @B ;
T Y]

T Y\ U
= @B ?
T Y\

= @B T Y[ U

T Y[

= @; A T YZ U T U VT VU WT WU XT

(b) Utility Boiler Firing Rate vs. time(min.)

= @; > : ; < = > ?: ?; ?< ?= ?> ;:
Fig. 5. Comparison of MPC and PID controllers for
(a) 900# Header Pressure(MPa) vs. time(min.) a 25.2kg/s steam demand change on the 900#
C IG D
header (solid: MPC; dashed: PID)
C IG
K
C IJ G The total steam load for the 900# header is
415kg/s.
C IJ F K
The total steam loads for the 600#, 150# and
50# headers are 88kg/s, 16.5kg/s, and 215kg/s,
C IJ E
respectively.
K
C IJ D The total electricity generated by the steam tur-
bines is 230.2MW.
C IJ
The responses resemble the simulation results we re-
C IF G C D E F G HC HD HE HF HG DC
ported in the preceding sections based on the linear
model. The DMC controller has very good perfor-
(b) Utility Boiler Firing Rate vs. time(min.) mance though the model used in DMC design is a
simple step response one.
Fig. 4. Comparison of MPC and PID controllers for a
6.3kg/s steam demand change on the 900# header Sudden large changes on the steam demand that we
(solid: MPC; dashed: PID) considered could come from failures found in some
subsystems in the plant. For example, in the current
plant operation, an OTSG or a steam turbine might
Figures 4 and 5 show the responses of the 900# header
trip due to some unknown reasons; since an OTSG
pressure for a step steam demand on 900# header
generates steam at 900# level, while a steam turbine
with small and large magnitudes, respectively. The
consumes 900# steam, a trip would cause a sudden
operating condition is specified as follows:
large change on 900# steam demand. More specifi-
K
Two utility boilers, three CO-type boilers, and cally, at the operating point, an OTSG trip amounts
two OTSG boilers are on-line, with each util- to a 12.8kg/s steam demand increase on 900# header.
ity boiler generating steam at 90.6kg/s, each To maintain system stability in such cases, we need
CO-type boiler at 69.4kg/s, and each OTSG at the utility boilers to increase or decrease the steam
12.8kg/s. generation as fast as possible. The response of the
utility boiler firing rate controller shows that the DMC We remark that the current PI firing rate controller
controller can take advantage of the full capacity of the used in the plant is unable to maintain the system
utility boiler; the time required for the DMC controller stability in case of a CO-type boiler trip or a steam
to bring back the 900# header pressure is about half turbine trip. The SYNSIM testing results clearly show
of that for the PID controller. The larger the sudden the advantages of the DMC controller over PID type
steam change is, the more advantageous the DMC of controllers in the extreme situations.
controller is.
We have conducted simulations on SYNSIM for an
6. CONCLUSIONS
OTSG trip and a CO-type boiler trip. The responses
of the 900# header pressure for the case of an OTSG
In this paper, we studied PID and MPC controllers for
trip and a CO-type boiler trip are similar to Figure 5,
firing rate control of an industrial boiler in Syncrude
hence they are omitted to save space.
Canadas utility plant in Fort McMurray, Alberta. Ex-
A steam turbine trip amounts to a steam demand de- tensive testing showed that both controllers can im-
crease on 900# header; Figures 6 show the case of one prove plant stability and performance due to sudden
steam turbine trip. Since the turbine extracts 83.8kg/s change in steam demand; but we highly recommend
steam, the loss of the steam load is so large that it the MPC controller which has superior performance
cannot be handled by two utility boilers. The vent and stability and handles controller constraints effec-
system at 50# header is invoked to send some steam tively.
to the atmosphere. As shown in Figure 6(b), using
the MPC controller, less steam is sent to the atmo-
sphere, compared with the PID controller; which is 7. REFERENCES
quite desirable for environmental considerations. We
Astrom, K. J. and R. D. Bell (2000). Drum-boiler
also observe that although there is less steam sent to
dynamics. Automatica 36(3), 363378.
the atmosphere, the MPC controller has a considerably
Camacho, E. F. and C. Bordons (1999). Model Predic-
better performance than the PID controller.
tive Control. Springer-Verlag. London.
Cutler, C. R. and B. C. Ramaker (1980). Dynamic Ma-
c d_
trix Control - a computer control algorithm. In:
Proc. Automatic Control Conference. San Fran-
c de _
cisco.
Dukelow, Sam G. (1991). The Control of Boilers. 2nd
c de
ed.. Instrument Society of America. Research
Triangle Park, NC.
c db _
Hanus, R. (1980). A new technique for preventing
control windup. Journal A 21(1), 1520.
c db
Maffezzoni, C. (1997). Boiler-turbine dynamics in
power-plant control. Control Engineering Prac-
c da _ tice 5(3), 301312.
Pellegrinetti, G. and J. Bentsman (1994). H con-
c da ^ troller design for boilers. Int. J. Robust and Non-
_ `^ `_ a^ a_ b^
linear Control 4, 645671.
(a) 900# header pressure (MPa) vs. time (min.) Poulin, E. and A. Pomerleau (1996). PID tuning for in-
mf tegrating and unstable processes. IEE Proc. Con-
trol and Applications 143(5), 429434.
lf
Rink, R., D. White, A. Chiu and R. Leung (1996).
gf SYNSIM: A computer simulation model for the
Mildred Lake steam/electrical system of Syn-
kf crude Canada Ltd.. Technical report. University
of Alberta.
jf
Tan, W., J. Z. Liu and P. K. S. Tam (1998). PID
if
tuning based on loop-shaping H control. IEE
Proc. Control and Applications 145(4), 485490.
hf Wang, L. and W. R. Cluett (1997). Tuning PID con-
trollers for integrating processes. IEE Proc. Con-
ff g hf hg if ig jf trol and Applications 144, 385392.

(b) Vent steam vs. time (min.)

Fig. 6. MPC (solid) and PID (dash) control for a steam
turbine trip