Nonlinear Coal Mill Modeling and Its Application To Model Predictive Control

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Control Engineering Practice 21 (2013) 308–320

Contents lists available at SciVerse ScienceDirect

Control Engineering Practice


journal homepage: www.elsevier.com/locate/conengprac

Nonlinear coal mill modeling and its application to model predictive control
Andrea Cortinovis a,n, Mehmet Mercangöz a, Tarun Mathur b, Jan Poland a, Marcel Blaumann c
a
ABB Switzerland Ltd., Corporate Research, Baden-Daettwil, Switzerland
b
ABB India Ltd., BU Metals, Bangalore, India
c
ABB Germany Ltd., Power Systems, Mannheim, Germany

a r t i c l e i n f o a b s t r a c t

Article history: Coal mills play an important role in the overall dynamic response of coal fired power plants and there is
Received 3 April 2012 significant potential to improve the load ramp rates of coal fired power plants through improvements
Accepted 27 October 2012 of coal mill control strategies. This potential can be employed to compensate for the power fluctuations
Available online 5 December 2012
generated by renewable but intermittent energy sources such as wind and solar in a more efficient way.
Keywords: In this paper a three state coal mill model based on heat and mass balances as well as a single step
Coal mill modeling coarse to fine particle grinding relationship is presented with the purpose of predicting the dynamic
Coal fired power plant behavior of coal mills during both start-up and in normal operation. The parameters of the model were
Parameter and state estimation identified and later validated with measurements obtained from a hard coal fired power plant. During
Nonlinear MPC
these studies several parameters were found to be time varying. In order to estimate the values of these
Improving load ramp rates
time varying parameters and the internal states of the coal mill an extended Kalman filter was
designed. The proposed solution is observed to achieve very good agreement with measurements and
can be used for various applications such as model-based control, performance monitoring, fault
detection, and maintenance scheduling. In order to demonstrate one of these use cases a nonlinear
model predictive control (NMPC) application was developed based on the coal mill model and the
performance of the NMPC was compared to a conventional coal mill control strategy for tracking load
change references and for rejecting disturbances caused by variations in coal moisture. The results
demonstrate that the coal mill control system performance can be significantly improved through the
use of the model presented in this paper.
& 2012 Elsevier Ltd. All rights reserved.

1. Introduction The present work focuses on vertical spindle roller pressure


coal mills, which are one of the most common arrangements for
The majority of the presently operating coal fired power plants coal grinding. The working principles of these mills are exten-
and their control systems were designed for base load operation sively explained in Niemczk and Andersen (2009), Shin and Li
with large safety margins from key operating constraints. How- (2009) and Flynn (2003). In general, several mill types exist which
ever current requirements due to an increasing share of renew- can be classified as vertical roller pressure, vertical ball and
able but intermittent power generation push these coal fired horizontal ball coal mills. The fundamental principles of operation
power plants to be more flexible in providing a fast dynamic of the different coal mill types are treated in Miller (2010) and
response for their power output (Niemczk & Andersen, 2009). European Commission (2006). It is worth to mention that due to
Therefore it is important to adapt the control system designs for similarities in the grinding process, the same models can typically
coal fired power plants to meet the present requirements of faster be used for vertical roller pressure and vertical ball coal mills.
availability. Several studies can be found in the literature (Franke Coal mills are used in coal fired power plants to grind crushed
& Vogelbacher, 2006; Krueger & Franke, 2003; Krueger, Franke, & coal lumps with several cm diameters to very fine powder in the
Rode, 2004) focusing on improving ramp rates on the steam and order of 50–100 mm diameter. A further task of coal mills is to dry
fire side of coal fired power plants, while fuel preparation and the coal particles avoiding wet coal conditions during combus-
delivery were not considered in as much detail. tion. This is achieved by hot air passing through the mill which
evaporates the coal moisture. The air dries the coal and also
carries the fine coal particles to the outlet of the coal mill. This
n
Corresponding author. mix of air and pulverized coal is called pulverized fuel (PF) and is
E-mail addresses: andrea.cortinovis@ch.abb.com (A. Cortinovis),
mehmet.mercangoez@ch.abb.com (M. Mercangöz),
combusted in the burners. Details about boiler systems can be
tarun.mathur@in.abb.com (T. Mathur), jan.poland@ch.abb.com (J. Poland), found in UNEP (2006), Leopold (2009) and Leimbach (2007). In a
marcel.blaumann@de.abb.com (M. Blaumann). typical coal fired power plant there are five to eight coal mills

0967-0661/$ - see front matter & 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.conengprac.2012.10.006
A. Cortinovis et al. / Control Engineering Practice 21 (2013) 308–320 309

List of symbols R measurement noise covariance (–)


Rpr roller pressure (t)
BR1 breakage rate parameter kg=ðt sÞ S1 flow parameter (1/kg)
BR2 breakage rate parameter (kg) T temperature (1C)
Csp classifier speed (1/s) TD time delay (s)
C0 classifier parameter (1/s) t time (s)
!
CM coal moisture parameter (–) u inputs (–)
Cp specific heat capacity kJ=ðkg 1CÞ uC cold valve position (–)
D feed forward matrix (–) uH hot valve position (–)
!
dP0 pressure drop parameter (mbar) x states (–)
dP1 pressure drop parameter (mbar/kg) x_ state derivative (–)
dP2 pressure drop parameter mbar s /kg xF fineness (–)
!
ER1 free water parameter (1/1C) y outputs (–)
!
F Jacobian matrix (–) z measurements (–)
F1 flow parameter (1/s) Dp pressure drop (mbar)
F2 flow parameter (kg) DT sampling time (s)
H measurement matrix (–) y parameter in EKF (–)
I identity matrix (–) o weight (–)
i index (–)
J cost function (–) Indicies
K Kalman gain matrix (–)
k discrete time index (–) amb ambient
khu hold up parameter (–) a air
kloss,F heat loss parameter (kJ/kg) c coal
kloss,T heat loss parameter kJ/(s1C) coarse coarse coal
M mass (kg) fine fine coal
m _ mass flow (kg/s) free free water term
N number of measurements (–) hu hold up
Nopt optimization window length (–) in inlet variable
P covariance matrix (–) meas measured value
P power (kW) out outlet variable
P0 power parameter (kW) sim simulated value
P1 power parameter (kW/kg)
Q process noise covariance (–)

which supply the burners continuously with pulverized fuel. In approach is presented in Wei and Wang (2007). This approach
intended design, one or two coal mills are allowed to be out of uses five different segments based on a nominal grinding model
operation to perform maintenance and the remaining mills can and subsequent modifications to match the changing conditions
cover the fuel preparation need for full load operation. It is during start up and shut down. More complex models with more
however not unusual at power plants to see all mills in operation than two particles sizes can be found in Flynn (2003), Fan and
well below their design capacity and with excessive grinding Rees (1994) and Zhou and Si (2000). The work presented in Flynn
power consumption in order to allow extra room for controll- (2003) shows also the difference between a constant and a
ability, mainly in the case where one of the mills needs to be run- distributed parameter model and discusses performance results
back and its load needs to be redistributed over the remaining for both approaches. Zhou and Si (2000) presents the most
mills. A schematic illustration of a typical coal fired power plant is complex model with more than 30 states. Finally, conventional
shown in Fig. 1. In these plants, pulverized coal is not stored, coal mill control is discussed in Lindsley (2000) and Liptak (2006).
which makes the operation with vertical spindle mills – having In the present work, a novel nonlinear coal mill model is
very limited internal pulverized fuel storage – a just in time task. derived for model-based real time applications based on a former
All these points highlight the importance of coal mill control for study (Mercangoez & Poland, 2011). One of the objectives is to
the operation of coal fired power plants. consider a trade-off between model complexity and model
Coal mill modeling has been actively studied in recent years accuracy based on real measurements from an actual hard coal
considering either simple models for process monitoring and fault burning power plant to validate the results. Parameter identifica-
detection or complex models for dynamic simulations. In Odgaard tion studies revealed that an on-line parameter estimation
and Mataji (2006) a simple energy balance model is used for fault approach is needed to estimate some parameters in real time.
detection by estimating an unknown energy input. Shin and Li Based on simulations with this model, a conventional coal mill
(2009) presents a model with two particle sizes and pressure drop control strategy from the same hard coal power plant is imple-
dynamics. A similar model is used and validated with an evolu- mented and a new model-based control strategy is developed.
tionary algorithm in Niemczk and Andersen (2009) and Niemczk, The conventional control is utilized as a reference to compare the
Bendtsen, and Ravn (2012). The validation also includes different performance improvement on ramp rates of coal fired power
coal mill types to investigate the parameter dependency on coal plants using nonlinear model predictive control (NMPC). The
mill types. Zhang and Wu (2002) discusses a simple model and paper is structured as follows: In Section 2 the nonlinear model
uses a genetic algorithm (GA) to optimize the parameters of the equations are derived and the control objectives are defined. The
model. To increase the accuracy of this model a multisegment model properties are investigated in Section 3 and used to choose
310 A. Cortinovis et al. / Control Engineering Practice 21 (2013) 308–320

Fig. 1. Working principle of a coal fired power plant (Niemczk & Andersen, 2009).

the optimization parameters of Section 4. After the validation, the


estimation technique is discussed in Section 5. Finally the model-
based control strategy and the simulation results are presented in
Sections 6 and 7.

2. Nonlinear coal mill model

Before introducing the model equations for the system depicted


in Fig. 2, the most important assumptions are listed as follows:

 Only two particle sizes are considered, fine and coarse coal
particles. This was found to be the simplest approach possible
which can account for some indication of pulverized fuel
fineness. It is also used in the majority of related works which
aim to have an on-line application for the model, as it was
described in Section 1. The complexity of such a model is
proportional to the number of coal particle size bins used,
because every particle size introduces an additional state
through its mass balance and an algebraic equation for
describing the grinding rate.
Fig. 2. Vertical roller pressure coal mill (Niemczk & Andersen, 2009).
 The incoming coal mass flow is composed of only coarse coal
particles. This is a simplification that does not restrict the
model and results in a good approximation. Furthermore, the written as
inlet coal moisture is assumed to be a parameter of the model. !
u ¼ ½m_ in _ in in
c m a T a C sp Rpr 
T
ð1Þ
 A lumped parameter assumption is used for the temperature,
meaning that Tout is the uniform temperature of the coal mill From start up measurements, it has been found that there is a
and also the coal outlet temperature. This is a common pure time delay for the input variable T in a . As it can be seen in
assumption for continuous processes carried out in compact Fig. 3, the time delay between T ina and Tout amounts approximately
geometries. Therefore the air, the coarse and the fine coal to T D ¼ 32 s. This is considered in the nonlinear coal mill model. In
particles all have the same temperature Tout at the outlet of addition, the simplified notation T in in
a is used instead of T a ðtT D Þ in
the mill. the following equations. The derivation of the model equations
 The coal drying is described by the term m _ water
free , which begins with the two mass balances for coarse and fine coal
corresponds to the amount of evaporated coal moisture inside particles, which can be written as
the mill. d
 Mill chocking has not been modeled in detail. The saturating _ in
M coarse ¼ m _ _ out _ water
c m BR m coarse m free ð2Þ
dt
behavior of the breakage rate and the coarse particle carry over
describe partially the effect of mill choking. d
M ¼m _ out
_ BR m ð3Þ
dt fine fine

The inputs of the coal mill are (i) the coal inlet mass flow, (ii) the
air inlet mass flow, (iii) the air inlet temperature, (iv) the class- It has to be emphasized that the only source of fine particles is the
ifier speed and (v) the roller pressure. The input vector can be grinding of coarse coal. The first term on the right hand side of
A. Cortinovis et al. / Control Engineering Practice 21 (2013) 308–320 311

0.35 written as
uC 
d 1
0.3 uH T out ¼ _ in
m in _ in in _ out out _ out out
coal  H coal þ m air  H air m coal  H coal m air  H air
dt C p,c  M hu
Tout
0.25 d 
Tain m _ water
free  H evap þ PQ loss C p,c  T out ðM coarse þ Mfine Þ ð10Þ
dt
Arbitrary Units [a.u.]

0.2
This equation is quite complex, also because of the couplings
to Eqs. (2) and (3). The first four terms on the right hand side
0.15
represent the enthalpy balances of water, coal and air respec-
tively. Then, the evaporation term appears followed by the power
0.1
supplied to the system and the heat loss term. The last term and
the fractional factor come from the chain rule of derivatives for
0.05
Tout and MTotal, because both of them are time dependent. Mhu and
0
the heat loss are defined as
Mhu ¼ khu  ðM coarse þ Mfine Þ ð11Þ
−0.05
2270 2280 2290 2300 2310 2320 2330 2340 2350
_ in
Q loss ¼ kloss,F  m a þ kloss,T  ðT out T amb Þ ð12Þ
Time [s]
The nonlinear model equations are concluded at this point. This
Fig. 3. Determination of the time delay TD.
model consists of three states and has a total number of 15
parameters. The heat capacity of water C p,w , the heat capacity of
coal C p,c , heat capacity of air C p,a , the ambient temperature Tamb,
Eq. (2) is u1, while the second one is the breakage rate defined as the inlet coal temperature T in c and evaporation enthalpy Hevap are
follows: assumed to be known constants. The measured outputs of the
  system are (i) the power consumed by the mill, (ii) the pressure
M coarse
m_ BR ¼ BR1  Rpr  tanh ð4Þ drop over the mill and (iii) the outlet temperature. Note that Tout
BR2
is both, a measured output and also a state of the model.
This term captures the grinding of coarse coal particles to fine
!
coal. It can be seen, that at a certain point the almost linear ymeas ¼ ½P Dp T out T ð13Þ
relationship to Mcoarse flattens out, because the breakage rate The air system is modeled as two joining pipes transporting cold
cannot be increased arbitrarily. This limit is imposed by the size and hot air respectively. The hot and cold air mass flow can be
of the coal mill and by the available grinding capacity of the influenced by two valves, which then determine the mixed air
rollers. The parameter BR1 can be used to simulate a deterioration temperature Tin _ in
a and mixed air mass flow m a . These two variables
of the roller performance (e.g. due to wear damage). The third define the air conditions at the inlet of the coal mill. For the valves
term in Eq. (2) describes the coarse particles that reached the a linear relationship is assumed, where uC and uH are the valve
coal mill outlet. Here a similar limitation is present as it can be positions which vary from 0 (closed) to 1 (open). Furthermore, it
seen in Eq. (5). This saturation behavior reflects the fact that is assumed that TH and TC are constant and that there is a
only a limited amount of coarse particles can be carried by the maximum air flow for hot and cold air. It has to be mentioned,
air flow. that in general TH is not necessarily constant. Fresh air is heated
 
M coarse 1 with the help of an exhaust gas heat exchanger, which then
m_ out _ in
coarse ¼ F 1  m a  tanh  ð5Þ
F2 C sp þ C 0 results in the hot air temperature TH. The effect of varying TH
would affect only the start up scenario, whereas TH remains
The last term in Eq. (2) describes the amount of evaporated coal
approximately constant after reaching nominal operation.
moisture based on the temperature of the inlet air and is defined
as m _ max
_ C ¼m C  uC ðtÞ
_ water
m _ in in
free ¼ ER1  ðm c  CMÞ  ðT a T out Þ ð6Þ _ max
_ H ¼m
m H  uH ðtÞ ð14Þ
The first term in Eq. (3) is again the breakage rate, while the The resulting andTin _ in
m
can be calculated considering a mass
a a
second one describes the fine particles carried to the outlet of the balance and an energy balance. For that, incompressible flow is
coal mill. This term is defined in Eq. (7) and no limitation is assumed and the air is approximated as an ideal gas. The specific
present, because the fine particles are easily suspended in air, heat capacity of air C p,a is also assumed to be constant for the
contrarily to coarse ones. present temperature range. With Eqs. (15) and (16) the mixed air
_ out
m _ in mass flow and temperature are determined.
fine ¼ S1  m a  M fine ð7Þ
_ C  T C þC p,a  m
C p,a  m _ H  TH
The power consumed by the coal mill for pulverizing the coal is T in
a ¼ ð15Þ
C p,a  ðm_ C þm_ HÞ
assumed to be linearly dependent on Mcoarse and is defined as
P ¼ P 0 þP 1  M coarse ð8Þ
_ in
m _ max  uC þ m
a ¼ mC
_ max  uH ð16Þ
The pressure drop over the coal mill is also dependent on Mcoarse H

and it contains furthermore a feed-through term on u1, as it can


be seen in Eq. (9). This means that Dp as an output is directly
2.1. Requirements for coal mill control
dependent on an input of the system.
_ in
Dp ¼ dP0 þ dP1  Mcoarse þ dP2  m a ð9Þ For completeness of the coal mill modeling part, the objectives
for the controller design are discussed. As mentioned in the
The third differential equation is obtained from an energy introduction, the most important requirement is a fast response
equation over the whole coal mill. The energy balance can be time for start up and load steps. This implies indirectly, that the
312 A. Cortinovis et al. / Control Engineering Practice 21 (2013) 308–320

controller should be able to track the outlet coal mass flow set through a 200 mesh sieve. The Set point of 0.7 means that if a
_ out
point m _ out is an
c,SP , which is set by the plant master controller. m c sieve analysis is performed, 70% of the coal particles pass through
estimated value, because typically a direct measurement of coal a 200 mesh sieve. A 200 mesh sieve, corresponds to a particle size
mass flow is not available. Therefore, the outlet coal mass flow is of 75 mm. Therefore, xF corresponds to the percentage of particles
defined with the help of Eqs. (5) and (7). that have a particle size smaller than 75 mm. This lower bound is a
soft constraint, because especially at the start up it is not possible
_ out
m _ out _ out
c ¼ m fine þ m coarse ð17Þ to fulfill this requirement.
A static Tout set point of 95 1C is desired, to avoid wet coal at the
outlet of the coal mill. This is a good trade off point between
unnecessary heating and avoiding high moisture content at the
outlet of the mill. An accuracy window of 73–5 1C is mandatory 3. Parameter sensitivity analysis
for Tout to maintain optimal conditions for the combustion and
safe operation of the mill. Another important specification is the In this section, the influence of different model parameters on
upper limit on the pressure drop. Beyond 90–95 mbar, it is likely the steady state values of the three output variables is investi-
that mill choking takes place. The pressure drop Dp should gated. According to these results a suitable parameter set is
therefore not exceed the choking limit Dpch of approximately selected for the optimization in Section 3.1.
90 mbar. It is not necessary that choking occurs if the pressure The parameters are always varied from 50% up to þ50% of their
drop is inside the choking range, although this has to be avoided reference values with a resolution of 10%. The reference values have
in any case to ensure a safe operation. The last objective is to been taken from the former study (Mercangoez & Poland, 2011), in
achieve a good coal fineness xF. If the pulverized particles are too which the parameters were optimized for the start up case. The
large, the combustion takes place only on the surface of the model is simulated for all combinations of analyzed parameters until
particles, while the core possibly remains unburned. This results a steady state is reached. Finally, the output variables are plotted in
in economic losses because more coal is needed for the same heat terms of the relative change of their reference steady state value.
output and also high levels of unburned carbon in the fly ash can Offset parameters like P0 are excluded from this analysis because they
reduce its value as a by-product. Moreover, if a staged combustion shift the steady state values by definition. For conciseness, only one
arrangement is employed, the larger particles can continue example is presented in which two parameters are varied.
releasing nitrogen compounds in the oxygen rich zones leading The results for P1 and CM variations are plotted in Fig. 4. As it is
to increased NOx formation. The fineness is defined as expected from the model equations, the influence of both para-
meters on the pressure drop is very small. Furthermore, it can be
_ out
m fine m_ out
fine seen that the power is highly influenced while the temperature
xF ¼ ¼ ð18Þ varies just slightly. Variations in CM have almost no influence on
_ out
m c m out
_ fine þ m _ out
coarse
the power for constant P1, whereas both parameters have a large
The last requirement can be translated into a lower bound of impact on the temperature. P1 causes a coupling between power
approximately 0.7 for xF. The unit of the fineness xF is in % passing and temperature. This coupling comes from the power term

Reference point Reference point

60 0.4
Relative Variation of Δp [%]
Relative Variation of P [%]

40
0.2
20

0 0

−20
−0.2
−40

−60 −50 −0.4 −50


50 50
0 0
0 0
ΔCM [%]
ΔP1 [%] −50 50
ΔCM [%] P1 [%] −50 50

Reference point
[%]

20
out
Relative Variation of T

10

−10

−20 −50
50
0
0
P1 [%] −50 50 ΔCM [%]

Fig. 4. Sensitivity of output variables with respect to P1 and CM.


A. Cortinovis et al. / Control Engineering Practice 21 (2013) 308–320 313

Table 1 4. Parameter optimization and validation


Parameter influence on outputs expressed as relative changes form the steady
state values.
The parameter optimization is based on the minimization of
Varied parameters Influence on the squared error between simulated and measured outputs. For
this purpose, several data sets containing measured inputs and
P in % Dp in % Tout in % outputs have been obtained from a hard coal fired power plant.
The parameter optimization can be formulated as a minimization
BR1 and BR2 r 7 100 r 7 21 r 7 15
P1 and CM r 7 50 r 7 0:3 r 7 15
problem with the following quadratic cost functions.
dP1 and dP2 r 70 r 7 50 r 70 For power and temperature
kloss,T and kloss,F r 70 r 70 r 7 40
N 
X  N  
CM and kloss,T r 7 1:2 r 7 0:3 r 7 14 Psim,i P meas,i 2 X T sim,i T meas,i 2
CM and kloss,F r 7 1:2 r 7 0:3 r 7 40 JP,T ðP0 ,P1 ,kloss,F ,kloss,T ,CMÞ ¼ þ
i¼1
Pmean i¼1
T mean
CM and khu r 7 1:5 r 7 0:4 r 78
S1 and C0 r 77 r 7 1:5 r 7 1:1 ð19Þ
ER1 and CM r 73 r 71 r 7 20
F1 and F2 r 7 40 r 78 r 76 For the pressure drop
N  
X Dpsim,i Dpmeas,i 2
JDp ðdP0 ,dP1 ,dP 2 Þ ¼ ð20Þ
i¼1
Dpmean
presence in the energy balance. On the other hand, there is no
coupling between power and temperature for varying CM. Keeping the considerations of Section 3.1 in mind, the pressure
This analysis has been done for almost all combinations of drop parameters are optimized separately from temperature and
parameters. The most important results are summarized in power. This has the advantage of faster convergence from having
Table 1. On the left side of the table the varied parameters are fewer optimization parameters. For solving the optimization
listed, whereas on the right hand side the maximum and mini- problem a direct search method has been used, which is described
mum influence on the output variables is shown. After having in Lagarias and Reeds (2007). It has to be emphasized that the
discussed the results, a strategy has been developed to choose the optimized function properties for the parameter estimation pro-
optimization parameters. This is explained and justified in the blem regarding convexity or local convexity have not been
next subsection but it is worth to mention that for simplified first established, but with physically reasonable initial estimates the
principles models steps such as selecting the parameters to solution of Eqs. (19) and (20) always converged and improved the
optimize and differentiating them for being time varying or initial model predictions providing reasonable results.
constant requires a good understanding of the system and For the optimization and validation procedure, three indepen-
practical experience. dent datasets have been used. Every dataset corresponds to 4 h of
data acquisition on the actual power plant with a sampling time
of Dt ¼ 2 s. Only datasets from one specific mill are used, because
3.1. Results of sensitivity analysis
of possible differences between the mills. Dataset 1 represents the
coal mill start up until reaching a nominal set point. Dataset 2 was
The present system is underdetermined1 and if all 15 para-
recorded in nominal steady state operation. During this dataset,
meters are used for parameter estimation, the procedure would
steps in various inputs have been performed. The last dataset
lead to overfitted parameters. Assuming that the start up behavior
corresponds also to an approximately steady state operating
is well captured with the nominal values from the previews study
condition and will be called Dataset 3 in the following. Also in
(Mercangoez & Poland, 2011), the following approach is used to
this dataset several steps in inputs were carried out to have
decide which parameters should be optimized: parameters are
transient behavior in the recorded data. Datasets 1 and 2 are used
chosen which have a large impact on one output, while at the
to identify the parameters and Dataset 3 for the model validation.
same time having a small influence on the other outputs. This
After independent parameter optimization with Datasets 1 and
approach can be justified, because the main underlying dynamic
2, the two parameter sets are compared and a trade off to a
behavior of the model is correct in the start up case, meaning that
possible global parameter set is suggested. Parameters that
the interactions between the states behave correctly. Even if the
spread significantly will not be included in this global parameter
reference parameter set is used for all measurements, the correct
set, but rather adapted on-line. This appears reasonable, because
behavior results in the system response, but it obviously shows
also in the literature several works (Flynn, 2003; Shin & Li, 2009)
some offset and some scaling differences.
found the necessity to have time-varying parameters. Finally, the
For the power output the parameters P0 and P1 are chosen,
global parameter set is applied to Dataset 3 to validate the model.
even if P1 has a weak coupling to the temperature. The pressure
Graphical summaries for the independent optimization results
drop is optimized using dP0, dP1 and dP2. These parameters have
can be found in Figs. 5 and 6, while the corresponding parameter
no influence on other outputs, hence the pressure drop can be
values are listed in Table 2.
treated separately. Finally, for temperature the parameters kloss,F ,
The optimized model for Dataset 1 shows an accurate fit to the
kloss,T , and CM are used. CM has an influence on all outputs, but the
measurements. The fit for Dataset 2 is also quite accurate, but for
temperature is influenced most significantly, therefore this choice
the temperature some deviations are observed. Even if it can be
is justified. Parameters with strong couplings between outputs
seen that the deviations are under 5 1C, further investigation
like F1, F2 or BR1 and BR2 as well as parameters of lesser influence
showed that a better result can be achieved if CM is allowed to be
like S1, C0 and khu are excluded from the optimization procedure.
time varying. A more detailed analysis on the terms of the energy
This is also because the durations of the available datasets used in
equation, indicated that CM is always present in terms with large
this study are limited and it is not possible include phenomena
magnitudes. Therefore, CM influences significantly the derivative
such as wearing out of the grinding surfaces having an influence
of Tout and is a candidate for on-line estimation. It is expected that
on e.g. the breakage rates.
the coal feed cannot have exactly the same moisture content since
it is a natural resource with varying character. It should also be
1
There are 3 measured outputs and 15 parameters. noticed, that the pressure offset parameter dP0 is not zero.
314 A. Cortinovis et al. / Control Engineering Practice 21 (2013) 308–320

400 400

P [kW]
P [kW]

200 300
Model Measurement
0 200
0 5000 10000 15000 0 5000 10000 15000
t [s] t [s]
80
80
Δ p [mbar]

Δ p [mbar]
60
40 60
20
0 40
0 5000 10000 15000 0 5000 10000 15000
t [s] t [s]
100 100

Tout [°C]
Tout [°C]

90
50 Model Measurement
80
0 5000 10000 15000 0 5000 10000 15000
t [s] t [s]

Fig. 5. Results with optimized parameters for Dataset 1. Fig. 7. Validation results with global parameter set for Dataset 3.

400
is observed. First it can be recognized that for the pressure drop
P [kW]

300 an almost constant bias of approximately  7 mbar is present.


Considering also the values of dP0 in Table 2, it seems appropriate
200 to choose dP0 as a time varying parameter to match the pressure
0 5000 10000 15000
t [s] drop response. A possible explanation to justify a variation of dP0
60 can be an effect of slowly changing depositing conditions of coarse
Δ p [mbar]

coal particles over the primary air flow path. This deposition
50 distribution can cause a higher or lower pressure drop offset as
Model Measurement the particles periodically coagulate and then blow away again.
40 The mismatch in the temperature predictions is a drift rather than
0 5000 10000 15000
t [s] a bias. As it was mentioned before when the results for Dataset 2
95 were discussed, the parameter CM is possibly varying. The
temperature drift observed in Fig. 7 can be explained with the
Tout [°C]

90 variation of CM. It has to be emphasized, that this is not a


modeling error but instead CM should be thought of as an
85 exogenous disturbance that cannot be predicted or modeled.
0 5000 10000 15000
t [s] Therefore, an on-line parameter correction is necessary to adapt
the two parameters CM and dP0. This adaptation will be discussed
Fig. 6. Results with optimized parameters for Dataset 2.
in the next section, where also the final results are shown.

Table 2
Optimized parameters and global parameter set. 5. On-line parameter adaptation with EKF
Optimal parameter Dataset 1 Dataset 2 Global
Before starting with the description of the parameter estima-
P0 0 0 0 tion, some considerations about Kalman filtering are given at this
P1 0.4329 0.4369 0.4356 point. The basics of Kalman filter equations and extended Kalman
CM 0.032 0.0404 – filtering can be found in Simon (2006) and Haykin (2001). When
kloss,T 3.44 2.535 2.8102
states of a model are estimated, the KF approach provides the
kloss,F 94.17 85.6845 90.8172
dP0 0.00001  2.9492 – optimal linear observer, which results in a trade-off between
dP1 0.0235 0.0233 0.023 model behavior and measured data. This is achieved by keeping
dP2 4.0884 4.1795 4.0536 track of mean and variance of the state errors, which then define
an appropriate weight-matrix K(k). The choice of K(k) minimizes
mean-squared error between actual states and estimated states.
It has to be clarified at this point, that the global optimality of Extending the states including also parameters, the KF can
the solution for the individual datasets is not directly relevant for estimate states and parameters at the same time. The additional
the final solution and is therefore not analyzed in the present states have usually no dynamics. In the present work, the model
work. The presented optimization step is used to improve the is significantly nonlinear and therefore an EKF is used. The EKF
response for individual datasets without over-fitting the para- linearizes the nonlinear model at each step evaluating the
meters and results in a good trade-off between accuracy for the Jacobian at the current states and applies then the KF equations
individual datasets and avoiding over-fitting for one specific in a slightly modified version.
dataset. Therefore, this approach cannot be rigorously proved to The inputs and outputs of the EKF are illustrated in Fig. 8. The
be optimal. Nevertheless consistent results and significant EKF runs at a sampling time of 2 s and it always uses the corrected
improvements were obtained by using it. parameters for the estimation. As it will be seen in the following,
The validation plots are illustrated in Fig. 7. The power shows a the parameters change slowly and as a consequence, it is not
good agreement, while for the other two outputs some mismatch necessary to update the parameters every 2 s but rather in the
A. Cortinovis et al. / Control Engineering Practice 21 (2013) 308–320 315

Fig. 8. Estimation structure for parameters and states.

order of tens of minutes to hours. This slow adaption functionality The parameter yCM has to be updated at every time step
for the parameters is incorporated in the EKF block. The dashed yCM ðkÞ ¼ x5 ðkÞ for the system update equations and also for the
line represents the slow model-parameter update for the math- Jacobian computation. The EKF performs in two steps providing
!
ematical model inside the control block. Note that the state prior and posterior estimates. x k9k1 is the a priori estimated
estimates can be used for model-based control and estimation state vector before the kth measurement is taken into account,
!
of additional outputs. whereas x k9k represents the a posteriori estimated state vector,
!
For the coal mill application, the three states of the system which was corrected to be the measurement z k . In the first step,
presented in Section 2 are extended with the two parameters the state vector and covariance matrix are predicted using the
which need to be estimated. This results in the following state model equations. The expressions for step one are defined as
vector for the EKF: ! ! ! !
x k9k1 ¼ x k19k1 þf ð x k19k1 , u k1 Þ  DT
!
x ¼ ½M coarse M fine T out dP0 yCM T ð21Þ
Pk9k1 ¼ F k1  P k19k1  F Tk1 þ Q ð26Þ
The two additional states for the parameters have no dynamics,
e.g. y_ CM ¼ 0 or equivalently yCM ðt k þ 1 Þ ¼ yCM ðt k Þ. Before coming to In the second step, the Kalman Filter gain Kk is computed and the
the EKF equations, the nonlinear model is discretized with a estimated state vector is corrected by the available measurements
!
simple first order approximation and the sampling time of z k ¼ ½P Dp T out T . The matrix Kk represents a trade off between
DT ¼ 2 s. The choice of the sampling time fulfills the requirement trust in the model equations and trust in the actual measure-
formulated in Haugen (2009).2 ments. This trade off can be tuned using the covariance matrix of
the process noise Q and the measurement covariance matrix R.
! !
_
! !! ! xð t k þ 1 Þ x ðt k Þ Also the initial conditions of P090 have a decisive impact on the
x ðt k Þ ¼ f ð x ðt k Þ, u ðt k ÞÞ  ð22Þ
DT filter performance. In Eq. (27) the second step is shown.
Rearranging Eq. (22) and using a simplified time index notation, K k ¼ Pk9k1  HT  ðH  P k9k1  HT þ RÞ1
the discrete approximation is obtained. ! ! ! ! !
x k9k ¼ x k9k1 þK k  ð z k H  x k9k1 D  u k1 Þ
! ! !! !
x ðkþ 1Þ ¼ x ðkÞ þ f ð x ðkÞ, u ðkÞÞ  DT ð23Þ Pk9k ¼ ðI5x5 K k  HÞ  P k9k1 ð27Þ
The vector equation for the model outputs is then defined as
2 3 After the implementation in Matlab/Simulink and tuning the
M coarse filter, the system has been simulated with the available datasets.
2 3 6 7
P1 0 0 0 0 6 M fine 7 For this analysis Dataset 3 as well as a new Dataset 4 are used.
! 6 7 6 7
Dataset 4 is a steady state dataset which has interesting char-
y ¼ 4 dP1 0 0 1 0 5  6 out 7 6 T 7
6 dP 7 acteristics for CM.
0 0 1 0 0 4 0 5
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} In Fig. 9 the results for Dataset 3 are shown. It can be seen that
H yCM
2 3 the EKF results are in good agreement with the measurements
_ in
m c and outperform predictions with constant parameters. Further-
2 3 6 in 7 more, it can be seen that the variation of CM has almost no
0 0 0 0 0 6m_ 7
6 a 7
6 7 6 in 7 influence on power and pressure drop. This is in agreement with
þ 4 0 dP2 0 0 0 5  6 T a 7 ð24Þ
6 7 what was found in the parameter sensitivity analysis part. dP0 is
0 0 0 0 0 6C 7
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} 4 sp 5 almost constant for the whole dataset, while CM decreases
D Rpr towards the end. This decrease causes the drift in Tout of the
! uncorrected model. The curves of the model are simulated with
After the initialization with x 090 , P 090 , R and Q, the standard EKF
constant CM, while the EKF result uses at each time step the
equations can be applied to the present system. The discrete
corresponding estimated CM. It can be observed, that both
Jacobian matrix is needed for the filter equations at each time
parameters vary quite slowly.
step and can be written as
The results for Dataset 4 are illustrated in Fig. 10. Also for this
2 !! !! 3
@f 1 ð x , u Þ simulation, the EKF results are in a good agreement with the
6 @Mcoarse    @f 1 ð x@y , u Þ 7
6 7 measured data. The pressure parameter increases slowly for less
F k1 ¼ I5x5 þ 6 ^ & ^ 7  DT ð25Þ than 1 mbar, whereas the coal moisture CM decreases in the first
6 7
4 !! !! 5
@f 5 ð x , u Þ @f 5 ð x , u Þ part of the dataset and increases significantly towards the end.
@M coarse  @y ! !
x ¼ x ðk1Þ Also here the same reasoning applies as before for Dataset 3.
The drifts of Tout in the constant model case are caused by the
2
changing parameter CM.
The sampling time is required to satisfy the inequality DT r 0:1=max9l9,
where max9l9 is the maximum absolute eigenvalue of the system matrix A after
Some considerations about the application of the suggested EKF
linearization around the nominal operating point. For the present case estimation structure to the present problem can be given at this
max9l9  0:0061 and the inequality DT r 16:39 s is satisfied. point. First the two main aspects of observability and structural
316 A. Cortinovis et al. / Control Engineering Practice 21 (2013) 308–320

400 350
P [kW]

P [kW]
300 300
Measurement EKF Model Measurement EKF Model
200 250
0 2000 4000 6000 8000 10000 12000 14000 0 2000 4000 6000 8000 10000 12000 14000
t [s] t [s]
60 80
Δ p [mbar]

Δ p [mbar]
50 60

40 40
0 2000 4000 6000 8000 10000 12000 14000 0 2000 4000 6000 8000 10000 12000 14000
t [s] t [s]
100 95
Tout [°C]

Tout [°C]
90 90

80 85
0 2000 4000 6000 8000 10000 12000 14000 0 2000 4000 6000 8000 10000 12000 14000
t [s] t [s]

10

dP0 [mbar]
2
dP0 [mbar]

0
−2 5
−4
−6 0
0 2000 4000 6000 8000 10000 12000 14000 0 2000 4000 6000 8000 10000 12000 14000
t [s] t [s]
0.06 0.06
θCM [−]

0.05 0.05
θCM [−]

0.04 0.04
0.03 0.03
0.02 0.02
0 2000 4000 6000 8000 10000 12000 14000 0 2000 4000 6000 8000 10000 12000 14000
t [s] t [s]

Fig. 9. Results and estimated parameters for Dataset 3. Fig. 10. Estimated dP0 and CM for Dataset 4.

dependency of the variables need to be addressed. If the linearized the simulations and from the analysis of Section 4, it has been found
system used in the EKF equations is not observable, the EKF cannot that the two parameters vary slowly. This is important for control
estimate the desired variables and diverges. To ensure observability, applications. If the parameters have slow dynamics, a constant
the nonlinear system equations have been analyzed and linearized parameter assumption is justified for model based control. Therefore,
around several realistic operating points. After that, the observability for the control part of the present work models with constant
of the linearized systems was investigated. It has been found, that the parameters are utilized. The model-parameter update takes place in
resulting linearizations are observable. This can also be seen by the order of hours, as it was mentioned before, even if the parameters
inspection of Eqs. (9) and (10) analyzing the structural dependencies are available at a rate of 2 s. How the parameters are updated for the
of the two parameters dP0 and yCM . The parameter dP0 is directly model-based control part is left open because this is a part of an
observed through the measured pressure drop and appears only in ongoing work including field trials. The model-parameters can be
linear terms. The coal moisture yCM is observed indirectly through the updated by averaging of the last N parameters, or based on the rate of
differential equation of the temperature. Note that the bilinear terms change of the parameters. This operation can be viewed as an output
yCM  T out do not represent a problem, because Tout is a measured state. filter in the EKF and is therefore inside the EKF block of Fig. 8.
This justifies the employment of an EKF for the parameter estimation To conclude this section, it is worth to mention that the
part. For the state estimation Eqs. (2), (3) and (10) are examined. The presented model captures the essential dynamics of the coal mill
temperature Tout is measured. The state Mcoarse is directly observed and can be used also for additional tasks. Whereas at the beginning
through pressure drop and power measurements, which results in of this section in Fig. 8 the application to control systems is shown,
strong observability. The estimation of the last state Mfine is more the present arrangement can also be used for control performance
difficult and represents a delicate point. As it can be seen from the monitoring, for fault detection and for maintenance services.
equations, Mfine cannot be observed directly and is therefore only
weakly observable. This state appears only in the differential equation
in mixed terms with yCM and Tout. If the parameter yCM varies slowly 6. Nonlinear MPC control
and the temperature is directly measured, Mfine can be observed
through Tout. This is the case as it was explained previously. Therefore, The standard problem formulation for Nonlinear Model Pre-
also for estimation of the states the suggested approach is suitable. dictive Control (NMPC) is captured by a nonlinear minimization
This is also underlined by the presented results. problem and is defined as
In the NMPC application presented in the next sections it has been
! X
N 1
! !
decided to use the EKF approach, because it results in good min JN ð x N Þ þ Jð x k , u k Þ
performance and needs minimal computational power. Furthermore, !
uk for k A ½0 N1 k¼0
it estimates both parameters and states at the same time without !! !
!
having to solve an on-line optimization problem. From the results of subject to x k þ 1 ¼ f ð x k, u kÞ
A. Cortinovis et al. / Control Engineering Practice 21 (2013) 308–320 317

! !^ Nopt, the occurring costs and the constraints on inputs and out-
x k ¼ x ðkÞ
! ! ! puts. For that, appropriate cost and bound-blocks were available
u min r u k r u max in the toolbox.
_
! _
! !_
u min r u k r u max ð28Þ Consultations with the plant operator helped to define the
constraints. The same upper and lower bounds used for the
Choosing a suitable cost function, the NMPC approach minimizes
! simulations are used for the NMPC problem formulation as well.
the objective with respect to the inputs u ðtÞ over an optimization
These constraints were also used in the simulation environment,
window of N steps, while satisfying the system equations and
to have fair comparison for other control strategies. The cost for
other constraints at the same time. This results in an input matrix
! ! ! the NMPC problem is composed of several terms containing
U ¼ ½ u t u t þ 1    u t þ N1 , which contains all optimal inputs for
measured signals as well as model outputs and is defined as
the next N1 steps. Only the first column of U is applied to the
 2  2
plant, such that measurements can be taken into account and JðtÞ ¼ o21  T out T out þ o22  m_ out _ out þ o23  J 3 ðxF Þ
SP c m c,SP
provide feedback in the next steps. Using this strategy, a so called ! !
receding horizon iteration is obtained. This means that the _ in
m Rpr
optimization window is shifted one step forward as time passes. þ o24  J 4 ðDpÞ þ o25  J5 a
þ o26  J 6 ð29Þ
_ in
m c m_ in
c
Also the concept of receding horizon is explained in Maciejowski
(2000). The basics of NMPC can be found in Grüne and Pannek It is a common method to choose quadratic terms, because of
(2011), while in Magni, Raimondo, and Allgoewer (2009) several faster convergence and smoother action. Furthermore, no abso-
applications are presented. lute value operation is needed. The first two terms of Eq. (29)
To implement a nonlinear MPC controller for the present represent deviations between actual values and set points for Tout
application, a research tool for graphical nonlinear optimization and m _ out
c . These two terms are the squared errors of outlet
has been used in Simulink. The toolbox directly translates the temperature and outlet coal mass flow. The functions J3 and J4
Simulink model into an NLP format which is suitable for the have been designed to meet the specifications for fineness and
Interior Point Optimizer (IPOPT), which is used for the optimiza- pressure drop respectively. The last two terms of Eq. (29)
tion. The implementation follows a standard scheme as it is represent the Loadline and the Roller Pressure Curve, which are
shown in Fig. 11. How the NLP is actually solved is reported in equations given by the mill manufacturer to define the relation-
Waechter and Biegler (2006). The estimation is performed using ship between the air flow and coal flow and between the roller
the same extended Kalman filter introduced in Section 5. pressure and coal flow. J5 and J6 describe the absolute deviation
As illustrated in Fig. 11, the NMPC controller is embedded into from the Loadline and the Roller Pressure Curve. These two terms
the existing plant master controller. In this way, the new have small weights to give enough flexibility to the system during
controller can be tested without replacing the whole power transients, while having well-defined values specified by these
demand control system. Note that the signal obtained from the curves for steady state operation.
plant master control loop is directly used as a set point for the In the following, the controller design is discussed. The
estimated coal mass flow at the outlet. Therefore, the input coal optimization horizon was chosen as N opt ¼ 40. With this choice
mass flow m _ in
c remains a manipulated input for the NMPC the optimization problem could be solved faster than the sam-
controller. This choice results in increased flexibility for the NMPC pling time of 2 s. For larger windows better results can be
approach and makes it possible to deviate temporarily from the obtained, but then the optimization is more time intensive and
plant master controller specified values to utilize the pulverized the limit of 2 s can be exceeded on typical hardware and software
fuel storage capacity of the coal mill. If the inlet coal mass flow is platforms used for NMPC applications. For the design of the
set by the outer loop, the NMPC approach would lose one degree weights o1 to o6 some effort had to be made to achieve good
of freedom and could use only four inputs. The calculation of the performance. Several differed cost function arrangements and
inlet coal mass flow required to match the set point at the outlet weightings have been investigated. The main challenge was to
is realized by optimizing the trajectories of the involved variables. find the balance between having a lot of terms in the cost function
The optimization is characterized by the optimization window restricting the optimization or having fewer terms leaving a lot of

Fig. 11. Full control and estimation structure embedded in the existing plant master controller.
318 A. Cortinovis et al. / Control Engineering Practice 21 (2013) 308–320

8 9
8
7
Conventional 7
6 NMPC 6
5 Reference 5
0 50 100 150 200 0 200 400 600 800 1000

80
14
70
12
60 10
50 8
0 200 400 600 800 1000 0 200 400 600 800 1000

95.5 300

95 250

94.5 200

94 150
0 200 400 600 800 1000 0 200 400 600 800 1000

160

0.8 140

0.75 120
0.7 100
0 200 400 600 800 1000 0 200 400 600 800 1000

2000
16
1500 14
Mcoarse Mfine Mcoarse Mfine 12
1000 10
8
500
0 200 400 600 800 1000 0 200 400 600 800 1000

Fig. 12. Simulated results of output variables (left side) and corresponding inputs (right side) for a load change scenario.

freedom to the system. Finally a good trade-off could be found, Similarly to Odgaard, Stoustrup, and Mataji (2007) and
while fulfilling the desired control objectives. The weights used Odgaard and Mataji (2008), the second scenario simulates a coal
for the NMPC approach are always a ratio between real weight moisture disturbance during a load step. This scenario simulates
and nominal value of the corresponding signal. an additional sewage sludge flow for a sludge co-combustion
process. These results are illustrated in Fig. 13. It has to be
emphasized, that this scenario is very challenging for the coal
7. Control results mill control system. The inlet air temperature is very near to the
saturation of 350 1C, meaning that the cold air valve is almost
In this section, first the results for a load change scenario are fully closed. This situation results in a control problem where
investigated by comparing the performances of NMPC and a constraints play an important role. A higher moisture disturbance
conventional SISO control strategy. This scenario represents a step would result in a condition where Tout could not be anymore
from nominal operation to full load conditions. The transient maintained at the set point of 95 1C. The conventional approach
responses from simulations are illustrated in Fig. 12. The conven- can follow the coal mass flow step, but it undershoots signifi-
tional control has a pulverized fuel response time of 40 s, while cantly after reaching the set point. The undershoot of m _ out
c just
keeping the outlet temperature in an acceptable range of 71.25 1C remains inside the acceptable tolerance. This undershoot could
from the set point. The reason for the poorer coal flow response of not be avoided by increasing the integral part of plant master
the conventional control is the slightly slower m _ in
a evolution controller, otherwise the response would overshoot in other
compared to the NMPC approach. This has a great impact on the control scenarios. Also the outlet temperature cannot be kept
outlet coal mass flows as it can be seen from the system equations. close to the set point and undershoots due to the higher moisture
The NMPC approach performs significantly better in all aspects. content. These results are not satisfactory for the plant operation,
Note the finer action on the temperature and also the rapid especially because the temperature deviation is very large. As a
decrease of the classifier speed during the step. By decreasing Csp consequence, less sewage sludge should be added to the coal
the outlet coal mass flow is increased almost instantaneously. This mass flow. Analyzing the inputs, it can be found that T in
a increases
results in a very fast coal flow response time of 18 s. quite slowly, causing this significant fall in outlet temperature.
A. Cortinovis et al. / Control Engineering Practice 21 (2013) 308–320 319

8 9
8
7
7
6 6
5 Conventional NMPC Reference 5
0 200 400 600 800 1000 0 200 400 600 800 1000

80
14
70
12
60 10
50 8
0 200 400 600 800 1000 0 200 400 600 800 1000

100 400

95 300

90 200

85 100
0 200 400 600 800 1000 0 200 400 600 800 1000

160

0.8 140

0.75 120

0.7 100
0 200 400 600 800 1000 0 200 400 600 800 1000

2000
16
1500 14
Mcoarse Mfine Mcoarse Mfine 12
1000 10
8
500
0 200 400 600 800 1000 0 200 400 600 800 1000

Fig. 13. Simulated results of output variables (left side) and corresponding inputs (right side) for a moisture disturbance during a load change.

The NMPC approach achieves again a fast response time for the instantaneously, this input is kept as constant as possible, while
outlet coal mass flow, while keeping the temperature close to the the inlet air temperature, which is affecting the outlet temperature
set point. A similar action as in the previous scenario is observed. more slowly, is used to compensate for the uncertainty and there-
Csp is used to increase m _ out
c almost instantaneously. This results in fore varies more significantly.
a temporary loss of fineness, which recovers immediately after These results prove the usability of the presented model for
the set point is reached. Due to the faster heating, the tempera- control and confirm the clear benefit of model-based control for
ture deviates only slightly from the set point. As a consequence of coal mills. The decisive difference between the conventional control
the faster response for the temperature and the outlet coal mass strategy and the NMPC approach is the embedding of constraints
flow, the pressure drop is increased compared to the conventional into the problem formulation. Furthermore, the control design
control strategy. However, NMPC allowed this to occur, because of process is simplified and made more intuitive through weighting
the large margin to the Dpch constraint. of the desired objectives. This is in contrast to the SISO controller
To show that the proposed arrangement is also usable in a more design, which is not straightforward and the interaction between
realistic environment, the NMPC simulation for the load step control loops have to be considered. Therefore, for coal mill control
scenario has been repeated taking into account measurement noise the application of NMPC is a recommended solution.
as well as noise on the inlet coal flow. The noise properties have
been identified from measurement data and added to the simula-
tion environment. For the two unmeasured states Mcoarse and Mfine 8. Conclusions
Gaussian distributed numbers around 75% of the nominal values
of the states were used to emulate an internal process noise. In In the present paper, the modeling of a vertical roller pressure
Fig. 14 the results of the realistic noise scenario are shown. As it can coal mill was addressed. As a result, a nonlinear model has been
be seen from the results, there are no major influences caused by developed and validated to predict the coal mill behavior during
measurement, process and input noise on the control performance. start up and normal operation. The parameter optimization and
Obviously the signals became more noisy and there is a larger the model validation have been conducted with real coal mill
action on the plant inputs to compensate for the uncertainty and measurements. It has been found that an adaptation is necessary
fulfill the control requirements. This effect can be seen mainly on for the coal moisture and for the pressure drop offset. This
the inlet air mass flow and temperature. Due to the fact that the parameter correction was performed with the help of an extended
inlet air mass flow affects the outlet coal mass flow almost Kalman filter (EKF). By using this approach, parameters and states
320 A. Cortinovis et al. / Control Engineering Practice 21 (2013) 308–320

8
8
7
6
6
Reference 5
4
0 50 100 150 200 0 200 400 600 800 1000

80
14
70
12
60
10
50
0 200 400 600 800 1000 0 200 400 600 800 1000

96 400

95 300

200
94 Reference
100
0 200 400 600 800 1000 0 200 400 600 800 1000

Fig. 14. Simulated results of output variables (left side) and corresponding inputs (right side) for a load change scenario with measurement, process and input noise.

are estimated at the same time. The parameter variations were Krueger, K., & Franke, R. (2003). On-line optimization of drum boiler start-up . ABB
Utilities GmbH, UTD/PAT, Mannheim, Germany.
found to be slow, such that a constant parameter assumption for
Krueger, K., Franke, R., & Rode, M. (2004). Optimization of boiler start-up using a
control studies can be made. The model could predict the nonlinear boiler model and hard constraints. ABB Utilities GmbH, UTD/PAT,
evolution of the main output variables in an accurate way, while Mannheim, Germany.
Lagarias, J. C., & Reeds, J. A. (2007). Convergence properties of the Nelder–Mead
still being simple enough for a real-time control environment.
simplex method in low dimensions. SIAM Journal of Optimization, 9(1),
Therefore, the model can be used for several on-line applications 112–147.
like coal mill control, performance monitoring and maintenance Leimbach, R. A. (2007). Drum Pressure the key to managing boiler stored energy.
services. Metso Automation on /www.powermag.comS.
Leopold, T. (2009). Boiler-tuning basics. ABB on /www.powermag.comS.
After the modeling studies, the coal mill model was used for an Lindsley, D. (2000). Power-plant control and instrumentation. Cambridge University
advanced control application. The performance of the conven- Press.
tional control strategy was compared to that of nonlinear model Liptak, B. G. (2006). Instrument engineers’ handbook, process control and optimiza-
tion. CRC, Taylor and Friends Group.
predictive control (NMPC). The design of the NMPC approach Maciejowski, J. M. (2000). Predictive control with constraints. Prentice Hall, Pearson
amounted to the choice of cost functions, the weighting of the Education.
costs and the determination of appropriate constraints. The Magni, L., Raimondo, D. M., & Allgoewer, F. (2009). Nonlinear model predictive
control: Towards new challenging applications. Springer Verlag.
simulation results showed that the conventional control is quite Mercangoez, M., & Poland, J. (2011). Coal mill modeling for monitoring and
conservative with respect to the pressure drop levels. This limits control. IFAC World Congress.
achieving a better performance for the coal load response time Miller, B. G. (2010). Clean coal engineering technology. Elsevier, Butterworth-
Heinemann.
and temperature control. Using the advanced controller, a sig- Niemczk, P., & Andersen, P. (2009). Derivation and validation of a coal mill model
nificantly faster coal load response time is achieved while main- for control. Department of Electronic Systems, Aalborg University, Denmark.
taining the temperature and pressure drop in the desired ranges. Niemczk, P., Bendtsen, J. D., & Ravn, A. P. (2012). Derivation and validation of a coal
mill model for control. Control Engineering Practice, 20(5).
The NMPC approach exploits the system characteristics in order
Odgaard, P. F., & Mataji, B. (2006). Fault detection in coal mills used in power
to fulfill the control objectives and safety requirements. This plants. In Proceedings of IFAC symposium on power plants and power systems.
results in a more efficient and safer operation of the coal mill Odgaard, P. F., & Mataji, B. (2008). Observer-based fault detection and moisture
compared to conventional coal mill control. estimating in coal mills. Control Engineering Practice, 16(8).
Odgaard, P. F., Stoustrup, J., & Mataji, B. (2007). Preventing performance drops of
coal mills due to high moisture content. In Proceedings of the European control
References conference.
Shin Hwi-Beom, & Li Xin-lan (2009). Modeling and parameter identification of coal
mill. Department of Electrical Engineering, Gyeongsang University, Korea.
European Commission (2006). Large combustion plants, integrated pollution Simon, D. J. (2006). Optimal state estimation: Kalman H1 and nonlinear approaches
prevention and control. European Commission. (1st ed.). John Wiley & Sons Inc..
Fan, G. Q., & Rees, N. W. (1994). Modeling of vertical spindle mills in coal fired power UNEP (2006). Boilers and thermic fluid heaters. United Nations Environment
plants. Sydney, Australia: School of Electrical Engineering UNSW. Programme (UNEP).
Flynn, D. (2003). Thermal power plant simulation and control. In IEE power & Waechter, A., & Biegler, L. T. (2006). On the implementation of an interior-point
energy series. filter line-search algorithm for large-scale nonlinear programming. Mathema-
Franke, R., & Vogelbacher, L. (2006). Nonlinear model predictive control for tical Programming, 106(1), 25–57.
Cost Optimal Startup of steam power plants. Automatisierungstechnik, 54(12), Wei Jian-Lin, & Wang Jihong (2007). Development of a multisegment coal mill
630–637, Oldenbourg-Link. model using an evolutionary computation technique. Department of Electrical
Grüne, L., & Pannek, J. (2011). Nonlinear model predictive control. Theory and Engineering, University of Liverpool, UK.
algorithms. London: Springer. Zhang, Y. G., & Wu, Q. H. (2002). Coal mill modeling by machine learning based on
Haugen, F. (2009). Lecture notes in models, estimation and control. Technical Teach, onsite measurements. Beijing: Electric Power Research Institute.
ISBN: 798-82-91748-14-6. Zhou, Guian, & Si, Jennie (2000). Modeling and simulation of C–E deep bowl
Haykin, S. (2001). Kalman filtering and neural networks. John Wiley and Sons, Inc.. pulverizer. Department of Electrical Engineering, Arizona State University.

You might also like