Professional Documents
Culture Documents
Abstract: Accurate control of the engine air system is essential to comply with emission
legislation. Typically the real emission targets, like N Ox , cannot be measured directly at
a production car and thus different intermediate candidates, like gas mass flows or oxygen
concentrations, are used. In general the candidate choice is a key factor for emissions, because
each of them can only be determined with certain accuracy and consequently a feedback control
based on a specific quantity is only capable of reaching a point within the accuracy range.
Moreover, a slight deviation of one candidate may, due to the nonlinear relations between
emissions and candidates, lead to high N Ox deviations. In a previous work four candidates for
EGR valve control were analyzed and a scheduling scheme according to their N Ox tracking
performance under uncertainties was proposed. Within this work the approach is extended by
the application of a multi reference online MPC control strategy which takes also dynamical
properties into account. To this end a scheduling between the candidates based on the static
and dynamic analysis in the previous work is applied, where the idea is to adapt the inherited
objective function during runtime. The proposed dynamic reference scheduling strategy was
implemented on a simulation model of a passenger car Diesel engine and satisfying results were
obtained.
with one candidate, where the reference was adapted in available, as it is typically the case in production standard
steady state to match the candidate with the minimal SISO control structures. The proposed methodology, to
N Ox deviation. find the best reference in each operating range, can be
used within a MIMO control approach. However, in this
In the following this idea is continued and the approach
work, all other quantities, like turbo charger guide vane or
is extended by the application of a predictive control
throttle positions, cannot be altered and are determined
approach. The idea is to identify a model of the air system
by separate controllers.
and then to use all reference candidates simultaneously,
where scheduling or weighting of the candidates, accord- As already stated above, different intermediate quantities
ing to the previous results, is performed directly in the can be used to control the amount of EGR and thus the
MPC objective function. To this end, the MPC problem obtained emission levels. Within this work four different
is formulated as QP and implemented with qpOASES, candidates are considered. These are the fresh air mass
Ferreau et al. (2008), with the possibility to adapt the flow (M AF ), the EGR rate (rEGR ) determined by the
objective function during runtime. qpOASES is an open mass fraction of recirculated exhaust gas, the exhaust
source online QP solver which uses an active set strategy manifold oxygen concentration (O2,ex ) and the intake
and is especially suited for real time MPC applications. manifold oxygen concentration (O2,in ). Two of these four
For example, the same solver was already used in Ferreau candidates can be measured directly with production
et al. (2007) to control the whole air system of a Diesel standard sensors, while the other two candidates (O2,in
engine by online MPC. and rEGR ) are determined based on model assumptions
and several measurements, like boost pressure or intake
This paper is organized as follows. In Section 2 the
temperature. A detailed description of the necessary steps
considered system and the previous work leading to the
for the estimation can be found in Waschl et al. (2012b).
selection strategy are presented. Section 3 describes the
However, for clarification also here the basic equations are
MPC approach and the dynamical scheduling strategy
presented. First the cylinder intake mass flow ṁcyl,in is
and in Section 4 evaluation results based on a MVEM
determined by simple mass balance equations and the ideal
simulation are presented and discussed. Finally in Section
gas law
5 the conclusions and future work directions are given.
M AP · v̇cyl,in
ṁcyl,in = (1)
2. PROBLEM AND SYSTEM DESCRIPTION Tintk · R
Veng · n · ηV ol
with: v̇cyl,in = (2)
A model of a 2l 4 cylinder Diesel engine, equipped with 2
common rail direct injection system, cooled high pressure where Veng is the engine displacement, ηV ol the volumetric
EGR and a variable geometry turbine turbo charger is con- efficiency, v̇cyl,in the volumetric flow rate from the intake
sidered, where a schematic representation is given in Fig. 1. manifold to the cylinders, R is the ideal gas constant and
It should be noted that the focus is on the EGR path of Tintk is the temperature in the intake manifold.
Finally the EGR rate can be calculated by
ṁegr
rEGR = (3)
M AF + ṁegr
with: ṁegr = ṁcyl,in − M AF (4)
and the intake oxygen concentration is given by
ṁegr · O2,ex + M AF · O2,air
O2,in = (5)
ṁegr + M AF
where O2,air is the mass based oxygen concentration of the
fresh air.
475
2012 IFAC E-CoSM (E-CoSM'12)
Rueil-Malmaison, France, October 23-25, 2012
Espd Q
O2,ex Reference R
wf Scheduling S
MAF
476
2012 IFAC E-CoSM (E-CoSM'12)
Rueil-Malmaison, France, October 23-25, 2012
current system state information for the MPC and addi- the used sensor principles. As stated above and in the
tionally to provide filtering for the current system outputs previous works, the optimal feedback reference candidate
which were then used for MPC. in steady state does not necessarily lead to a suitable
tracking performance during dynamic scenarios. In the
3.2 Online MPC multi reference scheduling approach this behavior should
be taken into account, first by the model based controller
For the multi reference EGR-valve control with the identi- and second by a smart weighting scheduling. To this end
fied model an online capable MPC implementation was the following scheme is proposed; during transient changes
used. In the considered linear discrete time case it is the reference candidate with the fastest dynamic response,
possible to condense the MPC problem as a QP, see according to the previous results rEGR , is selected and
e.g. Maciejowski (2002), which is solved online with the afterwards the weighting of the optimal reference in steady
realtime capable solver qpOASES . operation, determined by the static analysis, is increased
stepwise, while the weightings of the other candidates are
In MPC the applied control signal is calculated based on decreased.
the optimization of an objective function which utilizes
the system model to predict the future outputs and so The strategy to determine the weighting qµ,k for each
determines the optimal control inputs. In our SIMO case candidate µ can be formulated as:
the optimization problem can be stated as if setpoint change detected then
nP H qrEGR ,k = qrEGR ,max
1X T qµ,k = qµ,min ∀µ 6= rEGR
min (yk − yref,k ) Qk (yk − yref,k ) + ∆uk Rk ∆uk
u 2 else
k=0
+ uk Sk uk (9) if mod(k, nupdate ) = 0 then
s.t. −1 ∀Sµ,N Ox 6= min (Sµ,N Ox )
ζµ ∈
uk = uk−1 + ∆uk 1 ∀Sµ,N Ox = min (Sµ,N Ox )
xk+1 = Ap xk + Bp uk yk = Cp xk qµ,k = qµ,k + ζµ · α · qµ,k−1
u ≤ uk ≤ u k = 0 . . . nCH − 1 else
∆uk = 0 k = nCH . . . nP H . qµ,k = qµ,k
end if
The optimization is performed according to the objective end if
function and over the prediction horizon with the length qµ,k = min (qµ,k , qµ,max )
nP H samples, where control inputs are determined only qµ,k = max (qµ,k , qµ,min )
for the control horizon with nCH samples. The influence
of the different entries on the objective function and so the with the minimal qµ,min and maximal qµ,max allowed en-
optimization result is defined by the weighting matrices Q, tries for each qµ . Additionally two parameters α ∈ [0, 1]
R and S. The deviations from each of the four reference and nupdate are introduced. α defines the increase and
candidates yµ − yµ,ref are weighted by the matrix Q which decrease rate of the weightings to the maximal or min-
is defined by imal allowed entries of qµ . With nupdate the time scales
qM AF ,k 0 0 0
of the discrete time MPC and the weighting algorithm
0 qO2,ex ,k 0 0 can be separated. This is necessary because the in- and
Qk = . (10) decrease after a setpoint change should be adjusted to
0 0 qO2,in ,k 0
0 0 0 qrEGR ,k the specific rise and settle times of the used candidates
and these are typically slower than the used system sam-
Moreover, due to the structure of Q it is possible to
pling time. Furthermore, it should be mentioned that
reformulate the condensing problem in dependency of one
in the current approach a setpoint change is detected if
of the four weightings and aggregate the Hessian and
the gradient of engine speed or fuel amount exceeds a
the gradient of the uprising QP online by straightforward
specific threshold. An example for the weighting schedul-
matrix operations. The necessary operations for this online
ing performed by the algorithm is depicted in Fig. 4,
tuning approach of a similar MPC application in QP
where at the time instants t = 0 s and t = 5 s a set-
formulation and with the identical solver are described
point change was performed. The setpoint of the engine
more elaborately in Waschl et al. (2011). In addition, in
was changed from (Espd = 1750 rpm, wf = 5 mg/cyc) to
the objective the overall control effort can be weighted by
(Espd = 2000 rpm, wf = 5 mg/cyc). In this figure also the
S and the control advance is penalized by R, which are
optimal static reference to which the weighting should
both scalar values in the considered application.
converge, as given in Fig. 2 is presented.
3.3 Reference scheduling strategy 4. RESULTS
Although in steady conditions all four reference candidates For evaluation of the proposed strategy a simulation en-
can be transformed into each other, in the dynamic op- vironment based on a mean value model in combination
eration different responses arise, even in the perfect case with an emission model was used. Mean value models are
without sensor errors. More detailed information on the frequently used for control design orientated engine model-
different dynamics, like step responses or rise times, is ing and the particular model was calibrated with measure-
presented in Waschl et al. (2012a). These differences are ments of the real engine and provides reasonable results in
caused by physical properties, like system dynamics or comparison with real measurements over a driving cycle,
mixture and transport effects, and can also depend on see Alberer (2009). To determine the N Ox emissions a
477
2012 IFAC E-CoSM (E-CoSM'12)
Rueil-Malmaison, France, October 23-25, 2012
500
via calculation to determine the four reference candidates.
0
0 2 4 6 8 10 The given values are based on sensor properties provided
1000 by the manufacturers and all are given as multiplicative
qO2,ex
500
Sensor error
0
0 2 4 6 8 10 ∆M AF/M AF 5% ∆O2,ex /O2,ex 2.5%
1000 ∆M AP/M AP 2% ∆v̇cyl,in /v̇cyl,in 3%
qrEGR
478
2012 IFAC E-CoSM (E-CoSM'12)
Rueil-Malmaison, France, October 23-25, 2012
300
based on the sensor accuracies is similar to a worst case
250 analysis. Consequently, it may happen that the real sensor
200
Reference errors used to calculate a reference quantity cancel each
0 5 10 15 20MAF MPC25 30
60 Optimal scheduling MPC other out and thus lead to lower N Ox deviations than
40
the MPC based on the scheduling strategy. Nevertheless,
the real values are unknown at a production engine and
20
0 5 10 15 20 25 30 thus the reference scheduling provides a sensible choice to
40
improve the N Ox tracking performance under worst case
xEGR
30
assumptions.
20
0 5 10 15 20 25 30 In another test the specified sensor error boundaries from
NOx in g/h
4
Table 1 in all possible 32 combinations, i.e. positive or
3
2
negative relative error for each sensor, were used. The
0 5 10 15 20 25 30 scenario was identical to the test with the four different
t in s sets and as evaluation criterion the summed N Ox tracking
error for all combinations was chosen. The results are
Fig. 6. Detailed part of Fig. 5, during the first setpoint summarized in Table 4 and again with the multi reference
change between t = 5 s and 30 s scheduling the best result could be achieved.
one reference candidate. For example, in the first case Table 4. Overall N Ox tracking performance for
a MPC which focuses on M AF tracking was designed, 32 sensor error combinations
to this end the value for the M AF weighting qM AF was Control M AF O2,ex O2,in rEGR opt.scheduling
set to the maximum value of qµ and all other tracking 32
P
eN Ox ,i 27.9% 27.1% 30% 100% 13.5%
errors were weighted with qµ,min . In Fig. 5 a comparison i=1
of the tracking behavior for the solely on M AF based MPC
approach and the optimal scheduling is presented. The test 5. CONCLUSIONS AND FUTURE WORK
scenario in this example was a series of fuel amount steps
at constant engine speed and sensor errors from Set1 were A multi reference scheduling algorithm in combination
applied during the simulation. with MPC framework for the control of the EGR-Valve
As it can be seen in Fig. 5, the multi reference scheduling of a Diesel engine was presented. The aim was to achieve
leads to lower deviations of N Ox compared to a controller minimal N Ox tracking errors in the case of erroneous
focusing only on M AF . In this scenario in steady state sensor readings used for feedback control. This work was
first rEGR , afterwards M AF and finally O2,ex led to the an extension of previous analysis where it was shown that
smallest deviations, as determined by the selection strat- the choice of reference is crucial for emission control with
egy. To provide further insight into the dynamic behavior intermediate quantities. The strategy was applied on a
a detailed region of the same test, during the change from nonlinear simulation model and led to satisfactory results.
rEGR to M AF is presented in Fig. 6. Moreover, in this It was possible to achieve a better N Ox tracking perfor-
figure also xEGR is presented. mance than a control based on one reference candidate
under the assumption of different sensor errors. Although
An additional scenario was used to verify the dynamic the implementation was performed on a simulation model,
control performance of the scheduling strategy, where a later implementation on a real engine at a testbench
at a constant engine speed of Espd = 2150 rpm a se- was kept in mind, thus the real time capable MPC and
quence of fuel amount steps, occurring each 25 s, was scheduling strategy can be transferred to real engine with-
performed. The sequence was given by eight values wf = out substantial changes.
[7, 9, 2.5, 6, 10, 8, 15, 4] mg/cyc. In order to evaluate the
performance under different sensor error combinations, In future work we will focus on the evaluation of the
four different sensor error sets listed in Table 2 were strategy on a real engine, under different test scenarios
examined and the results are presented in Table 3. As and the combination with a MIMO model predictive
performance criterion (11) was used, whereas all values control approach for the whole air system with respect
are given in percent and normalized to the worst value for to N Ox tracking performance. An additional task will
each set. be the reduction of the necessary measurements for the
determination of the N Ox sensitivity and the extension
Table 3. N Ox tracking performance for test of the operating range, e.g. by multiple linear models or
scenario nonlinear MPC.
Set1 Set2 Set3 Set4
eN Ox ,M AF 100.0% 8.3% 32.0% 22.3%
ACKNOWLEDGEMENTS
eN Ox ,O2,ex 91.4% 9.4% 61.1% 41.2%
eN Ox ,O2,in 50.4% 13.7% 73.7% 68.6% The authors gratefully acknowledge the sponsoring of
eN Ox ,rEGR 73.3% 100.0% 100.0% 100.0% this work by the COMET K2 Center ”Austrian Center
eN Ox ,opt.scheduling 32.2% 5.8% 19.4% 13.5% of Competence in Mechatronics (ACCM)” and the JKU
Hoerbiger Research Institute for Smart Actuators (JHI).
These results demonstrate that for each sensor combina- The COMET Program is funded by the Austrian Federal
tion it was possible to achieve the minimal N Ox tracking Government, the Federal State Upper Austria and the
error with the multi reference scheduling MPC strategy. Scientific Partners of ACCM.
479
2012 IFAC E-CoSM (E-CoSM'12)
Rueil-Malmaison, France, October 23-25, 2012
350 1000
MAF in mg/cyc
300
qMAF
500
250
200 0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
20 Reference
100
MAF MPC
O2,ex in %
qO2,ex
15
50
10
0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
22 1000
O2,in in %
20
qO2,in
500
18
16 0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
60 1000
rEGR in %
50
qrEGR
500
40
30 0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
Espd in rpm
wf in mg/cyc
2200 10 4
NOx in g/h
7.5
2100 3
Espd 5
2000 wf 2.5 2
0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
t in s t in s
Fig. 5. Comparison between M AF based and optimal scheduling MPC for sensor errors specified in Set1
480
2012 IFAC E-CoSM (E-CoSM'12)
Rueil-Malmaison, France, October 23-25, 2012
481