The Engineering Meetings Board has approved this paper for publication.

It has successfully completed SAEs peer review process under the supervision of
the session organizer. This process requires a minimum of three (3) reviews by industry experts.
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic,
mechanical, photocopying, recording, or otherwise, without the prior written permission of SAE.
ISSN 0148-7191
Positions and opinions advanced in this paper are those of the author(s) and not necessarily those of SAE. The author is solely responsible for the content of
the paper.
SAE Customer Service: Tel: 877-606-7323 (inside USA and Canada)
Tel: 724-776-4970 (outside USA)
Fax: 724-776-0790
Email: [email protected]
SAE Web Address: http://www.sae.org
Printed in USA
2009-01-0733
Multi-Variable Air-Path Management for a Clean Diesel Engine Using Model
Predictive Control
Mitsuhiro Iwadare, Masaki Ueno
Honda R&D Co., Ltd.
Shuichi Adachi
Keio University
Copyright 2009 SAE International
ABSTRACT
Recently, emission regulations have been strict in many
countries, and it is very difficult technical issue to reduce
emissions of diesel cars.
In order to reduce the emissions, various combustion
technologies such as Massive EGR, PCCI, Rich
combustion, etc. have been researched. The combustion
technologies require precise control of the states of in-
cylinder gas (air mass flow, EGR rate etc.). However, a
conventional controller such as PID controller could not
provide sufficient control accuracy of the states of in-
cylinder gas because the air-pass system controlled by
an EGR valve, a throttle valve, a variable nozzle turbo,
etc. is a multi-input, multi-output (MIMO) coupled system.
Model predictive control (MPC) is well known as the
advanced MIMO control method for industrial process.
Generally, the sampling period of industrial process is
rather long so there is enough time to carry out the
optimization calculation for MPC. However, due to the
progress of the computer in the last decade and
improvement of the optimization algorithm, the MPC can
be applied to fast process such as mechanical systems.
The air-pass management system for the diesel engine
is assumed a two-input, two-output system in this
research. The inputs of the system are the throttle valve
and the EGR valve, and the outputs air mass flow and
EGR rate. Consequently, MPC was applied to air-pass
management system and was modified by adding the
disturbance observer to eliminate steady-state error and
the compensator for nonlinear characteristics of
actuators.
The performance of the proposed control system was
examined by using an actual testing vehicle. From the
experiments, it was shown that an accurate decoupled
control of two outputs, i.e. air mass flow and EGR rate,
was accomplished by the proposed MPC control.
INTRODUCTION
Diesel engines display higher thermal efficiency and emit
less CO
2
than gasoline engines, and have therefore
attracted attention as a form of internal combustion
engine that is effective in combating global warming.
However, diesel engines are basically lean combustion
engines, and the simple after-treatment systems
involving three-way catalysts with close to 100%
reduction efficiency that are employed in gasoline
engines cannot be employed in them. This makes it
necessary to reduce the amount of NOx, HC, soot and
other emissions that are found in the exhaust gas in the
combustion process in the cylinders.
The reduction of NOx in diesel engines is particularly
important, and exhaust gas recirculation (EGR) systems
are well known as an effective means of doing so. The
higher the maximum temperature of combustion gas is
the more NOx increases. Consequently, EGR reduces
NOx by mixing the inert gases present in the exhaust
with the intake air, reducing the oxygen concentration,
and thus reducing the maximum temperature of
combustion. However, as EGR is steadily increased,
there is a rapid increase in the volume of soot produced
above a certain excess air factor. In order to effectively
reduce NOx while controlling soot production, therefore,
it is necessary to accurately control the state of


combustion (parameters such as the excess air factor,
etc.) to within a desired range by controlling EGR and
fuel injection.
PCCI combustion has been researched as a technology
to reduce both NOx and soot emissions. PCCI
combustion is premixed one achieved by increasing
EGR amount and extending the ignition delay period.
However, as indicated by the results of tests on steady-
state PCCI combustion (Figures 1, 2), the state of the
intake EGR produces significant variations in the
characteristics of emissions. Therefore, the tracking
response characteristics of fuel injection control and
EGR control should be improved to achieve PCCI
combustion during transient operating conditions.
There are, however, limits of further reductions in
engine-out emissions. So, NOx exhaust after-treatment
system, for example a NOx catalyst, is recently
researching to enable further emissions reductions. The
NOx catalyst system requires a regenerative combustion
control known as rich control, which reduces the NOx
that is adsorbed during lean combustion [1]. The rich
control requires high-accuracy fuel injection control
tracking air-fuel ratios on target values in order to keep
combustion noise performance and drivability.
Furthermore, high-accuracy intake EGR control is also
important to realize this combustion control.
The various requirements discussed above indicate that
the necessity for more precise control of the state of
intake EGR is increasing.

1. DEVELOPMENT AIMS
The EGR valves, throttle valves, etc., are the intake
actuators to directly operate the controlled object in
intake control. The controlled outputs of the controlled
object are oxygen and inert gases in the cylinder. The
amounts or ratios meanings these two parameters also
used as the controlled output.
However, in general, if one of the intake actuators is
changed, both controlled outputs also change. The
system indicates this behavior is known as a coupled
system, and in such cases to design feedback control
systems using conventional methods, such as PID
controllers, is known to represent a challenge. Figure 3
shows oscillations caused by interference when a basic
PID controller is employed in an engine simulator
demonstrating behaviors of the controlled object
discussed above.
Model predictive control (MPC), which enables multiple
outputs to be controlled to target values, is known as a
control algorithm that can be applied to multi-input, multi-
output systems. Because MPC requires optimization
calculations to be performed in real-time for each
sampling period, it is mainly employed in applications in
which the sampling period is long, such as in
petrochemical plants.
However, increases in calculation speed with higher-
speed processing of ECU and improvements in
calculation algorithms [2] have increased the potential for
the application of MPC to comparatively fast systems [3].
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
25 30 35 40 45
EGR rate(%)
N
O
x

(
g
/
k
W
h
)
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
S
o
o
t

(
g
/
k
W
h
)
Convential-NOx
PCCI-NOx
Conventinal-Soot
PCCI-Soot
Figure 1 Emission characteristics of conventional
and PCCI combustion

0.0
0.2
0.4
0.6
0.8
1.0
1.2
25 30 35 40 45
EGR rate (%)
d
P
/
d

m
a
x

(
M
p
a
/
d
e
g
)
0.0
1.0
2.0
3.0
4.0
5.0
6.0
7.0
8.0
C
O
V

[
%
:
]
PCCI-dPdt
Conventinal-dPdt
PCCI-COV
Conventinal-COV
d
P
/
d

m
a
x

(
M
P
a
/
d
e
g
)
C
O
V

(
C
o
e
f
f
i
c
i
e
n
t

o
f

v
a
r
i
a
t
i
o
n
)
o
f

I
M
E
P

(
%
)
PCCI-dP/d
Conventional-dP/d
PCCI-COV
Conventional-COV
0.0
0.2
0.4
0.6
0.8
1.0
1.2
25 30 35 40 45
EGR rate (%)
d
P
/
d

m
a
x

(
M
p
a
/
d
e
g
)
0.0
1.0
2.0
3.0
4.0
5.0
6.0
7.0
8.0
C
O
V

[
%
:
]
PCCI-dPdt
Conventinal-dPdt
PCCI-COV
Conventinal-COV
d
P
/
d

m
a
x

(
M
P
a
/
d
e
g
)
C
O
V

(
C
o
e
f
f
i
c
i
e
n
t

o
f

v
a
r
i
a
t
i
o
n
)
o
f

I
M
E
P

(
%
)
PCCI-dP/d
Conventional-dP/d
PCCI-COV
Conventional-COV

Figure 2 Noise and vibration characteristics of
conventional and PCCI combustion
65
70
75
80
85
0 2 4 6 8
P
r
e
s
s
u
r
e
[
k
p
a
]
Turget pressure
Actual pressure
10
20
30
40
0 2 4 6 8
M
a
s
s

f
l
o
w
[
g
/
s
]
Turget mass f low
Actual mass f low
5
15
25
35
0 2 4 6 8
time [sec]
V
a
l
v
e
[
%
]
EGR
Thrott le

Figure 3 PI control simulation results


The project discussed here applied MPC in the
construction of an air mass flow/EGR feedback control
system for a diesel engine. This paper will mention the
basic configuration of the system and discuss the results
of tests conducted to verify its effectiveness.

2. MODELING
Figure 4 shows the configuration of the intake and
exhaust system of a diesel engine. The project
discussed here assumed a two-input, two-output
controlled object, with input and output parameters
defined as follows:
Inputs: EGR valve and throttle valve
Outputs: Intake chamber pressure and mass flow
rate of new air intake (air mass flow)

When employing MPC or any other model-based control
theory such as optimal control or adaptive control, the
dynamic characteristics of the controlled object must be
modeled by using the simultaneous equations of
differential or difference equations. In other word, the
dynamic characteristics are expressed using state-space
equations.
A model based on physical equations was therefore first
constructed to enable the configuration of the controlled
object to be understood, and the order of the state-space
equations was estimated. Next, system identification was
employed to determine the physical parameters in the
state-space equations in order to obtain the state-space
equations employed in the control theory incorporated in
a real engine.

2.1. PHYSICAL MODELING - As Figure 4 shows, the
intake and exhaust system of a diesel engine is made up
of valve and chamber elements. Here, an example of
modeling based on physical principles will be provided,
focusing on the valve and chamber in its wake shown in
Figure 5.
The mass flow rate past the throttle valve is expressed
using a convergent nozzle equation [4], [9].
( )

\
|
+

|
|

\
|
+
|

\
|
+
|

\
|
+
>
|
|

\
|

|
|

\
|

|
|

\
|

=
=

+
1
0
i
1
1
0
i
1
0
i
2
0
i
0 0 th th th
1
2
,
1 1
2
1
2
,
1
2

p
p
p
p
p
p
p
p
p A m
(1)
where:
th
m : Mass flow in throttle valve section, A
th
: Effective
opening area,
th
: Valve opening, p
0
: Upstream
pressure,
0
: Upstream density, : Specific heat ratio,
p
i
: Downstream pressure
From the equation of state for an ideal gas in the cylinder
and the principle of conservation of mass [10]
i i i i i
T R M V p = (2)
cyl th
i
m m
dt
dM
= (3)
where:
cyl
m : Mass flow in cylinder inflow part
Pressure variation in the intake chamber can be
expressed by the following equation:
( )
cyl th
i
i i
m m
V
RT
n
dt
dp
= (4)
where:
n : Polytrophic index
Input2: EGR valve
Output 1:
Intake chamber
pressure
Exhaust
chamber
Cylinder
Intercooler
Input1: Throttle valve
Output 2:
Air mass flow rate
Turbo-
charger
Catalyst
Input2: EGR valve
Output 1:
Intake chamber
pressure
Exhaust
chamber
Cylinder
Intercooler
Input1: Throttle valve
Output 2:
Air mass flow rate
Turbo-
charger
Catalyst


Figure 4 Air-path system of diesel engine
and definition of the controlled object
Mass flow rate
Valve opening
Pressure
Temperature
Volume
th
m
cyl
m
i
p
i
V
0
p
i
T
th

Mass flow rate

Valve opening
Pressure
Temperature
Volume
th
m
cyl
m
i
p
i
V
0
p
i
T
th

Figure 5 Element model of valve
and chamber system


In addition, because the cylinder intake mass flow rate,
m
cyl
, is expressed using a function almost exactly
proportional to the chamber pressure, p
i
, the following
equation
i e cyl
v
p n k m

= (5)
where:
n
e
: Engine speed,
v

k : Factor including volume

efficiency
From equations (1), (4) and (5), enabling expression as a
first-order lag system.
( )
th th i
i
A b p a
dt
dp
+ = (6)

= =
0 0
i
i
e
i
i
2 ,
v
p
V
RT
n b n k
V
RT
n a
However, because both the lag parameter a and gain
parameter b are expressed as nonlinear functions, the
system is a nonlinear system in which dynamic behavior
varies depending on the point of the system under
consideration. A linear control theory incorporating MPC
could not be applied to the nonlinear system.
Linearization of the system was therefore attempted
using a method of linearizing input [4]. Figure 6 shows
the control system formed by linearizing input. By
positioning the linearization of input using convergent
nozzle equation (1) in the controller output, the controlled
object from the perspective of the controller can be
regarded as linear. It is therefore possible to construct a
linear controller with the equations above as the
equations of state.
The controlled object in this paper discussed here
incorporated element systems other than the example
offered above, but they were all able to be expressed as
dead time or first-order lag systems, as follows:
Actuator : first-order lag system with dead time
From position of air flow meter via chamber to
throttle valve: first-order lag system.

Therefore, the controlled object system for this project is
described by lag systems with dead time.

2.2. SYSTEM IDENTIFICATION - Using aforementioned
techniques was possible to understand the
characteristics of the controlled object, however the
parameters incorporated in the controlled object model
were made up of a large number of physical constants,
making physical determination of their value a challenge.
System identification [6] was therefore conducted in
order to obtain the equations of state used in the control
theory incorporated in a real engine based on the
following input-output data.
Input: Command value for mass flow in throttle
valve section (m
th_cmd
[g/s]) and Command value for mass
flow in EGR section (m
egr_cmd
[g/s])

Output: Intake chamber pressure (p
i_act
[kPa]) and
air mass flow (m
a_act
[g/s])

The method employed to measure input and output data
for system identification was to apply pseudo random
binary signal (PRBS) [7], and to measure the resulting

Linear
Controller
Non-linear
Linear
Controlled
object
Valve
non-linear
compensator
i_cmd
p
th
m b
) (
th th
A
i
p
Linear
Controller
Non-linear
Linear
Controlled
object
Valve
non-linear
compensator
i_cmd
p
th
m b
) (
th th
A
i
p



Figure 6 Control structure with non-linear
compensator
I
n
t
a
k
e

p
r
e
s
s
u
r
e
[
k
P
a
]
M
a
s
s

f
l
o
w

r
a
t
e
i
n
a
i
r

f
l
o
w

m
e
t
e
r

[
g
/
s
]
M
a
s
s

f
l
o
w

r
a
t
e
i
n
t
h
r
o
t
t
l
e

v
a
l
v
e

[
g
/
s
]
M
a
s
s

f
l
o
w

r
a
t
e
i
n
E
G
R
v
a
l
v
e

[
g
/
s
]
5
6
7
8
9
13
14
15
16
17
70
75
80
85
90
12
13
14
15
16
0 5 10 15
Time [sec]
I
n
t
a
k
e

p
r
e
s
s
u
r
e
[
k
P
a
]
M
a
s
s

f
l
o
w

r
a
t
e
i
n
a
i
r

f
l
o
w

m
e
t
e
r

[
g
/
s
]
M
a
s
s

f
l
o
w

r
a
t
e
i
n
t
h
r
o
t
t
l
e

v
a
l
v
e

[
g
/
s
]
M
a
s
s

f
l
o
w

r
a
t
e
i
n
E
G
R
v
a
l
v
e

[
g
/
s
]
5
6
7
8
9
13
14
15
16
17
70
75
80
85
90
12
13
14
15
16
0 5 10 15
Time [sec]

Figure 7 Input and output measurement data


two output values. Figure 7 shows the input and output
data.
The dead time of the controlled object, T, was
determined from the measured input and output time
series data using a method of estimating impulse
response [6].
Next, system identification was conducted using the
measured input and output data, with the model
assumed to be a two-input, two-output multivariable ARX
model, expressed by the equations as follows.
) 1 ( ) (
) 1 ( ) ( ) 1 (
2 1
2 1
+ =
+ + +
T k u B T k u B
k y A k y A k y
(7)
where
(

=
(

=
) (
) (
) ( ,
) (
) (
) (
2
1
2
1
k u
k u
k u
k y
k y
k y (8)
(

=
(

=
(

=
(

=
222 221
212 211
2
122 121
112 111
1
222 221
212 211
2
122 121
112 111
1
,
, ,
a b
b b
B
b b
b b
B
a a
a a
A
a a
a a
A
(9)
In order to verify the accuracy of the identified two-input,
two-output ARX model, the output of the model was
compared with the output of an actual air mass flow for
time series data (Figure 8) and step response (Figure 9).
In results for time series data, a fit rate of 85.8% was
obtained for air mass flow (output 1) and a fit rate of
80.27% was obtained for pressure (output 2), indicating
that the ARX model displays excellent accuracy. This is
believed to be due to the fact that an ideally-functioning
nonlinear compensator enabled input and output data
with a high S/N ratio to be obtained.
The frequency characteristics obtained from the ARX
model and the spectrum actually measured using
spectral analysis [6] are almost same over the entire
frequency range shown in Figure 10, demonstrating that
the ARX model displays excellent accuracy even in the
frequency domain. The results of frequency analysis also
indicated the following regarding this controlled object:
72
76
80
84
88
I
n
t
a
k
e

p
r
e
s
s
u
r
e

[
k
P
a
]
Measured output
ARX model
12
13
14
15
16
0 5 10 15
Time [sec]
A
i
r

m
a
s
s

f
l
o
w

r
a
t
e

[
g
/
s
]

Figure 8 Time series result


-1
0
1
2
3
-1
0
1
2
3
4
5
Measure
ARX model
-0.5
0
0.5
1
1.5
-1 0 1 2 3
time [sec]
-0.4
-0.3
-0.2
-0.1
0
0.1
-1 0 1 2 3
time [sec]
Input : m
th_cmd
[g/s] Input : m
egr_cmd
[g/s]
p
i
_
a
c
t

[
k
P
a
]
m
a
_
a
c
t

[
g
/
s
]
p
i
_
a
c
t

[
k
P
a
]
m
a
_
a
c
t

[
g
/
s
]
-1
0
1
2
3
-1
0
1
2
3
4
5
Measure
ARX model
-0.5
0
0.5
1
1.5
-1 0 1 2 3
time [sec]
-0.4
-0.3
-0.2
-0.1
0
0.1
-1 0 1 2 3
time [sec]
-1
0
1
2
3
-1
0
1
2
3
4
5
Measure
ARX model
-0.5
0
0.5
1
1.5
-1 0 1 2 3
time [sec]
-0.4
-0.3
-0.2
-0.1
0
0.1
-1 0 1 2 3
time [sec]
Input : m
th_cmd
[g/s] Input : m
egr_cmd
[g/s]
p
i
_
a
c
t

[
k
P
a
]
m
a
_
a
c
t

[
g
/
s
]
p
i
_
a
c
t

[
k
P
a
]
m
a
_
a
c
t

[
g
/
s
]


Figure 9 Step response


Phase lags as the frequency becomes higher, and
the controlled object is therefore a lag system
The inclination of gain in the high-frequency range
other than the path of air mass flow from the EGR
flow rate is -40 dB/dec, and the relative degree of the
controlled object is therefore 2
Because the relative degree of the controlled object
is 2, the phase lag in the high-frequency range
should be 90*2 = -180 deg, but higher phase delays
also exist

These characteristics show that the controlled object is a
non-minimum phase system with dead time. The fact
that the controlled object is basically a lag system
containing dead time indicates the validity of the
characteristics obtained from physical analysis in the
previous section.

2.3. STATE-SPACE MODELING OF AN AUGMENTED
SYSTEM - The ARX model obtained from system
identification was transformed into state-space equations
to enable it to be used as the internal model in MPC.
Expressing equations (7), (8) and (9) as a matrix from
| | O O I k C
I
O
B
B
O O O
O O I
B A A
A
k u
k y
k y
k X
=
(
(
(

=
(
(
(

=
(
(
(

=
) (
, , ,
) 1 (
) 1 (
) (
) (
1 2 2 1
(10)
In an augmented state assuming constant output
disturbance,
| |

+
(

=
(

+
+
) (
) (
) (
) (
0 ) (
) (
0
0
) 1 (
) 1 (
k d
k X
I C k y
T k u
B
k d
k X
I
A
k d
k X
(11)
Using a disturbance observer, the constant disturbance
term, d (k), is successively corrected in relation to the
augmented system model.
Now, the input and output vectors are defined below.
(

=
(

=
) (
) (
) ( ,
) (
) (
) (
a
i
egr_cmd
th_cmd
k m
k p
k y
k m
k m
k u (12)
The state-space equations can be expressed as follows.


0.01
0.1
1
10
A
m
p
r
i
t
u
d
e
-500
-400
-300
-200
-100
0
0.001 0.01 0.1 1 10
Frequency [Hz]
P
h
a
s
e

[
d
e
g
]
0.1
1
10
A
m
p
r
i
t
u
d
e
ARX model
Measure
-400
-300
-200
-100
0
0.001 0.01 0.1 1 10
Frequency [Hz]
P
h
a
s
e

[
d
e
g
]
0.01
0. 1
1
10
A
m
p
r
i
t
u
d
e
-600
-400
-200
0
0.001 0.01 0.1 1 10
Frequency [Hz]
P
h
a
s
e

[
d
e
g
]
0.01
0. 1
1
A
m
p
r
i
t
u
d
e
-800
-600
-400
-200
0
0.001 0.01 0.1 1 10
Frequency [ Hz]
P
h
a
s
e

[
d
e
g
]
Input : m
th_cmd
Output : m
a_act
Input : m
th_cmd
Output : p
i_act
Input : m
egr_cmd
Output : m
a_act
Input : m
egr_cmd
Output : p
i_act
0.01
0.1
1
10
A
m
p
r
i
t
u
d
e
-500
-400
-300
-200
-100
0
0.001 0.01 0.1 1 10
Frequency [Hz]
P
h
a
s
e

[
d
e
g
]
0.1
1
10
A
m
p
r
i
t
u
d
e
ARX model
Measure
-400
-300
-200
-100
0
0.001 0.01 0.1 1 10
Frequency [Hz]
P
h
a
s
e

[
d
e
g
]
0.01
0. 1
1
10
A
m
p
r
i
t
u
d
e
-600
-400
-200
0
0.001 0.01 0.1 1 10
Frequency [Hz]
P
h
a
s
e

[
d
e
g
]
0.01
0. 1
1
A
m
p
r
i
t
u
d
e
-800
-600
-400
-200
0
0.001 0.01 0.1 1 10
Frequency [ Hz]
P
h
a
s
e

[
d
e
g
]
Input : m
th_cmd
Output : m
a_act
Input : m
th_cmd
Output : p
i_act
Input : m
egr_cmd
Output : m
a_act
Input : m
egr_cmd
Output : p
i_act

Figure 10 Bode diagram

(
(
(
(
(
(
(
(
(
(
(

=
(

(
(
(
(

+
(
(
(
(
(
(
(
(
(
(
(

(
(
(
(

=
(
(
(
(
(
(
(
(
(
(
(

+
+
+
+
) (
) (
) 1 (
) 1 (
) 1 (
) 1 (
) (
) (
) (
) (
) (
) (
) (
) (
) 1 (
) 1 (
) 1 (
) 1 (
) (
) (
) 1 (
) 1 (
) (
) (
) (
) (
) 1 (
) 1 (
a_Afm
i
a_Afm
i
a_Afm
i
egr_cmd
th_cmd
a
i
a
i
a
i
egr_cmd
th_cmd
1
egr_cmd
th_cmd
a
i
a
i
2 2 1
egr_cmd
th_cmd
a
i
a
i
k d
k d
k m
k m
k m
k p
k m
k p
C
k m
k p
T k m
T k m
O
I
O
B
T k d
T k d
T k m
T k m
T k m
T k p
T k m
T k p
I O O O
O O O O
O O O I
O B A A
k d
k d
k m
k m
k m
k p
k m
k p
m
p
m
p
m
p
(13)
| |
(

=
(

= =
0 0
0 0
,
1 0
0 1
, O I O O O I C (14)
In this paper these equations are used in designing an
MPC controller.

3. CONTROL SYSTEMS DESIGN
3.1. CONTROL REQUIREMENT The requirements in
the design of a control system for the controlled object
that formed the subject of this research can be
summarized as follows:
1. Increase tracking speed of outputs to the changes in
target values
2. Control of the effect of mutual interference between
input and output systems
3. Correction of nonlinearity of valve flow rate
characteristic
4. Compensation for lag characteristic of pressures and
flow rates generated by actuators or volume of
chambers
5. Maintenance of control performance and stability in
relation to manufacturing variations and decline in
flow rate characteristic of actuators, etc.
6. Maintenance of stability in relation to disturbances
originating in environmental changes, etc.

3.2. CONFIGURATION OF CONTROL SYSTEM -
The design of the control system responded to the
requirements outlined in the section above using the
following elements. Figure 11 shows the configuration of
the control system.
Element Requirement
Model predictive controller 1, 2, 4
Internal model: Multivariable ARX model
Reference trajectory 1, 6
Disturbance observer 5, 6
Static nonlinear compensator for valve section
3

3.2.1. Model predictive control - The fact that MPC is
able to deal entirely naturally (effortlessly) with multi-
input/multi-output systems represents a considerable
advantage from the perspective of practical use. In
addition, the facts that its basic operating concepts are
easy to understand, and that adjustment of the control in
a real engine is simple and intuitive, are advantageous
factors from the perspective of use in practical
applications. Focusing on these advantages, MPC is
applied as a controller.
Figure 12 shows a conceptual diagram of the basic
operation of the controller. The order of procedures is as

MPC
EGR
valve
opening
Disturbance
observer
Controller
EGR valve
non-linear
compensator
Pressure
command
Reference
trajectory Pressure
Throttle
valve
opening
Throttle valve
non-linear
compensator
Mass flow
command
Mass flow
EGR
mass flow
Throttle
Mass
flow
Reference
trajectory
th b i b
, , , A T p p
egr e i e
, , , A T p p
MPC
EGR
valve
opening
Disturbance
observer
Controller
EGR valve
non-linear
compensator
Pressure
command
Reference
trajectory Pressure
Throttle
valve
opening
Throttle valve
non-linear
compensator
Mass flow
command
Mass flow
EGR
mass flow
Throttle
Mass
flow
Reference
trajectory
th b i b
, , , A T p p
egr e i e
, , , A T p p



Figure 11 Block diagram of controller

Reference trajectory r
Input u(k)
Output
y(k)
Set point s(k)
k k + H
u
Current time
k + H
p
Prediction horizon
Discrete time
Error e(k)
Prediction trajectory y
^
Control horizon
Input trajectory u
Reference trajectory r
Input u(k)
Output
y(k)
Set point s(k)
k k + H
u
Current time
k + H
p
Prediction horizon
Discrete time
Error e(k)
Prediction trajectory y
^
Control horizon
Input trajectory u


Figure 12 Basic concept of MPC


follows:
1. Formulation of reference trajectory: The speed at
which the plant converges on the set value is
specified using the real time error, e(k).
2. Calculation of prediction trajectory: Behavior of
output in the prediction horizon is predicted from the
input to the internal model and control horizon.
3. Setting input trajectory: Using optimization
calculations, an optimum combination of inputs
(input trajectory) is calculated that will achieve as
close a match as possible between the reference
trajectory, 1, and prediction trajectory, 2.
4. Application of input value: The first element of the
input trajectory, 3, u(k), is applied to the actual
controlled object.

The entire cycle from steps 1 to 4 is repeated after one
sampling cycle.

3.2.2. Disturbance observer - Because MPC is a model-
based control theory, modeling error in the controlled
object model can result in deviations from the intended
behavior of the system. In particular, if the steady-state
gain in the model is inaccurate, steady-state deviation
will occur in the controlled object output. As a result,
performance may decline due to variations or
deterioration in device characteristics or unexpected
changes in the environment, etc. A disturbance observer
was therefore introduced to the system in order to control
effects from disturbances that would be challenging to
incorporate in the model.
DMC [8], as shown in the procedures below, is known as
a method enabling a disturbance observer to be
employed in MPC. Using DMC, this project incorporated
a disturbance observer in the system that would exclude
the effect of constant disturbance.
The procedures conducted by a disturbance observer in
MPC are as follows for time k:
1. Measures actual output, y(k)
2. Estimates disturbance as the difference between
actual output and estimated output
3. Uses estimated value of disturbance to predict
output between prediction horizons

This method requires data for all of the state variables of
the controlled object, necessitating estimation of
variables that cannot be directly measured, i.e., the use
of a state observer.
Using the augmented state model incorporating the
output disturbance model introduced via equation (13),
state vectors that include constant disturbance can be
estimated. Terming the state estimation gain matrix L,
the standard observer equations can be expressed as
follows:
| |
) ( ) (
0
) 1 | (

) 1 | (
0
0
) | 1 (

) | 1 (
k y
L
L
k u
B
k k d
k k x
I C
L
L
I
A
k k d
k k x
d
x
d
x
(

+
(

+
(

|
|

\
|
(

=
(

+
+
(15)
The entire configuration formed from the MPC and the
observer is expressed as shown in Figure 13.

4. PERFORMANCE VERIFICATION USING
SIMULATION
Simulations of responses to target values were
conducted under conditions of constant engine speed
and amount of fuel injection.
Multiple calibration parameters exist for MPC. These
calibration parameters were estimated on the basis of
generally known MPC tuning parameters [8], and were
ultimately determined by trial and error in the simulation.
The effect of the coincidence point parameter is
discussed here as an example. Figures 14 and 15 show
the results of simulations of responses to target values.
As Figure 14 shows, when coincidence points are set
across the entire range between the predictions horizons,
output response is rapid, but input is unstable.
Figure 15 shows that when coincidence points are set
only at the beginning and end of the prediction horizon,
results display a comparatively good balance, with good
response and stability for both input and output. In
addition, minimizing the number of coincidence points
reduces the calculation load on the ECU.
The simulation results show rapid convergence of two
output values on two independent changes in target
values without oscillation due to interference,
demonstrating that the MPC behaves as required.

B
A
1
z I

C 1
z I

+
L
d
Reference
trajectory
MPC
Observer
K
MPC

+
+

+

) ( k u
) (k u
I
z
z
1
I z
1
) 1 ( k u
) | ( k k x
) | ( k k x
) 1 | ( k k x
) 1 | 1 ( k k x
) 1 ( k u
) 1 | ( k k y
) (k y ) (k y
) 1 | ( k k d
) (k s
B
A
1
z I

C 1
z I

+
L
d
Reference
trajectory
MPC
Observer
K
MPC

+
+

+

) ( k u
) (k u
I
z
z
1
I z
1
) 1 ( k u
) | ( k k x
) | ( k k x
) 1 | ( k k x
) 1 | 1 ( k k x
) 1 ( k u
) 1 | ( k k y
) (k y ) (k y
) 1 | ( k k d
) (k s


Figure 13 Block diagram of MPC with disturbance
observer


5. PERFORMANCE VERIFICATION IN REAL
ENGINE
Tests of responses to target values were conducted
using MPC and a conventional control method in a real
vehicle. Steady-state operating conditions were
employed, with engine speed and the amount of fuel
injection held constant.
Figure 16 shows the results of a test of step response to
intake chamber pressure target values with the target
value for air mass flow held constant. Despite a certain
degree of overshoot, MPC converges on the target value
in around 300 ms. The conventional control method was
able to converge on the target value in 500 ms when the
value was reduced, but a long convergence time
approximately 2.5 sec was required when the value
was increased. In addition, the conventional control
method displayed a certain amount of fluctuation in
actual output in relation to the target value for air mass
flow, and exact convergence was not achieved.
Figure 17 shows the test results of step response to air
mass flow target values with the target value for intake
chamber pressure held constant. MPC converged on the
target value in around 300 ms, while convergence in the
conventional control method took longer at
approximately 800 ms. In addition, the intake chamber
target pressure was constant, but the conventional
control method produced one significant reduction in the
pressure, which is believed to have been a result of
interference.
As in the case of the engine simulator results, these
results indicate that the use of MPC has enabled the
achievement of rapid convergence on two outputs for
independent target values.
Under the test conditions used here, the time required for
the conventional control method to converge on the
target values was 2.5 to 8 times that required by MPC,
and this phenomenon was particularly conspicuous when
the intake chamber pressure target value was increased,
and when the air mass flow target value was changed.
As indicated in Chapter 1, deviation in the intake state
can negatively affect engine performance, for example
by increasing emissions and combustion noise. The
application of intake control using MPC, which is able to
increase the speed of response of the intake state, can
be expected to result in a significant improvement of
emissions, combustion noise and other performance
parameters under transient operating conditions.

6. CONCLUSION
An intake control system using MPC has been applied to
the intake system of a diesel engine, a multi-input, multi-
output coupled system. The following results have been
achieved:
1. The internal model used in MPC was constructed
using a multivariable ARX model produced by means
of system identification and correction of the
nonlinearity of the actuator flow rate characteristic
based on a convergent nozzle equation. This
enabled an extremely accurate internal model to be
obtained.

I
n
t
a
k
e
p
r
e
s
s
u
r
e
[
k
P
a
]
M
a
s
s
f
l
o
w
r
a
t
e
[
g
/
s
]
V
a
l
v
e

o
p
e
n
i
n
g

[
%
]
65
70
75
80
Target
Ref erence trajectory
Actual
4
10
16
22
0 1 2 3 4 5 6 7
time [sec]
Throttle v alv e
EGR v alv e
14
16
18
20
22
I
n
t
a
k
e
p
r
e
s
s
u
r
e
[
k
P
a
]
M
a
s
s
f
l
o
w
r
a
t
e
[
g
/
s
]
V
a
l
v
e

o
p
e
n
i
n
g

[
%
]
65
70
75
80
Target
Ref erence trajectory
Actual
4
10
16
22
0 1 2 3 4 5 6 7
time [sec]
Throttle v alv e
EGR v alv e
14
16
18
20
22

Figure 14 Simulation results for variable
coincidence point
65
70
75
80
Target
Ref erence trajectory
Actual
4
10
16
22
0 1 2 3 4 5 6 7
time [sec]
Thrott le v alve
EGR valv e
14
16
18
20
22
I
n
t
a
k
e
p
r
e
s
s
u
r
e
[
k
P
a
]
M
a
s
s
f
l
o
w
r
a
t
e
[
g
/
s
]
V
a
l
v
e

o
p
e
n
i
n
g

[
%
]
65
70
75
80
Target
Ref erence trajectory
Actual
4
10
16
22
0 1 2 3 4 5 6 7
time [sec]
Thrott le v alve
EGR valv e
14
16
18
20
22
I
n
t
a
k
e
p
r
e
s
s
u
r
e
[
k
P
a
]
M
a
s
s
f
l
o
w
r
a
t
e
[
g
/
s
]
V
a
l
v
e

o
p
e
n
i
n
g

[
%
]

Figure 15 Simulation results with variation of
target value


2. MPC has been applied to intake control in a diesel
engine. The system is able to achieve rapid
convergence of two outputs on independent target
values, and convergence is 2.5 to 8 times faster than
when a conventional control method is employed.

REFERENCES
1. Wada, K., Suzuki, N., Satoh, N., Morita, T.,
Yamaguchi, S., Ohno, H. : Study on Emission
Reducing Method with New Lean NOX Catalyst for
Diesel Engines, SAE 2007-01-1933
2. Toshiyuki Otsuka :Calculation way of nonlinear
Receding Horizon control, Measurement and control
SICE, Vol. 41, No. 5, pp.366-371 (2002)
3. Masanori Hamamatsu, Hiroaki Kagaya, Yukinobu
Kono : Application to the automatic navigation
system of nonlinear Receding Horizon control, SICE,
Vol.44, No.8, pp.685-691 (2008)
4. Japan Society of Mechanical Engineers : A5 Fluid
engineering, Maruzen Publishing Co., Ltd., 1986,
p.58
5. Masami Masubuchi, Seiichi Kawada : Modeling of a
system and nonlinear control, Corona Publishing
Co., Ltd., pp.230-233 (1996)
6. Shuichi Adachi : Advanced System Identification for
Control, Tokyo Denki University Press (2004)
7. Toru Katayama : System identification - Approach
from partial subspace method -, Asakura Publishing
Co., Ltd., pp.69-74 (2004)
8. Maciejowski, J.M. : Predictive Control with
Constraints, Prentice-Hall (2002)
9. JSAE : Control technique of the vehicle, Asakura
Publishing, pp.17-19 (1997)
10. Tsutomu Saito, Mutsuo Koizumi : Industrial
thermodynamics, Kyoritsu Publishing Co., Ltd.
(1978)







0
5
10
15
20
25
65
70
75
80
85
0
10
20
30
40
0
10
20
30
0 2 4 6 8 10
Time [s]
A
i
r

m
a
s
s

f
l
o
w

r
a
t
e

[
g
/
s
]
I
n
t
a
k
e

p
r
e
s
s
u
r
e

[
k
P
a
]
E
G
R

v
a
l
v
e

o
p
e
n
i
n
g

[
%
]
T
h
r
o
t
t
l
e

v
a
l
v
e

o
p
e
n
i
n
g

[
%
]
Conventional
MPC
Target
0
5
10
15
20
25
65
70
75
80
85
0
10
20
30
40
0
10
20
30
0 2 4 6 8 10
Time [s]
A
i
r

m
a
s
s

f
l
o
w

r
a
t
e

[
g
/
s
]
I
n
t
a
k
e

p
r
e
s
s
u
r
e

[
k
P
a
]
E
G
R

v
a
l
v
e

o
p
e
n
i
n
g

[
%
]
T
h
r
o
t
t
l
e

v
a
l
v
e

o
p
e
n
i
n
g

[
%
]
Conventional
MPC
Target
0
5
10
15
20
25
65
70
75
80
85
0
10
20
30
40
0
10
20
30
0 2 4 6 8 10
Time [s]
A
i
r

m
a
s
s

f
l
o
w

r
a
t
e

[
g
/
s
]
I
n
t
a
k
e

p
r
e
s
s
u
r
e

[
k
P
a
]
E
G
R

v
a
l
v
e

o
p
e
n
i
n
g

[
%
]
T
h
r
o
t
t
l
e

v
a
l
v
e

o
p
e
n
i
n
g

[
%
]
Conventional
MPC
Target
Conventional Conventional
MPC MPC
Target Target


Figure 16 Test results at intake chamber pressure
target values with the target value
for air mass flow held constant


0
5
10
15
20
25
60
65
70
75
80
85
0
10
20
30
40
50
60
0
10
20
30
10 11 12 13 14
Time [s]
A
i
r

m
a
s
s

f
l
o
w

r
a
t
e

[
g
/
s
]
I
n
t
a
k
e

p
r
e
s
s
u
r
e

[
k
P
a
]
E
G
R

v
a
l
v
e

o
p
e
n
i
n
g

[
%
]
T
h
r
o
t
t
l
e

v
a
l
v
e

o
p
e
n
i
n
g

[
%
]
Conventional
MPC
Target
0 2 1 3 4
0
5
10
15
20
25
60
65
70
75
80
85
0
10
20
30
40
50
60
0
10
20
30
10 11 12 13 14
Time [s]
A
i
r

m
a
s
s

f
l
o
w

r
a
t
e

[
g
/
s
]
I
n
t
a
k
e

p
r
e
s
s
u
r
e

[
k
P
a
]
E
G
R

v
a
l
v
e

o
p
e
n
i
n
g

[
%
]
T
h
r
o
t
t
l
e

v
a
l
v
e

o
p
e
n
i
n
g

[
%
]
Conventional
MPC
Target
0
5
10
15
20
25
60
65
70
75
80
85
0
10
20
30
40
50
60
0
10
20
30
10 11 12 13 14
Time [s]
A
i
r

m
a
s
s

f
l
o
w

r
a
t
e

[
g
/
s
]
I
n
t
a
k
e

p
r
e
s
s
u
r
e

[
k
P
a
]
E
G
R

v
a
l
v
e

o
p
e
n
i
n
g

[
%
]
T
h
r
o
t
t
l
e

v
a
l
v
e

o
p
e
n
i
n
g

[
%
]
Conventional
MPC
Target
Conventional Conventional
MPC MPC
Target Target
0 2 1 3 4


Figure 17 Test results at air mass flow target values
with the target value for intake chamber
pressure held constant