You are on page 1of 8

5th IFAC Conference on

5th IFAC Conference on


Engine
5th IFACand Powertrain
Conference onControl, Simulation and Modeling
Engine and Powertrain Control, Simulation and Modeling
Available
Changchun,
Engine
5th IFACand China, September
onControl,20-22,
Powertrain
Conference 2018 and online
Simulation at www.sciencedirect.com
Modeling
Changchun, China, September 20-22, 2018
Changchun,
Engine China, September
and Powertrain Control,20-22, 2018 and Modeling
Simulation
Changchun, China, September 20-22, 2018
ScienceDirect
IFAC PapersOnLine 51-31 (2018) 846–853
Energy-Optimal
Energy-Optimal Adaptive
Adaptive Cruise
Cruise Control
Control
Energy-Optimal
in Time Adaptive
Domain Cruise
based on Control
Energy-Optimal
in Time Adaptivebased
Domain Cruiseon Control
in TimePredictive
Model Domain based Controlon
in Time Domain based
Model on
Model Predictive
Predictive Control
Control
Model
Yanzhao Predictive
Jia ∗∗ Tomohiro Saito ∗∗Control
Yutaro Itoh ∗∗∗
∗∗∗
Yanzhao Jia ∗∗ Tomohiro ∗∗ Saito ∗∗ Yutaro Itoh ∗∗ ∗∗∗
∗∗∗
Yanzhao
Yernar Jia Tomohiro ∗∗ Saito Yutaro Itoh
Yernar Nukezhanov
Nukezhanov
∗ ∗∗
Daniel
Daniel ∗
Görges
Görges ∗∗
∗∗
∗∗ ∗∗∗
Yanzhao
Yernar Tomohiro ∗∗
JiaNukezhanov SaitoDaniel Yutaro
Görges Itoh
∗∗

∗ DENSO AUTOMOTIVEYernar Nukezhanov Deutschland Daniel
GmbH, Aachen∗∗Engineering
Görges
∗ DENSO AUTOMOTIVE Deutschland GmbH, Aachen Engineering

DENSO Centre, Wegberg, Germany
AUTOMOTIVE (e-mail:GmbH,
Deutschland y.jia@denso-auto.de)
Aachen Engineering
∗ ∗∗ Centre, Wegberg, Germany (e-mail: y.jia@denso-auto.de)
∗∗ Juniorprofessorship
DENSO AUTOMOTIVE
Centre, Wegberg, for
Germany Electromobility,
Deutschland
(e-mail: GmbH, Technical
Aachen
y.jia@denso-auto.de) University
Engineering of
∗∗ Juniorprofessorship
∗∗
Juniorprofessorship
Kaiserslautern,
Centre, Wegberg,
for Electromobility, Technical
for Electromobility,
Germany
Germany (e-mail:
(e-mail: goerges@eit.uni-kl.de)
y.jia@denso-auto.de)
University
Technical University of
of
∗∗ Kaiserslautern,
∗∗∗ Germany (e-mail: goerges@eit.uni-kl.de)
∗∗∗ DENSO
Juniorprofessorship
Kaiserslautern, CORPORATION,
for Electromobility,
Germany Kariya,
Technical
(e-mail: goerges@eit.uni-kl.de) Japan University of
∗∗∗ DENSO CORPORATION, Kariya, Japan
∗∗∗
Kaiserslautern,
DENSO (yuutarou
Germany
CORPORATION,itou@denso.co.jp)
(e-mail:
(yuutarou itou@denso.co.jp) goerges@eit.uni-kl.de)
Kariya, Japan
∗∗∗ (yuutarou itou@denso.co.jp)
DENSO CORPORATION, Kariya, Japan
(yuutarou itou@denso.co.jp)
Abstract: This
Abstract: This paper
paper presents
presents aa novelnovel energy-optimal
energy-optimal adaptive adaptive cruise cruise control
control (EACC)
(EACC) strategystrategy
based
Abstract: on model
This predictive
paper control
presents a (MPC).
novel The
energy-optimalEACC has three
adaptive
based on model predictive control (MPC). The EACC has three main objectives: a) reduce the main
cruise objectives:
control (EACC)a) reduce
strategythe
based
real
Abstract: on model
driving fuel
This predictive
consumption
paper control
presents of
a (MPC).
the
novel The EACC
controlled
energy-optimal car; b) hasthe three
adaptive main control
controlled
cruise objectives:
car has a) reduce
good
(EACC) ability
strategythe
of
real driving fuel consumption of the controlled car; b) the controlled car has good ability of
tracking
based
real on
driving its preceding
model
fuel car;
predictive
consumption c) the
control
of controlled
(MPC).
the The
controlled car strictly
EACC
car; b) haskeeps
the
tracking its preceding car; c) the controlled car strictly keeps the minimum safety distance tothree the main
controlled minimum
objectives:
car hassafety a)
good distance
reduce
ability to
the
of
its preceding
real preceding
driving
tracking
its car. consumption
fuel
its preceding
car. For this
For this multi-objective
of the
car;multi-objective
c) the controlled optimal
controlled car;control,
car strictly
optimal b) the
control, it is
keeps
it is important
controlled
the
importantminimum tohas
carto build good
safety
build updistance
up aaability
suitable
suitable of
to
problem
tracking
its precedingformulation
its preceding
car. For to describe
car;
this c) the various
controlled
multi-objective controlcar
optimal targets
strictly
control, properly.
keeps
it
problem formulation to describe various control targets properly. In this paper, three different is the In
important this
minimum paper,
to three
safety
build up different
distance
a suitable to
problem
methods
its precedingformulation
of problem
problem to describe
formulation
car. For formulation
this various control
for MPC-based
multi-objective MPC-basedoptimaltargets
EACC
control, properly.
init time
time In this to
domain
is important paper,
arebuild three
up a different
proposed, deeply
suitable
methods of for EACC in domain are proposed, deeply
analysed
problem
methods and
of compared
formulation
problem to in real
describe
formulation driving
various
for situations.
control
MPC-based In
targets
EACCaddition
properly.
in
analysed and compared in real driving situations. In addition to the importance of the problem to
time the In
domainimportance
this paper,
are of the
three
proposed, problem
different
deeply
formulation,
methods
analysed
formulation, and the
of problem prediction
compared
the prediction information
formulation
in real driving
information such as
for MPC-based
situations.
such as the
theIn future
EACCaddition
future speed
inspeed
time
to theofdomain
of the preceding
are proposed,
importance
the preceding car
ofcar also
thealso has aa
deeply
problem
has
formulation,
considerable
analysed and the prediction
influence
compared onin information
MPC-based
real driving such as
EACC.
situations. theInfuture
Therefore, how
addition speed the
to ofprediction
the the preceding
importance car
information
of thealso has a
affects
problem
considerable influence on MPC-based EACC. Therefore, how the prediction information affects
considerable
the performance
performance
formulation, influence
the of EACCon MPC-based
prediction ininformation EACC.
different driving
driving
such as Therefore,
situations
the future ishow theofprediction
investigated.
speed information
Furthermore,
the Furthermore,
preceding the
carthe affects
impact
alsoimpact
has a
the of EACC in different situations is investigated.
of
the the prediction
considerable
performance ofhorizon’s
influence EACCon inlength
MPC-based
differenton MPC
EACC.
driving
of the prediction horizon’s length on MPC is analysed in this work. is analysed
Therefore,
situations in
is this
how work.
the
investigated. prediction information
Furthermore, the affects
impact
the performance
of the predictionofhorizon’sEACC inlengthdifferent driving
on MPC is situations
analysed in is investigated.
this work. Furthermore, the impact
© the
of 2018,prediction
IFAC (International
horizon’s Federation
length onofMPCAutomatic Control)inHosting
is analysed this work. by Elsevier Ltd. All rights reserved.
Keywords: Model Predictive Control, Energy-Optimal
Keywords: Model Predictive Control, Energy-Optimal Adaptive Cruise Adaptive Cruise Control,
Control, Real Real Driving
Driving
Fuel Reduction,
Keywords: Model Problem
Predictive Formulation,
Control, Influence
Energy-Optimal
Fuel Reduction, Problem Formulation, Influence of Prediction Information. of Prediction
Adaptive Information.
Cruise Control, Real Driving
Keywords:
Fuel Reduction, ModelProblem
Predictive Control, Energy-Optimal
Formulation, Influence of Prediction AdaptiveInformation.
Cruise Control, Real Driving
Fuel Reduction,
1. Problem Formulation, Influence car’s
INTRODUCTION of Prediction
speed to Information.
a dynamic driving driving environment
environment for
1. INTRODUCTION car’s speed to a dynamic for better
better
1. INTRODUCTION energy
car’s economy.
speed
energy economy. to a dynamic driving environment for better
1. INTRODUCTION car’s
energy speed to a dynamic driving environment for better
Besideseconomy.
Besides the target
the target of of saving
saving energy,
energy, ecological
ecological speed speed con-
con-
In automotive industry, the topic of connected vehicles energy economy.
In automotive industry, the topic of connected vehicles Besides trol functions
the often
target have
of another
saving two
energy,
trol functions often have another two objectives: a) strictly objectives:
ecological a)
speedstrictly
con-
with
In
with predictiveindustry,
automotive
predictive strategiesthe
strategies is topic
is attracting
attracting more and
of connected
more and more Besides
vehicles
more guarantee
trol thedriving
functions often of
driving
target have
safety;another
saving b)energy,twothe
keep objectives:
trip time
ecological timea) strictly
speed of con-
the
attention.
In
withautomotiveThe industry,
predictive big potential
strategies the of the of
topic
is of real
attracting driving
connected
more and emission
vehicles
more trol guarantee safety; b) keep the trip of the
attention. The big potential the real driving emission guarantee
controlled
functions driving
host car
often safety;
similar
have b)
to
another keep
itstwo the
preceding trip car.
objectives: timea)In oforder
the
strictly
attention.
with andThe
(RDE)predictivefuelbig potential
reduction
strategies ofattracting
the real
isisexpected
expected by driving
utilizing
more emission
and various controlled host car similar to its preceding car. In order
more guarantee
(RDE) and fuel reduction is by utilizing various controlled
to solve
solve thethehostcontrol
driving car safety;
similar b)
problem towith
its multiple
keep preceding
the trip car.
timeInoforder
objectives, two
the
information
attention.
(RDE) and (sensor
The
fuel big data,
potential
reduction Car-to-X
is of the
expected communication,
real
by driving
utilizing emissionroad
various to control problem with multiple objectives, two
information (sensor data, Car-to-X communication, road controlled tomajor
solve categories
the host
control carof optimization
similar
problem to its
with
major categories of optimization approaches were investi- approaches
preceding
multiple were
car.
objectives,Ininvesti-
order
two
topography
(RDE)
information
topography and
and fuel
(sensor
and speed
reduction
speed data,limits,
Car-to-X
limits, etc.) communication,
is expected
etc.) with
with optimal various
by utilizing
optimal control
road to
control gated
major solve in recent
the control
categories years. One approach
of problem approach
optimization is based
with approaches
multiple based on dynamic
objectives,
were dynamictwo
investi-
approaches (Kamal
information
topography (Kamal
(sensor
and speedet al.,
data, 2013;etc.)
Car-to-X
limits, Nakagawa et al.,
communication,
with optimal al., 2016).
2016). road major
control gated in recent years. One is on
approaches et al., 2013; Nakagawa et gated in
programming recent(DP)
categories years.
of One et
(Luu approach
optimization al., 2010; is based
approaches Lin et onal.,
were dynamic
2014)
investi-
approaches (Kamal
topography and speed et al., 2013;etc.)
limits, Nakagawa et al., 2016).
with optimal programming (DP) (Luu et al., 2010; Lin et al., 2014)
control programming
Among different
different application
application domains,
domains, the the ecological driv- driv- gated and the
and the other (DP)
in recent
other is
years.
is model
(Luu
model predictive
Onepredictive
approach
et al., 2010; control
is based (MPC)
etonal.,
Lin (MPC)
control (Luo
dynamic
2014)
(Luo
Among
approaches (Kamal et al., 2013; Nakagawaecological
et al., 2016). et
and al.,
programming
the 2010;
other Stanger
(DP)
is and
(Luu
model del
et Re,
al.,
predictive 2013).
2010; DP
Lin
control is
et
(MPC) a classical
al., 2014)
(Luo
ing assistance
Among
ing assistance
differentsystem (Kamaldomains,
application
system (Kamal et al.,
et al., 2010)
2010) has been
the ecological
has been driv-
pro- et al., 2010; Stanger and del Re, 2013). DP is a classical
pro-
ing
posed
Amongassistance
in recent
differentsystem
years (Kamal
with
application an et al., 2010)
anticipation
domains, the has
of a pro- and
been driv-
significant
ecological etmethod
al.,the
method to
2010;
to find
other
find isaa model
Stanger globally
globally optimal
and predictive
deloptimal
Re, 2013). solution
controlDP (MPC)
solution fora complex
isfor complex
(Luo
classical
posed in recent years with an anticipation of a significant et method
non-linear
al., to control
2010; find
Stanger a problems.
globally
and del optimal
Although
Re, solution
2013). some
DP fora complex
work
is showed
classical
reduction
ing
posed in in
assistance powertrain’s
recent system
years (Kamal
with energy
an et consumption
al., 2010)
anticipation
reduction in powertrain’s energy consumption under real method has
of a under
been real
pro-
significant non-linear control problems. Although some work showed
drivinginconditions.
posed
reduction
driving conditions.
recent With
yearsWith
in powertrain’s the
withthe
energy general
an anticipation concept
consumption
general concept of
of a of
underecologi-
significant
ecologi- aa big
real non-linearbig potential
potential
to control
findofofaimproving
improving fuel
globally optimal
problems. fuel economy
Although
economy solution
someby adopting
by adopting
for complex
work DP,
showed
DP,
driving
reduction conditions.
cal driving,
driving, new
in new With the
functionsenergy
powertrain’s general
of optimal
optimal concept of
vehicle speed
consumption speedecologi-
under con- con- a it is
bighard
non-linear to
potential implement
controlof DP
problems.
improving for
real it is hard to implement DP for on-line control because DP on-line
Although
fuel economy control
someby workbecause
adopting DP
showed
DP,
cal functions of vehicle often
aoften
big requires
potential excessive
of improving computational
fuel economy effort.
by In
adopting contrast,
DP,
cal
trol driving,
appeared,
driving newsuch
conditions. functions
as
With of general
cooperative
the optimal vehicle
adaptive
concept speed
cruise
of con- it
control
ecologi- is hard to implement DP for on-line
requires excessive computational effort. In contrast, control because DP
trol appeared, such as cooperative adaptive cruise control it
(CACC)
cal
trol (Ploeg
driving,
appeared,
(CACC) new
(Ploeg etfunctions
such
et al., 2011;ofAlam
as cooperative
al., 2011; Alam et al.,
optimal al.,
adaptive
et 2015)
vehiclecruise
2015) and
speed
and con- often
look-
control
look-
linear
linear
requires
is hardMPC
MPC with
to with excessive
implement computational
aa quadratic
quadratic
DP cost
forcost function
on-line
function
effort.
controlis
In contrast,
is considerably
considerably
because DP
headappeared,
trol
(CACC) cruise control
(Ploeg such (Hellström
as
et al., cooperative
2011; Alam et adaptive
al.,
et al.,2010; Alam
cruise
2015) and etlook-
control more
often
linear promising
requires
MPC with for
excessive
a on-line control.
computational
quadratic
al., more promising for on-line control. With suitable numer- cost With
function suitable
effort.
is In numer-
contrast,
considerably
head cruise control (Hellström et al., 2010; Alam et al., ical solvers,
solvers,
head
2013).
(CACC) cruise
Althoughcontrol
(Ploeg these
et (Hellström
al., functions
2011; Alamet al.,
haveet 2010;
different
al., 2015)
2013). Although these functions have different names, their more
Alam
names,
and al., linear
etlook-
their more
ical MPC with
promising aa large-scale
large-scale
fora quadratic quadratic
cost function
on-line quadratic
control. Withprogramming
is considerably
suitable
programming (QP)
numer-
(QP)
2013).
key
head Although
purpose
cruise is these
similar.
control functions
All these
(Hellström have
et different
functions
al., aim
2010; names,
at
Alam their
planning
et al., ical
problemsolvers,
can
promising abelarge-scale
solved
for within
on-line quadratic
milliseconds
control. Withprogramming
on
suitable an (QP)
embed-
numer-
key purpose is similar. All these functions aim at planning ical problem can be solved within milliseconds on an embed-
an optimal
key optimal
2013).
purpose speed
Although trajectory
these
is similar. for improving
functions
All these improving
have energy
different
functions aimnames, efficiency problem
ded system can
solvers, be
(Bemporad
a solved
large-scale within
and
their ded system (Bemporad and Patrinos, 2012). However, the
at efficiency
planning milliseconds
Patrinos,
quadratic 2012). on
programming an
However, embed-
the
(QP)
an speed trajectory for energy problem
ded system of
can ecological
be
(Bemporad solved speed
within
and control is
milliseconds
Patrinos, often
2012). inherently
on an
However, not
embed-
the
through
key
an optimal
through exploiting
purpose speed
exploiting various
is similar.
trajectory
various information
All these functions
for improving
information outside
aim at
energy
outside the vehicle.
planning
theefficiency
vehicle. ded problem of ecological speed control is often inherently not
through
The
an newly
optimalexploiting
developed
speed various
trajectory information
ecological
for driving
improving outside
energy
The newly developed ecological driving functions, to dis- problem
theefficiency
functions, vehicle.
to dis- aa simple
simple
problem system linear
of
linear problem
(Bemporad
ecological
problem withcontrol
and
speed
with aa quadratic
quadratic
Patrinos, 2012).
is often cost.
cost. Therefore,
However,
inherently
Therefore,the
not
The newly
tinguish
through developed
themselves
exploiting from ecological
various conventionaldriving
information ACC, functions,
outside which
the to dis- afor
enables
vehicle. forsimple
MPC linear
ofit is problem
essential
ecological with
to
speed find a the
quadratic
control
MPC it is essential to find the most suitable problem most
is often cost.
suitable Therefore,
inherently problem
not
tinguish themselves from conventional ACC, which enables afor MPC linear
formulation,
simple it iswhich
essential
problem to
accurately
withfindadescribes
the most control
quadratic suitable problem
objectives
cost. Therefore,
the host
The
the host
newly
tinguish cardeveloped
car to either
themselves
to eitherfrom drive
ecologicalat aadriving
conventional
drive at pre-set
ACC,
pre-set speed
functions,
which
speed or enables
or simply
to dis- formulation,
simply which accurately describes control objectives
follow
tinguish
the the
host preceding
themselves
car to either fromcar’s behaviour,
conventional
drive at a pre-set can speed
ACC, adjust
which the and
for still
MPC
formulation, keeps
it is the
essential
which optimization
host and still keeps the optimization problem simple.
enables
orthe
simply to
accurately find problem
the most
describes simple.
suitable
control problem
objectives
follow the preceding car’s behaviour, can adjust host
follow
the thecar
host preceding
to eithercar’s drive behaviour,
at a pre-set can speed
adjustorthe host formulation,
simply and still keeps which accurately describes
the optimization problemcontrol simple.objectives
follow the preceding car’s behaviour, can adjust the host and still keeps the optimization problem simple.
2405-8963 © 2018
Copyright 2018, IFAC
IFAC (International Federation of Automatic Control)
890 Hosting by Elsevier Ltd. All rights reserved.
Copyright
Peer review© 2018 IFAC 890 Control.
Copyright © under responsibility of International Federation of Automatic
2018 IFAC 890
10.1016/j.ifacol.2018.10.113
Copyright © 2018 IFAC 890
IFAC E-CoSM 2018

Changchun, China, September 20-22, 2018 Yanzhao Jia et al. / IFAC PapersOnLine 51-31 (2018) 846–853 847

In terms of the control domain, MPC-based speed control 2. CONTROL OBJECTIVES AND APPROACHES
can be formulated either in space domain or time domain.
In space domain, control variables change every sample 2.1 Control Objective
distance ∆s. The advantage is that all position-dependent
information (e.g., road gradient, road curvature) can be Figure 1 shows different control objectives of EACC and
pre-stored for calculating the optimal speed trajectory, their relationships. Objective 1 is to minimize the energy
while non-linearity is the main drawback of control in consumption of the host car by exploiting the benefit of
space domain. For instance, when one control variable is a varying inter-vehicle distance within given constraints.
the traction force at the wheels Fw and the vehicle speed However, if the control only has Objective 1, the host car
v is a state variable, the discrete-time system equation
2 Objec ve 1: Con ict Objec ve 2:
2Fw,i ∆s = mv (vi+1 − vi2 ) is non-linear, where mv is Energy saving for host car Good driving comfort
vehicle mass. Although some work proposed alternative Con ict
Support
ways to dodge non-linear problems in space domain, either
considerable effort in the design of control parameters, Objec ve 4: Con ict Objec ve 3:
Inter-vehicle distance not too small Inter-vehicle distance not too big
such as speed references (Schwickart et al., 2016) or the
affine approximation of the original problem (Johannesson Support
et al., 2015) is often needed for sub-optimal performance. Objec ve 5:
Assured Clear Distance Ahead
In time domain, where control variables vary with time,
linear equations can describe the system dynamics and Fig. 1. Multiple control objective of EACC
control constraints well. Unlike in space domain, exact
road information is unavailable for optimizing the future would have no incentive to move fast because less fuel is
speed trajectory because road information is decided by consumed when the vehicle speed is lower. When the host
the vehicle position, while the vehicle’s future position car drives at lower speed, the distance interval between the
depends on its future speed trajectory. To find a solution host and the preceding car will gradually become larger,
for this dilemma, the future vehicle position needs to which conflicts with Objective 3. Additionally, Objective 1
be estimated before doing speed optimization. Although and Objective 2 also conflict with each other because
such estimation has unavoidable errors, the estimation’s frequent oscillations in the control variables like traction
accuracy is continuously improved with MPC-based ap- force could be an optimal solution for fuel minimization, at
proaches, because the estimation is updated at every new the expense of driving comfort. Contrary to Objective 3,
sampling instant. Objective 4 is to avoid too small inter-vehicle distance. The
Objective 4 is helpful for achieving the goal of reducing
Several ideas of the MPC problem formulation in time fuel consumption of the host car, namely, Objective 1,
domain were proposed in recent work, e.g., in (Schmied because when the inter-vehicle distance is too small and
et al., 2015a; Weißmann et al., 2017). Although these the preceding car brakes, the host car can do nothing
papers indicate a considerable fuel reduction for a car only but apply the same braking to avoid collision, even if the
with a fixed gear, there is no conclusion on which problem braking is harmful for fuel economy. The Objective 5 of the
formulation method works better in which condition, as assured clear distance ahead (ACDA) is the lower bound
no comparison of these methods under the same test of the inter-vehicle distance, which is defined by driving
conditions has been made. regulations and is a hard control constraint for safety.
The main contribution of this paper is to deeply analyse Due to the existence of multiple objectives with some
the working principles of various problem formulations conflicting relations, an appropriate problem formulation
and evaluate them in the same daily driving situations. is essential for having the best trade-off, which is explained
Furthermore, previous problem formulations are enhanced in the following subsections.
in this paper, with considering other criteria, such as driv-
ing comfort and vehicle’s tracking ability. The application 2.2 Vehicle Dynamics
target is extended from a fixed-gear system in previous
work to a car with a multi-speed transmission. For sim- The discrete-time vehicle dynamics is described by
ulation, the real specification data of a gasoline car in vp,i + vp,i+1 vh,i + vh,i+1
the B-segment with a 5-speed transmission and its engine di+1 = di + ( − )Ts (1a)
2 2
efficiency map are used. In addition to analysing differ-
vh,i+1 = vh,i + ah,i Ts (1b)
ent problem formulations, important influencing factors of
MPC-based strategies, such as the prediction information 1
ah,i = (Ft,i − Fb,i − Froll,i − Fg,i − Fair,i ) (1c)
and the prediction horizon, are also revealed in this paper. meq
This paper is organized as follows. The objectives of where the inter-vehicle distance, speed of the host car and
energy-optimal adaptive cruise control (EACC) are in- preceding car are denoted as d, vh and vp . The variable i
troduced in Section 2. In the same section, the system is an index of the discrete-time step. Ts represents the
dynamics and various designs of EACC are explained. Sub- sample time whose value is 1 s in this work. The host
sequently, the test environment and comparison of MPC car’s acceleration ah,i at step i is calculated with (1c),
problem formulations are presented in Section 3. A deeper where the variable Ft,i and Fb,i represent the traction and
analysis of vehicle behaviours and the influence of predic- braking forces generated by the engine and braking system
tion information are investigated. Finally, conclusions and separately. The equivalent vehicle mass meq includes the
remarks on future work are given in Section 4. vehicle mass mv and the effect of all rotational mass of
the powertrain. The driving resistance consists of three

891
IFAC E-CoSM 2018
848
Changchun, China, September 20-22, 2018 Yanzhao Jia et al. / IFAC PapersOnLine 51-31 (2018) 846–853

parts: a) rolling resistance Froll,i = mv gcr cosθi with road The minimization problem (4) is solved by an off-line
angle θi ; b) road gradient resistance Fg = mv gsinθi with convex programming. Figure 3 demonstrates the percent
gravitational acceleration g; c) aerodynamic resistance error in the power map approximation. The percent error
Fair,i = 12 ρa Af ca vh,i
2
, with air density ρa , drag coefficient is smaller than 5% within a large operation area, while
ca and frontal area Af . a bigger error only appears when the traction force is
very small. The power map approximation is done for each
The inter-vehicle distance should be controlled for driving different gear separately.
safety and comfort. The minimum safety distance dsafety is
defined with (2a), while (2b) defines the maximum comfort 3000 35

Traction Force (N)


distance dcomfort . dmin,0 and dmax,0 are two constants, 30
representing the minimum and maximum inter-vehicle 2000 25
distance when the car is at standstill, while the coefficients 20
of hmin and hcomfort enable dsafety and dcomfort to vary with Percent Error (%) 15
the speed of the host car. 1000
10
dsafety,i = dmin,0 + hmin vh,i (2a) 5
dcomfort,i = dmax,0 + hcomfort vh,i (2b) 0
5 10 15 20 25 30 35 40
Speed (m/s)
2.3 Approximation of the Power Consumption Map
Fig. 3. Error in power map approximation for one gear
The stationary power consumption Pc of an internal com-
bustion engine (ICE) car is a non-linear mapping of the 2.4 Variants of MPC Problem Formulations
vehicle speed v, the gear selection and the traction force
Ft . The mapping function ϕ is defined with (3). Figure 2 To reduce the fuel consumption in daily driving, some
shows a measurement result of the stationary power map recent studies (Schmied et al., 2015b; Lin et al., 2017)
for one specific gear. proposed a type of cost function, which is based on the
ϕ : Ft × v → Pc (3) intuition that avoiding frequent hard acceleration and
braking is beneficial for saving energy. Although these
studies showed some good results, an exact correlation
between the consumed energy and the control variables
Power Consumption (Watt)

5
×10
4 was not defined in the cost functions. Because the power
map approximation function (4b) depicts the real con-
2 sumed power of an ICE car with good accuracy, this
function is used in this paper to explicitly describe how fuel
0
consumption varies with the control and state variables. As
3000 illustrated in Fig. 1, the control objectives of EACC do not
2000 40
1000 20
30 only include fuel reduction but also other aspects, which
10
Traction Force (N) 0 0 Vehicle Speed (m/s) need to be depicted with a suitable problem formulation.
Three problem formulation methods are discussed below.
Fig. 2. Measured power consumption map for one gear A. Method I The first problem formulation is a variant
and an extension of the method proposed by (Schmied
Finding the best function for approximating the non-linear et al., 2015a). The mathematical formulation is given by
power map ϕ is a minimization problem: k+H−1

2 2 2
N
M  min f (Ft,i , vi ) + ω1 δ1,i + ω2 δ2,i + ω3 Fb,i (5a)
 Ft,i ,Fb,i
min (f (Ft,m , vn ) − ϕ(Ft,m , vn ))2 (4a) i=k
p00 ,...,p11
m=1 n=1
Ts
s.t. vh,i+1 = vh,i + (Ft,i − Fb,i − mv gcr cosθi
where f (Ft , v) = p00 + p01 Ft + p10 v + p02 Ft2 meq
(4b) 1
+ p20 v 2 + p11 vFt 2
− mv gsinθi − ρa Af ca vh,i )
  2
2p02 p11 (5b)
s.t. A = 0 (4c)
p11 2p20 vp,i + vp,i+1 vh,i + vh,i+1
di+1 = di + ( − )Ts (5c)
The polynomial function (4b) with six coefficients is de- 2 2
signed to approximate the non-linear power map. The di ≥ dmin,0 + hmin vh,i (5d)
optimization problem (4) is used to find the most optimal di ≤ dmax,0 + hcomfort vh,i + δ1,i (5e)
values of the six coefficients (i.e., p00 , p01 , p10 , p02 , p20 , p11 ),
with which the discrepancy between the polynomial func- 0 ≤ Ft,i ≤ a10 + a11 vh,i
tion and the real map is minimized. The indices of m and Ft,i ≤ a20 + a21 vh,i (5f)
n in (4a) represent discrete feasible operation points of Ft,i ≤ a30 + a31 vh,i
the traction force and the vehicle speed for one specific |Ft,i − Ft,i−1 | ≤ ∆Ft,max + δ2,i (5g)
gear. The optimization constraint (4c) is to ensure that
0 ≤ Fb,i ≤ Fbrake,max
the Hessian matrix of this approximation function (4a) is (5h)
positive semi-definite for having a global optimal solution. vmin,i ≤ vh,i ≤ vmax,i

892
IFAC E-CoSM 2018

Changchun, China, September 20-22, 2018 Yanzhao Jia et al. / IFAC PapersOnLine 51-31 (2018) 846–853 849

where the index of time i is defined as i = k, ..., k + H − 1.


Here, k is the current time step and H represents the
length of the prediction horizon. The total cost function
(5a) consists of four items. The first item represents the
power consumption with the approximation function (4b).
Only with this item, the host car would tend to move
slowly, as discussed in Section 2.1. Therefore, the second
2
item δ1,i is added to the cost function, as a penalty
when the inter-vehicle distance is bigger than the comfort
distance (2b). δ1,i is a slack variable in the soft constraint
(5e). With the second item, the host car is forced to Fig. 5. MPC-based EACC with Method I
move faster. The newly introduced third item with the
variable δ2 is used to penalize the change in traction safety, i.e., when the inter-vehicle distance dk+i|k within
force Ft between two sampling instants for better driving the prediction horizon H will become smaller than the
comfort, which is defined with (5g). The last item of this minimum safety distance dsafety,i without braking. Simi-
cost function is used to penalize unnecessary braking Fb,i , larly, when the host car stays in Zone 2, it will not rapidly
which causes a dissipation of the vehicle’s kinetic energy. accelerate in line with the preceding car, until it enters into
The cost function is minimized with using vehicle dynamic 2
Zone 1, where the second item ω1 δ1,i in (5a) will increase if
equations (5b) and (5c), which are deduced in Section 2.2. the host car drives continuously slower than the preceding
The necessity of having the third and the fourth item in car. With such control design, the host car applies both
(5a) is further explained in Section 3.4. driving and braking force in a more ecological way.
During the optimization process, several constraints are
considered: (5d) is the safety constraint, which guarantees B. Method II The inspiration of the second problem
that the inter-vehicle distance is strictly bigger than the formulation is from a recent work (Lu and Müller, 2015),
minimum safety distance (2a). The upper limit of the although this work was done in space domain. The cost
traction force depends on the vehicle speed and the gear function and constraints are formulated with
position, as shown in Fig. 4. The maximum traction force k+H−1

2 2 2 2
for each gear is represented with three linear lines, which min f (Ft,i , vi ) + ω1 δ1,i + ω2 δ2,i + ω3 Fb,i + ω4 δ4,i
Ft,i ,Fb,i
are shown with dash lines in Fig. 4 and mathematically i=k
expressed with (5f). The constraints (5h) limits the brak- (6a)
ing force Fb and restricts the vehicle speed vh according s.t. (5b), (5c), (5d), (5f), (5g), (5h)
to driving regulations.
di ≥ dmax,0 + hcomfort vh,i − δ1,i (6b)
k+H
 k+i

1st gear vh,k+i|k ≥ vp,k+i|i − δ4 (6c)
a20+a21x
a11 x

a 30
Traction Force (N)

+a

i=k i=k
a10+

31

2nd gear
x

The key idea of Method II is to track the preceding


3rd gear car (i.e., Objective 3 in Fig. 1.) by defining an explicit
4th gear requirement on the minimum average speed of the host car.
5th gear The slack variable δ4 in the soft constraint (6c) quantifies
the difference between the average speed of the host car
Speed (km/h) and the preceding car in the prediction horizon H. The
lower the speed of the host car is, the bigger the penalty
Fig. 4. Upper limit of the traction force for different gears δ4 becomes, which makes the value of cost function (6a)
rise. On the one hand, the slack variable δ4 propels the host
The general control structure is schematically depicted in car to drive faster, on the other hand, it could push the
Fig. 5. The traction force Ft,k and braking force Fb,k at the host car too close to the preceding car, especially when the
current time step k are optimized according to the future preceding car brakes continuously. As discussed in Section
speed of the preceding car vp,k+i|k , as well as other driving 2.1, when the inter-vehicle distance becomes too small, the
information (e.g., road speed limit vlimit,k+i|k , inter-vehicle free space left for speed optimization is compressed, which
distance dk+i|k and road gradient θk+i|k ). The notation is harmful for improving fuel economy.
(·)k+i|k represents the prediction of the variable (·) at
future time step k + i, calculated at the current time k. Therefore, in order to have more space for regulating
the speed of the host car, another soft constraint (6b)
is introduced. In contrast to the soft constraint (5e) in
The inter-vehicle distance Zone 1 and 2 in Fig. 5 are Method I, (6b) has an opposite purpose. The slack variable
designed for avoiding unnecessary braking and too early δ1 increases when the inter-vehicle distance di is smaller
acceleration. For instance, when the host car is in Zone than the comfort distance (2b), which causes a higher cost
2 and the preceding car suddenly brakes, the host car of control. So, with applying the penalty δ12 in cost function
will probably do coasting first with the ICE off, instead (6a), the host car is reluctant to move too fast when it
2
of braking, because braking makes the fourth item ω3 Fb,i enters into Zone 2, which is shown in Fig. 6. The host
in the cost function (5a) increase but brings no benefit for car would prefer slowing down within Zone 2, often with
others. So, braking is used only when it is necessary for coasting, in order to reduce the total cost (6a).

893
IFAC E-CoSM 2018
850
Changchun, China, September 20-22, 2018 Yanzhao Jia et al. / IFAC PapersOnLine 51-31 (2018) 846–853

car will speed up again to reduce the deviation from vref,i .


The first item f (Ft,i , vi ) in (7) is used to make the transient
behaviour of re-tracking the speed reference vref,i smoother
for less fuel consumption.

3. SIMULATION AND EXPERIMENT RESULTS

In this section, the three different methods for the MPC


problem formulation, which are introduced in Section 2,
are compared in a real driving environment. Moreover, the
Fig. 6. MPC-based EACC with Method II vehicle behaviours with EACC, the effects of additional
Figure 6 describes the general control loop with the second items in the cost function and the prediction information’s
problem formulation. With comparison to Fig. 5, Method impacts on MPC are revealed. The computation time for
I and II have a close affinity, while their main difference solving a MPC problem is also investigated.
is the design of the soft penalty zones and the way of
impelling the host car to move faster. 3.1 Test Environment

C. Method III The idea of the third problem formulation In automotive industry, a vehicle’s fuel economy is often
is based on the previous work (Weißmann et al., 2017). As calculated through testing in standard cycles whose speed
discussed in Section 1, MPC is capable of on-line optimal profiles are pre-defined. However, real driving emission
control, while it normally requires that the system is linear (RDE) and daily fuel consumption can often differ from
and the cost can be accurately approximated with a convex test results in standard cycles. In order to investigate real
function. However, the control system in the real world is driving fuel reduction by using EACC, test cycles are
inherently often non-linear, for which dynamic program- designed with recorded daily driving data under real traffic
ming (DP) can provide a globally optimal solution, while conditions. Figure 8 shows three speed profiles of recorded
it is hard to be implemented for on-line control due to its data on the same driving route of 17.8 km, which are used
unaffordable computation effort. So the third method is to as test cycles. The upper speed limit varies from 30 km/h
combine the strengths of DP with the merits of MPC. to 130 km/h as the driving route includes urban and rural
areas, as well as highways. The trajectories of these cycles
are treated as the speed profiles of a preceding car, which
are used as a baseline in Section 3.2 to study how EACC
optimally regulates the host car’s speed, with utilizing
future information of the preceding car.

140 Long Test Cycle 1 Highway


Long Test Cycle 2
120 Long Test Cycle 3
Road Speed Limit Rural Area
Velocity (km/h)

100

80
Fig. 7. MPC-based EACC with Method III Urban Area
60

40
Figure 7 shows the general control structure, where vref
is the globally optimal result of DP through an off-line 20
calculation. It represents the most economical speed profile 0
from the start point to the final destination of a trip, with 0 100 200 300 400 500 600 700 800
Time (sec)
consideration of the road speed limits, road gradient and
expected trip time. The second item in cost function (7a) Fig. 8. Profiles of test cycles under real traffic conditions
penalizes the deviation from vref because the fuel economy
of the host car in a pre-defined trip is the most optimal,
when its speed trajectory is the same as DP’s result. 3.2 Test Result and Performance Comparison
k+H−1

min f (Ft,i , vi ) + ω5 (vh,i − vref,i )2 (7a) MPC-based EACC with the three problem formulation
Ft,i ,Fb,i approaches whose working principles are explained in Sec-
i=k
s.t. (5b), (5c), (5d), (5f), (5h) tion 2.4, is investigated in different real driving situations.
Table 1 shows the performance of the three methods in the
However, in real driving situations a slower preceding car
three test cycles. For all methods, it is assumed that the
often hinders the host car from driving at vref,i . When the
exact future speed trajectory of the preceding car within
minimum safety distance is to be violated in the prediction the prediction horizon of 14 s is pre-known.
horizon, the host car has to brake, even if its speed would
differ from vref,i . Therefore, the hard constraint (5d) is Among the three methods, Method III shows more fuel
applied in problem formulation (7) to ensure that the host reduction in test cycle 2 and 3, while it performs re-
car can drive safely. Additionally, constraints (5f) and (5h) markably worse in cycle 1. The reason is that Method
are used to ensure a feasible behaviour of the host car. III provides a globally optimal solution only when the
When a slower preceding car accelerates or leaves, the host host car can follow the speed reference trajectory of DP

894
IFAC E-CoSM 2018

Changchun, China, September 20-22, 2018 Yanzhao Jia et al. / IFAC PapersOnLine 51-31 (2018) 846–853 851

vref , which is pre-calculated off-line without considering 3.3 Behaviours of Cars with using Method I
real traffic situations. However, in real driving conditions,
a slower preceding car would force the host car to slow The simulation results of Method I are analysed in this
down and deviate from vref . For instance, as shown in Fig. subsection. The vehicle speed trajectories of both the host
8, the preceding car drives slowly on the highway in test car and the preceding car, as well as the road speed
cycle 1. In this case, following the optimal trajectory vref limit, are shown in Fig. 10. The preceding car drives with
becomes impossible, which results in a deterioration of fuel frequent strong deceleration and acceleration, especially
economy. In addition, it should be noted that the host car before it drives onto the highway. In the urban area, the
does not track the preceding car well with Method III in preceding car violates the upper road speed limit of 50 km/h
test cycle 2 and 3. Figure 9 shows that the inter-vehicle for a while. With Method I, the host car’s speed trajectory
distance d becomes too large at the end of cycle 3 as the is smoother without applying unnecessary braking. In
host car tends to drive at vref for better fuel economy, while addition, the host car complies with the road speed limit
vref is considerably lower than the preceding car’s speed vp strictly. Due to this more eco-driving style, the host car
on the highway. saves 6.1% fuel in the whole trip, with a priori knowledge
150
of the preceding car’s future speed in the horizon of 14 s.
Speed (km/h)

Host Car
Preceding Car 150
100 Speed Limit Host Car Highway
DP Reference 120 Preceding Car
50 Road Speed Limit Rural Area

Speed (km/h)
0 90
1000
Inter-Vehicle Distance Urban Area
Distance (m)

Safety Distance 60
500 Comfort Distance
30
0
0 100 200 300 400 500 600 700 800 0
Time (sec) 0 100 200 300 400 500 600 700 800
Time (sec)
Fig. 9. Simulation results for Method III in test cycle 3, a
priori information in the prediction horizon of 14 s Fig. 10. Speed trajectories of two cars with Method I in test
cycle 1, a priori information in the prediction horizon
of 14 s
Table 1. Comparison of the fuel reduction for
the three methods with a priori information in It is observable in Fig. 10 that the host car’s behaviour is
the prediction horizon of 14 s relatively more similar to the preceding car on the highway,
 while its speed differs more from the preceding car in
 Test Cycle urban and rural areas. The two cars’ behaviours are further
 1 2 3
Method  examined in a short test cycle, as shown in Fig. 11. Here, it
Method I: 6.1 % 5.1 % 4.3 %
4
Gradient (degree)

150 Host Vehicle


Method II: 5.3 % 4.9 % 4.1 % Road Gradient
Speed (km/h)

Preceding Vehicle
Method III: 3.4 % 6.0 % 8.8 % 100 Speed Limit 2

50 0
In contrast, with Method I and II the host car tracks the -2
0
preceding car better with a smaller average inter-vehicle 200 Inter Distance 2000 Traction Force
Distance (m)

distance, which is noticeable in Table 2. In terms of fuel Safety Distance Braking Force
Force (N)

Comfort Distance
economy, Method I is slightly better than Method II in 100 1000

the three test cycles, as shown in Table 1. Although the


0 0
main working principles of Method I and II are similar, 0 40 80 120 160 200 240 0 40 80 120 160 200 240
as explained in Section 2.4, the slight difference in fuel Time (sec) Time (sec)

economy is caused by the tuning factors. Method II has


Fig. 11. Simulation results for Method I in the short test
more weighting factors in cost function (6a), for which
cycle, a priori information in the prediction horizon
bigger calibration effort is needed for Method II.
of 14 s
Table 2. Comparison of the average inter- is noticeable that the host car drives with more moderate
vehicle distance with three methods, a priori speed changes and applies the coasting strategy (i.e.,
information in the prediction horizon of 14 s traction force is zero.) more often. Meanwhile, the inter-
 vehicle distance is well controlled: it varies around the soft
 Test Cycle
 1 2 3 constraint on the comfort distance (5e) and strictly fulfils
Method 
Method I 87m 90m 87m the hard constraint on the minimum safety distance (5d).
Method II 87m 89m 87m In this short test cycle, where the speed of the preceding
Method III 87m 111m 137m car varies more dynamically within a large range, Method
I brings about a bigger fuel reduction of 13.1%.
Due to the robuster fuel reduction, better performance of
tracking the preceding car and less required calibration 3.4 Effects of Additional Items in the Cost Function
work, Method I is considered as the best problem formu-
lation. A more detailed analysis of Method I is given in As mentioned in Section 2.4, Method I is an extension
the subsections below. of a recent work (Schmied et al., 2015a), in which the

895
IFAC E-CoSM 2018
852
Changchun, China, September 20-22, 2018 Yanzhao Jia et al. / IFAC PapersOnLine 51-31 (2018) 846–853

driving comfort was essentially not considered. When the applied, namely prescient MPC (PMPC) and frozen-time
cost function (5a) only includes the first two items like MPC (FTMPC). PMPC works with a priori knowledge of
the previous work, the traction force can change freely the preceding car’s exact future speed within a short pre-
between two sampling instants. This could result in a diction horizon, while FTMPC works without the future
strong oscillation of the traction force, which reduces the information and assumes that the preceding car keeps a
driving comfort. Figure 12 shows how the traction force constant speed from the current time step till the end of
could frequently oscillate when there is no penalty for it. one prediction horizon. Figure 13 illustrates how PMPC
This oscillation can be suppressed by using the third item and FTMPC perform in three long test cycles and one
2
of ω2 δ2,i in (5a) as a penalty. short cycle. The following interesting points are notable.
2000
Traction Force
2000
Traction Force
(1) PMPC reduces the fuel consumption more than
1500 Braking Force 1500 Braking Force FTMPC as it exploits the accurate future speed in-
Force (N)
Force (N)

1000
formation of the preceding car.
1000
(2) Fuel economy with PMPC in general increases with
500 500
Oscillation
the length of the prediction horizon. However, when
0
0 50 100 150 200 250
0
0 50 100 150 200 250
the horizon is longer than a certain point, the perfor-
Time (sec) Time (sec) mance tends to become worse, due to the fact that
(a). Without Penalty (b). With Penalty a longer horizon could cause a bigger error in the
estimation of road information. This is a drawback of
Fig. 12. Suppressing oscillation of the traction force by a control in time domain, as mentioned in Section 1.
new penalty in the cost function (3) Without having any future information, the longer
the prediction horizon is, the worse FTMPC per-
Although this penalty is good for driving comfort, it could forms, because a longer horizon in this case means
cause an unnecessary braking as a side effect. For instance, a bigger error in the predicted future speed of the
under a certain condition, reducing the vehicle speed could preceding car.
be beneficial for cost function (5a). To slow down the host (4) The influence of future information is obviously big-
car, reducing the traction force is one choice. However, ger in the short test cycle than in the three long
a quick reduction in the traction force would result in test cycles. In three long cycles, the perfect future
a big penalty by δ2,i . To avoid such penalty, applying information brings a relatively small fuel reduction
brakes seems to be a better choice in this situation as it on the average because a big part of the long cycles
causes no cost. However, braking always causes a waste of is on the highway, where vehicle’s speed changes less
kinetic energy, which should be avoided unless it is needed dramatically, as shown in Fig. 8. In this case, a simple
for keeping the minimum safety distance. Therefore, the assumption that the speed in one horizon is the same
2
fourth item ω3 Fb,i is added into (5a), with the goal of as the current speed would result in a small error.
avoiding dispensable braking. In contrast, when the preceding car’s speed changes
more dramatically in the short test cycle, estimating
3.5 Influence of the Prediction Information its speed with assuming a constant speed within one
horizon could have a bigger error. Figure 14 shows the
All three methods of the problem formulation, explained cars’ behaviours with applying FTMPC in the short
in Section 2.4 and tested in Section 3.2, need the future test cycle. Compared to the performance of PMPC
speed of the preceding car as an input. On the one hand, shown in Fig. 11, with FTMPC the host car increases
it is hard to perfectly predict the preceding car’s future its speed more during the acceleration phase, exceed-
behaviour (Rezaei and Burl, 2015), on the other hand, ing the peak point of the speed profile of the preceding
MPC could still provide a good result with inaccurate car. The higher speed of the host car results in more
information (Lin et al., 2017) because MPC on-line up- braking later on for keeping the safety distance to the
dates its control inputs and corrects wrong assumptions preceding car, which causes a waste of the energy. So,
at the next sampling instant, according to the new mea- in this test situation, PMPC brings out around 7%
surement data. Therefore, it is worth studying how MPC- extra fuel reduction, compared to FTMPC.
based EACC performs when the perfect future information 150
Gradient (degree)

Host Car 4
is unavailable. So, in this work, two types of MPC are Road Gradient
Speed (km/h)

Preceding Car
Road Speed Limit
100 2
15
50 0
Fuel Reduction (%)

10
0 -2
200 Inter-Vehicle Distance
Distance (m)

5 Safety Distance Traction Force


Force (N)

Comfort Distance 2000 Braking Force

FTMPC, Three Long Cycles 100


0 1000
PMPC, Three Long Cycles
FTMPC, Short Cycle
0 0
-5 PMPC, Short Cycle 0 40 80 120 160 200 240 0 40 80 120 160 200 240
4 6 8 10 12 14 16 18 Time (sec) Time (sec)
Prediction Horizon (sec)

Fig. 13. Impact of the prediction information in different Fig. 14. Simulation results of FTMPC for Method I in the
test cycles with the problem formulation Method I short test cycle, without a priori information

896
IFAC E-CoSM 2018

Changchun, China, September 20-22, 2018 Yanzhao Jia et al. / IFAC PapersOnLine 51-31 (2018) 846–853 853

3.6 Computation Time Kamal, M.A.S., Mukai, M., Murata, J., and Kawabe,
T. (2010). On board eco-driving system for varying
Required computation time for MPC, which varies with road-traffic environments using model predictive con-
the length of the prediction horizon, is shown in Table trol. In IEEE International Conference on Control
3. The simulation is executed in Matlab R2016a on a Applications, 1636–1641.
computer with Windows 7 and Intel Core i5-2450M CPU Kamal, M.A.S., Mukai, M., Murata, J., and Kawabe, T.
with clock frequency 2.5GHz. The optimization problem (2013). Model predictive control of vehicles on urban
(5) is solved by the solver of quadprog with the interior- roads for improved fuel economy. IEEE Transactions
point method. It is noticeable that the average required on control systems technology, 21(3), 831–841.
computation time is remarkably less than the sample time Lin, X., Görges, D., and Liu, S. (2014). Eco-driving
of 1s, which shows the capability of MPC-based EACC for assistance system for electric vehicles based on speed
on-line control. profile optimization. In IEEE Conference on Control
Table 3. Computation time for Method I with Applications, 629–634.
different prediction horizons Lin, X., Görges, D., and Weißmann, A. (2017). Simplified
energy-efficient adaptive cruise control based on model
Horizon Length H=5 H=10 H=20 predictive control. IFAC Proceedings Volumes, 50(1),
Average Computation Time 16ms 33ms 62 ms 4794–4799.
Lu, L. and Müller, S. (2015). Model predictive speed and
headway control with energy optimization for a series
4. CONCLUSIONS AND FUTURE WORK hybrid vehicle with range extender. In 4th ICMCE, 35:
02003.
In this paper different control objectives and new ap- Luo, L.h., Liu, H., Li, P., and Wang, H. (2010). Model
proaches of applying EACC on an ICE car with multiple predictive control for adaptive cruise control with
gears have been discussed. Three different problem formu- multi-objectives: comfort, fuel-economy, safety and car-
lations are compared and analysed for MPC-based EACC following. Journal of Zhejiang University SCIENCE A,
in time domain. The best problem formulation method is 11(3), 191–201.
revealed with the consideration of the robustness of fuel Luu, H.T., Nouveliere, L., and Mammar, S. (2010). Dy-
reduction, driving comfort and requirements on calibra- namic programming for fuel consumption optimization
tion work. Around 4% to 5% fuel reduction is observed on light vehicle. IFAC Proceedings Volumes, 43(7), 372–
in three real driving test cycles. Influences of both the 377.
future information and the prediction horizon on EACC’s Nakagawa, M., Yano, K., Yamauchi, I., Stapelbroek, M.,
performance are revealed in different test conditions. Hülshorst, T., and Nase, A. (2016). New degrees of
In future, the specific driving situations, in which a priori freedom for dimensioning of hybrid powertrains by using
speed information of the preceding car significantly affects predictive energy management. In 37th International
MPC’s performance, will be further investigated. Predic- Vienna Motor Symposium, 248–264.
tion technology will be developed in order to estimate Ploeg, J., Scheepers, B.T., Van Nunen, E., Van de Wouw,
the preceding car’s future behaviour more accurately in N., and Nijmeijer, H. (2011). Design and experimental
these specific situations. The MPC-based controller will evaluation of cooperative adaptive cruise control. In
be adapted to predictive information in order to generate 14th IEEE ITSC, 260–265.
a bigger benefit in the fuel reduction and driving comfort. Rezaei, A. and Burl, J.B. (2015). Prediction of vehicle
EACC will be further tested in various geographical areas velocity for model predictive control. IFAC Proceedings
in order to ensure that it performs robustly on different Volumes, 48(15), 257–262.
roads. Schmied, R., Waschl, H., and del Re, L. (2015a). Extension
and experimental validation of fuel efficient predictive
REFERENCES adaptive cruise control. In American Control Confer-
ence, 4753–4758.
Alam, A., Martensson, J., and Johansson, K.H. (2013). Schmied, R., Waschl, H., Quirynen, R., Diehl, M., and del
Look-ahead cruise control for heavy duty vehicle pla- Re, L. (2015b). Nonlinear MPC for emission efficient
tooning. In 16th IEEE ITSC, 928–935. cooperative adaptive cruise control. IFAC Proceedings
Alam, A., Mårtensson, J., and Johansson, K.H. (2015). Volumes, 48(23), 160–165.
Experimental evaluation of decentralized cooperative Schwickart, T., Voos, H., Hadji-Minaglou, J.R., and
cruise control for heavy-duty vehicle platooning. Control Darouach, M. (2016). A fast model-predictive speed
Engineering Practice, 38, 11–25. controller for minimised charge consumption of electric
Bemporad, A. and Patrinos, P. (2012). Simple and certi- vehicles. Asian Journal of Control, 18(1), 133–149.
fiable quadratic programming algorithms for embedded Stanger, T. and del Re, L. (2013). A model predictive co-
linear model predictive control. IFAC Proceedings Vol- operative adaptive cruise control approach. In American
umes, 45(17), 14–20. Control Conference, 1374–1379.
Hellström, E., Åslund, J., and Nielsen, L. (2010). Design Weißmann, A., Görges, D., and Lin, X. (2017). Energy-
of an efficient algorithm for fuel-optimal look-ahead optimal adaptive cruise control based on model predic-
control. Control Engineering Practice, 18(11), 1318– tive control. IFAC Proceedings Volumes, 50(1), 12563–
1327. 12568.
Johannesson, L., Nilsson, M., and Murgovski, N. (2015).
Look-ahead vehicle energy management with traffic pre-
dictions. IFAC Proceedings Volumes, 48(15), 244–251.

897

You might also like