Publications of the DLR elib

This is the author’s copy of the accepted publication as archived with the DLR’s electronic
library at http://elib.dlr.de . There is no publicly available version as of November 2022.

IEEE VTS Motor Vehicles Challenge 2023: A Multi-

physical Benchmark Problem for Next Generation
Energy Management Algorithms
J. Brembeck, R. de Castro, J. Tobolar, and I. Ebrahimi

This article presents a benchmark problem that researchers can use to evaluate the performance of energy
management algorithms for multi-energy source and multimotor electric vehicles. The model makes use of
Modelica, an open source, a-causal, object-oriented language for modeling cyber-physical systems in multi-
domains (e.g. electrical, mechanical, thermal) of the vehicle components. The model also includes the aspect of
three-dimensional mechanics, which enables completely new degrees of freedom in the controller design in
comparison to one-dimensional approaches. To support interoperability among multiple design tools, the
Modelica vehicle model is provided as a Functional Mockup Unit, an industry standard for exchange of simulation
models. A set of standardized input-output interfaces and key performance metrics is also provided in the
benchmarking problem, enabling the systematic ranking of multiple energy management strategies.

Copyright Notice
©2022 IEEE. Personal use of this material is permitted. Permission from IEEE must be
obtained for all other uses, in any current or future media, including reprinting/republishing
this material for advertising or promotional purposes, creating new collective works, for
resale or redistribution to servers or lists, or reuse of any copyrighted component of this
work in other works.

J. Brembeck, R. de Castro, J. Tobolar, and I. Ebrahimi, " IEEE VTS Motor Vehicles
Challenge 2023: A Multi-physical Benchmark Problem for Next Generation Energy
Management Algorithms", 2022 IEEE Vehicle Power and Propulsion Conference (VPPC),
Merced, California, USA, 2022.
 IEEE VTS Motor Vehicles Challenge 2023:
A Multi-physical Benchmark Problem for Next Generation
Energy Management Algorithms

Jonathan Brembeck Ricardo de Castro Jakub Tobolář Iman Ebrahimi

Institute of System Dynamics Dept. of Mechanical Institute of System Dynamics Dept. of Mechanical
and Control (SR) Engineering and Control (SR) Engineering
German Aerospace Center (DLR) University of California German Aerospace Center (DLR) University of California
Oberpfaffenhofen, Germany Merced Oberpfaffenhofen, Germany Merced
jonathan.brembeck@dlr.de rpintodecastro@ucmerced.edu jakub.tobolar@dlr.de iebrahimi@ucmerced.edu
ORCID: 0000-0002-7671-5251 ORCID: 0000-0002-5546-3999 ORCID: 0000-0002-4888-4664 ORCID: 0000-0003-0130-8712

recuperation in both front and rear axle [9] and also

Abstract—This article presents a benchmark problem that decreasing tire slip losses [10]. It improves motion control of
researchers can use to evaluate the performance of energy the vehicle by extending the maximum lateral acceleration
management algorithms for multi-energy source and multi- [11] and decreases response of inner (yaw-rate) control loops.
motor electric vehicles. The model makes use of Modelica, an This last feature is particularly attractive to improve tracking
open source, a-causal, object-oriented language for modeling performance and vehicle dynamics stability of path/trajectory
cyber-physical systems in multi-domains (e.g. electrical, following algorithms of autonomous vehicles [12]. It also
mechanical, thermal) of the vehicle components. The model also
includes the aspect of three-dimensional mechanics, which
offers redundant traction actuators that can be exploited by
enables completely new degrees of freedom in the controller fault tolerant controllers to improve vehicle reliability [13].
design in comparison to one-dimensional approaches. To As reference vehicle for this year’s competition we
support interoperability among multiple design tools, the employed an adapted version of the DLR ROboMObil ( [14]
Modelica vehicle model is provided as a Functional Mockup [15] [16] [17]), which is extended here with a hydrogen fuel
Unit, an industry standard for exchange of simulation models. A cell. To model this vehicle, we make use of Modelica [18], an
set of standardized input-output interfaces and key performance open source, a-causal, object-oriented language that allows
metrics is also provided in the benchmarking problem, enabling the modeling of cyber-physical systems in multi-domains,
the systematic ranking of multiple energy management e.g. electrical, thermal and mechanical. It also offers powerful
strategies.
Keywords—energy management, vehicle dynamics control,
model inversion capabilities, which facilitate the design of
electro-mobility, trajectory control, energy management, battery non-linear motion control algorithms [19].
model, Modelica, Functional Mockup Interface, hydrogen fuel Since there is a manifold of simulators on the market, we
cell, multi-physical modelling, drivetrain models decided to export the vehicle model through the functional
mockup interface (FMI) [18] technology. This allows the
I. INTRODUCTION competitors to develop a control strategy using their tool of
In order to stimulate advances in vehicular energy and choice. Our simulation model gives the benefit of a very high
power management, the IEEE Vehicular Technology Society computational efficiency, with an average real-time factor
(VTS) initiated the VTS Motor Vehicles Challenge in 2016. larger than twenty on a standard PC, which can be particularly
The competition’s themes have focused on energy useful to accelerate the development of learning-based control
management of hybrid energy storage systems (fuel algorithms (e.g. [20]).
cells/batteries/supercapacitors [1] [2] [3] [4]) and their The paper is arranged as follows: Section II gives a general
applications to cars ( [5] [6]), trucks [7] and trains [8]. These overview of the challenge; Section III introduces the
competitions provide benchmark problems where researchers Modelica model of the ROboMObil; Section IV describes the
can evaluate and compare the performance of their energy evaluation process and ranking of energy management
management algorithms against other research groups. algorithms. Section V summarizes all development steps and
In this year’s competition we bring a new dimension to the provides a look into future research.
challenge: torque allocation in multi-motor electric vehicles.
II. OVERVIEW OF THE CHALLENGE
In addition to managing a hybrid energy storage system, the
competitors are asked to develop torque allocation for a four- As depicted in Figure 1, the challenge considers the
wheeled vehicle with three traction motors: two (rear) in- ROboMObil vehicle with a hybrid energy storage system
wheel motors and one front motor. This multi-motor (fuel cell & hydrogen tank and battery), two in-wheel electric
configuration offers various advantages. It can enhance motors installed in the rear axle (𝜏𝑅𝐿 , 𝜏𝑅𝑅 ), one central front
energy efficiency of the vehicle by enabling energy motor (𝜏𝐹 ), and a front steer-by-wire actuation (𝛿). These
actuators are manipulated by the vehicle’s motion controller
in order to track a pre-defined reference velocity (𝑣 ∗ ) and the wheel radius, which is then divided between the front and
track curvature (𝜌∗ ). rear axle
𝜏𝐹∗ = 𝜏 ∗ 𝛼𝐴𝐷 , 𝜏𝑅∗ = 𝜏 ∗ (1 − 𝛼𝐴𝐷 ), (1)
where 𝛼𝐴𝐷 indicates the variable front/rear axle distribution.
Afterwards, the rear axle torque is allocated into left and right
motor torques using a normalized torque vector ratio 𝛼 𝑇𝑉 ∈
[0, 1]:
∗
𝜏𝑅𝑅 = 𝜏𝑅∗ 𝛼 𝑇𝑉 , ∗
𝜏𝑅𝐿 = 𝜏𝑅∗ (1 − 𝛼 𝑇𝑉 ). (2)
When 𝛼 𝑇𝑉 = 0.5, both motors receive the same torque;
𝛼 𝑇𝑉 = 1 allocates all the torque solely to the right motor and
𝛼 𝑇𝑉 = 0 solely to the left motor. Additionally, all electric
motors are subject to torque constraints:
𝜏𝑚𝑖𝑛.𝑖 ≤ 𝜏𝑖∗ ≤ 𝜏𝑚𝑎𝑥,𝑖 , (3)
where 𝑖 ∈ {𝐹, 𝑅𝑅, 𝑅𝐿} is the motor index, 𝜏𝑚𝑖𝑛.𝑖 the minimum
allowed torque and 𝜏𝑚𝑎𝑥,𝑖 the maximum allowed torque.
The reference curvature 𝜌∗ is tracked through the
manipulation of the steering angle 𝛿 ∗ , while considering the
additional yaw-moment generated by the torque vectoring.
Note that, throughout this document, the superscript 𝑥 ∗ is
used to denote the desired value for the variable 𝑥.
Figure 1: Block diagram of the benchmark problem.
C. Energy Management Algorithm (EMA)
The competitors are invited to develop the energy
The EMA is responsible for computing four control
management algorithm (EMA) for the vehicle (Figure 1 – top
variables, cf. the red variables in Figure 1 top:
red block). The EMA determines the operating conditions for
1. the normalized fuel cell current, 𝛼𝐹𝐶 ∈ [0,1], which
the two energy storage devices and the three electric motors;
minimization of the energy consumption and battery affects the power split between the battery and the fuel cell
∗
degradation are some of the main goals of the EMA. (note 𝐼𝐹𝐶 = 𝛼𝐹𝐶 𝐼𝐹𝐶,𝑚𝑎𝑥 , where 𝐼𝐹𝐶,𝑚𝑎𝑥 is the maximum
allowed fuel cell current);
A. Energy Storage and Vehicle Actuators 2. the axle torque distribution ratio, 𝛼𝐴𝐷 ∈ [0,1], to
The hybrid energy storage is composed of a 20 kW fuel cell determine the front and rear distribution of the desired
and 20 kWh Li-ion battery. A DC/DC converter is connected torque 𝜏 ∗ , see eq. (1).
to the fuel cell in order to regulate its current (𝐼𝐹𝐶 ) and power 3. the torque vectoring ratio, 𝛼 𝑇𝑉 ∈ [0,1], to determine the
flow between the hybrid storage elements. The hybridization torque allocation between right and left motors, see eq. (2).
with a hydrogen range extender (REX), being composed of 4. the velocity derating factor, 𝛼𝑣 ∈ [0,1], which decreases
the fuel cell and a hydrogen tank, enables the reduction of the reference velocity 𝑣 ∗ to 𝛼𝑣 𝑣 ∗ (see Figure 1); it offers an
usage of rare elements like Lithium while enhancing the additional degree of freedom to prevent violation of safety
vehicle’s range. Both the fuel cell & H2 tank and the battery constraints in the system (e.g. over-discharge of the
provide energy to the two in-wheel motors and the central battery).
front motor. Because of their decoupled architecture, the in- The EMA provided by the competitors will be evaluated
wheel motors can generate non-symmetric torques in the left using a wide range of performance metrics, including energy
and right rear wheels and, thus induce additional yaw-moment consumption, violation of safety constraints, battery
to the vehicle’s chassis. This yaw-moment can be exploited to degradation and fulfillment of the vehicle’s mission (also
enhance vehicle handling and safety [21]; it can also decrease referred as velocity derating in the sequel). These metrics will
the amount of front steering, thus reducing tire slip be defined in more detail in Section IV.
losses [10]. Additionally, a steer-by-wire actuator is installed The EMA will have access to several states of the vehicle
in the front axle to modify the steering angle 𝛿. and energy storage, such as battery state of charge,
temperature, etc. Additionally, it will also receive a short
B. Mission Planning and Vehicle Control preview of future values of the velocity and curvature
We assume that the planning of the vehicle’s mission is references:
defined in advance, e.g. using trajectory planning methods
such as [16]. The mission is characterized by a reference 𝑣̂(𝑡 + Δ𝑡), 𝜌̂(𝑡 + Δ𝑡), (4)
vehicle velocity 𝑣 ∗ (𝑡) and the track curvature 𝜌∗ (𝑡) over a where Δ𝑡 ∈ [0, Δ𝑡𝑚𝑎𝑥 ] and Δ𝑡𝑚𝑎𝑥 is the preview window
given time horizon 𝑡 ∈ [0, 𝑡𝑒𝑛𝑑 ]. which is available in map data systems with a virtual horizon.
The motion controller tracks 𝑣 ∗ through manipulation of This preview information can be generated by the trajectory
the reference traction force 𝐹 ∗ . This reference force is planning module of the vehicle.
converted into a reference torque 𝜏 ∗ = 𝐹 ∗ 𝑟𝑤 , with 𝑟𝑤 being
 III. COMPONENTS MODELING braking/accelerating are neglected and the wheel is bounded
In this section we briefly describe the Modelica model of to the track-plane (holonomic-constraint). While operating at
the ROboMObil ( [16], [17]) vehicle with a slightly modified constant load 𝑓𝑁 , the slip velocity 𝑣𝑠𝑙𝑖𝑝 at the contact patch
architecture for this challenge, depicted in Figure 1. The aim determines the slip forces according to Coulombs’s law for
is to describe the most relevant physical laws and effects with dry-friction characteristics,
the focus on a numerical efficient simulation, which is 𝑣𝑠𝑙𝑖𝑝,𝑖
necessary for a quick assessment of the EMAs or a machine 𝑓𝑖 = −𝑓𝑁 µ(𝑣𝑠𝑙𝑖𝑝 ) . (5)
𝑣𝑠𝑙𝑖𝑝
learning process to synthesize the controller.
Eq. (5) can be used for both the longitudinal (𝑖 = 𝑙𝑜𝑛𝑔) and
A. ROboMObil’s Chassis Model the lateral (𝑖 = 𝑙𝑎𝑡) direction, resolved in the wheel
To capture the fundamental dynamics of the ROboMObil, coordinate system. The friction coefficient µ depends on the
we use a double track model implemented with the Modelica slip velocity 𝑣𝑠𝑙𝑖𝑝 . To construct this dependency, we follow
planar mechanics library [22]. It offers three degrees of the approach developed in [23] and utilize two pairs of
freedom: the vehicle can move in longitudinal and lateral parameters: (𝑣𝑎𝑑ℎ𝑒𝑠𝑖𝑜𝑛 , µ𝐴 ) and (𝑣𝑠𝑙𝑖𝑑𝑒 , µ𝑆 ). The former pair
direction and rotate about the vertical axis. The chassis model, determines maximum friction µ𝐴 at 𝑣𝑠𝑙𝑖𝑝 = 𝑣𝑎𝑑ℎ𝑒𝑠𝑖𝑜𝑛 . The
depicted in Figure 2, consists of a front and rear axle, both latter pair specifies a sliding area by friction µ𝑆 at slip
with wheels, a car body and air resistance. velocities 𝑣𝑠𝑙𝑖𝑝 ≥ 𝑣𝑠𝑙𝑖𝑑𝑒 .
In addition, the model contains three types of interfaces: The rolling motion of the wheel can be actuated by the
mechanical (gray and white circles), thermal (red squares) and driving or braking torque input, cf. “flangeWheelXY” in
control (yellow connectors). The mechanical interfaces are Figure 2, with X denoting front (“F”) of rear (“R”) position of
used to exchange position and torque/force of mechanical the wheel and Y denoting its left (“L”) or right (“R”) side.
elements. The heat interface captures the heat flow between The tire losses comprise losses at tire/road contact area due
components, and allows to quantify the energy losses that are to the dry-friction contact and are expressed with the power
eventually dissipated to the environment. The control loss
interfaces contain signals that are generated by sensors (e.g. 𝑃𝑙𝑜𝑠𝑠 = 𝑣𝑠𝑙𝑖𝑝 √𝑓𝑙𝑜𝑛𝑔 + 𝑓𝑙𝑎𝑡 . (6)
acceleration signal recorded by an inertial measurement unit)
and control modules (e.g., steering actuator demand). C. Battery Model with Aging Degradation
flangeDriveFront flangeSteering
To obtain a good tradeoff between simulation speed and
modeling accuracy, we modeled the battery with an
equivalent electrical circuit model. This circuit consists of an
flangeWheelFL flangeWheelFR
ideal voltage source (𝑈𝑂𝐶𝑉 ) in series with an internal
resistance (𝑅𝑖 ). The terminal voltage of the cell is described
axleFront airResistance
as
carBody 𝑢𝑐𝑒𝑙𝑙 = 𝑈𝑂𝐶𝑉 (𝑆𝑜𝐶𝑏 ) − 𝑅𝑖 (𝑆𝑜𝐶𝑏 ) ⋅ 𝑖𝑐𝑒𝑙𝑙 , (7)
heatPort
controlBus

where 𝑖𝑐𝑒𝑙𝑙 is the current in the battery cell.

Both the internal voltage and resistance depend on the
flangeWheelRL
axleRear
flangeWheelRR
battery state of charge (𝑆𝑜𝐶𝑏 ), a normalized indicator for the
amount of charge stored in the battery. Lookup tables are
employed to characterize this variation, but are limited to
Figure 2: Structure of the Modelica planar vehicle model room temperature values. A more complex implementation
with temperature dependency is given in [24].
The rigid front axle is assembled from an open differential In Figure 3 the functional model of the battery pack is
that distributes the torque of the central front motor – provided shown. On the left side we have the electrical connectors (in
by “flangeDriveFront” connector in Figure 2 – to the left and blue), which propagate the electric signals (voltage and
right wheel, which utilizes a slip-based tire model, see current) to the other components in the vehicle, while on the
Section III.B. Moreover, a steering mechanism that equally right side we have the 𝑆𝑜𝐶𝑏 calculation realized by an
transmits the steering input onto both left and right front integrator (𝑐𝑎𝑙𝑐_𝑆𝑜𝐶𝑏 ), in dependency of the nominal cell
wheels is implemented. The rigid rear axle is, in contrast,
capacity 𝐶𝑐𝑒𝑙𝑙,0 . The pack is scaled by the number of in serial
composed only of two slip-based tire models. They can be
𝑛𝑜𝑠 and in parallel 𝑛𝑜𝑝 connected cells. The battery model
directly driven by torque sources connected from outside to
connectors “flangeWheelRL” and “flangeWheelRR” for left also contains a simple thermal model that captures heat flow
and right wheel, respectively. In this way, an in-wheel drive between the battery cell and a (constant-temperature)
can be realized. environment using a thermal resistor.

B. Slip-based Tire Model with Losses

ROboMObil’s four wheels are modelled with a slip-based
tire model [23] which is extended with energy loss effects. To
simplify the model, tire load fluctuation during cornering or
 fundamental electric machine equations 𝑑𝛹𝑑 ⁄𝑑𝑡 and d𝛹q ⁄d𝑡
voltage of the stator flux are given as:
voltageScale lookup_SOC_OCV

cellVoltage
p
voltage 𝑑𝛹𝑑
= 𝑈𝑑 − 𝑅𝑠 𝐼𝑑 + 𝛺𝐿 𝐿
⏟1 𝐼q , (11)
𝑑𝑡
voltageSensor

k=no_s
𝛹q
V

variableResistor
packScale lookup_Ri

d𝛹q
= 𝑈𝑞 − 𝑅s 𝐼q − 𝛺L (𝛹
⏟ PM + 𝐿1 ⋅ 𝐼d ).
heatFlowSensor

k=no_s/no_p SoC_b
d𝑡 (12)
Q_flow
𝛹d
W

calc_SoC_b
sensorCurrent

current These equations are easily implemented in Modelica using

A

I
u1 / u2 the acausal equation environment, i.e. all equation hold for the
n degC C_aged
four quadrants of operation of the electric machine and are not
DeltaC_dN dependent on any signal flow direction [26].
batteryBus
R=20 K/W

agingCalc
mechanical
pin_p airGap_torq
airGap_torque Power inertiaRotor flange
current
M_Mi
tau W
K heatPort controlBus J=J_r
I_d
T=25 °C lossesFriction
i_DC ElectricPower

signalCurrent
u1 / u2 P_total
Figure 3: The equivalent circuit battery model in Modelica
I_q
The battery model also includes an aging model (agingCalc sensorFilter

Voltage
Sensor
in Figure 3) for predicting the capacity degradation of the fixed

V
PT1
pin_n
battery due to the charge/discharge events. This aging is T=0.01 s
prescribedHeatFlow
computed using the average current (𝑖𝑎𝑣𝑔 ) and temperature heatPort
LossPower
P_loss

(𝑇𝑎𝑣𝑔 ) of the battery over a discharge cycle, as described in electricMotorBus

[25] p. 1232. The normalized cell capacity loss during the

Figure 4: Quasi-stationary PMSM model
drive cycle Δ𝐶𝑐𝑒𝑙𝑙 is calculated as
To prevent a slow simulation, the electric motor model and
Δ𝐶𝑐𝑒𝑙𝑙 (𝑖𝑎𝑣𝑔 , 𝑇𝑎𝑣𝑔 , 𝑁) = its controller are designed in the rotating d/q-frame and the
𝜃4 𝜃5 reverse transformation to the a/b/c phases is not explicitly
= 𝜃1 exp (− + (𝜃2 + 𝑖 ))
⏟ 𝑇𝑎𝑣𝑔 𝑇𝑎𝑣𝑔 𝑎𝑣𝑔 (8)
implemented. The closed-loop response of the motor current
(𝐼𝑞 ) controller is approximated by a second-order transfer
start of battery lifetime − slow aging
+ 𝜃8 exp(𝑁 − 𝑁𝑘𝑛𝑒𝑒 𝜃7 ) ,
⏟ function with cut-off frequency 100 Hz.
"late" battery lifetime − fast aging Table 1: PMSM variables description
where 𝜃𝑖 are aging parameters taken from experiments carried Quantity Unit Description
out in [25]. The parameter 𝑁 is the number of discharge cycles
𝑈𝑑/𝑞 V Voltage in d-/q-axis
and 𝑁𝑘𝑛𝑒𝑒 is a parameter point where the battery aging 𝐼𝑑/𝑞 A Current in d-/q-axis
accelerates. The remaining cell capacity, in dependency of the 𝐿1 H Inductance in d- & q-axis
initial cell capacity 𝐶𝑐𝑒𝑙𝑙,0 is defined as 𝑅𝑠 Ω Warm resistance per phase
𝑝 - Pole pair number
𝐶𝑎𝑔𝑒𝑑 = (1 − Δ𝐶𝑐𝑒𝑙𝑙 )𝐶𝑐𝑒𝑙𝑙,0 . (9) 𝜓𝑃𝑀 Wb Magnetic flux of permanent magnets
𝜓𝑑⁄𝑞 Wb d-/q- component of stator flux
To facilitate the evaluation of the EMA, we focus on the
𝜔𝐿 rad/s (Normed) Electrical angular velocity of rotor
rate of aging of the battery during the driving cycle. It is
determined via a linearization of Δ𝐶𝑐𝑒𝑙𝑙 around the current The model of the electric machine neglects reluctance
number of cycles 𝑁 ̅ as follows influences (𝐿𝑑 = 𝐿𝑞 = 𝐿1 ). Its air-gap torque is calculated as
dΔ𝐶𝑐𝑒𝑙𝑙 (𝑖𝑎𝑣𝑔 , 𝑇𝑎𝑣𝑔 ) follows:
𝑑𝑁 𝜏𝑗𝑘 = 𝑀𝑀𝑖 = 3⁄2 ⋅ 𝑝 ⋅ 𝛹𝑃𝑀 ⋅ 𝐼𝑞 . (13)
dΔ𝐶𝑐𝑒𝑙𝑙 (𝑖𝑎𝑣𝑔 , 𝑇𝑎𝑣𝑔 , 𝑁)
= | (10) Besides this quasi-stationary electric machine model, we
dN ̅
𝑁=𝑁 also consider the energy losses within the PMSM, summed up
𝜃4 𝜃5
̅ 𝜃3−1 . in 𝑃𝑙𝑜𝑠𝑠 = 𝑃𝑙𝑜𝑠𝑠,𝑖𝑛𝑣 + 𝑃𝑐𝑜𝑝 + 𝑃𝑖𝑟𝑜𝑛 + 𝑃𝑓𝑟𝑖𝑐 in Figure 4:
= 𝜃1 exp (− + (𝜃2 + 𝑖 )) 𝜃3 𝑁
𝑇𝑎𝑣𝑔 𝑇𝑎𝑣𝑔 𝑎𝑣𝑔 Inverter losses (switching and basic load)

D. Quasi-Stationary Electric Motor Model 𝑃𝑙𝑜𝑠𝑠,𝑖𝑛𝑣 = 𝑃𝑖𝑛𝑣,𝑐𝑜𝑛𝑠𝑡 + 𝑘𝑖𝑛𝑣 ∙ 𝐼𝑞 . (14)

The three traction motors installed in the vehicle (cf. Figure Copper losses (coil resistance)
1) rely on permanent magnet synchronous machines 2
𝑃𝑐𝑜𝑝 = (√3/2 ∙ 𝐼𝑞 ) ∙ 𝑅𝑠 . (15)
(PMSM). They are represented using a quasi-stationary
model (Figure 4) with the variables listed in Table 1. The Iron losses (also known as core losses)
 2
𝑃𝑖𝑟𝑜𝑛 = 𝑘ℎ𝑦𝑠𝑡 ∙ 𝜔𝑚 + 𝑘𝑒𝑑𝑑𝑦 ∙ 𝜔𝑚 . (16) 𝑥̇𝑆𝑇𝑀 = 𝐴(𝑣 ∗ )𝑥𝑆𝑇𝑀 + 𝐵(𝑣 ∗ )𝛿 ∗
+ 𝐸(𝑣 ∗ )𝑀𝑧,𝑇𝑉 (18)
Mechanical losses (friction effects, e.g., in bearings)
𝛽 ∗̇ = 𝑣 ∗ 𝜌∗ − 𝑟 ∗ ,
𝑃𝑓𝑟𝑖𝑐 = 𝑘𝑓𝑟𝑖𝑐 ∙ 𝜔𝑚 . (17)
where 𝑥𝑆𝑇𝑀 = [𝛽 ∗ , 𝑟 ∗ ]𝑇 is vector with the reference side-slip
E. Hydrogen Fuel Cell Model angle and yaw rate; 𝐴(𝑣 ∗ ) and 𝐸(𝑣 ∗ ) are matrices that depend
The fuel cell model relies on a quasi-stationary model as on the vehicle velocity; 𝑀𝑧,𝑇𝑉 is the torque vectoring
proposed in [27] and depicted in Figure 5. generated by the in-wheel motors:
𝑐 𝑐𝜏 ∗
p 𝑀𝑧,𝑇𝑉 ≈ (𝜏𝑅∗ − 𝜏𝐿∗ ) = (2𝛼 𝑇𝑉 − 1), (19)
H2_dynamics switchH2on H2_activation
2𝑟𝑤 2𝑟𝑤
R
PT1 B where 𝑟𝑤 is the wheel radius and 𝑐 the vehicle’s track width.
generatorCurrent

I_dem
T=0.1 s H2_deactivated
Assuming slow variations in vehicle curvature we can obtain
0 ≈ 𝐴(𝑣 ∗ )𝑥𝑆𝑇𝑀 + 𝐵(𝑣 ∗ )𝛿 ∗ + 𝐸(𝑣 ∗ )𝑀𝑧,𝑇𝑉
voltageSensor

k=0
0 ≈ 𝑣 ∗ 𝜌∗ − [0, 1]𝑥𝑆𝑇𝑀 , (20)
V

H2Consumption H2_tank
H2SoC
H2power
which represents a system of three linear equations with three
I
unknowns (𝛿 ∗ , 𝑥𝑆𝑇𝑀 ) = (𝛿 ∗ , 𝛽 ∗ , 𝑟 ∗ ). The steering angle
sensorCurrent

H2Power
rexBus
2Efficiency powerLoss
applied to the vehicle is computed from the solution of these
A

add
n
-1 equations:
+
inv 1 𝛿 ∗ = 𝑓𝛿 (𝑣 ∗ , 𝑀𝑧,𝑇𝑉 , 𝜌∗ ). (21)
H2Losses
ground
k=1 This represents a feedforward control law, which allows the
heatPort controlBus
vehicle to follow the reference curvature (𝜌∗ ) when the model
REXparameters
uncertainty is reduced.
Figure 5: Functional hydrogen range extender model
IV. ENERGY MANAGEMENT ALGORITHM
Whenever a demanded current 𝐼𝑑𝑒𝑚 exceeds a minimum
threshold (1 A in our model), the functional model This section provides a brief overview of the requirements
approximates the fuel cell’s output current using first order for the EMA that the competitors will need to develop, as well
system dynamics. The fuel consumption relies on two tables. the scoring and ranking assessment of the competition.
The first table maps the necessary hydrogen mass flow that is A. Input/Output Interfaces
taken from the tank (Figure 6 – blue line), modelled as an The EMA will have access to the following vehicle states
integrator, whereas the second table maps the efficiency
(Figure 6 – green line) depending on the point of operation. ̂ ] ∈ 𝑋𝐸𝑀𝐴 ,
𝑥 = [𝑣, 𝑎, 𝑆𝑜𝐶𝑏 , 𝑇𝑏 , 𝑆𝑜𝐶𝐹𝐶 , 𝑊 (22)

0.6
which contain the current velocities 𝑣, accelerations 𝑎, state
of charge 𝑆𝑜𝐶𝑏 and temperature of the battery 𝑇𝑏 , and the state
0.5 of charge of the fuel cell 𝑆𝑜𝐶𝐹𝐶 . The variable 𝑊̂ is a vector
H2 consumption [kg/kW] with a short preview information about the reference velocity
0.4
Max. fuel cell efficiency [0,1] and track curvature
0.3 𝑁𝑝𝑟𝑒
̂ = (𝑣 ∗ (𝑡 + 𝑘Δ𝑡𝑝 ), 𝜌∗ (𝑡 + 𝑘Δ𝑡𝑝 ))
𝑊 , (23)
𝑘=0
0.2
where Δ𝑡𝑝 is the sample time and 𝑁𝑝𝑟𝑒 are the samples of the
0.1 preview window. We denote 𝑋𝐸𝑀𝐴 as the set of all possible
combinations of states that the EMA might receive.
0.0
0.0 5.0 10.0 15.0 20.0 The EMA generates four output control signals, see Section
Power [kW] II.C,
Figure 6: Efficiency and hydrogen consumption map of the fuel cell
based on [27] 𝑢 = [𝛼𝐹𝐶 , 𝛼𝐴𝐷 , 𝛼 𝑇𝑉 , 𝛼𝑣 ] ∈ 𝑈 = [0,1]4 , (24)
F. Vehicle Motion Controller where 𝑈 represents the set of allowed control actions.
The longitudinal velocity control relies on a linear PI B. Safety Constraints
controller [28]. The lateral controller uses a model inversion
technique to follow the reference curvature (𝜌∗ ) generated by The competitors will provide a control policy 𝜋 for the
the mission planner. It makes use of a simplified single-track energy management that maps the states into the control
model [29] actions
𝜋(𝑥): 𝑋𝐸𝑀𝐴 → 𝑈. (25)
 This policy will need to fulfill two type of safety 𝛼̃𝐴𝐷 (𝑥) = 1/2 (31)
constraints. The first are state of charge constraints of the
energy storage devices: 1 (32)
𝛼̃ 𝑇𝑉 (𝑥) = + 𝑘 𝑇𝑉 𝜌∗ (𝑣 ∗ )2
2
𝑋𝑠𝑎𝑓𝑒,𝑆𝑂𝐶 = {𝑥𝐸𝑀𝐴 ∈ 𝑋𝐸𝑀𝐴
𝑆𝑜𝐶𝑏,𝑚𝑖𝑛 ≤ 𝑆𝑂𝐶𝑏 ≤ 𝑆𝑜𝐶𝑏,𝑚𝑎𝑥 (26) 𝛼̃𝑣 (𝑥) (33)
1 if 𝑥(𝑡) ∈ 𝑋𝑠𝑎𝑓𝑒,𝑆𝑂𝐶
𝑆𝑜𝐶𝐹𝐶,𝑚𝑖𝑛 ≤ 𝑆𝑂𝐶𝐹𝐶 ≤ 𝑆𝑜𝐶𝐹𝐶,𝑚𝑎𝑥 },
= { 𝑠𝑎𝑡(𝑚0 𝑆𝑜𝐶𝑏 + 𝑏0 ), 𝑖𝑓 𝑆𝑜𝐶𝑏 (𝑡) ≤ 𝑆𝑜𝐶𝑏,𝑚𝑖𝑛
where 𝑆𝑜𝐶𝑏,𝑚𝑖𝑛 , 𝑆𝑜𝐶𝐹𝐶,𝑚𝑖𝑛 represent the minimum 𝑆𝑜𝐶 levels 𝑠𝑎𝑡(𝑚1 𝑆𝑜𝐶𝑏 + 𝑏1 ), 𝑖𝑓 𝑆𝑜𝐶𝑏 (𝑡) ≥ 𝑆𝑜𝐶𝑏,𝑚𝑎𝑥
for the energy storage devices, and 𝑆𝑜𝐶𝑏,𝑚𝑎𝑥 , 𝑆𝑜𝐶𝐹𝐶,𝑚𝑎𝑥 their
This policy enforces
maximum values. The second set of constraints is the battery
• constant usage of the fuel cell (with ratio 𝑘𝐹𝐶 ) if
temperature
safety constraints are fulfilled (eq. (30)); it disables
𝑋𝑠𝑎𝑓𝑒,𝑇 = {𝑥𝐸𝑀𝐴 ∈ 𝑋𝐸𝑀𝐴 : 𝑇𝑏 ≤ 𝑇𝑏,𝑚𝑎𝑥 }. (27) the fuel cell whenever violation of 𝑆𝑜𝐶 constraints
occur, eq. (33),
This set of constraints can be temporarily violated;
however, these violation increases the risk of failure of this • constant front-rear torque distribution ratio, eq. (31),
component (e.g. thermal runaway [30]) and are penalized in • a torque allocation policy proportional to the
the EMA performance score. expected lateral acceleration of the vehicle
(𝜌∗ (𝑣 ∗ )2 ); where 𝑘 𝑇𝑉 is a constant, eq. (32),
C. Performance Metrics • a simple derating strategy that reduces the maximum
The EMA will be evaluated using the following vehicle velocity whenever the battery 𝑆𝑜𝐶 are
performance metrics (see also Table 2). violated; see eq. (33) for details, where 𝑚0 , 𝑏0 , 𝑚1 ,
• 𝐽𝐸 : Total energy consumption of the vehicle. It is 𝑏1 are parameters and 𝑠𝑎𝑡() a saturation function
computed by integrating the power delivered by the that enforces the range [0,1].
battery (𝑝𝑏𝑎𝑡 ) and the fuel cell (𝑝𝐹𝐶 ). This baseline policy generates baseline metrics, which are
• 𝐽𝑆𝑜𝐶 : timespan that SoC constraints 𝑋𝑠𝑎𝑓𝑒,𝑆𝑂𝐶 are denoted as 𝐽̃𝐸 , 𝐽̃𝑆𝑜𝐶 , 𝐽̃𝑇𝐶 , 𝐽̃𝐷𝑒𝑔 , 𝐽̃𝑣 .
violated. Figure 7 shows an example of vehicle states and control
• 𝐽𝑇𝐶 : maximum temperature violation. inputs that were generated by the baseline EMA policy 𝜋̃(𝑥)
• 𝐽𝑑𝑒𝑔 : battery capacity that is lost during the vehicle during an urban driving cycle.
mission due to battery cycle aging.
• 𝐽𝑣 : derating metric that captures ability of the vehicle
to track the velocity profile defined by the mission
planning.
Table 2: Summary of performance metrics

Metric Formula
𝑡𝑒𝑛𝑑
𝐽𝐸
∫ 𝑝𝑏𝑎𝑡 (𝑡) + 𝑝𝐹𝐶 (𝑡)𝑑𝑡
0
𝐽𝑆𝑜𝐶
∫ 𝕝{𝑥(𝑡)∉𝑋𝑠𝑎𝑓𝑒,𝑆𝑂𝐶} 𝑑𝑡
𝐽𝑇𝐶 max max(0, (𝑇𝑏𝑎𝑡 (𝑡) − 𝑇𝑏𝑎𝑡,𝑚𝑎𝑥 ))
𝑡
𝐽𝑑𝑒𝑔 dΔ𝐶𝑐𝑒𝑙𝑙 (𝑖𝑎𝑣𝑔 , 𝑇𝑎𝑣𝑔 )
𝑑𝑁
𝐽𝑣
∫(1 − 𝛼𝑣 (𝑡))𝑑𝑡
𝕝{𝑐𝑜𝑛𝑑} is an indicator function that returns 1 if 𝑐𝑜𝑛𝑑 = 1, and zero
otherwise
Note that the value of these performance metrics will be
dependent on the EMA provided by the algorithm 𝜋(𝑥) and
the mission profile:
𝑁
𝑊 = (𝑣 ∗ (𝑡𝑘 ), 𝜌∗ (𝑡𝑘 ))𝑘=1 . (28)

D. EMA Baseline Policy

The simulation model provided to the competitors contains
a simple example policy, which is called baseline policy. This Figure 7. Example of a subset of vehicle states and control inputs
baseline policy is implemented as follows generated by the vehicle model and EMA baseline policy

𝜋̃(𝑥) = [𝛼̃𝐹𝐶 (𝑥), 𝛼̃𝐴𝐷 (𝑥) 𝛼̃ 𝑇𝑉 (𝑥), 𝛼̃𝑣 (𝑥)] (29) In this example, the battery temperature exceeds the upper
limit at 1000 s, which might compromise battery safety. The
𝑘 if 𝑥(𝑡) ∈ 𝑋𝑠𝑎𝑓𝑒,𝑆𝑂𝐶 (30) derating strategy becomes active after 1800 s, decreasing 𝛼𝑣
𝛼̃𝐹𝐶 (𝑥) = { 𝐹𝐶
0 otherwise and the maximum velocity (and maximum power) that the
vehicle can reach. The competitors are invited to develop fulfillment of safety constraints, battery degradation and loss
better EMAs that can significantly avoid violation of the of vehicle performance (derating).
safety constraints, extend vehicle operation without loss of The simulation framework will be available in the GitHub
performance and reduced energy consumption. repository https://github.com/DLR-VSDC/IEEE-MVC-2023
by November 2022. The repository will contain the vehicle
E. Evaluation and Ranking
model (FMU), the baseline energy management algorithm
The score of the EMA over a given mission profile is (MATLAB/Simulink) for an easy point-of-entry, and scripts
computed as a weighted summation of the performance to generate performance reports. The value of all vehicle and
metrics, normalized with respect to the EMA baseline policy component parameters will be included in the repository to
(𝜋̃). Mathematically, this means: support the competitor’s controller synthesis process.
𝐽𝐸 𝐽𝑆𝑜𝐶 Future points of interest in research are the development of
𝐽 = 𝑘𝐸 + 𝑘𝑆𝑜𝐶 more complex vehicle components and architectures which
𝐽̃𝐸 𝐽̃𝑆𝑜𝐶
𝐽𝑇𝐶 𝐽𝐷𝑒𝑔 𝐽𝑣 (34) will be tuned and validated with real world experiments.
+𝑘 𝑇𝐶 + 𝑘𝐷𝑒𝑔 + 𝑘𝑣 ,
̃𝐽𝑇𝐶 ̃𝐽𝐷𝑒𝑔 𝐽̃𝑣 VI. AUTHOR CONTRIBUTIONS

where 𝑘𝐸 , 𝑘 𝑇𝐶 , 𝑘𝐷𝑒𝑔 , 𝑘𝑣 are known weights (defined by the Author Contributions: Conceptualization, J.B., R.C., and J.T.;
methodology, J.B., R.C., and J.T. Modelica vehicle model and components,
organizers) and 𝐽𝐸 , 𝐽𝑆𝑜𝐶 , 𝐽𝑇𝐶 , 𝐽𝐷𝑒𝑔 , 𝐽𝑣 are the performance J.B., and J.T.; controller and trajectories, R.C., and J.B.; EMA assessment
metrics obtained with the EMA provided by the competitor, and baseline algorithm: I.E., R.C. and J.B., library management and testing
𝜋(𝑥). Note that the overall score of the EMA depends not only J.T., I.E.,J.B.; writing—original draft preparation, R.C., J.B., J.T. and I.E.;
All authors have read and agreed to the published version of the manuscript.
on the policy 𝜋(𝑥), but also on the mission profile 𝑊, i.e., Acknowledgement: We would like to thank Basilio Lenzo for testing of
challenge data and models.
𝐽(𝜋, 𝑊). (35) Funding: internal institutional funding of the DLR granted by the
Helmholtz Association of German Research Centres HGF.
The evaluation will consider a bank of mission profiles
𝑊1 , 𝑊2 , … , 𝑊𝑀 , which contain typical operating conditions
VII. REFERENCES
for the vehicle. Each profile can be selected with probability
𝑝(𝑊𝑗 ) = 𝑝𝑗 , 𝑗 = 1, … , 𝑀. (36) [1] E. Amaya, H. Chiacchiarini, C. De Angelo und M.
Asensio, „The Energy Management Strategy of FC/Battery
Some driving cycles (but not all) will be provided to the
Vehicles Winner of the 2017 IEEE VTS Motor Vehicles
competitors; the driving cycle probability (𝑝𝑗 ) is also Challenge,“ in The Energy Management Strategy of
unknown to the competitors. The performance of an EMA FC/Battery Vehicles Winner of the 2017 IEEE VTS Motor
will be performed based on the average cost over all mission Vehicles Challenge, 2017.
profiles: [2] A. Ferrara und H. C., „Rule-Based Energy Management
𝑀 Strategy of Fuel Cell/Ultracapacitor/Battery Vehicles:
winner of the IEEE VTS Motor Vehicles Challenge 2020,,“
𝔼{𝐽(𝜋, 𝑊)} = ∑ 𝑝(𝑊𝑗 )𝐽(𝜋, 𝑊𝑗 ), (37) in 2020 IEEE Vehicle Power and Propulsion Conference,
𝑗=1 2020.
where 𝔼{. } is the expected value operator (over 𝑊). The final [3] B. Nguyen; J. P. F. Trovão; S. Jemei; L. Boulon; A.
EMA ranking will be performed as follows. We will collect Bouscayrol, „IEEE VTS Motor Vehicles Challenge 2021 -
all the energy management algorithms provided by the Energy Management of A Dual-Motor All-Wheel Drive
Electric Vehicle,“ in IEEE Vehicle Power and Propulsion
competitors, 𝜋1 , … , 𝜋𝐿 , where 𝐿 is the number of received
Conference (VPPC), 2021.
submissions. The competitor that provides the lowest average
cost will be the winner, i.e.: [4] C. Depature, S. Pagerit, L. Boulon, S. Jemei, A. Jemei
und A. Bouscayrol, „IEEE VTS Motor Vehicles Challenge
𝐿𝑤𝑖𝑛𝑛𝑒𝑟 = arg min 𝔼{𝐽(𝜋𝑙 , 𝑊)}. (38) 2018 - Energy Management of a Range Extender Electric
l∈{1,…,L} Vehicle,“ in IEEE Vehicle Power and Propulsion
Conference (VPPC), 2017.
V. SUMMARY AND OUTLOOK [5] H. Pereira, R. de Castro, R. E. Araújo, „How to Win the
2021 IEEE VTS Motor Vehicles Challenge With a
This article presents the conceptualization, modeling and
Pragmatic Energy Management Strategy,“ in IEEE Vehicle
setup for the IEEE Motor Vehicle Challenge 2023. It relies on Power and Propulsion Conference (VPPC), 2021.
a multi-domain modeling approach based on Modelica, FMI
[6] T. Vo-Duy, J. P. F. Trovão, S. Jemeï, L. Boulon, M. C.
technology for a seamless model exchange between different
Ta und A. Bouscayrol, „IEEE VTS Motor Vehicles
simulation tools, as well as an integrated vehicle control with Challenge 2022 - Sizing and Energy Management of
trajectory following. Prospective competitors are invited to Hybrid dual-Energy Storage System for a Commercial
submit algorithms that can efficiently perform the energy Electric Vehicle,“ in 2021 IEEE Vehicle Power and
management of the multiple electric motors and energy Propulsion Conference (VPPC), 2021.
storage devices available in the reference vehicle. A wide [7] J. Solano, S Jemei, L Boulon, L Silva, D Hissel, M-C
range of performance metrics will be used to rank the Pera, „IEEE VTS motor vehicles challenge 2020-energy
submitted strategies, including energy consumption, management of a fuel cell/ultracapacitor/lead-acid battery
 hybrid electric vehicle,“ in IEEE Vehicle Power and [22] D. Zimmer, „A Planar Mechanical Library for Teaching
Propulsion Conference (VPPC), 2019. Modelica,“ in Proceedings of the 9th International
[8] W. Lhomme; T. Letrouve; L. Boulon; S. Jemei; A. Modelica Conference, Munich, Germany, 2012.
Bouscayrol; F. Chauvet; F. Tournez, „IEEE VTS Motor [23] D. Zimmer und M. Otter, „Real-time models for wheels
Vehicles Challenge 2019 - Energy Management of a Dual- and tyres in an object-oriented modelling,“ Vehicle System
Mode Locomotive,“ in IEEE Vehicle Power and Dynamics, 2009.
Propulsion Conference (VPPC), 2018. [24] J. Brembeck, „A Physical Model-Based Observer
[9] R. de Castro, M. Tanelli, R. E. Araújo und S. M. Framework for Nonlinear Constrained State Estimation
Savaresi, „Design of safety-oriented control allocation Applied to Battery State Estimation,“ Sensors, Vol. 19, Nr.
strategies for overactuated electric vehicles,“ Vehicle 20, p. 4402, 2019.
System Dynamics, Vol. 52, pp. 1017-1046, 2014. [25] R. de Castro, H. Pereira, R. E. Araújo, J. V. Barreras und
[10] J. Torinsson, M. Jonasson, D. Yang und B. Jacobson, H. C. Pangborn, „qTSL: A Multilayer Control Framework
„Energy reduction by power loss minimisation through for Managing Capacity, Temperature, Stress, and Losses in
wheel torque allocation in electric vehicles: a simulation- Hybrid Balancing Systems,“ IEEE Transactions on
based approach,“ Vehicle System Dynamics, Vol. 60, pp. Control Systems Technology, Vol. 30, pp. 1228-1243,
1488-1511, 2022. 2022.
[11] B. Lenzo, „Torque Vectoring Control for Enhancing [26] Wikipedia, „Modelica,“ 08 2022. [Online]. Available:
Vehicle Safety and Energy Efficiency,“ in Vehicle https://en.wikipedia.org/wiki/Modelica.
Dynamics: Fundamentals and Ultimate Trends, B. Lenzo, [27] D. Feroldi, M. Serra und J. Riera, „Energy Management
Pub., Cham, Springer International Publishing, 2022, p. Strategies based on efficiency map for Fuel Cell Hybrid
193–233. Vehicles,“ Journal of Power Sources, Vol. 190, pp. 387-
[12] C. Chatzikomis, A. Sorniotti, P. Gruber, M. Zanchetta, 401, 2009.
D. Willans und B. Balcombe, „Comparison of Path [28] K. J. Åström und R. M. Murray, Feedback Systems: An
Tracking and Torque-Vectoring Controllers for Introduction for Scientists and Engineers, Second Edition,
Autonomous Electric Vehicles,“ IEEE Transactions on Princeton University Press, 2021.
Intelligent Vehicles, Vol.. 3, pp. 559-570, 2018.
[29] J. Ackermann, P. Blue, T. Bünte, L. Guvenc, D.
[13] R. de Castro und J. Brembeck, „Lyapunov-based fault Kaesbauer, M. Kordt, M. Muhler und D. Odenthal, Robust
tolerant control allocation,“ Vehicle System Dynamics, Control: The Parameter Space Approach, 2002.
Vol. 0, pp. 1-26, 2021.
[30] X. Feng, D. Ren, X. He und M. Ouyang, „Mitigating
[14] J. Brembeck, L. M. Ho, A. Schaub, C. Satzger, J. Thermal Runaway of Lithium-Ion Batteries,“ Joule, Vol.
Tobolar, J. Bals und G. Hirzinger, „ROMO - The Robotic 4, pp. 743-770, 2020.
Electric Vehicle,“ in 22nd IAVSD International
Symposium on Dynamics of Vehicle on Roads and Tracks,
Manchester Metropolitan University, 2011.
[15] J. Brembeck und P. Ritzer, „Energy optimal control of
an over actuated Robotic Electric Vehicle using enhanced
control allocation approaches,“ in Proceedings of the 2012
IEEE Intelligent Vehicles Symposium, Alcala de Henares,
Spain, 2012.
[16] J. Brembeck, „Model Based Energy Management and
State Estimation for the Robotic Electric Vehicle
ROboMObil,“ Ph.D. dissertation, Technical University of
Munich, Munich, Germany, 2018.
[17] DLR Vehicle System Dynamics Blog, „10 years
ROboMObil,“ 08 2022. [Online]. Available:
https://vsdc.de/en/robomobil-timeline/.
[18] Modelica Association, „Modelica,“ 2022. [Online].
Available: http://www.modelica.org.
[19] R. de Castro, T. Bünte und J. Brembeck, „Design and
Validation of the Second Generation of the Robomobil's
Vehicle Dynamics Controller,“ in 24th Symposium of the
International Association for Vehicle System Dynamics,
Graz, Austria, 2016.
[20] J. Ultsch, J. Mirwald, J. Brembeck und R. de Castro,
„Reinforcement Learning-based Path Following Control
for a Vehicle with Variable Delay in the Drivetrain,“ in
31st IEEE Intelligent Vehicles Symposium, 2020.
[21] T. Bünte, L. M. Ho, C. Satzger und J. Brembeck,
„Central Vehicle Dynamics Control of the Robotic
Research Platform ROboMObil,“ ATZelektronik
worldwide, June 2014.