Sie sind auf Seite 1von 6

Intelligent Power Management in SHEV based on

Roadway Type and Traffic Congestion Levels


Zhihang Chen, Leonidas Kiliaris, Yi L. Murphey

M. A. Masrur

Department of Electrical and Computer Engineering


University of Michigan-Dearborn
Dearborn, MI 48103, U.S.A
yilu@umich.edu

U.S. Army RDECOM-TARDE


Warren, MI, USA

Series hybrids are considered most suitable for heavier


vehicles like city busses or small cargo trucks.

Abstract This paper presents a machine learning approach to


train an intelligent power controller for a series hybrid electric
vehicle. The proposed machine learning approach exploits the
best efficiency of the components associated with the roadway
type and traffic congestion level to reduce the overall fuel
consumption. [Given certain non changeable parameters such as
the generator efficiency, the battery parameters, and the engine
efficiency, the optimal system point can be calculated.] The
algorithm itself will be able to exploit the road conditions at a
given time, but only an average value of the road conditions. It is
the goal of this paper to further refine the standard best
efficiency control schemes by utilizing the road type prediction
and dynamically controlling the engine/generator power to best
match not only the best efficiency calculations but also an optimal
prediction of the road conditions, not just the average.

I.

II.

MACHINE LEARNING ABOUT OPTIMAL POWER CONTROL


IN A SHEV

Figure 1 illustrates the interaction between the proposed


Intelligent Power Controller (IPC) and the major power
components in a general series hybrid electric vehicle (SHEV)
system. The propulsion power of a SHEV is provided only
by an electric drive made of an electric motor. The electric
power, in turn, comes from the storage system, the battery and
a dc source. The dc source can be a gasoline or diesel engine
or a bank of fuel cells.

INTRODUCTION

Growing environmental concerns coupled with the complex


issue of global crude oil supplies drive automobile industry
towards the development of hybrid fuel-efficient vehicles.
Due to the multiple-power-source nature and the complex
configuration and operation modes, the control strategy of a
hybrid vehicle is more complicated than that of a conventional
vehicle. The power management involves the design of the
high-level control algorithm that determines the proper power
split between the motor and the engine to minimize fuel
consumption and emissions, while satisfying constraints such
as drivability, charge sustaining and component reliability[7].
It is well recognized that the energy management strategy of a
hybrid vehicle has high influences over vehicle performances.
In this paper we present our research in real-time intelligent
power management for a series hybrid electric vehicles.
Series hybrid electric vehicle consist of an engine, which has a
sole purpose of powering a generator which creates electrical
power for the electric motor of the vehicle, and also a battery
to supply many of the transient power needs to the electric
motor that the generator cannot. A series hybrid can be
described as a pure electric vehicle fitted with on board
electricity generation. Series hybrid of the components might
be easier as there are no mechanical couplings and that, given
sophisticated power electronics, the system acts as an electric
continuously variable transmission (CVT). This allows the
system to be operated at maximum efficiency at all powers.

Ps: DC source power


Ps : Desired value of Ps
Pb: Battery power
Pd - driver power demand
V(t): vehicle speed at time t R[i]: road type prediction result
Psa: Average value of Ps
Figure 1. Intelligent Power Controller in a SHEV

In the above figure Pd is drive power, Ps is dc source power,


Ps* is the desired value of Ps , and Pb is battery terminal power.
Pd constitutes of the load of the electric sources (i.e., the
combined power of generator and the battery).
IPC_SHEV, is an intelligent power controller trained to
optimizing power flow in a SHEV in real time. Its main
control objective is to make the electric source supply this
power Pd , while managing the power sources so that the
corresponding fuel consumption is as low as possible.

This work is supported in part by a research contract from the US Army


RDECOM-TARDEC and a research grant from Michigan 21st Jobs Fund.

978-1-4244-2601-0/09/$25.00 2009 IEEE

915

best efficiency can be represented by the envelop of the SFC


curves at different engine speeds. Figure 2 shows the envelop
of the SFC curves generated by Ford Explorer series hybrid
electiric vehicle and the quadratic curve obtained by the least
MSE (Mean Square Error) method.

At any given time during a drive cycle, based on the current


vehicle state, which is represented by V(t), the current vehicle
speed, driver power demand Pd , and SOC of the battery, the
IPC calls the neural network, NN_RT&TC, to predict the
current road type and traffic congestion (RT&TC) level, and
calculates the electric power setpoint to the battery controller
and a feed forward torque compensation to the engine
controller. Pb, representing the power actually to be
charged(Pb > 0) or discharged(Pb < 0) from the battery, is set
by the IPC with the aim of minimizing fuel consumption. The
optimal engine power Ps, calculated based optimal setpoint of
Ps, and is used to find the optimal feed forward torque
compensation through the engine fuel efficiency map.
The machine learning algorithm ML_SHEV builds upon the
analytical optimization model presented in [1] by Barsali, et
al.. The following subsections give a brief overview the
model and our machine learning process that generates the
optimal power settings for various roadways and traffic
congestion levels.
A.

An analytical optimization model for SHEV

Figure 2. Envelop of SFC curves and its quadratic fitting curve.

Power management in a series HEV considers a drive cycle as


a sequence of on-off cycles of dc sources defined as follows.
An on-off cycle of the dc source has a time period of Tc,
during which engine is in operation during time in [0, Tr ] to
charge the battery up to a maximum level, and is off during
time [Tr , Tc]. During the engine time off, battery power is
discharged to supply the electric drive until a minimum level
is reached. The non-operating time is Tc-Tr. Our optimization
algorithm attempts to make the engine stay at its best fuel
efficiency points while minimizing the battery losses.

Let the coefficients of the quadratic curve be,


envelop of the SFC curves is obtained as follows:

0 <1

( 2)

C1 = C0 + aPs + bPs2 .
The total fuel consumption in a given time T has the
expression

C (T ) = N su C0Tr + a Ps (t )dt +b Ps2 (t )dt + Csu (3)


Tr
Tr

In [1] , Barsali, et al. proposed an analytical model for


optimizing the power flow in a SHEV. The optimization
model was developed based on the following hypothesis: the
power supplied by the dc source Ps(t) can be predicted based
on an average value Psa and a time-varying term proportional
to the instantaneous difference between the current drive
power demand and its average value Pda.

Ps (t ) = Psa + ( Pd (t ) Pda )

, a, b, the

where Nsu=T/Tc is the total number of start-ups in T, Csu is a


fixed amount fuel for each start-up. In [1], if we consider
as a constant during a drive cycle, combining the equations (1)
and (3) we have
C(T ) = T

(1)

Tc
Rb 2
Csu
Tr
2
2 Tc
2
C0 + aPda + a 2 Psa 2Psa Pda + Req + Pda + bPsa + ( 4)
Tr
Tc
Tr
Tr
Vb

where Rb is battery internal resistance and Vb is battery open


circuit voltage and Req is equivalent ripple [1].

It is worth noting here that the dc source is comprised of the


generator and the prime mover. In this paper we focus on the
dc source that contains an engine and a generator. The
optimization algorithm is to find out the optimal values of Psa
and .
The optimization is based on the specific fuel consumption
(SFC) curves of the engine, which describes the fuel efficiency
of an engine with respect to a mechanical output. The SFC
measures the mass of fuel needed to provide a given power for
a given period. In metric units this is measured in kilograms
per kilo-watt hour, kg/kWh. The SFC curves over various
engine speed can be derived directly from the fuel map and the

Then in (4) only Psa is the control variable. The optimization


becomes solving the following equation:

C (T )
=T
Psa

916

Pda +

Rb 2
Req + Pda2
Vb2
K0

) 1 2 VR

b
2
b

( Psa Pda )
K0

Csu

K1 + E K2 + K3 = 0

(5)

where

K 0 = Psa

Step 5. Calculate battery energy level variation


the following formula

Rb 2
( Psa 2 Psa Pda )
Vb2

( 6)

R
R
K1 = C0 + ( a b2 + b) Psa2 2a b2 Psa Pda
Vb
Vb

K 2 = 2Psa Pda

K 3 = 2a
B.

Rb 2
2
Req + (Psa Pda ) + Psa2 2 Psa Pda
2
Vb

Rb
( Psa Pda ) + 2bPsa
Vb2

=(SOCmax-SOCmin)*Battery_Capacity.

Step 6: Apply the LOS drive cycle to vehicle model V and get
the power demands for the entire drive cycle DC.
Step 7: Calculate the average value of the power requested by
the drivetrain, Pda , in the entire drive cycle time T and
calculate the equivalent ripple Req by

(7 )

E by using

Req =

(8)

1
( Pd (t ) Pda ) 2 dt
T T

(10)

Step 8: Calculation Psa from the derivative of C(T) equation.

(9)

The results of the machine learning algorithm is presented in


Table I.

Machine Learning of Optimal Control in a SHEV

TABLE I.

We model the road environment of a driving trip as a


sequence of different roadway types such as local, freeway,
arterial/collector, etc. augmented with different traffic
congestion levels. Sierra Research Inc. has shown that fuel
efficiency and emissions are connected to roadway types as
well as traffic congestion levels. They developed a set of 11
standard drive cycles presented in [2][11] called facilityspecific (FS) cycles, to represent passenger car and light truck
operations over a broad range of facilities and congestion
levels in urban areas. The 11 drive cycles are divided into four
categories of roadway types, freeway, freeway ramp, arterial,
and local. The two categories, freeway and arterial are further
divided into subcategories based on a qualitative measure
called level of service (LOS) that describes operational
conditions within a traffic stream based on speed and travel
time, freedom to maneuver, traffic interruptions, comfort, and
convenience. Six types of LOS are defined with labels, A
through F, with LOS A representing the best operating
conditions and LOS F the worst. Each level of service
represents a range of operating conditions and the drivers
perception of those conditions; however safety is not included
in the measures that establish service levels [5][11] In this
research we use this set of 11 FS cycles as the standard
measure of roadway types and traffic congestion levels. For
the convenience of description we label these 11 FS cycles as
RT1, ., RT11. The proposed machine learning algorithm is to
find the optimal control value Psa for every LOS drive cycle.
The major computational steps are described as follows.

III.

THE RESULTS OF THE MACHINE LEARNING ALGORITHM


Drive
Cycle
Freeway_A

Psa (kW)

Freeway_B
Freeway_C

21
20.81

Freeway_D

20.25

Freeway_E

15.04

Freeway_F
Ramp

9.3
8.42

22.29

ART_AB

9.5

ART_CD

9.7

ART_EF
Local

9.8
9.8

IPC_SHEV: AN ONLINE POWER CONTROLLER FOR A


SHEV

The online intelligent power controller, IPC_SHV


dynamically predicts the roadway type and traffic congestion
level using the neural network, NN_RT&TC. The output of
NN_RT&TC is one of the 11 LOS drive cycles that best
matches the current driving environment. The optimal Psa for
the predicted LOS is used to derive the optimal Ps. The online
SHEV power control algorithm, IPC_SHEV, is described as
follows.
IPC_SHEV algorithm
Step 1: Load in the optimal Psa's of 11 LOS drive cycles RT1,

ML_SHEV algorithm
Step 1: Build a vehicle model V using a high fidelity vehicle
Simulation software
Step 2: Generate the specific fuel consumption Curves for a
broad range of engine speed using V
Step 3; Calculate coefficients of the quadratic SFC curve, C0, a
and b by the least MSE method.
Step 4: Get battery parameters, battery open circuit voltage Vb
and battery internal resistant Rb, from vehicle model V.

., RT11, denote them as

Psai , i=1,11.

Step 2: At any time t during a real time drive cycle, obtain the
vehicle state, V_state (t) = {vs(t), Pd(t), SOC(t)}.
Step 3: Predict the current roadway type by calling the neural
network NN_RT&TC with V_state (t) as input and denote its
output as R(t), which in an index to the LOS's shown in Table
I.

917

Step 4: Let the optimal

TABLE II.

Psa (t ) be PsaR (t ) , which is the optimal

Psa associated with the predicted road type R(t).


Step 5: Calculate the optimal DC source power Ps by the
following formula,

Ps (t ) = PsaR (t ) (t ) + ( Pd (t ) Pda (t ))

(11)

where Pda(t) is the average value of Pd in the filtering window


[t-Tf, t], Tf > 0, and is a constant, 0< <1.
Step 6: Send Ps(t) to the engine controller in the vehicle
system.
Step 7 Continue the process at the next time step by going to
step 2.
IV.

EXPERIMENT RESULTS

We have completely implemented the machine learning


algorithm as well as the online controller, IPC_SHEV. The
vehile used in the following experiments is a Ford Explorer
series hybrid electiric vehicle built using the PSAT software.
The vehicle engine is a 4 liter 120 kW gasoline engine, the
propulsoin motor is a permemnaent magnet Unique Mobility
motor that has a peak power of 75 kW and is scaled to 170
kW. The generator has a peak and continous power of 35 kW
and is scaled to 110 kW. The battery pack is an Ovonic NiMH
28 Ah, 340 nominal volt battery. Figure 3 shows the engine
efficiency map in the Explorer vehicle. Other parameters used
in the simulation runs are shown in Table II.

VEHICLE PARAMETERS USED IN SIMULATION RUNS

Vehicle Total Mass

2508 kg

Frontal Surface Area

2.46 m2

Aerodynamic Resistance Factor

.41

Rolling friction Factor

.009

ICE Consumtion Curve: Co

352.99 g/h

ICE Consumtion Curve: a

-20.11 g/kWh

ICE Consumtion Curve: b

3.35 g/kWh/kW

Start-up Consumption: Csu

50g

Battery Open Circuit Voltage @SOC=70%

340 V

Battery Internal Resistance @SOC=70%

292.4 m

Battery Capacity

9.52 kWh

SOCmin Desired

60%

SOCmax

70%

The intelligent online power controller, IPC_SHEV is


implemented with Simulink inside the Ford Explorer in PSAT.
As an example, Figure 4. shows the roadway prediction results
generated by the neural network inside IPC_SHEV,
NN_RT&TC, on the drive cycle UDDS.

Figure 4. Roadway predictions made by the neural network in IPC_SHEV


on UDDS cycle

Simulation runs have been conducted on four benchmark drive


cycles, UDDS, HWFET, New York, and JPN 10-15. In these
experiments, the online controller IPC_SHEV used Tf = 300
seconds as the filtering window, the weight =0.21. In the
final paper, we will present more experiment results generated
by various parameter settings.
Figure 5 and Figure 6 show the Psa and Pda generated by the
IPC_SHEV for seven UDDS drive cycles. For the purpose of
comparison, we use the controller built in the SHEV model in
PSAT as the benchmark. Figure 7 illustrates the SOC
generated by the benchmark controller in the PSAT for the
UDDS drive cycle, and Figure 8 by the IPC_SHEV.

Figure 3. Efficiency Map of Ford Explorer Engine

918

Pda for UDDS

SOC generated by IPC SHEV

12

0.74
0.72

10

0.7
0.68

SOC

Power (kW)

0.66
0.64
0.62
0.6

0.58
0

1000

2000

3000

4000
5000
Time (Sec)

6000

7000

8000

9000

Figure 5. Pda generated by IPC_SHEV for drive cycle UDDS

2000

3000

4000 5000 6000


Time (sec)

7000

8000

9000 10000

Figure 8. SOC generated by the IPC_SHEV controller throughout the drive


cycle UDDS

Two sets of experiments were conducted to study the fuel


economies. Table III shows the results generated using the
SOC range of 60%-70%, Table IV shows the results
generated using a wider SOC range, 60%-80%.
The
"Algorithm Results" are the results generated by the offline
machine learning algorithm, which uses the knowledge of the
entire drive cycle to derive the optimal Psa, results shown
under "IPC_SHEV" are those generated by online IPC_SHEV
described in section III. The "Benchmark Controller" is the
controller built in the SHEV model in PSAT.
UDDS is mostly comprised of a local LOS. The HWFET
cycle is mostly comprised of a variation of the highway LOS
cycles. The New York cycles is mostly comprised of local
LOS with a few instances of arterial LOS. JPN 10-15 is
mostly comprised of local LOS with variants of arterial and
HWY LOS. For all four drive cycles, the IPC_SHEV
controller consumed less fuel than the benchmark controller in
both SOC ranges. It appears that more fuel economy can be
achieved when the SHEV is operated within the SOC range of
60%~70% except for the UDDS drive cycle. The most fuel
savings by the IPC_SHEV is on the New York cycle.

Figure 6. Ps throughout the drive cycle UDDS


Benchmark SOC
0.74
0.72
0.7

TABLE III.

0.68

SOC

1000

FUEL CONSUMPTION COMPARISON UNDER THE SOC RANGE


60%-70%

0.66

Algorithm Results

IPC_SHEV

Benchmark
Controller
C

g/kWh

g/kWh

g/kWh

0.64

Driving
Schedule

0.62
0.6
0.58

1000

2000

3000

4000 5000 6000


Time (sec)

7000

8000

9000 10000

Figure 7. SOC generated by the benchmark controller throughout the drive


cycle UDDS

919

Mode

sa

kW

s actual

UDDS

on/off

9.56

.21

297

315

321

HWFET

cont.

9.16

.21

258

265

272

New York

on/off

9.25

.21

346

363

398

JPN 10-15

on/off

7.96

.21

302

319

324

TABLE IV.

FUEL CONSUMPTION COMPARISON UNDER THE SOC RANGE


60%-80%
Algorithm Results
C

Benchmark
controller
C

g/kWh

g/kWh

g/kWh

.21

299

317

324

9.59

.21

260

269

273

on/off

9.85

.21

339

351

362

on/off

8.88

.21

303

320

324

Driving
Schedule

Mode

UDDS

on/off

9.83

HWFET

cont.

New York
JPN 10-15

sa

s actual

kW

IPC_SHEV
s

[5]

Highway Capacity Manual 2000, Transportation Res. Board, Wash., DC,


2000.
[6] Nashat Jalil, Naim A. Kheir, Mutasim Salman, A Rule-Based Energy
Management Strategy for a Series Hybrid Vehicle, Proceedings of the
American Control Conference, 1997.
[7] C.-C. Lin, Z. Filipi, L. Louca, H. Peng, D. Assanis and J. Stein,
Modelling and control of a medium-duty hybrid electric truck , Int. J.
of Heavy Vehicle Systems, Vol. 11, Nos. 3/4, pp. 349370, 2004
[8] W. Lhomme, A. Bouscayrol, P. Barrade, Simulation of a Series Hybrid
Electric Vehicle based on Energetic Macroscopic Representation, IEEE
International Symposium on Industrial Electronics, 2004.
[9] Matthew Alan Merkle, Variable Bus Voltage Modeling for Series
Hybrid Electric Vehicle Simulation, Thesis in Electrical Engineering
Virginia Polytechnic Institute and State University, Blacksburg, Virginia,
1997.
[10] Yi L. Murphey, ZhiHang Chen, Leonidas Kiliaris, Jungme Park, Ming
Kuang, Abul Masrur, Anthony Phillips, Neural Learning of Predicting
Driving Environment, Proceedings of International Joint Conference on
Neural networks, June, 2008
[11] Sierra Research, SCF Improvement Cycle Development, Sierra
Report No. SR2003-06-02, 2003.

In these above two tables, no post processing SOC correction


is needed. Since the specific fuel consumption is used for
comparison, than it is automatically accounted for. The total
fuel consumption at the end of the cycle is used in the
calculation and also the total engine energy is used, so the
energy and fuel that is used for calculation directly reflects the
state of SOC.

V.

CONCLUSION

We have presented an intelligent power controller,


IPC_SHEV, for real-time power control in a SHEV. The
IPC_SHEV uses a neural network to dynamically predict
roadway types and traffic congestion levels, and applies the
optimal control parameter value associated with the predicted
roadway type. The optimal control parameter values are found
via a machine learning algorithm,
ML_SHEV.
Our results show that the IPC_SHEV improved fuel
consumptions for four cycles.

REFERENCES
[1]

[2]

[3]

[4]

Stefano Barsali, Carmine Miulli, and Andrea Possenti, A Control


Strategy to Minimize Fuel Consumption of Series Hybrid Electric
Vehicles, IEEE TRANSACTIONS ON ENERGY CONVERSION,
VOL. 19, NO. 1, MARCH 2004.
T. R. Carlson and R. C. Austin, Development of speed correction
cycles, Sierra Research, Inc., Sacramento, CA, Report SR97-04-01,
1997.
Morten Hemmingsson, A Powerow Control Strategy to Minimize
Energy Losses in Hybrid Electric Vehicles, Thesis to Department of
Industrial Electrical Engineering and Automation, Lund University,
Sweden, 1999.
B. He and M. Yang, Optimisation-based energy management of series
hybrid vehicles considering transient behavior, Int. J. Alternative
Propulsion, Vol. 1, No. 1, 2006.

920

Das könnte Ihnen auch gefallen