'..fic-!,d NetLrd Ncv.mrk coaa-olaad hi 42 V DC ULr...
GUALOUS Hwud
cicr.y m,,uikgcincnt
Mil3.clid NCLLEa1 Nktwork conuol and eneqy rnaaa.cincnt iii 42 V DC link GUALOUS Hamid
Artificial Neural Network control and energy management in
42 V DC link
J. N. MARWE-FRANCOISE, H. GUALOUS, A. BERTHON
LABORATORY OF ELECTRICAL ENGINEERING AND SYSTEMS (L2ES)
Joint Research Unit UFC-UTBM EA3898, rue Thierry Mieg
Belfort, France
Tel.: +33 (0)3 84 58 36 00
Fax: +33 (0)3 84 58 36 36
E-Mail: hamid.gualousauniv-fcomte.fr, alain.berthon@univ-fcomte.fr
URL: http://l2es.utbm.fr
Keywords
Neuronal control, Energy storage, Automotive application, DC power supply, Energy system
management.
Abstract
This paper deals with an experimental realization of a 42V hybrid power sources for automotive
applications. It's composed by a battery which provides the power in constant mean power and a
supercapacitor tank in order to supply power in tnsient state. Two DC/DC converters are used to adapt
voltage and current levels between the 42V DC link, battery and supercapacitor tank. Voltage is regulated
by using Artificial Neural Networks (ANNs).
Introduction
Supercapacitors have a high power density and low internal resistance. They can be charged and
discharged at high current. They are ideally placed for peak power requirements. They can be used in
panllel with battery to start the internal combustion engine in order to reduce the size of the battery, in
hybrid vehicle to provide power during acceleration with the aim to reduce the power of the internal
combustion engine.
The study presented here is about experimental realization of a hybrid power source and its voltage
control by ANNs. These last are used in several applications. They are introduced as an alternative of
analytical model for non-linear systems when the formulation of the mathematical model is very difficult.
There principle is based on programmed computing approach. ANNs are used in a wide range of
engineering and non-engineering applications (identification, modelling, control...) [1-7]. They are
inspired from the human brain and can be considered like a black box where the system outputs are
predicted as a response of inputs combination without any physical relationship between outputs and
inputs. ANNs have the advantage to be able to perform non-linear mapping of multi-dimensional
Functions, and to predict outputs system from limited training expenmental data. ANN can determine the
output parameters outside their range of training experience. They have a high capacity to generate output
with high uncertainly input level, or with a perturbation noise data.
The high capability of ANN prediction is determined by the structure (number of the hidden neuron) and
by the learning process.
LOB 2005 -- L)i--;c11i
Dn.Jcii
0.1
DEl Own5 ISBSN :9(U1548tf-5 I,. I
1SB1.l
'..fic-!,d NetLrd N cL.mrh. coaa-ol aad nL"tgcnlm. in 42 V DC link GUALOUS Hwud
AAcilicial NCLLEa1 Neuworkanuci and enefyy
cncrgy nianagcnwn: in 42 V DC link GUALOUS 1-lamid
Structure of the hybrid power source and ANNs control
Design and ANN control
Figure 1 presents the system proposed. It consists of a 12V battery and a pack of supercapacitors
composed with 8 cells of 2600F in series. This last pack is sized to provide about lkW during transient
behaviour, while the battery gives the constant mean power requirement.
42V DC rail
Vet-
Vsc -
hysteresis
control
Fig. 1: 42V hybrid power source for automotive applications
Supercapacitors energy is used as auxiliary source in order to feed DC bus during the load current
an
peaks. When there is no current fluctuation, battery (12V) provides the voltage level on the DC bus (42V).
At the same time, the supercapacitors are charged with constant current until 20V. The reversible DC/DC
converter operates in buck mode. When there is a current fluctuation, the battery stops providing energy
and the voltage level on the DC bus is provided by the supercapacitors pack. These last are charged and
discharged between 50% and 100 % of the maximum voltage (20V). In this case the converter operates in
boost mode.
The control voltage of the DC/DC converters between battery and DC bus and between supercapacitors
and DC bus is realized by a Artificial Neural Network (ANN) regulator. For the buck converter between
the DC bus and supercapacitors; which is used to charge the pack at constant current; the current
regulation is based on a hysteresis with constant frequency
The ordered system by the neural network is composed by a source ( battery, supercapacitor, fuel cell), a
boost DC/DC converter and the load (DC bus) (cf. Figure 1).
DEl Own5
LOB 2005 -- L)i--;c11i
Dn>tii ISBSN: 9X-75815-08-5
.J-75815-O8-5 0.21i9.2
'..fic-!,d NetLrd Ncv.mrk coaa-olaad m,"tgement in 42 v DC liak GUALOUS Hwud
Arcificlid NCLLEa1 Nktwork ecanci and c=ryy
cncr.y manalcinent in 42 V DC link GUALOUS MaccUd
This approach has to learn the network, to reproduce the control signal u(t) recommended by the first
controller, starting from the desired output y(t). The network training is done in parallel with PI controller.
Therefore, the system is initially controlled by a taditional commandin order to obtain a data base.
The PI controller's behaviour is reproduced by training using the hypothesis model NARX.
Just the traditional command, the neural network's inputs (cf. Figure 1) are the measurement of the
as
output voltage and the input current of the converter. For output, the control signal is deduced from the
comparison of carrying frequency (fe) and a modulating wave (duty cycle), calculated by the neural
network.
The hypothesis model reproduces accurately the controller's behaviour, according to the following
equation:
a (k+1) =f [a (k),,...,a (k-nl+l); Ve (k+l),...,Ve (k-n±+1); le (k+1),...,Ie (k-n3+l)] (1)
where:
f is non map; k is discrete time step nl,n2,n3 are the number of past values' variables
linear
corresponding; input current of the chopper; Ve: Supercapacitor or battery supply voltage; a is the
Je:
duty cycle of the PWM signal.
The performance measurement used is the quadratic error:
This method is called direct method by reproduction of the existing controller.h identification domain,
they can be used as black box model. Any system modelling is possible without any knowledge about its
internal working procedure. You only need to have a training process made of input and output vectors
unit, resulting from experiments and representative of the task to be modelled. The quality of training is
very important, if the data is not representative of the relation to be approached, that means, if the
examples are not representative of the application domain, the training will be able to give only a partial
approximation of the problem (under training). In the opposed case, if there is a lot of parameters, the
network will memorize the examples and will be unable to generalize with examples, which do not belong
to the training data (over training). The quality of the training is evaluated in term of generalization
capacity, which is quantified by calculating the average quadratic error (AQE):
2(2
EAP=N A k
E
= 1m
YP ))(
It gives the error made on the whole NAPP examples of the training process between the wished outputs
yp(k) and the output model Ym (k).
The performance of generalization depends on the relevance of the training data used, of the network
parameters number and the state of the training algorithm convergence. Like every identification
procedure, the neuronal model obtained after training must be validated. It must thus be subjected to a
certain number of tests, which will make it valid. The validation has two levels, one conceming the
capacity of generalization and the other the tests of correlation. To detect a bad generalization quality, it is
enough to estimate the performances of the model on one or more tests data which are different from the
training ones. This is done by comparing the average quadratic error test (AQET) of the data test with the
1taining process error Eapp. If the two values are closed, the model has a good capacity of generalization.
MPh 2005 -- L)i--;c11i
Dn>cdcii
P.3
DEl Own5 ISBSN :9(U1548tf-5
lSB1.l X3-75S15-0S-5
i,.3
'..fic-!,d NetLrd Ncv.mrk coaa-olaad hi 42 V DC ULr... GUALOUS Hwud
eicr.y m,,uikgcincnt
Mil3.clid NeLeal Nervork conuol and eneqy rnaaa.eincat iii 42 V DC link GUALOUS Hamid
NTEST= ki YYP(k)Ym(k))
The control system has been implemented using MATLAB/SIMIULINKP' as an analytic model (source and
DC/DC converter). Simulation results show the adaptation between the voltage of a supercapacitor tank
using DC/DC converter and a 42V DC link.
Fack of Suner apa hom --LCWE VoIt:'j
34 T- - - - - - - - -
-T - - - - - - - - -, - - - - - - - - - - - - - - - - - -- _-_______
7> 3-l - - - - - - - -- -- - - - - -- -- - r -- -- - - m-- - - - - --
--
---- -- r - - - -- -- -- -
1-1 I-
-100 1 201f 4 0 DICI 6Fi111] 71 113
Ch;rg eiciDI ha ge ci reer-it -
1-1
-LI 100 201] 3 -L- 4 0 tOLl B1-II] 71 1IJ
Tin--ne(is)
Fig. 2: simulation results
Power management
The power management of the system is realised by an algorithm procedure. The principle operation is as
follows. In the algorithm, a threshold current is fixed in the steady state ('pci), the minimum and the
mmum supercapaciotrs pack voltage levels re fixed respectively at Vsciin and Vsn-x. In the first, when
the system is on, the pack of supereapacitors voltage V. is tested. If V,, is less than V,,,, supercapacitors
are charged until Vscmax. After that, if the current demad at the DC bus is high higher than Iper (transient
state), battery is disconnected, the power requirement is given by the pack of supercapacitors. In the study
state, the battery provides the power requirement ad the supercapcitors pack is charged.
If the current demand is high than Iper and the VscVj, the system can not provide the power demand.
Supereapacitors are charged form from the battery.
Experimental results
The experimental test of the 42V Power Net for automotive electrical system is realized (fig. 3).
hUE
P.4
DEl 2005 -- L)i--;c11i ISBSN :9(U1548tf-5 i,.4
ISBN
'..fic-!,d NetLrd Ncv.mrk coaa-olaad hi 42 V DC ULr... GUALOUS Hwud
cicr.y m,,uikgcincnt
Mil3.clid NCLLEa1 Nktwork conuol and eneqy rnaaa.cincnt iii 42 V DC link GUALOUS Hamid
The battery (12V, 105 Ah), DC bus is loaded by variable resistors. L = 50jH input inductor and C = 10
mF output capacitance.
The control and the power management of the system are realized by the Siemens SAB167
microcontroller with CAN bus to obtain the database . Of these results, the control system has been
implemented using MATLAB / SMUILLNK * program package ad its Real-Time-Workshop and DSpace
software. For the control and the power management, different experimental values are necessary: the
supercapacitors current and voltage, the load current at the DC bus and its voltage. The measured voltage
value at the DC/DC converters output is compared with a voltage reference, which is equal to the DC bus
voltage (42V). The generated error is corrected by a PI regulator in the two cases (when supercapacitors or
battery provide the power requirement). The PI compensator and the hysteretic current regulation
described previously are realized by the microcontroller. Currents and voltages are measured by LEM
sensors ad monitored by the microcontroller.
Filtarge capacitor
(C = 10iF; V =
Self (L 0.6inH ; R - M32Q;
=
=,_ 1SOA) Resistive load
T.!
_; \\;-Iil\\v;-^i
X uI'\ll;t ...................... i;
battery
(V= 12V; Q= .GBT: SKM 400GB
124D
Electronic control and sensors Supereapacitor pack
processing (C 32SF; Vmax 20V ; Pmax
= = =
1)LAAf
Fig. 3 : Experimental test bench
LOB
Its
DEl 2005 -- L)i--;c11i
Dn.Jcii ISBSN :9(U1548tf-5 I,.f
1SB1.l X3-75S15-0S-5
icial N cund Network oma-ol xid I 42 V DC La& G UALOLIS Haa-md
Artificial Neural Network control and energy
energy mmutgcmca.
management in 42 V PC link GUALOUS 1-lantid
In fig. 4, we have plotted the first expenmental results obtained by using ANNs control. This figure shows
the DC link voltage evolution as a function of time for three cases.
The first curve in blue is realized by a PI controller implemented in a microcontroller, the second in green
obtained by ANN control which uses experimental data for ANN training process and the last, is obtained
with ANN control which uses simulation data for ANN training. The ANN regulator are implemented in
MATLAB/SIMULINK and DSpace system. These results show that the DC link is regulated at 42V. At
the load current commutation, the regulated DC link voltage with a PI controller presents a higher peak
voltage than the ANN regulation. In the final paper, these results wil be analyzed and the other
experimental results with ANN energy management will be presented.
Voltage (V), load current (A)
60
60
-".
A
40 ---- -i
30
VbusIccnmnndeR)
VtscrI It4sri
- I
vius (oommande rF imu)
.
20
10
0D 60D 1ir 200 0o 3E!
Time (s)
Fig. 4: Experimental results obtained with ANN for DC bus control
The experimental test of the 42V Power Net for automotive electrical system is realized in the same
configuration of the simulation conditions. In figure 5, the evolution of the supercapacitors current and
voltage and the DC bus voltage as a function of time has been plotted.
During the control of the process, the behaviour ofthe system is described as follows:
0 t E [17;350]: supercapacitors are charging under constant current (IOA), The DC/DC converter
between supercapacitors and DC bus operates in buck mode. The DC bus voltage is controlled at 42V
from the battery.
lilt 20O5
Ell- 2005 -- Dredei
Etraden ISBN : 1)-758t5,-0&-l:
P.6
i,.6
'..fic-!,d NetLrd Ncv.mrk coaa-olaad hi 42 V DC ULr... GUALOUS Hwuid
cicr.y m,,uikgcincnt
Mil3.clid NCLLEa1 Nktwork conuol and eneqy rnaaa.cincnt iii 42 V DC link GUALOUS Hamid
Iot?' i6Lij!:E FA c:k of FS iJFiOe r c :ri rC:Itrs
4-A
0A e r rI:it : it c ~e
LI-
IO 7 -O 4 II III 6--il 0 CO 91 -I I
CIO I7 ~ IDC -03 D ID 7 1- I 8 9 IDC IIC
Ti mcnin (E-;)
Fig. 5 Experimental results of the 42V power Net for automotive electrical system applications
t the supercapacitors are charged, then starts to discharge when the current on the
cz [350;504] :
DC bus varies. They discharge with various current values. The DC bus voltage remains constant to 42V
using supercapacitors energy adthe DC/DC converter operates in boost mode.
t c [5 10;730] : the supercapacitors are discharged, they have to charge, the DC bus voltage is
again controlled by the battery.
t E- [735;762] The supercapacitors are discharged by a hihcurrent
t e [774;9 14]: The supercapacitors are not entirely charged but there is a current fluctuation on
the DC bus, then, the supercapacitors are solicited in priority to maintain the DC bus voltage constant.
The experimental voltage remains constant to 42V when the current load
results show that the DC bus
varies. However with each change of cycle of the supercapacitors charge and discharge, very fast over
voltages appear (trnmsient effects). This last is due to the fact that: when the supercapacitors are charged,
the battery provides the power to the 42V Power Net. The DC/DC converter between supercapacitors and
the DC bus is controlled by current. When the current load changes, battery is disconnected and
supercapacitors give the power requirement. This commutation causes perturbations to the controller, and
DC bus voltage varies. However, this transient effect is very fast and has no influence in operating system.
12K 2005 - L)iv--;dc,-ii
Pn.Jcii
It'
ISBN :9(.-758tf,-W-5
1SBt. X3-75815-O8-5
Ellh P.7
N.SE. Nm,rL cu..
Amf'bSilt I I . 42 V L ILi HW.d
GUALOUS
Conclusion
As a conclusion this paper shows Artificial Neural Network as a well adapted method to control voltage
on a 42V DC bus and to manage energy transfer between a pack of supercapacitors and batteries. A 2kW
test bench has been realised and comparisons between simulated and experimental results validate the new
methodology presented. It can be shown that higher dynarnic and adaptable regulation can be obtained. It
is noted that the adjustment of the neuronal network is more obvious than that PI correctors. The
performances of the system depend mainly on the algorithm of the energy management. So this new
approach is proposed for auxiliary power unit in automotive applications.
References:
[1] Paulo E. M. Almeida and Marcelo Godoy Sim6es,'Neural Optimal Control of PEM Fuel Cells with
palametric CMAC Networks' IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 41, NO.
1, JANUARY/FEBRUARY (2005) 237.
[2] T. Seijyu, H. Miyazato, S. Yokoda, K. Uezato, Speed Control of Ultrasonic Motols usmig Neural Network,
IEEE transactions on Power Electronics, Vol.13, N°3, May 1998.
[3] E. Lavretsky, N. Hovakimyan. 'Reconstruction of Continuous-Time Dynamics Usmig Delayed Outputs and
Feedforward Neural Networks', IEEE Transactions on Automatic Control, 2002.
[4] N. Hovakirnyan, F. Nardi, N. Kim, A.J. Calise, Adaptive Output Feedback Control of Uncertain Systems
using Single Hidden Layer Neural Networks, IEEE Transactiolns on Neuml Networks, 2001.
[5] R.P. Jones, A.S. Cherry, S.D. Farral, Application of Intelligent Control in Automotive Vehicles. IEE
International Conference on Control. Coventry, UK. Vol.389, pp 159-164, March 1994.
[6] Shaoduan Ou, Luke E.K Achenie 'A hybrid neural network model for PEM Fuel cells ' Journa of power
sources (2004)
[7] Jaime Arriagada, Perrilla Olausson, Azra Selimovic 'Artificial neural network simulator for SOFC'
performance precdiction' Journal of Power Sources 112, 2002, pp. 54-60.
[8] Jet P.H. Shu 'The Development of the Hybrid Propulsion System for the Light-Duty Vehicle Applications'.
International Electric Velicle Symposium and Exposition, EVS-20: Powering Sustainable Transportation.
Long Beach. Califomia (USA). November 15-19; 2003
[9] E.J. Dowgiallo and A.F. Biurke, 'Ultracapacitors for Electric and Hybrid Vehicles'. Electric Vehicle
Conference. Florence, Italy. 1993
[10] J. Lott, Helmut Spath 'Double layer capacitors as additional power source in electric vehicles' 18'h
International Electric Vehicle Symposium and Exhibition. Berlin, Germanuy. 2001, CD ROM.
[11] T. Dietrich 'UltraCaps-A new energy Storage Device for Peak Power Applications', 181' Intemational
Electric Vehicle Symposium and Exhibition. Berlin, Germany. 2001, CD ROM.
[12] L. Bertoni, H. Gualous, D. Bouquain, H., Hissel, D., Pea, M.C., Kauffmann, J.M., "Hybrid auxiliary power
unit (APU) for automotive applications", in Proc. of the IEEE Vehicular Technology VTC'02 Conference,
CD-ROM, ISBN 0-7803-7468-1, Vancouver, Canada, 2002.
kl3H 00 D"Wdc
- ISBN 90758t15X1--5 1'.8
