Professional Documents
Culture Documents
without focus on the parametrisation time. However, the 2.1 Battery experiments
speed of the battery model identification process can be cru-
cial as well, particularly in online applications. This topic has As a case study, a six-cell pack of NiMH batteries was tested
been touched by Hu et al. (2011b) where a control-oriented using a low-cost test bench that was proposed by Propp et al.
approach was used for battery model identification using sub- (2015). The NiMH battery pack was selected due to its sim-
space method. Based on the previous results in the literature, ple and save handling as well as its convenient output volt-
it is clear that the battery model parametrisation’s accuracy age. Specifications of the battery pack are listed in Table 1.
and speed depend on the model’s structure and fitting algo- The experiment was conducted at 25°C by applying consecu-
rithm both and a trade-off is needed between accuracy and tive discharge current pulses to the battery and measuring the
simplicity. The main contribution of this study which distin- battery’s terminal voltage. Data is saved in time domain with
guishes it from the previous works is an “application- a sampling rate of one second. Fig. 2 illustrates the battery
oriented” view on this trade-off. In this study, new discus- measurements during an experiment. The test started from
sions are presented which prove that the final target (the final fully charged state (8.5 V) and continued until the terminal
application) of the battery parametrisation problem can sig- voltage dropped below the cut-off voltage (6 V) which means
nificantly affect the choice of fitting algorithm or model depleted charge state. The discharge rate is 1C that is 2.4A
structure. This application-oriented point of view helps to and length of each pulse is 40 seconds with a relaxation time
prevent any unnecessary complexity while achieving the of 60 second in between.
specified targets as simply as possible.
Table 1. NiMH battery pack specifications
For this purpose, the battery model parametrisation problem
is analysed using a different visual approach firstly by plot- Parameter Value
ting battery parametrisation surfaces. Although the surfaces Rated capacity per cell 2400 mAh
seem simple, they contain quite useful information which is No. of cells 6
discussed and utilised in this study. Three standard fitting Rated voltage 7.2 V
algorithms are used and analysed including gradient descent Full-Charged voltage 8.5 V
(GD), genetic algorithm (GA) and prediction error minimisa- Cut-off voltage 6V
tion (PEM). There are many other algorithms in the literature
and this paper does not aim at reviewing them. GA and PEM
are selected just because they are two standard techniques in
the literature. The GD algorithm is selected to demonstrate
that simpler algorithms may also be applicable for battery
parametrisation. For each level of complexity, the algorithms’
performance is investigated using experimental data from a
small NiMH battery pack. The results of this study are not
limited to NiMH battery chemistry and the proposed contri-
bution can be utilized in other applications as well.
Fig. 1. Online battery state estimation based on parameter Equivalent circuit network (ECN) battery model structures
identification are used in this study. The ECN battery models are construct-
ed by putting resistors, capacitors and voltage sources in a
2. BATTERY MODEL IDENTIFICATION circuit. The simplest form of an ECN battery model is inter-
nal resistance model (R model). The R model includes an
Generally, a system identification procedure consists of three ideal voltage source ( VOC ) and a resistance ( RO ) as depicted
main parts, Ljung (1987), (i) experiment design, (ii) model
in Fig. 3(a) in which Vt is the battery terminal voltage and I L
structure selection, and (iii) fitness criterion selection. The
same parts are considered for battery model identification in is the load current. Adding one RC network to the R model
this study. Different model structures and data fitting algo- increases its accuracy by considering the battery polarisation
rithms are used. The algorithms’ computational effort and characteristics as discussed by Salameh et al. (1992). This
precision have been assessed using experimental data for model, called Thevenin Model (1RC model), is illustrated in
different model structures. Fig. 3(b) in which Vt is cell’s terminal voltage, VOC is open
circuit voltage (OCV), RO is internal resistance, RP and CP
are equivalent polarisation resistance and capacitance respec-
51
IFAC AAC 2016
50
June 19-23, 2016. Norrköping, Sweden Abbas Fotouhi et al. / IFAC-PapersOnLine 49-11 (2016) 048–053
tively. Adding more RC networks to the battery model (such this purpose, a fitness function, i.e. the root mean square error
as 2RC in Fig. 3(c)) may improve its accuracy but it increases (RMSE) function, is used as follows:
the complexity too. Other elements can be added to the model 1
52
IFAC AAC 2016
June 19-23, 2016. Norrköping, Sweden Abbas Fotouhi et al. / IFAC-PapersOnLine 49-11 (2016) 048–053 51
Fig. 4. Battery model identification surface (R model) at (a) Fig. 5. Identified values of battery OCV and resistance vs.
10%, (b) 50% and (c) 90% SOC SOC using GA, PEM and GD algorithms
53
IFAC AAC 2016
52
June 19-23, 2016. Norrköping, Sweden Abbas Fotouhi et al. / IFAC-PapersOnLine 49-11 (2016) 048–053
54
IFAC AAC 2016
June 19-23, 2016. Norrköping, Sweden Abbas Fotouhi et al. / IFAC-PapersOnLine 49-11 (2016) 048–053 53
fication algorithm selection is the simplicity of programming are mainly usable in predictive models rather than descriptive
and being easily embeddable. Regarding this aspect, GD al- models. A good example of the predictive modelling applica-
gorithm is the simplest one that can be easily programmed tion is real-time estimation of battery SOC and/or SOH. The
and embedded comparing to PEM and GA. This advantage is choice of the parametrisation algorithm and the model struc-
more important for real-time application of the algorithm on ture depends on the required accuracy, required speed, com-
less powerful electronic boards. This also confirms im- putational effort and hardware limitations, etc. For the pre-
portance of the contribution of this study, which is applica- sented case, a NiMH battery, SOC estimation is possible just
tion-oriented trade-off between accuracy and simplicity. by using OCV which is available using the simplest model
structure that is R model. Both PEM and GD algorithms are
4.3 Study of the trade-off between accuracy and simplicity of suggested in this case depending on the computational effort
the battery model structures and hardware limitations. Considering another battery type,
Lithium-Sulfur for example, OCV is not enough for SOC
In addition to the identification algorithm selection, the bat- estimation and other parameters are required as well. Same
tery model structure can also be selected using an applica- discussions can be presented for battery SOH estimation
tion-oriented approach. In order to address this, we have fo- which is another application for battery model parametrisa-
cused on this question that how the identification algorithms tion.
can handle an additional complexity to the model? Here, the
battery model’s complexity is added gradually by putting REFERENCES
more RC networks in the model which increases the number
Brand, J., Zhang, Z., Agarwal, R.K. (2014). Extraction of
of unknown parameters from the identification point of view.
battery parameters of the equivalent circuit model using
Considering the results presented in Fig. 6 and Fig. 7, it is
a multi-objective genetic algorithm. Journal of Power
concluded that GD algorithm’s speed decreases significantly
Sources, 247, 729-737.
by adding model’s complexity while keeping the accuracy in
Fotouhi, A., Propp, K., Auger, D.J. (2015). Electric Vehicle
a comparative level. In GD algorithm, the identification time
Battery Model Identification and State of Charge Esti-
that is needed for a battery model with k unknown parame-
mation in Real World Driving Cycles, 7th Computer
ters is proportional to 2k . For example, the time needed to Science and Electronic Engineering Conference, UK.
identify 1RC model (which has 4 parameters) is 4 times more Fotouhi, A., Auger, D.J., Propp, K., Longo, S., Wild M.
than R model (which has 2 parameters), and the time needed (2016). A review on electric vehicle battery modelling:
to identify 2RC model (which has 6 parameters) is 16 times From Lithium-ion toward Lithium–Sulphur, Renewable
more than R model and so on. Therefore, GD method is not and Sustainable Energy Reviews, 56, 1008-1021.
suggested for very complex models however it is a good Goldberg, D.E. (1989). Genetic Algorithms in Search, Opti-
choice when a simple and easily embeddable identification mization & Machine Learning, Addison-Wesley.
algorithm is needed for online applications. Hu, X., Li, S., Peng, H. (2012). A comparative study of
On the other hand, GA algorithm may not be successful in a equivalent circuit models for Li-ion batteries. Journal of
short period of time however the complexity can be handled Power Sources, 198, 359-367.
if there is no time limitation such as in offline applications. In Hu, Y., Yurkovich, S., Guezennec, Y., Yurkovich, B.J.
this case, GA algorithm outperforms GD because of its addi- (2011a). Electro-thermal battery model identification for
tional capabilities like jumping out of local minima. So, GA automotive applications. Journal of Power Sources, 196,
method is suggested to be used for very complex models 449-457.
when time is not an important factor. In addition, GA algo- Hu, Y., Yurkovich, S. (2011b). Linear parameter varying
rithm is difficult to be embedded because of its programming battery model identification using subspace methods.
complexity. Referring to PEM method’s results in Table 3 Journal of Power Sources, 196, 2913-2923.
and Fig. 8, there is an improvement in RMSE by adding the Ljung, L. (1987). System Identification - Theory for the User,
first RC network. However, this achievement is obtained by Prentice Hall, New York.
increasing the identification time to four times more. Adding Pattipati, B., Sankavaram, C., Pattipati, K.R. (2011). System
the second RC network causes a small improvement of error Identification and Estimation Framework for Pivotal
while the identification time approximately doubles. So, it is Automotive Battery Management System Characteris-
concluded that 2RC battery model structure is not suggested tics. IEEE Trans. Syst. Man Cybern. C: Appl. Rev.,
in this case. All in all, PEM method is suggested when bat- 41(6), 869-884.
tery model’s complexity is needed in online applications Propp, K., Fotouhi, A., Auger, D.J. (2015). Low-Cost Pro-
however; its embeddable program is not as simple as GDs. grammable Battery Dischargers and Application in Bat-
tery Model Identification, 7th Computer Science and
Electronic Engineering Conference, UK.
5. CONCLUSION
Salameh, Z.M., Casacca, M.A., Lynch, W.A. (1992). A
In this study, a new approach to the battery parametrisation mathematical model for lead-acid batteries. IEEE
problem was presented which is called application-oriented Transactions on Energy Conversion, 7(1), 93-98.
trade-off between accuracy and simplicity. The main ad- Snyman, J.A. (2005). Practical Mathematical Optimization:
vantage of this approach is avoiding unnecessary complexi- An Introduction to Basic Optimization Theory and Clas-
ties by using the most efficient model structure and para- sical and New Gradient-Based Algorithms, Springer
metrisation algorithm in each case. The results of this study Publishing.
55