You are on page 1of 12

Green Energy and Intelligent Transportation 1 (2022) 100029

Contents lists available at ScienceDirect

Green Energy and Intelligent Transportation


journal homepage: www.journals.elsevier.com/green-energy-and-intelligent-transportation

Full length article

A branch current estimation and correction method for a parallel connected


battery system based on dual BP neural networks
Quanqing Yu a, *, Yukun Liu b, Shengwen Long a, Xin Jin a, Junfu Li a, Weixiang Shen c
a
School of Automotive Engineering, Harbin Institute of Technology, Weihai, Shandong 264209, China
b
School of New Energy, Harbin Institute of Technology, Weihai, Shandong 264209, China
c
School of Science, Computing and Engineering Technologies, Swinburne University of Technology, Hawthorn, Victoria 3122, Australia

H I G H L I G H T S G R A P H I C A L A B S T R A C T

 A branch current estimation method for


parallel battery packs is proposed.
 A method to eliminate the current esti-
mation error caused by dynamic excita-
tion is proposed.
 The influence of inputs to the dual neu-
ral networks on the estimated results is
discussed in depth.
 The effectiveness of the proposed
method is validated by a battery pack
with two parallel cells.

A R T I C L E I N F O A B S T R A C T

Keywords: In the actual use of a parallel battery pack in electric vehicles (EVs), current distribution in each branch will be
BP neural Network different due to inconsistence characteristics of each battery cell. If the branch current is approximately calculated
Branch current estimation and correction by the total current of the battery pack divided by the number of the parallel branches, there will be a large error
Electric vehicles
between the calculated branch current and the real branch current. Adding current sensors to measure each
Lithium-ion battery pack
branch current is not practical because of the high cost. Accurate estimation of branch currents can give a safety
warning in time when the parallel batteries of EVs are seriously inconsistent. This paper puts forward a method to
estimate and correct branch currents based on dual back propagation (BP) neural networks. In the proposed
method, one BP neural network is used to estimate branch currents, the other BP neural network is used to reduce
the estimation error cause by current pulse excitations. Furthermore, this paper makes discussions on the selection
of the best inputs for the dual BP neural networks and the adaptability of the method for different battery capacity
and resistence differences. The effectiveness of the proposed method is verified by multiple dynamic conditions of
two cells connected in parallel.

cycle life and light weight. Energy storage power stations are more inclined
to choose large-capacity batteries to be connected in series to provide more
1. Introduction energy, while EVs have high demands on both types of batteries due to
various vehicle types. Most pure EVs adopt the structure of connecting
Lithium-ion batteries are widely used in electric vehicles (EVs) and large-capacity batteries in series due to space constraints, while a small
energy storage power station because of their high energy density, long

* Corresponding author.
E-mail address: qqyu@hit.edu.cn (Q. Yu).

https://doi.org/10.1016/j.geits.2022.100029
Received 12 July 2022; Received in revised form 4 September 2022; Accepted 5 September 2022
Available online 4 October 2022
2773-1537/© 2022 The Author(s). Published by Elsevier Ltd on behalf of Beijing Institute of Technology Press Co., Ltd. This is an open access article under the CC BY
license (http://creativecommons.org/licenses/by/4.0/).
Q. Yu et al. Green Energy and Intelligent Transportation 1 (2022) 100029

Nomenclatures and abbreviations bI 2c Corrected current of branch 2

Nomenclatures Abbreviations
OCV1 Open circuit voltage of branch 1 EVs Electric vehicles
OCV2 Open circuit voltage of branch 2 BP Back propagation
R1 Internal resistances of branch 1 SOC State of charge
R2 Internal resistances of branch 2 OCV Open circuit voltage
I1 Current of branch 1 CC Constant current
I2 Current of branch 2 CV Constant voltage
V Branch voltage CCCV Constant current constant voltage
I Total current of parallel battery pack DST Dynamic stress test
bI 1 Estimated current of branch 1 UDDS Urban dynamometer driving schedule
bI 2 FUDS Federal urban driving schedule
Estimated current of branch 2
RNN Recurrent neural network
Cap Accumulated electricity
LSTM Long short term memory
EOB1 Error of branch 1
P-LiBs Parallel lithium-ion batteries
EOB2 Error of branch 2
ANN Artificial neural network
E OBb 1 Estimated error of branch 1 NN Neural network
E OBb 2 Estimated error of branch 2 MSE Mean square error
bI 1c Corrected current of branch 1 CNN Convolutional neural network

number of pure EVs and most hybrid EVs choose the structure of con- neural network (CNN) to estimate branch current by using terminal
necting small-capacity batteries in parallel and then in series in recent voltage, total current and the accumulated charge/discharge capacity
years. For example, Tesla Model S first connected 74 cells in parallel, and since the beginning of charge/discharge. The experimental results indi-
then connected 96 such structures in series to form a battery system, which cate a good estimation accuracy under CC working conditions. The
is also referred to as 74p96s battery pack. EVs with a similar battery system verification is only carried out when the total current is completely
structure also include Tesla Model 3 (46p96s), Nissan Leaf (2p96s), BAIC constant, and the situation under sudden changes in current, constant
EU260 (3p90s), Renault Zoe (2p96s) and etc [1]. voltage (CV), and dynamic discharging working conditions is not
For such a battery system including parallel lithium-ion batteries (P- considered. In the actual situation, most EVs are driven under dynamic
LiBs), even after rigorous screening of a large number of single cells with conditions. Moreover, there is a lack of in-depth discussion on the se-
differences in capacity and internal resistance caused by the production lection of inputs to the RNN.
process, there is still a certain initial consistency difference between cells. This paper proposes an accurate branch current estimation and
During the charging and discharging process, the battery management correction method under dynamic working conditions based on dual
system (BMS) often regards this structure as a “big battery” and applies a Back propagation (BP) neural networks. The first BP neural network can
total current excitation [2–5]. The initial differences in the internal estimate branch current and the second BP neural network can reduce
resistance, capacity and contact resistance of each single cell lead to un- the estimation error caused by high current pulse excitations. The con-
equal currents in each branch, which in turn affects the charging and tributions of this paper are summarized as follows: (1) one BP neural
discharging depth and heat generation of each single cell. The consistency network is used to accurately estimate the branch current of parallel
between cells gradually deteriorates under cycling conditions, which in battery under CC charging and dynamic working conditions; (2) the
turn exacerbates the branch current distribution differences [6–8]. In estimation result can be corrected based on the other BP neural network
addition, battery cells with poor consistency may experience different under dynamic working conditions; (3) the inputs of the two BP neural
degrees of overcharge and overdischarge, which not only accelerates networks are discussed respectively, and the optimal inputs of the two
battery aging, but also may lead to battery safety hazards [9,10]. neural networks are found; (4) the adaptability of the dual BP algorithm
For P-LiBs, in addition to implementing a more stringent battery for different capacity and resistence differences is discussed.
screening strategy, the solution of adding a current sensor in each branch The remainder of this paper is organized as follows: Section 2 in-
to collect the current of each branch is rarely considered due to the high troduces the branch current estimation and correction scheme. In Section
cost of current collection. There are also few reports on the estimation of 3, the main experimental process is explained. Results and discussions
branch currents for a parallel battery pack. Huang et al. [11] adopted are given in Section 4. Finally, the conclusions are drawn in Section 5.
second-order RC equivalent circuits to model the battery cells connected
in parallel. The branch currents can be estimated accurately under 2. Branch current estimation and correction scheme
Constant Current (CC) working condition and Dynamic Stress Test (DST)
working conditions. However, the parameters of this model are closely Artificial neural network (ANN) has the ability of self-learning, self-
related to the temperature and state of charge (SOC) of the battery, and adaption and nonlinear mapping [14]. BP neural network is one of the
the accuracy of the second-order RC model is poor in some SOC regions. common networks of ANN and has been widely used in all aspects of
Khalid et al. [12] proposed a recursive weighted covariance-based engineering [15,16]. This paper applies BP neural network to estimate
method to estimate the branch current of a parallel battery pack. Re- branch currents and compensate estimation errors in real time. The ar-
sults show that the proposed method can accurately estimate battery chitecture (one hidden layer in this case) of BP neural network is shown
parameters even in the presence of faults and random-noises. However, in Fig. 1, x1, xi, xk and xm represent the inputs. y1, yk and yl represent the
after introducing the relationship between temperature and battery pa- outputs. wi,j and wk,i represent the weights between different neurons.
rameters, it is difficult to establish an accurate battery model when θ1;1 , θ1;2 , θ1;i , θ1;j , θ1;q , θ2;1 , θ2;k and θ2;l represent the bias of neurons. f1
ambient temperature changes. Cui et al. [13] adopted a recurrent neural and f2 represent the activation functions, which calculate how much the
network (RNN) with long short term memory (LSTM) and convolutional neuron is ‘stimulated’ [17]. The Levenberg Marquardt algorithm is

2
Q. Yu et al. Green Energy and Intelligent Transportation 1 (2022) 100029

improvement in the performance of the training set [18]. The goal of


training set mean square error (MSE) is 0.000,001 A2 and the maximum
number of iterations is 100. In such a way, over-fitting can be avoided in
the training process. Then, the trained BP neural network is tested by the
testing set. The estimated branch currents bI 1 and bI 2 are obtained under
CCCV charging working condition.
To estimate branch currents under dynamic working conditions, the
training set includes the experiment data set of V, I1, I2 and I measured
under DST and UDDS, and the testing set includes the experiment data set
of V and I under FUDS. The goal of training set MSE is 0.000,001 A2 and
the maximum number of iterations is 100;
In step 4, the second BP neural network is used to compensate the
estimation error caused by high current pulse. The number of neurons
in the input layer becomes three, which are total current I, voltage V
and Cap. The Cap is the accumulated electricity and it can be calculated
by

X
t
CapðtÞ ¼ IðtÞΔt (4)
i¼1

where I, t and Δt represent the total current, current moment and time
step, respectively. The output layer contains two neurons, which are the
estimation error of branch 1 (EOB1) and the estimation error of branch 2
(EOB2). The mathematical expressions of EOB1 and EOB2 are shown in
Eqs. (5) and (6).
Fig. 1. Schematic of BP neural network with one hidden layer.
EOB1 ¼ bI 1  I1 (5)

adopted for neural network (NN) training, which is a variation of EOB2 ¼ bI 2  I2 (6)
gradient descent. The mathematical relationship between outputs and
The training set includes the experiment data set of V, I, Cap, EOB1
inputs is shown in Eq. (1).
and EOB2 under DST and UDDS, and the testing set includes the exper-
X
q Xm iment data set of V, I and Cap under FUDS. The goal of training set MSE is
yk ¼ f2 ðwk;i f1 ð wi;j xi þ θ1;j Þ þ θ2;k Þ k ¼ 1; 2; :::l (1) 0.000,001 A2 and the maximum number of iterations is 100. Then the
j¼1 i¼1
trained BP neural network is tested by the testing set. The estimated
The estimation and correction scheme are illustrated in Fig. 2, where E OB b 2 are obtained to correct the estimated currents in step 3,
b 1 and E OB
a battery pack with two parallel cells is taken as an example. respectively. The mathematical expressions of bI 1c and bI 2c are shown in
In step 1, a parallel battery pack model is established. The voltage Eqs. (7) and (8).
and current of P-LiBs are measured by the voltage and current sensors;
In step 2, off-line tests are carried out under capacity and dynamic b 1
bI 1c ¼ bI 1  E OB (7)
discharging working conditions. Capacity test includes the constant
current constant voltage (CCCV) charging and CC discharging conditions b 2
bI 2c ¼ bI 2  E OB (8)
and the dynamic discharging working condition includes DST, Federal
Urban Driving Schedule (FUDS), Urban Dynamometer Driving Schedule The MSE expressed in Eq. (9) is chosen to evaluate the estimation and
(UDDS). The voltage V, branch current I1 and I2, and total current I of the correction performance of the BP neural network [14]. Generally, the low
P-LiBs are measured; MSE value represents high accuracy estimation. The N expresses the
In step 3, the first BP neural network is adopted to estimate the number of the points in the testing set.
branch currents. The structure with one hidden layer is selected for this
BP neural network. The input layer contains two neurons, which are total 1 XN
MSE ¼ ðxi  yi Þ2 (9)
current I and voltage V. The output layer contains two neurons, which are N i¼1
the two branch currents I1 and I2. The number of neurons in the hidden
layer is 10. The activation functions of the hidden layer and output layer where xi and yi are the estimated and measured values.
are logsig and purelin, respectively. The mathematical expressions of
activation functions can be seen in Eqs. (2) and (3). 3. Experiment
For EVs, it may be more likely to use CCCV conditions for charging
and dynamic conditions for discharging. To estimate branch currents This study takes the parallel structure of two Li(NiCoMn)O2 cells as an
under CCCV charging working condition, the total experimental data set example to verify the proposed scheme. The nominal capacity of the
is divided into training set and testing set. The training set includes the selected cells is 30.0 Ah at 25  C, The lower/upper cutoff voltage of each
experiment data set of V, I1, I2, and I measured under CCCV charging and cells is 2.75 V/4.15 V. Considering the inconsistency of the battery sys-
discharging working condition, and the testing set includes the experi- tem, three group battery packs with different capacity and resistence
ment data set of V and I under CCCV charging working condition. The BP differences are studied, and the actual parameters and of each cell 25  C
neural network is trained by the training set until there is no significant are given in Table 1.

3
Q. Yu et al. Green Energy and Intelligent Transportation 1 (2022) 100029

Fig. 2. Branch current estimation and correction process.

2
logsig : f ðxÞ ¼ (2)
1 þ ex

purelin : f ðxÞ ¼ x (3)

Table 1
The parameters of three battery pack groups.
Cells Capacity (Ah) Resistence (mΩ)

Branch 1 Branch 2 Difference Branch 1 Branch 2 Difference

Group 1 22.5 27.1 4.6 2.9 2.6 0.3


Group 2 31.2 31.6 0.4 1.9 1.8 0.1
Group 3 31.7 29.3 2.6 1.7 1.6 0.1

4
Q. Yu et al. Green Energy and Intelligent Transportation 1 (2022) 100029

The experimental devices as illustrated in Fig. 3 consists of a XINWEI with single NN. In the CC charging stage where the total current is
CT-4004 test machine for charging and discharging the battery pack, a completely constant, the estimation results of the LSTM algorithm are
GDBELL thermal chamber to control the ambient temperature of battery better than that of the BP and CNN algorithms, which is consistent
packs, a branch current collection device to collect the branch current, with Ref [13]. It should be noted that from the resting state to the
and a host computer to restore and analysis the testing data. The current sudden application of the charging current, BP has a smaller estima-
sensor uses CZ3457 with a range of 100 A and an accuracy of 0.5%. The tion error than LSTM and CNN. From the MSE index, BP is better than
current was sampled at an interval of 50 ms. LSTM and CNN in the entire CCCV working condition. From the re-
For the vehicle battery system, the charging condition is relatively sults of using dual NNs to estimate and correct the branch currents in
fixed, but the discharging condition is more complicated. This study selects Fig. 5 (c) and (d), the correction effect of LSTM–LSTM is the best and
the CCCV and three dynamic working conditions to conduct experiment. the correction effect of BP-BP is the worst. However, it should be
The measured branch current I1, I2, and the difference between them of pointed out that for CCCV working conditions, the use of BP for
group 1 are illustrated in Fig. 4. It can be seen that the difference is quite branch current estimation is better in the terms of current accuracy
large with the maximum difference of 8 A. If it is still assumed that the than that of LSTM, CNN and the three dual NNs. It can also be seen
branch currents are equal when perform various functions in the BMS, the from the estimation results that the estimation error of the branch
reliability and safety of the battery pack will be seriously affected. current has a certain positive correlation with the rate of change of

Fig. 3. (a) Experimental device. (b) Schematic diagram of branch current collection device.

4. Results and discussions the current for some NNs. Therefore, It is necessary to explore the
effect of dual NNs under dynamic conditions.
4.1. Results The results under dynamic working conditions are shown in Figs. 6
and 7. To compare with the estimated results, a curve named ‘Compared’
This section only takes group 1 as an example to demonstrate the calculated by Eq. (10) has been added to the figures.
effectiveness of the proposed method. In the experiment, dual BP (BP-
BP), dual CNN (CNN–CNN) and dual LSTM (LSTM–LSTM) algorithms IFUDS
Compared ¼ (10)
are selected to estimate and correct the branch current. Fig. 5 illus- 2
trates the current results under CCCV charging condition. Statistical where IFUDS represents the total current of the parallel battery pack under
results of branch currents of group 1 under CCCV condition are given FUDS working condition.
in Table 2. As shown in Figs. 6 and 7, BP is the best algorithm and LSTM is
Fig. 5 (a) and (b) shows the estimated branch currents and errors the worst algorithm when only one NN is used to estimated the

5
Q. Yu et al. Green Energy and Intelligent Transportation 1 (2022) 100029

Fig. 4. The branch currents and difference of parallel battery pack of Group 1 under (a) DST. (b) UDDS. (c) FUDS and (d) CCCV working conditions.

Fig. 5. Results of group 1 under CCCV condition: (a) branch 1 and (b) branch 2 before correction. (c) branch 1 and (d) branch 2 after correction.

6
Q. Yu et al. Green Energy and Intelligent Transportation 1 (2022) 100029

Table 2
Statistical results of branch currents of group 1 under CCCV condition.
Cells MSE (A2) Decrease (%)

Before correction After correction

BP-BP Branch 1 0.032,7 0.037,3 14.07


Branch 2 0.032,6 0.036,9 13.19
LSTM–LSTM Branch 1 0.081,8 0.054,0 33.99
Branch 2 0.085,4 0.043,9 48.59
CNN–CNN Branch 1 0.110,4 0.117,3 6.25
Branch 2 0.109,1 0.117,6 7.79

Fig. 6. Estimated error of different algorithms under FUDS working condition: (a) measured current of branch 1. (b) estimated error of CNN–CNN. (c) estimated error
of LSTM–LSTM. (d) estimated error of BP-BP.

7
Q. Yu et al. Green Energy and Intelligent Transportation 1 (2022) 100029

Fig. 7. Estimated error of different algorithms under FUDS working condition: (a) measured current of branch 2. (b) estimated error of CNN–CNN. (c) estimated error
of LSTM–LSTM. (d) estimated error of BP-BP.

Table 3
Statistical results of branch current of group 1 under dynamic conditions.
Algorithm Cells MSE (A2) Difference (A2) Decrease (%)

Before correction After correction

CNN–CNN Branch 1 0.115,0 0.095,0 0.020,0 17.39


Branch 2 0.099,6 0.090,8 0.008,8 8.84
LSTM–LSTM Branch 1 0.290,6 0.065,6 0.225,0 77.43
Branch 2 0.341,1 0.073,5 0.267,6 78.45
BP-BP Branch 1 0.027,2 0.016,6 0.010,6 38.97
Branch 2 0.025,8 0.015,5 0.010,3 39.92

8
Q. Yu et al. Green Energy and Intelligent Transportation 1 (2022) 100029

branch currents under dynamic conditions. The correction effect of 4.2. Discussion
the LSTM is most obvious when the second NN is used to correct the
estimated currents. However, in terms of the corrected branch current 4.2.1. Select the best input for the dual BP neural network
accuracy, the current estimation accuracy using the BP-BP algorithm This section still takes group 1 as an example to select the best model
is still higher than that of CNN–CNN and LSTM–LSTM. Table 3 il- input. The input and output data sets of the dual BP neural network is
lustrates the MSEs of the branch current before and after correction obtained from the above experimental data. According to a few observ-
with different NNs under FUDS working conditions. When using BP to able parameters (V, I, Cap), the alternative scheme for the input of the
estimate the branch current, the MSEs of two branch currents are two BP neural network is shown in Table 4. The parameters of the first BP
0.027,2 and 0.025,8, respectively, which are 76.3% and 74.1% neural network used in the experiment are as follows: the number of
respectively lower than when CNN is used, and 90.6% and 92.4% neurons in the hidden layer is 10, the activation functions of the hidden
respectively lower than when LSTM is used. After correction by BP-BP layer and output layer are logsig and purelin, respectively. The goal of
algorithm, the MSEs of branch 1 and branch 2 decrease by 38.97% training set MSE is 0.000,001 A2 and the maximum number of iterations
and 39.92%, respectively. However, the correction effect of is 100. The parameters of the second BP neural network are the same as
CNN–CNN is not obvious, only 17.39% and 8.84%. Therefore, from the first one.
the perspective of the correction effect and accuracy of the branch In order to ensure the reliability of the experimental results, the first
currents, BP-BP is more suitable than only using a single NN and BP neural network is set with the same initial weight and bias in each
other dual NNs in dynamic conditions. scheme. The result of scheme 2 has already illustrated in Figs. 6 and 7,
and the results of the other schemes under FUDS condition are shown in
Fig. 8. The statistical analysis is shown in Table 5. As can be seen, when V
Table 4 and I are input of the first BP neural network, the total MSE of the esti-
Input schemes for the first BP neural mated branch currents is the minimum, which is 0.053 A2.
network. After obtaining the branch current estimation results, in order to
Scheme Inputs further eliminate the estimation error caused by dynamic current pulse
1 I excitations, the second BP neural network is also given the same initial
2 VI weight and bias in each scheme to find the best input. The result of
3 I Cap scheme 5 has already shown in Fig. 8 (d) and the results of the other
4 V Cap schemes under FUDS condition are shown in Fig. 9. The statistical
5 I V Cap
analysis is shown in Table 6. As can be seen, when I, V and Cap are set as

Fig. 8. The estimation results of different inputs scheme. (a) Scheme 1. (b) Scheme 3. (c) Scheme 4. (d) Scheme 5.

9
Q. Yu et al. Green Energy and Intelligent Transportation 1 (2022) 100029

Table 5
Statistical results of different input schemes.
Scheme Inputs MSE (A2) Total MSE (A2)

Branch 1 Branch 2

1 I 0.092,3 0.093,8 0.186,1


2 VI 0.027,2 0.025,8 0.053,0
3 V Cap 9.743,3 12.155,3 21.898,6
4 Cap I 0.069,7 0.065,9 0.135,6
5 V I Cap 0.030,2 0.030,0 0.060,2

Fig. 9. The estimation results of different inputs scheme. (a) Scheme 1. (b) Scheme 2. (c) Scheme 3. (d) Scheme 4.

Table 6
Statistical analysis of different input schemes.
Inputs MSE (A2) Total MSE (A2)

Branch 1 Branch 2

I 0.026,3 0.026,1 0.052,4


VI 0.024,2 0.023,3 0.047,5
I Cap 0.029,3 0.026,9 0.056,2
V Cap 0.020,5 0.019,2 0.039,7
I V Cap 0.016,6 0.015,5 0.032,1

10
Q. Yu et al. Green Energy and Intelligent Transportation 1 (2022) 100029

the input of the second BP neural network, the total MSE is the minimum,
which is 0.032,1 A2.

4.2.2. Analysis of algorithm adaptability


The other two groups of batteries, group 2 and group 3, are also used
to further verify the effectiveness and limitations of the proposed
method. The estimation and correction results are shown in Figs. 10 and
11, respectively. The statistical analysis is shown in Table 7.

Fig. 11. Results of group 3 under FUDS working condition: (a) branch 1 and (b)
branch 2.

Table 7
Statistical analysis of experimental results.
Cells MSE (A2) Decrease (%)

Before correction After correction

Fig. 10. Results of group 2 under FUDS working condition: (a) branch 1 and (b) Group 1 Branch 1 0.027,2 0.016,6 38.97
branch 2. Branch 2 0.025,8 0.015,5 39.92
Group 2 Branch 1 0.008,0 0.002,6 67.5
Branch 2 0.007,9 0.002,3 70.89
It can be seen from Table 7 that the estimation and correction results Group 3 Branch 1 0.009,3 0.002,8 69.89
of groups 2 and 3 are significantly better than those of group 1. The Branch 2 0.009,1 0.002,8 69.23
reasons for this may be that the capacity difference between groups 2 and
3 is smaller than that of group 1, or that the internal resistance difference
between groups 2 and 3 is smaller than that of group 1. In addition, it than the reason of capacity difference. Therefore, the smaller the dif-
should be pointed out that the estimated and corrected results of group 2 ference between battery capacity and internal resistance, the more ac-
were similar to those of group 3. The large probability for this is due to curate the estimation result of BP algorithm and the more obvious in the
the small difference in internal resistance between groups 2 and 3, rather correction effect.

11
Q. Yu et al. Green Energy and Intelligent Transportation 1 (2022) 100029

5. Conclusion [2] Brand MJ, Hofmann MH, Steinhardt M, Schuster SF, Jossen A. Current distribution
within parallel-connected battery cells. J Power Sources 2016;334:202–12.
[3] Yu Q, Dai L, Xiong R, Chen Z, Zhang X, Shen W. Current sensor fault diagnosis
In this paper, branch currents estimation and correction for a parallel method based on an improved equivalent circuit battery model. Appl Energy 2022;
battery pack is realized by dual BP neural networks. The experimental 310:118588.
results verify the effectiveness of the proposed method. Compared with the [4] Rumpf K, Naumann M, Jossen A. Experimental investigation of parametric cell-to-
cell variation and correlation based on 1100 commercial lithium-ion cells. J Energy
ideally calculated branch current which is expressed as the total current Storage 2017;14:224–43.
divided by the branch numbers, the use of NN can effectively improve the [5] Xiong R, Li Z, Yang R, Shen W, Ma S, Sun F. Fast self-heating battery with anti-aging
accuracy of branch current. For the CCCV condition, using a single BP NN awareness for freezing climates application. Appl Energy 2022;324:119762.
[6] Khalid HM, Ahmed Q, Peng JCH, Rizzoni G. Pack-level current-split estimation for
is better than using CNN, LSTM, BP-BP, CNN–CNN and LSTM–LSTM. For health monitoring in Li-ion batteries. IEEE Transactions on Transportation
dynamic conditions, using dual BP neural networks is better than using BP, Electrification 2016:1506–11.
CNN, LSTM, CNN–CNN and LSTM–LSTM. Furthermore, the experimental [7] Xiong R, Huang J, Duan Y, Shen W. Enhanced lithium-ion battery model
considering critical surface charge behavior. Appl Energy 2022;314:118915.
results suggest that, the total current I and voltage V are suitable as inputs [8] Yu Q, Xiong R, Yang R, Pecht MG. Online capacity estimation for lithium-ion
for the first BP neural network while the total current I, voltage V, and Cap batteries through joint estimation method. Appl Energy 2019;255:113817.
are suitable as inputs for the second BP neural network. In addition, the [9] Xiong R, Sun W, Yu Q, Sun F. Research progress, challenges and prospects of
fault diagnosis on battery system of electric vehicles. Appl Energy 2020;279:
smaller the difference between battery capacity and internal resistance, the 115855.
more accurate the estimation result of BP algorithm and the more obvious [10] Yu Q, Li J, Chen Z, Wan C, Pecht M. Multi-fault diagnosis of lithium-ion battery
in the correction effect. In future, the validity and error bounds of the systems based on improved correlation coefficient and similarity approaches. Front
Energy Res 2022;10:891637.
estimated branch currents for BMS modeling, state estimation, and fault
[11] Xin H, Xuning F, Xuebing H, Languang L, Minggao O. Study on modelling and
diagnosis could be explored. analysis of imbalanced current inside parallel-connected lithium-ion batteries for
electric vehicle. J Mech Eng 2019;55(20):44–51.
Declaration of competing interest [12] Khalid HM, Ahmed Q, Peng JCH, Rizzoni G. Current-Split estimation in Li-ion
battery pack: an enhanced weighted recursive filter method. IEEE Transactions on
Transportation Electrification 2015;1:402–12.
The authors declare that they have no known competing financial [13] Cui Z, Cui N, Rao J, Li C, Zhang C. Current distribution estimation of parallel-
interests or personal relationships that could have appeared to influence connected batteries for inconsistency diagnosis using long short-term memory
networks. IEEE Transactions on Transportation Electrification 2021:1013–25.
the work reported in this paper. [14] Yang F, Cho H, Zhang H, Zhang J, Wu Y. Artificial neural network (ANN) based
prediction and optimization of an organic Rankine cycle (ORC) for diesel engine
Acknowledgements waste heat recovery. Energy Convers Manag 2018;164:15–26.
[15] Tian J, Xiong R, Shen W, Lu J, Yang X. Deep neural network battery charging
curve prediction by using 30 points collected in 10 minutes. Joule 2021;5(6):
This work was jointly supported by the Natural Science Program of 1521–34.
Shandong Province (Grant No. ZR2020ME209), National Natural Science [16] Liu J, Huang J, Sun R, Yu H, Xiao R. Data fusion for multi-source sensors using GA-
PSO-BP neural network. IEEE Trans Intell Transport Syst 2021;22:6583–98.
Foundation of China(Grant No. 52177210) and China Postdoctoral Sci- [17] Tong Shijie, Lacap Joseph H, Park Jae Wan. Battery state of charge estimation using
ence Foundation (Grant No. 2021M690740). a load-classifying neural network. J Energy Storage 2016;7:236–43.
[18] Chen C, Xiong R, Yang R, Li H. A novel operational data-driven battery open-circuit
voltage characterization mining method for large-scale Applications. Green Energy
References
and Intelligent Transportation 2022;1(1):100001.

[1] Fill A, Koch S, Birke KP. Algorithm for the detection of a single cell contact loss
within parallel-connected cells based on continuous resistance ratio estimation.
J Energy Storage 2020;27:101049.

12

You might also like