Energy 223 (2021) 120116

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

State of charge prediction of EV Li-ion batteries using EIS: A machine

learning approach
Iman Babaeiyazdi*, Afshin Rezaei-Zare, Shahab Shokrzadeh
Department of Electrical Engineering and Computer Science, York University, Toronto, ON, M3J 1P3, Canada

a r t i c l e i n f o a b s t r a c t

Article history: Due to the significantly complex and nonlinear behavior of li-ion batteries, forecasting the state of charge
Received 6 October 2020 (SOC) of the batteries is still a great challenge. Therefore, accurate SOC estimation is essential for the
Received in revised form proper operation of batteries while the battery is monitored by the battery management system (BMS).
29 December 2020
To this end, this paper employs informative measurements of electrochemical impedance spectroscopy
Accepted 12 February 2021
(EIS) in machine learning models (ML), i.e., linear regression model and Gaussian process regression
Available online 16 February 2021
(GPR), to accurately predict the SOC of li-ion batteries. First, a feature sensitivity analysis of the data is
conducted to extract the most reliable features, i.e., the EIS impedances which are highly correlated with
Index Terms:
Electric vehicle
SOC, from EIS measurements. Then, the models are fed by the chosen features. The models are designed
Electrochemical impedance spectroscopy to train the input features and establish the mapping relationship between the selected features and the
Li-ion batteries SOC. The results demonstrate that the error of the GPR model was found to be less than 3.8%. Considering
Machine learning onboard EIS measurements, this method can be practically embedded in the battery management system
for accurate measurements of SOC of li-ion batteries and ensure the proper and efficient operation of
battery-powered electric vehicles.
© 2021 Elsevier Ltd. All rights reserved.

1. Introduction estimation [5].

Myriads of literature propose various SOC estimation methods
Lithium-ion batteries have been widely utilized in Plug-in such as coulomb counting method, data-driven, and model-based
Electric Vehicles (PEVs), by virtue of their high energy and long estimation methods [5]. Coulomb counting or ampere-hour
life duration [1]. However, the performance of PEVs may be counting is one of the widely used methods in the laboratory, but
impacted by low performance of the battery management system due to cumulative current measurement errors, this method is not
(BMS) due to the unpredictability of battery’s chemical reactions. considered a highly accurate method for SOC estimation [5]. In the
Estimation of the battery state of charge (SOC) as one of the main model-based methods, equivalent circuit models (ECMs) and
functions of the BMS contributes greatly to proper operation, electrochemical impedance models (EIMs) are the major models
charging/discharging cycling, and lifespan of the EV batteries [2]. derived from empirical data to predict the SOCs. ECMs and EIMs are
Also, Reliable and accurate SOC estimation can have other impor- combined with various adaptive algorithms such as Kalman filter
tant applications such as baseload power generation of intermittent [6,7], extended and unscented Kalman filter [8,9], and Particle filter
sources in the electric grid [3], and the safe operation of EV fast- [10] to calculate the SOC of the battery based on the charge and
charging stations integrated with battery storage system [4]. SOC discharge voltage and current. Moreover, electrochemical imped-
is defined as the capacity of the battery at the current state ance spectroscopy (EIS) measurements are utilized in ECMs to es-
compared to the battery’s capacity at fully charged state [2]. timate SOC. In Refs. [2,11], and [12], the authors identify the
Moreover, SOC cannot be measured directly from inner quantities parameters of ECMs based on the EIS measurements. In Ref. [2], the
such as internal resistance and capacitance of the battery, and as a EIS data is derived only at one SOC, which prevents the model from
result, the external quantitative indices are utilized for SOC being an inclusive model, but on the other hand, a wide range of
temperature is considered for modeling the battery based on ECM.
The EIS data at above-zero temperature and SOCs between 10% to
* Corresponding author. 90% and 10%e100% has been derived in Refs. [11,12], respectively,
E-mail addresses: [email protected] (I. Babaeiyazdi), [email protected] which, however, decrease the accuracy and reliability of the
(A. Rezaei-Zare), [email protected] (S. Shokrzadeh).

https://doi.org/10.1016/j.energy.2021.120116
0360-5442/© 2021 Elsevier Ltd. All rights reserved.
I. Babaeiyazdi, A. Rezaei-Zare and S. Shokrzadeh Energy 223 (2021) 120116

estimation for SOCs at sub-zero temperatures and SOCs below 10%. batteries with the state of health (SOH) between 60% and 100%. The
Therefore, due to the internal complex chemical reaction process reason that the battery’s degradation was not considered in this
and uncertain external operating conditions of batteries, modeling paper is due to the unavailability of EIS measurements for different
the batteries based on the ECM methods is challenging for esti- SOHs for the dataset utilized in the paper. In contrast to many other
mating the battery characteristics in real-life operation [13,14]. studies that only take into account the EIS data obtained at above-
Physic-based models (PBMs) demonstrate insights through chem- zero temperatures, this study considers the EIS data for both above-
ical and electrochemical dynamics, such as li-ion diffusion and zero and sub-zero temperatures, i.e., as low as
20  C. The results
Ohmic effects [13]. However, to estimate SOC using PBMs, partial demonstrate an error of less than 3.8% for the GPR model.
derivatives equations should be solved by the BMS controller, Considering the online and on-board EIS measurement [22e24],
which is a highly intensive computational burden [15]. Data-driven this method can be practically embedded in the BMS for accurate
models are only dependent on historical data, and they do not need measurements of SOC.
complicated equivalent or mathematical models. However, the The paper is organized as follows: in section 2, the electro-
challenge of data-driven models is the acquirement of informative chemical impedance spectroscopy measurement is fully explained.
inputs to construct a robust model for predicting the battery In the next section, the methodology for extracting reliable features
characteristics. Additionally, effective extraction of the features and building the prediction models based on linear regression and
from historical data still remains a challenging task [16]. In Ref. [17], GPR algorithm is discussed. Section 4 introduces the result of the
the SOC of the battery is predicted by a neural network (NN) which built-up models for predicting the SOC, and the last section is
utilizes voltage, current, temperature, and power of the battery as dedicated to the conclusions.
the input features. Prediction of SOC also has been conducted in
Ref. [18] employing NN and random forest/tree. Voltage, current, 2. Electrochemical impedance spectroscopy
and cycling number contribute as the inputs of the machine
learning (ML) black box in the mentioned paper. In another study EIS is a non-destructive and information-rich test which is
[19], support vector machine along with Gaussian methods esti- conducted by galvanostatic or potentiostatic excitation signal over
mate the SOC of the battery and extracts feature variables based on a wide range of frequency to obtain the impedance of the battery
the charging curve. However, all of these data-driven models that during charging and discharging [25]. The excitation signals in
use terminal voltage as input feature may lose the accuracy as the galvanostatic and potentiostatic methods are commonly sinusoidal
terminal voltage of battery suddenly drops at the end of discharge current and voltage and the corresponding response will be voltage
which accordingly does not provide reliable data for low SOCs [20]. and current, respectively. Based on these waveforms, the electro-
Thus, identifying and extraction of reliable features become the chemical impedance of the battery can be calculated. The imped-
main bottleneck of the adoption of the data driven approaches and ance of the battery is obtained based on the following equations in
thus, more research is required in this regard. galvanostatic mode [26]:
On the other hand, EIS measurements over a wide range of
frequency provide rich information about the dynamic character- DI ¼ Imax sinð2pftÞ; (1)
istics of the battery and pave the way for precise estimation of the
battery status. Nevertheless, none of the reviewed papers have DV ¼ Vmax sinð2pft þ ∅ Þ; (2)
adopted the EIS measurements directly as input data for machine
learning models to predict SOC except for [21], in which the EIS data Vmax j∅
obtained for SOCs above 30% and at room temperature have been Zðf Þ ¼ e ; (3)
Imax
utilized in a deep NN. The model does not employ the EIS data in a
wide range of temperature and at different SOC points [21], while where DI is a sinusoidal current at frequency f, which is super-
such a data exclusion decreases the accuracy and reliability of the imposed on the dc charging or discharging current and results in
model. Also, the reported error of the model of [21] is less than 5%. DV and phase angle ∅. Accordingly, Eq. (3) shows that the battery’s
This study investigates the effectiveness of the EIS measurement impedance is frequency-dependent and characterized by its
data for estimating the SOC of the li-ion batteries using machine magnitude and phase angle. Fig. 1 indicates a typical EIS spectrum.
learning techniques. In opposition to Ref. [21], which uses the The horizontal axis indicates the real part of the impedance, and
whole EIS impedances from the EIS spectrum to estimate SOC, only the vertical axis shows the negative of the imaginary part of the
highly correlated EIS impedances with SOC are used in this paper. impedance. The EIS spectrum is drawn over a wide range of
The proposed method’s advantages are higher accuracy of the
models and lower computational burden by eliminating irrelevant
input features, i.e., EIS impedances with low correlations. There-
fore, highly correlated impedances are first identified and then
extracted from EIS spectrum measurements obtained at SOCs from
0% to 100%. The chosen impedances are utilized as input features
for the linear regression model and Gaussian process regression
(GPR). The models are designed to train the input features and
establish the mapping relationship between the selected fre-
quencies and the SOC. Finally, the trained models are employed to
achieve SOC prediction.
Moreover, since the machine learning algorithm is neither
dependent on the model of the battery nor the method that the
battery is charged/discharged, and only the input and output of the
dataset matter here, the model can predict the SOC by interpolating
or extrapolating the dataset, regardless of charging or discharging
mode of the battery. The SOC can be precisely estimated for aged
batteries if the EIS measurements dataset is available for degraded Fig. 1. Typical EIS spectrum of li-ion battery.

2
I. Babaeiyazdi, A. Rezaei-Zare and S. Shokrzadeh Energy 223 (2021) 120116

Table 1 impedance spectroscopy at different temperatures for the given

Panasonic 18650 PF cell parameters. SOC of 50%. It can be observed from Fig. 2(a) that the semicircle
Nominal Open Circuit Voltage 3.6 V enlarges as the temperature decreases. It is due to the fact that the
Min/Max Voltage 2.5 V/4.2 V charge transfer rate decreases in the solid electrolyte interface (SEI)
Mass/Energy storage 48 g/9.9 Wh layer and electrode-electrolyte interface at lower temperatures
Capacity 2.75 Ah
Cycles to 80% Capacity 500 (100% DOD, 25  C)
[29]. Fig. 2(b) shows that as the temperature decreases, the polar-
Minimum Charging Temperature 10  C ization resistance of the battery (where eIm Z ¼ 0) increases due to
a decrease in ion transfer in the electrolyte at lower temperatures
[29]. Fig. 3(a), Fig. 3(b) and 3(c) plot the EIS measurements of the
frequency. The low-frequency tail indicates the diffusion processes battery at different SOCs and temperatures of þ25  C, 0  C,
inside the active material of the battery, the mid-frequency semi- and
20  C, respectively. This figure manifests that at above zero
circle indicates the double-layer capacitance effect, and in the high- temperatures, the SOC is less effective on impedance spectra, while
frequency region, the intercept of the EIS curve with the real axis is at zero and sub-zero temperatures, SOC significantly affects the
the indicator of Ohmic resistance of the battery. impedance curve, especially in the mid-frequency regions. Thus,
the effects of SOC on the impedance spectra at different tempera-
2.1. EIS experimental data tures are different.

In this study, the experimental data from Ref. [27] have been 3. Methodology
utilized, where a Panasonic NCR18650PF lithium-ion battery, an
NCA chemistry cell similar to the cells used in Tesla’s electric cars This section is dedicated to the feature sensitivity analysis to
[28], was tested. The battery specifications have been presented in capture the highly correlated EIS features, i.e., highly correlated EIS
Table 1. In the test, EIS measurements were conducted over SOCs impedances with SOC of the battery, and then the selected reliable
from 0% to 100% and temperature range of
20  C e 25  C for features are utilized for training and testing of the machine learning
frequency sweep of 1 mHze6 kHz. Fig. 2 shows the battery’s models.

3.1. Feature sensitivity analysis

Extracting highly relevant features to the machine learning

models’ output, i.e., SOC, is essential for accurately predicting the
output. To this end, the correlation matrix, which indicates the
dependency of two or more variables on each other, is calculated
and then shown on a color-coded image plot. The calculation of the
correlation is performed by the Pearson correlation coefficient as
follows:

EðX; YÞ
EðXÞEðYÞ
rX;Y ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  ffi; (4)
E X 2
EðXÞ2 : E Y 2
EðYÞ2

where E is the expected value operator, and X and Y are two random
variables. Fig. 4 shows the heatmap of the dataset [27] that we used
for this study at different temperatures. Fig. 4 is a 2-D graphical
representation that indicates the dependency of the features of the
dataset. The features in the used dataset are electrochemical im-
pedances at the corresponding frequencies that they were
measured. In this case, the number of the features is 54 since the
impedances were measured at 54 frequencies, sweeping from
1 mHz to 6 kHz. In Fig. 4, the correlation between the features
varies from
1 to 1. The positive correlation has been shown in the
spectrum of light to dark red, and the negative correlation has been
shown in the spectrum of light to dark blue. The positive and
negative correlations mean that the output varies in the same or
opposite direction of the input variables’ variations. The heatmap is
a symmetric figure; thus, the last row or the last column represents
the relation of the input variables, i.e., impedances at different
frequencies with the output, i.e., SOC of the battery. Fig. 4(a)e(e)
show the heatmap of the dataset at temperatures of (25  C) e
(
20  C), respectively. Fig. 4 (a) shows that the first few features are
highly and negatively correlated with the SOC and these features lie
in the high and mid-frequency regions of the EIS spectrum. As also
depicted in Fig. 4 (a), it is apparent from Fig. 4(b)e(e) that as the
temperature decreases, some other features from mid-frequency
appear to be positively correlated with the SOC. Another remark-
Fig. 2. EIS spectrum of the battery at (a) SOC of 50% and different ambient tempera- able result deduced from Fig. 4 is that the low frequencies are
ture, (b) zoomed-in version of (a). significantly less correlated with the SOC. The reason is that
3
I. Babaeiyazdi, A. Rezaei-Zare and S. Shokrzadeh Energy 223 (2021) 120116

according to Figs. 2 and 3, the number of features, i.e., EIS imped-

ances at lower frequencies, is significantly less than the number of
features at higher frequencies. In other words, most of the EIS
impedances which have lower magnitudes exist in all of the con-
ditions in Figs. 2 and 3, and they lie in high and mid-frequency
regions; however, the EIS impedances with larger magnitudes
only exist in extreme conditions, i.e., at the temperature of
20  C
in Fig. 2 and SOC of 20% in Fig. 3.
Therefore, heatmaps are useful for extracting reliable features to
build an accurate and fast model for prediction purposes. Moreover,
pairgrid plot is also utilized to show the relationships of highly
correlated features obtained from heatmap with the SOC and their
distribution in the dataset. In Fig. 5, only four high-correlation
features f1, f2, f4, and f5 have been shown as examples at 25  C.
The diagonal plots represent the distribution of single-variable and
off-diagonal plots, which are the mirror image of each other and
show the relations between every two variables of the
dataset along with the regression plots. It is worthy to note that
Fig. 5 shows only the negatively correlated features with the SOC. It
is observed that with the increase of features f1, f2, f4, f5 in
magnitude, the SOC decreases. Therefore, feature analysis helps
identify the most influential features that contribute to the varia-
tion of the output variable.

3.2. Linear regression algorithm

A linear regression algorithm is used for identifying the rela-

tionship between a dependent variable and one or more inde-
pendent variables. In this case, the impedances at different
frequencies are the independent variables, and the SOC is the
dependent variable. The basic multiple regression model of a
dependent variable Y on a set of k independent variables (xk) can be
expressed as [30]:
8
> y ¼ b0 þ b1 x11 þ … þ bk x1k þ e1
< 1
y2 ¼ b0 þ b1 x21 þ … þ bk x2k þ e2
(5)
:«
>
yn ¼ b0 þ b1 xn1 þ … þ bk xnk þ en

Therefore:

yi ¼ b0 þ b1 xi1 þ … þ bk xik þ ei (6)

where yi is the i-th case of the dependent variable Y, xij is the value
of the j-th independent variable (Xj) for the i-th case of the
dependent variable, b0 is the Y-intercept of the regression surface,
each bj is the slope of the regression surface with respect to variable
Xj, and finally ei is the random error component for the i-th case. In
each equation in Eq. (5) the error is distributed with zero mean and
standard deviation, and it is independent of the errors in the other
equations. Since the variables are fixed quantities, the randomness
of Y results from the randomness of error terms in each equation;
although, in terms of correlation, the input variable are taken into
account random variables, and the input variables are independent
of the error terms. In matrix notation, Eq. (5) can be written as [30]:

Y ¼ Xb þ e (7)

where:

Fig. 3. EIS spectrum of the battery at (a) 25  C, (b) 0  C, and (c)
20  C and at different Y ¼ ½y1 y2 / yn T
SOC levels. b ¼ ½b1 b2 /bkþ1 T
2 3
x11 / x1ðkþ1Þ (8)
X ¼ 4« 1 « 5
xn1 / xnðkþ1Þ

4
I. Babaeiyazdi, A. Rezaei-Zare and S. Shokrzadeh Energy 223 (2021) 120116

Fig. 4. Heatmap for feature sensitivity analysis of EIS spectrum at (a) 25  C. (b) 10  C, (c) 0  C, (d)
10  C, (e)
20  C.

and Y is the target vector, e is the error vector which is a column pairs of inputs xi , which may have one or more than one features,
vector of length n, and b is the vector of parameters, which is a and output yi , the GPR model computes the predictive distribution
column vector of length k þ 1. Matrix X is the input matrix, which is of unobserved test datasets with y* as output and x* as input [31]. In
n by k þ 1 matrix. To do prediction, b and e should be calculated. this study, X and Y are defined as X ¼ ½x1 ; …:; xn T and Y ¼
The structure of the regression model has been shown in Fig. 6.
½y1 ; …:; yn T , respectively. In this case xi ¼ ½EIS impedances is the
EIS impedances and the output yi is the SOC of the cells. It is also
3.3. Gaussian process regression (GPR) assumed that yi ¼ f ðxi þεi Þ where εi  N ð0; s2 Þ is an independent
and identically distributed Gaussian noise. The outputs F ¼
For a given training dataset of T ¼ fðxi ; yi Þ; i ¼ 1; 2; …:; ng with n ðf ðx1 Þ þ… þf ðxn ÞÞ are modeled as Gaussian random field F
5
I. Babaeiyazdi, A. Rezaei-Zare and S. Shokrzadeh Energy 223 (2021) 120116

 
1
D2 ¼ Kðx* ; x* Þ
Kðx* ; XÞ KðX; XÞ þ s2 I KðX; x* Þ (11)

3.4. Data partitioning

To avoid the malfunction of the ML models over a new dataset,

the dataset should be split into two partitions, i) training set and ii)
test set. The training set is used to construct the models and con-
tains known output. The ML models are fitted to the training data
and calculate the regression coefficients. Then, the test set is uti-
lized to observe the performance of the ML models over unknown
data. This implies that the test set serves as a criterion for the
evaluation of the model predictions. In contrast to many other
studies, the effect of different test-size (TS) is observed in this study
for the dataset with the selected features and different correlation
values.

Fig. 5. PairGrid for reliable features of EIS spectrum at 25  C. 3.5. Accuracy evaluation

The indices used for the evaluation of the performance of the

proposed models are:

3.5.1. R-squared
Goodness-of-fit R-squared (R2 ) is defined as [32]:
 2
Pn
i¼1 yi
b
yi
R2 ¼ 1
P 2
(12)
n
i¼1 yi
y

where yi is the actual value and b

y i is the predicted value for the i-th
case. R2 ranges between 0 and 1. The closer to 1, the better the
prediction.

3.5.2. Mean absolute error (MAE)

MAE is used to compare the precision of the prediction and is
defined as follows [32]:

100 Xn
yi
b yi
MAE ¼ (13)
n b
y i
i¼1

The smaller the MAE, the better the prediction.

Fig. 6. Linear regression structure.

3.5.3. Root mean squared error (RMSE)

 N ð0; KÞ where Kij ¼ kðxi ; xj Þ is the covariance kernel. In this case, The root square of the above equation provides the standard
the radial basis function is utilized for covariance kernel as well. deviation of the random error term. Root mean square error is an
estimate of the standard deviation of the random component in the
The joint distribution of the training dataset fðxi ; yi Þ; i ¼ 1; 2; …:; ng
data and is defined as follows:
and the predicted test output ðx* ; y* Þ is [31]:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u  2
uPn
u i¼1 yi
b y
     t i
KðX; x* Þ RMSE ¼
0; KðX;* XÞ þ s I
Y 2 (14)
* ¼ N (9) n
y Kðx ; XÞ Kðx* ; x* Þ
Conditioning on the training set yields the predicted mean on
x* : 4. Prediction results and discussion

In this study, Python and MATLAB have been harnessed to

 
1 perform the statistical analysis and prediction of the SOC. Fig. 7
* *
y ¼ Kðx ; XÞ KðX; XÞ þ s I 2
Y (10) shows the flowchart of the proposed methodology for predicting
the SOC of the battery using EIS measurements. In this section, the
And its predicted variance is: models’ performance at different temperatures is discussed,
6
I. Babaeiyazdi, A. Rezaei-Zare and S. Shokrzadeh Energy 223 (2021) 120116

considering the effect of reliable features extraction based on their Table 2

correlation value and portioning of the dataset. The statistical Evaluation indices under different conditions.

evaluation indices discussed in the previous section have been ML model Corr-value Temp(C) R_squared MAE RMSE
tabulated in Table 2 in different conditions. Linear Regression 0.5 25 (TS ¼ 0.2) 0.978 4.8456 5.9924
As mentioned earlier, some datasets have been introduced to 10 (TS ¼ 0.4) 0.74245 8.9317 20.098
the ML models with different portioning. The default portioning is 0 (TS ¼ 0.4) 0.62 17.444 20.098
that 80% of the dataset is dedicated to the training set, and 20%
10 (TS ¼ 0.2) 0.9894 2.9056 3.3387

20 (TS ¼ 0.2) 0.9825 4.230 4.3
(test_size (TS) ¼ 0.2) is dedicated to the test set. However, to see the
Linear Regression 0.7 25 (TS ¼ 0.2) 0.975 4.8456 5.9924
effect of TS, other values for this variable have been also taken into 10 (TS ¼ 0.4) 0.6983 9.5468 17.913
account. For corr_value of 0.5, different evaluation indices have 10 (TS ¼ 0.6) 0.9429 5.5629 8.4434
been obtained at different temperatures as presented in Table 2. 0 (TS ¼ 0.4) 0.7348 13.664 16.797
0 (TS ¼ 0.5) 0.8764 9.7318 11.481
The linear regression model can perfectly predict the battery’s SOC

10 (TS ¼ 0.2) 0.9877 3.0525 3.602
based on the values of R_squared, MAE, and RMSE at the mentioned
20 (TS ¼ 0.2) 0.977 4.9398 4.8428
temperatures except for 10  C and 0  C. The MAE for 25  C,
10  C, Linear Regression 0.9 25 (TS ¼ 0.2) 0.9524 6.2756 8.275
and
20  C temperatures is less than 4.9%, but for 10  C and 0  C, 10 (TS ¼ 0.2) 0.78241 15.237 17.694
the MAEs are 8.9% and 17.5%, respectively. Moreover, it is clear that 10 (TS ¼ 0.6) 0.899 6.9951 11.216
0 (TS ¼ 0.4) 0.4345 15.094 24.528
for TS of 0.4 the highest accuracy is achieved for temperatures 10  C
0 (TS ¼ 0.5) 0.8785 9.0877 11.382
and 0  C. As for the corr_value of 0.7, we can observe that for the 0 (TS ¼ 0.6) 0.9259 6.9159 8.877
temperatures of 25  C,
10  C, and
20  C, the evaluation indices
10 (TS ¼ 0.3) 0.9828 3.2369 3.5628
values have not changed, significantly. However, in the cases of
20 (TS ¼ 0.5) 0.875 6.5511 9.9834
10  C and 0  C temperatures, the improvement of evaluation
20 (TS ¼ 0.6) 0.93 6.11 6.8883
GPR 0.9 25 (TS ¼ 0.2) 0.998 1.3409 1.4602
criteria is noticeable such that the MAEs have reduced to 5.5% and 10 (TS ¼ 0.4) 0.9817 3.8178 4.4068
9.7%, respectively. Moreover, one may observe the influence of TS 0 (TS ¼ 0.3) 0.9036 8.6819 9.9630
on the mentioned temperatures, as the TS increases, an increase in
10 (TS ¼ 0.3) 0.9838 2.7223 3.4537
R_squared, and a reduction in MAE and RMSE are observed.
20 (TS ¼ 0.2) 0.9880 2.7493 3.5530
Considering the corr_value of 0.9, The MAE for all the cases is
achieved with a value of less than 7%. Since the extracted features
are reliable, it is expected that the MAE and RMSE decrease, but on Fig. 8 showcases SOC’s predicted values versus f1 and f2, among
the contrary, they increase. This is because when highly correlated the highly correlated features.
features are selected, most of the other features are lost, and the The study results demonstrate that in addition to identifying
machine learning model may lose accuracy if the dataset is not big and extracting reliable features, the learning ability of the model
enough. Thus, the performance of the model over a dataset is of and partitioning of the data for training are highly crucial for pre-
importance. Although the linear regression model functions prop- cise prediction. Considering the above-mentioned elements’ ef-
erly for corr_values of 0.9, with a maximum error of 7%, a more fects, we also observed that the GPR model outperforms the linear
accurate and reliable model, i.e., GPR, is used for this corr_value. regression model. The proposed method will be implemented in
The GPR model results for corr_value of 0.9, and the best TS have BMS for online measurement of EIS and SOC prediction utilizing the
been presented in Table 2. The MAE for 25  C,
10  C, and
20  C potential approaches from Refs. [22e24] such as fractional-order
temperatures is less than 2.8%, but for 10  C and 0  C, the MAEs are equivalent circuit model (FOECM) and pseudo-random sequences
3.8% and 8.7%, respectively. As an example, the training and test (PRS), which are fast and easily implementable for measuring EIS at
data and their predicted values at different temperatures have been low measurement time and low complexity.
shown in Fig. 8.
5. Conclusions

In this investigation, the prediction of li-ion battery SOC using

EIS measurements was performed based on an ML approach. This
study was conducted based on extracting reliable features ac-
cording to their correlation value with the SOC of the battery. The
features are the impedances of EIS measurements over the range of
desired frequencies, i.e., from 1 mHz to 6 kHz. After selecting the
reliable features of different datasets at various temperatures and
different SOCs, the linear regression model and GPR were trained,
and the prediction was performed by the trained models over the
test set. Statistical indices such as R_squared, MAE, and RMSE were
used to evaluate the accuracy and robustness of the models. The
results indicated that the proposed models are able to precisely
predict the SOC of the battery using the reliable features. The
models trained by the features with corr_value of above 0.9 indi-
cated the best performance among the others, such that the error of
the GPR model was found to be less than 3.8%. Furthermore, the
impact of test-size on the model precision was evaluated. It was
observed that for some cases, the larger test-size results in higher
accuracy. Therefore, considering the online and onboard EIS mea-
surement, this method can be practically embedded in the BMS for
Fig. 7. Flowchart of the proposed model for predicting the SOC of the battery using EIS accurate measurements of SOC of li-ion batteries and ensure the
measurements. proper operation of PEVs.
7
I. Babaeiyazdi, A. Rezaei-Zare and S. Shokrzadeh Energy 223 (2021) 120116

Fig. 8. SOC prediction of the proposed model at temperatures of (a) 25  C, (b) 10  C, (c)
10  C, (d)
25  C.

Credit author statement p. 1e6.

[5] Sidhu MS, Ronanki D, Williamson S. State of charge estimation of lithium-ion
batteries using hybrid machine learning technique. In: Iecon 2019 - 45th
Iman Babaeiyazdi: Conceptualization, Methodology, Software, annual conference of the IEEE industrial electronics society, vol. 1; 2019.
Validation, Investigation, Writing e original draft, Visualization, p. 2732e7.
Afshin Rezaei-Zare: Supervision, writing e Reviewing & editing. [6] Sepasi S, Ghorbani R, Yann B. A novel on-board state-of-charge estimation
method for aged Li-ion batteries based on model adaptive extended Kalman fi
Shahab Shokrzadeh: Supervision, writing e Reviewing & editing. lter. J Power Sources 2014;245:337e44.
[7] Baccouche I, Jemmali S, Manai B, Omar N, Essoukri N, Amara B. Improved OCV
model of a Li-ion NMC battery for online SOC estimation using the extended
Declaration of competing interest
kalman filter. Energies 2017;10(6):764.
[8] Lee J, Nam O, Cho BH. Li-ion battery SOC estimation method based on the
The authors declare that they have no known competing reduced order extended Kalman filtering. J Power Sources 2007;174(1):9e15.
[9] Wang W, Wang X, Xiang C, Wei C, Zhao Y. Unscented kalman filter-based
financial interests or personal relationships that could have
battery SOC estimation and peak power prediction method for power distri-
appeared to influence the work reported in this paper. bution of hybrid electric vehicles. IEEE Access 2018;6:35957e65.
[10] Xia B, et al. A comparative study of three improved algorithms based on
Particle filter algorithms in SOC estimation of lithium ion batteries. Energies
References 2017;10(8):1149.
[11] Xu J, Mi CC, Cao B, Cao J. A new method to estimate the state of charge of
[1] Cai L, Meng J, Stroe D, Luo G, Teodorescu R. An evolutionary framework for lithium-ion batteries based on the battery impedance model. J Power Sources
lithium-ion battery state of health estimation. J Power Sources 2019;412: 2013;233:277e84.
615e22. July 2018. [12] Waag W, Kabitz S, Sauer DU. Experimental investigation of the lithium-ion
[2] Chen N, Zhang P, Dai J, Gui W. Estimating the state-of-charge of lithium-ion battery impedance characteristic at various conditions and aging states and
battery using an H-infinity observer based on electrochemical impedance its influence on the application. Appl Energy 2013;102:885e97.
model. IEEE Access 2020;8:26872e84. [13] Ng M, Zhao J, Yan Q, Conduit GJ, Seh ZW. Predicting the state of charge nad
[3] Shokrzadeh S, Jafari M, Bibeau E, Molinski T. A statistical algorithm for pre- health of batteries using data-driven machine learning. Nat. Mach. Intell.
dicting the energy storage capacity for baseload wind power generation in the 2020;2:161e70.
future electric grids. Energy 2015;89:793e802. [14] Xiong R, Cao J, Yu Q, He H, Sun F. Critical review on the battery state of charge
[4] Babaeiyazdi I, Rezaei-zare A, Shokrzadeh S. Fast charging systems to enable estimation methods for electric vehicles. IEEE Access 2017;6:1832e43.
electrification of Transportation : an operational constrained based analysis. [15] Lacroix M, Jokar A, Rajabloo B, Martin D. Review of simpli fi ed Pseudo-two-
In: 2019 IEEE transportation electrification conference and expo (ITEC); 2019.

8
I. Babaeiyazdi, A. Rezaei-Zare and S. Shokrzadeh Energy 223 (2021) 120116

Dimensional models of lithium-ion batteries. J Power Sources 2016;327: SPRINGER; 1999.

44e55. [31] Zhang Y, Tang Q, Zhang Y, Lee AA, Wang J, Stimming U. Identifying degra-
[16] Wu Y, et al. State of health estimation for lithium-ion batteries based on dation patterns of lithium ion batteries from impedance spectroscopy using
healthy features and long short-term memory. IEEE Access 2020;8:28533e47. machine learning. Nat Commun 2020;11(6e11).
[17] Zahid T, Xu K, Li W, Li C, Li H. State of charge estimation for electric vehicle [32] Makridakis R, Wheelwright S, Hyndman S. Forecasting methods and appli-
power battery using advanced machine learning algorithm under diversi fi ed cations. NewYork: Wiley; 1998.
drive cycles. Energy 2018;162:871e82.
[18] Donato THR, Quiles MG. Machine learning systems based on xgBoost and MLP
neural network applied in satellite lithium-ion battery sets impedance esti- Iman Babaeiyazdi was born in Yazd, Iran. He received B$Sc. from Amirkabir University
mation. Adv. Comput. Intell 2018;5:1e20. of Technology, Tehran, Iran, and his M.Sc. from York University, Toronto, Canada, both
[19] Hu C, Jain G, Schmidt C, Strief C, Sullivan M. Online estimation of lithium-ion in electrical engineering, in 2017 and 2019, respectively. Currently, he is perusing his
battery capacity using sparse Bayesian learning. J Power Sources 2015;289: Ph.D. in electrical engineering at York University.
105e13.
[20] Chang W-Y. The state of charge estimating methods for battery: a review. Int.
Sch. Res. Not; 2013. Afshin Rezaei-Zare (M0 08eSM0 10) received the B.Sc., M.Sc., and Ph.D. degrees (Hons.)
[21] Messing M, Shoa T, Ahmed R, Habibi S. Battery SoC estimation from EIS using in electrical engineering from the University of Tehran, Tehran, Iran, in 1998, 2000,
neural nets. In: 2020 IEEE transportation electrification conference & expo and 2007, respectively. He was a Postdoctoral Fellow from 2007 to 2009 at the Depart-
(ITEC). IEEE; 2020. p. 588e93. ment of Electrical and Computer Engineering, University of Toronto, Toronto, ON, Can-
[22] Guha A, Patra A. Online estimation of the electrochemical impedance spec- ada. From 2010 to 2017, he was with the Department of Special Studies, Hydro One
trum and remaining useful life of lithium-ion batteries. IEEE Trans. Instrum. Networks Inc., Toronto. In 2017, he joined the Department of Electrical Engineering
Meas. 2018;67(8):1836e49. and Computer Science, York University, Toronto, as associate professor. His research in-
[23] Crescentini M, et al. Online EIS and diagnostics on lithium-ion batteries by terests include power system resilience to geomagnetic disturbance, electromagnetic
means of low-power integrated sensing and parametric modeling. IEEE Trans. transients in power systems, and renewable and energy storage systems. Dr. Rezaei-
Instrum. Meas. 2020;9456(1). Zare is a registered Professional Engineer in the Province of Ontario, Canada, and an
[24] Sihvo J, Stroe D, Messo T, Roinila T. Fast approach for battery impedance Associate Editor of the IEEE Transactions on Power Delivery and IEEE Power Engineer-
identification using pseudo-random. IEEE Trans Power Electron 2020;35(3): ing Letters.
2548e57.
[25] Din E, Schaef C, Member S, Moffat K, Member S, Stauth JT. A scalable active
battery management system with embedded real-time electrochemical Shahab Shokrzadeh (M 0 15) received the B$Sc. degree in mechanical engineering from
impedance spectroscopy. IEEE Trans Power Electron 2017;32(7):5688e98. Iran University of Science and Technology, Tehran, Iran, in 2004, the M.Sc. degree in
[26] Huet F. A review of impedance measurements for determination of the state- production engineering and management from the Royal Institute of Technology
of-charge or state-of-health of secondary batteries. J Power Sources 1998;70: (KTH), Stockholm, Sweden, in 2009, and his PhD in mechanical engineering from the
59e69. University of Manitoba, Winnipeg, Canada in 2015. He was the co-founder and Asso-
[27] Kollmeyer P. Panasonic 18650PF Li-ion battery data. 2018. ciate Director of the International Renewable Energy Academy at York University in
[28] Tesla model S 18650 cell test data (March 30, 2015), Tesla Motors Club Dis- Toronto, Canada. He is currently a Project Manager for Alternative Energy in the Stan-
cussion Forum.”. dards Division at CSA Group in Toronto, Canada. Dr. Shokrzadeh has several years of
[29] Wang Q, He Y-J, Hu X, Ma Z-F. State of charge-dependent polynomial research and industry experience in the energy and automotive sectors, with a focus
equivalent circuit modeling for electrochemical impedance spectroscopy of on electrified transportation, renewable energy, and advanced energy storage techno-
lithium-ion batteries. IEEE Trans Power Electron 2018;33(10):8449e60. logies.Dr. Shokrzadeh is a registered Professional Engineer in Ontario.
[30] Brandt BS. Linear and polynomial regression. In: Data analysis. New York: