Received 1 May 2023, accepted 28 May 2023, date of publication 31 May 2023, date of current version 7 June 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3281552

Research on Electric Vehicle Charging Safety

Warning Based on A-LSTM Algorithm
XIAOHONG DIAO1 , LINRU JIANG1 , TIAN GAO 2 , LIANG ZHANG 3, (Member, IEEE),
JUNYU ZHANG2 , LONGFEI WANG2 , AND QIZHI WU2
1 BeijingElectric Vehicle Charging Engineering Technology Research Center, China Electric Power Research Institute, Beijing 100192, China
2 Schoolof Electrical Engineering, Northeast Electric Power University, Jilin 132012, China
3 Key Laboratory of Modern Power System Simulation and Control and Renewable Energy Technology, Ministry of Education, Northeast Electric Power

University, Jilin 132012, China

Corresponding author: Liang Zhang ([email protected])
This work was supported by the 2022 Beijing Electric Vehicle Charging Engineering Technology Research Center Open Fund Project
‘‘Research on Electric Vehicle Adaptive Charging Safety Early Warning Model Based on Deep Learning’’ under Grant YDB51202101417.

ABSTRACT Accidents involving electric vehicle fires have increased as the number of electric vehicles
has grown recently. The issue of charging safety is a key barrier to the growth of the electric vehicle sector
because these accidents have resulted in large financial losses for car owners and charging facility operators.
The approach for resolving the issue of electric car charging safety through an electric vehicle charging
safety warning system is suggested in this research. The suggested solution uses an adaptive optimization
of long short-term memory neural network (A-LSTM) to forecast voltage changes throughout the whole
charging process by using the vehicle’s daily historical charging data. The warning threshold adjustment
method is established by the difference between the predicted voltage data and the actual voltage data, which
is dynamically adjusted as the charging process progresses. Finally, a real-time warning model for vehicle
charging alert is developed. The daily charging data of electric vehicles is used in the paper to verify the
precision of data prediction and the accuracy and timeliness of the model. The study’s findings demonstrate
that the early warning model suggested in this paper can quickly send out early warning signals to safeguard
the safety of car charging and can identify aberrant charging data.

INDEX TERMS Electric vehicle, charging safety, early warning, A-LSTM algorithm, daily charging data.

I. INTRODUCTION of digitalization and intelligence in the automotive industry

The ecological and energy crises are becoming increasingly is becoming more and more obvious. The use of big data
prominent on a global scale. Compared with traditional fuel research to solve the safety issues of EVs has become an
vehicles, electric vehicles (EVs) have significant advantages important pathway. The deep integration of EVs and big data
in saving oil resources and reducing carbon emissions. They will undoubtedly accelerate the transformation of automotive
have received attention from governments and automotive safety regulatory technology, thereby further promoting the
companies worldwide, and the number of EVs in use has high-quality development of China’s EV industry [5], [6].
continued to grow [1], [2]. However, the frequent occurrences At present, the research on the safety warning and fault
of spontaneous combustion and fires in EVs have caused diagnosis of EV charging process has just started. Generally,
serious economic losses to car owners and charging facility the traditional research for EV fault diagnosis and warning is
operators, and charging safety issues have hindered the devel- done by constructing the electrochemical model of the bat-
opment of EVs and related industries [3], [4]. With the full tery. Seo et al. used a recursive least squares method to detect
implementation of the national big data strategy, the trend internal short-circuit faults in batteries based on an equivalent
circuit model with the battery open-circuit voltage and state
The associate editor coordinating the review of this manuscript and of charge (SOC) as inputs [7]. Zhang et al. electrochemical
approving it for publication was Junho Hong . model of the power battery is constructed, and the battery
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 11, 2023 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 55081
 X. Diao et al.: Research on EV Charging Safety Warning Based on A-LSTM Algorithm

status is judged by comparing the charge response informa- early warning of thermal runaway [18]. The aforementioned
tion simulated by the battery model and the battery charge sta- methods based on big data and machine learning are able
tus information for charging fault monitoring and early warn- to handle time-series EV charging data, but they are mostly
ing [8]. Tran et al. comprehensive consideration of SOC, tem- based on laboratory data, and less research has been con-
perature and state of health (SOH) establishes an equivalent ducted on daily EV charging data.
circuit model for lithium-ion batteries, which is capable of In summary, big data research and machine learning algo-
estimating the battery state of charge, temperature and health rithms are widely used in the field of fault diagnosis and
status with high accuracy and can be effectively monitored early warning, however, electric vehicle charging data is time-
by the battery management system (BMS) [9]. In addition, series data, and the data volume is huge, the input dimension
some scholars have used the expert system approach for fault is many, and the coupling is strong, some deep learning
detection and early warning. Song designed a comprehensive algorithms are not well applied in this scene. The time series
evaluation index system for charging safety based on expert warning method based on Long Short-Term Memory (LSTM)
scoring and other methods from three aspects: power battery, can predict the future development trend of the data based
charging equipment and distribution network, and use the on current and historical data features, which is suitable for
gray correlation degree method to determine the weight of vehicle charging data warning requirements. Therefore, this
each index [10]. Qian et al. designed a safety and protection paper is based on LSTM algorithm to build the vehicle charg-
monitoring device for electric vehicle charging, and then ing safety warning model. Firstly, we analyze and summarize
established an early warning model based on the operation the factors affecting vehicle charging safety by combining
status of charging equipment by fuzzy comprehensive evalu- the characteristics of vehicle charging data and the charging
ation method [11]. EV battery systems are strongly nonlinear characteristics of lithium-ion battery, and filter the charging
and it is difficult to build an accurate model for them. Fearing safety warning factors according to the vehicle charging data
that a model is usually only applicable to a specific fault characteristics by using Pearson correlation analysis to cal-
type, it requires a lot of modeling work and is inconvenient culate the correlation coefficients between different warning
in engineering. factors; then we construct the A-LSTM prediction algorithm
In recent years, the widespread use of big data and machine and design the prediction model of charging data, and get the
learning has also led to new approaches to fault monitoring predicted values of electric vehicle charging data by analyz-
and early warning. Zhao et al. proposed a fault diagnosis ing and learning from the historical data; finally, we combine
method for electric vehicle battery system based on big the A-LSTM prediction algorithm and design the prediction
data statistical method to construct a more complete bat- model of charging data. Finally, the dynamic warning thresh-
tery system fault diagnosis model based on machine learn- old method is constructed by combining the predicted value,
ing algorithm and multi-level screening strategy to detect actual value and charging standard, and the real-time warning
abnormal changes of voltage [12]. Zhang et al. Constructing model of EV charging safety is designed. Finally, the dynamic
a charging warning model based on improved grey wolf warning threshold method is constructed by combining the
optimization-back propagation neural network (IGWO-BP) predicted values, actual values and charging standards to
to accurately identify abnormal EV charging voltage condi- design a real-time warning model for EV charging safety.
tions for diagnosis and warning [13]. Xia et al proposed a The daily charging data are analyzed and verified to prove
short-circuit fault diagnosis method based on voltage profile the accuracy and timeliness of the charging safety warning
correlation coefficients to detect short-circuit faults by cap- model constructed in the paper.
turing the decreasing trend of voltage correlation coefficients
of two adjacent cells individually and using recursive sliding II. EV CHARGING SAFETY ANALYSIS
windows to maintain the sensitivity of fault detection during A. ANALYSIS OF EV ACCIDENTS
operation. [14] Gao et al. constructed an electric vehicle There are significant differences between combustion acci-
charging process fault warning method based on adaptive dents in new energy vehicles and those in traditional fuel
deep belief networks, which can accurately warn of faults dur- vehicles. Combustion accidents in traditional vehicles are
ing the charging process of electric vehicles [15]. Zhang et al. typically caused by non-standard modifications, aging cir-
proposed a new method for quantifying electrical safety index cuits, arson, and other factors, whereas most combustion
system based on comprehensive weights, which is suitable for accidents in new energy vehicles are caused by thermal run-
application to electrical safety protection of electric vehicle away [19]. The fire scenarios caused by thermal runaway
charging equipment [16]. Yu combines BP neural networks in EVs mainly include five types: lithium-ion battery charg-
with outlier testing to achieve charging safety monitoring ing combustion, natural parking combustion, driving com-
and fault warning for electric vehicles, and is capable of bustion, collision combustion, and immersion combustion.
alerting when charging faults occur in electric vehicles [17]. According to statistics [20], the accident ratios for these
Yang et al. proposed a current prediction method based on scenarios are shown in Fig. 1, with the highest proportion of
back propagation neural network to estimate the current of combustion accidents occurring when EVs are being charged,
cells with external short circuit faults in the battery pack by parked, or driven. Generally, there are fewer supervisory
neural network using only voltage information to achieve personnel around vehicles in the charging and parking states
55082 VOLUME 11, 2023
X. Diao et al.: Research on EV Charging Safety Warning Based on A-LSTM Algorithm

than during driving. Therefore, it is necessary to improve the period of time, causing the internal temperature of the battery
safety warning system for EV charging and protect the safety to rise sharply, which will lead to charging accidents. From
of charging EVs. the perspective of the triggering mechanism, there are three
main types of internal short circuits in batteries: internal short
circuits caused by overcharge or over discharge, internal short
circuits caused by mechanical damage and self-triggered
internal short circuits [23].

2) BATTERY OVERCHARGE
Overcharging of a battery can be a dangerous condition where
the charging equipment continues to supply energy to the
battery for an extended period of time. This can lead to
safety accidents. There are several factors that can cause
battery overcharging, including high ambient temperatures
and incorrect charging methods [24].

3) BATTERY SEPARATOR AND ELECTROLYTE MATERIAL

FIGURE 1. EV accident scene diagram.
The battery separator serves the critical function of isolating
the positive and negative electrodes to prevent short circuit
B. ANALYSIS OF EV CHARGING METHOD accidents that may arise from the penetration of the separator
There are two main charging methods used for EVs in the during the electrochemical reaction. However, as the battery
market today, one is Direct Current (DC) fast charging and undergoes cycles of charge and discharge, the electrolyte
the other is Alternating Current (AC) slow charging. DC fast degrades, leading to a decline in the battery’s overall perfor-
charging mainly uses the high-power DC charger of the mance over its life cycle [25].
charging pile to directly charge the EV power battery, which
has a higher charging current and requires a short charging 4) BATTERY PACK CONSISTENCY
time. AC slow charging relies on the on-board charger inside Due to differences in production technology and daily usage,
the EV to convert the AC power provided by the charging individual battery parameters of EV battery packs may
equipment to DC power for charging the power battery. This become inconsistent, and there may be differences between
charging method has a lower charging current but requires a individual batteries due to internal decay effects of the battery
longer charging time, which generally causes less chemical pack. The inconsistency of battery packs can lead to differ-
damage to the battery and can prolong the battery’s service ences in SOC, voltage, SOH, etc. between individual cells,
life [21], [22]. The voltage and temperature trends of an which can seriously affect normal use [26].
on-board battery pack change when a vehicle is charged using In summary, it is evident that EV spontaneous combustion
different charging techniques. Therefore, in order to develop accidents are primarily caused by thermal runaway issues
effective charging warning models, it is important to study resulting from internal battery short circuit, battery over-
fast charging and slow charging vehicles separately. charge, and battery pack inconsistency. Therefore, this paper
integrates vehicle charging data and charging safety influ-
C. ANALYSIS OF THE FACTORS AFFECTING encing factors and selects real-time battery pack SOC, initial
THE SAFETY OF EV CHARGING SOC, battery pack SOH, charging current, maximum temper-
Numerous factors are implicated in the safety of EV charging. ature of single battery, and maximum voltage of single battery
Typically, mistreatment of the battery can cause a sudden as EV safety warning factors. Battery pack SOC and charging
rise in its internal temperature [7], which if exceeds a critical voltage current temperature can represent the charging status
threshold, may result in the melting of the separator, decom- of EV, SOH can represent the aging condition of the battery
position of the positive electrode material, and electrolyte pack, and the aforementioned early warning factors can be
oxidation within the battery, ultimately leading to a violent used to determine whether the EV is malfunctioning or not.
combustion event and subsequent thermal runaway. While
the phenomena may vary slightly depending on the specific III. MODE DESIGN
trigger, the underlying mechanisms are largely analogous. The safety warning model is developed based on the charac-
In the following, we will provide a comprehensive overview teristics of EV charging data, comprising five distinct steps.
of the commonly observed influential factors. The first four steps involve offline processing to establish
EV charging warning thresholds, while the fifth step involves
1) INTERNAL SHORT CIRCUIT IN BATTERY online comparison of EV charging data with the warning
When a battery is internally short-circuited for some reason, thresholds to achieve status detection and warning. The model
a huge current will be generated inside the battery in a short operates as follows:

VOLUME 11, 2023 55083

 X. Diao et al.: Research on EV Charging Safety Warning Based on A-LSTM Algorithm

Step 1: EV charging data processing. Firstly, the collected TABLE 1. EV charging safety warning factor.
EV historical charging data are subjected to de-hybridization
operation to remove missing values and great outliers from
the data, and then the data are normalized according to equa-
tion (1) to make the EV historical charging data vector EVhis
a standard data mapped to [−1, 1]. This can prevent the
subsequent calculation errors caused by data changes, and at
the same time can improve the operation speed and prediction
accuracy of the early warning model.
2 (EVhis − EVhis-min )
EVhis-input = −1 (1)
EVhis-max − EVhis-min
where EVhis-input is the normalized historical data value,
to obtain the EV warning factor prediction vector EVpre
which is used as the standard input data in the subsequent
and EV warning factor pressure residual vector EVre , which
steps; EVhis is the original charging data; EVhis-max and
lay a good foundation for the subsequent charging warning
EVhis-min are the maximum and minimum values of the cor-
threshold setting.
responding data in the original charging data.
Step 2: EV charging warning factor calculation. First, EVpre = fA-LSTM (EVinput ) (4)
the EV charging safety warning factor rwf is determined by
EVre = EVpre − EVnor (5)
considering the EV charging history data types and the EV
charging safety influencing factors summarized in the intro- Step 4: EV dynamic warning threshold setting. Accord-
duction section, and then the corresponding data in EVhis , ing to the EV prediction data residual EVre obtained
the EV charging history data set, are extracted according in step 3 and the ‘‘Electric Vehicle Safety Requirements
to the warning factor, and the correlation coefficient ck (GB 18384-2020)’’, the vehicle charging voltage threshold
among the influencing factors is calculated by the Pearson array EVthr is established with the data characteristics. then
correlation coefficient formula shown in equation (2) and the vehicle charging process is divided into 4 different warn-
equation (3), and finally the correlation coefficients among ing regions according to the SOC value, I, II, III and IV, con-
the factors are compared, and the one with the strongest sidering the changing characteristics of the vehicle charging
correlation with other factors is selected as the charging safety data, and each region is set up with the warning The threshold
warning factor rwf . adjustment factor w, w is dynamically adjusted according to
n
P different regions, and equation (6) constructs the dynamic
(xi − x)(yi − y) warning threshold EVthr-dy for EV charging warning factor.
i=1
cxy−i = s (2) Xn
n n EVthr-dy = wj EVthr−j (6)
− x)2 − y)2
P P
(xi (yi j=1
i=1 i=1
Xm EVthr = ±λ · EVre (7)
ck = cxy−i (3)
i=1 where j is the charging warning region serial number, taking
where i is the data number of the warning factor, n is the the value of [1] and [4], wj is the warning threshold adjust-
number of input data, x and y are two different warning data. ment factor of region j, and λ is the adjustment factor, taking
cxy−i shown in Table 1 indicates the correlation coefficient the value between [−1.15, 1.15] according to the specific data
between warning data x and y, x and y are the average value variation.
of the corresponding warning data set, k is the warning factor Step 5: EV charging safety warning. Input EV real-time
serial number, m is the number of warning factors, and m is charging data EVrt , record the initial SOC of EV charging
taken as 6 according to Table 1. as SOC0 and the highest voltage of vehicle battery as V0 ,
Step 3: EV charging data prediction. Extract EV standard determine the vehicle warning area according to the initial
charging data EVhis-input to select suitable parameters as input state of vehicle charging, then select the dynamic warning
EVinput , take the predicted data of charging factor as output threshold of the corresponding area and monitor the vehicle
EVpre , and build A-LSTM deep network for regression cal- charging status, when the vehicle charging status is abnor-
culation of the data. Observe the prediction results, gradu- mal, i.e., the changes of real-time charging voltage, charging
ally adjust the hyperparameters of the algorithm, explore the current and temperature are different from the safety model
balance between the amount of input data and the prediction constructed based on the historical normal charging data,
accuracy of the algorithm, while ensuring the optimal pre- consider that there is a safety problem of EV charging at this
diction accuracy and prediction time, determine the optimal time, implement the corresponding warning rules in time and
hyperparameters of the algorithm, form the A-LSTM deep take measures to deal with it. When the vehicle charging state
learning network model, and then use the A-LSTM algorithm is abnormal, i.e., the real-time charging voltage, charging
to fit the data, as shown in equation (4) and equation (5) current and temperature changes are different from the safety

55084 VOLUME 11, 2023

X. Diao et al.: Research on EV Charging Safety Warning Based on A-LSTM Algorithm

model built based on the historical normal charging data, it is has proven effective in analyzing and learning input vehicle
considered that there is a safety problem in EV charging at historical charging time series data and forecasting the trend
this time, and the early warning is carried out in time and of subsequent charging data based on the historical and cur-
measures are taken to deal with it. rent state. Therefore, this study adopts LSTM to address the
The specific EV charging safety warning rules are as fol- data regression prediction challenge and further optimizes the
lows: problem solution by constructing an A-LSTM model.
The first level is EV normal charging state, the vehicle
charging data in this state is below the warning threshold, A. LSTM ALGORITHM
which is the most ideal charging state and no alarm will be The LSTM neural network is a variant of Recurrent Neural
made; Network (RNN) that addresses the problems of gradient van-
The second level is EV warning state, the EV SOC in this ishing and explosion. It replaces the hidden layer of the orig-
state is below 60% and the maximum battery voltage is higher inal RNN with LSTM units, which contain input, output, and
than the warning threshold, the state of this level indicates that forget gates. The forget gate regulates the amount of historical
the EV charging is abnormal and protective measures should input by controlling which information is retained and which
be taken in time to make it converge to normal charging; is discarded. The activation function of all three gates is the
The third level for the electric vehicle alert state, the state sigmoid function, which produces values between 0 and 1.
of the vehicle SOC in 60%-80%, the vehicle single battery The gates learn to weight the historical input, current input,
maximum voltage is higher than the warning threshold, at this and historical output, thereby achieving the memory function
time the vehicle state may have risk, should be alarmed and of historical input and output.
timely measures to prevent the vehicle from danger; The LSTM unit is constructed as shown in (8) to (13), with
The fourth level is EV dangerous state, the state of the the candidate LSTM memory cell state value represented by:
vehicle SOC is above 80%, the highest voltage of the vehicle
C̃ (t) = tanh (ωx − cx (t) + ωhc h (t − 1) + bc ) (8)
single battery is higher than the warning threshold and lasts
for a long time, the EV continues to charge in this state is where: x(t) is the input data of the historical charging of the
very likely to burn the car accident, should promptly cut off EV at the current moment, h(t − 1) is the output of the LSTM
the power and stop charging. unit at the previous moment, ωx and ωhc are the connection
In summary, the EV charging safety warning model is weights corresponding to the two inputs x(t) and the output
constructed, and the specific model operation block diagram h(t − 1), C̃ (t) is the memory unit reference, and bc is the
is shown in Figure 2. bias of the network. Values of the input gates of the LSTM
network:
ωxi x (t) + ωhi h (t − 1)
 
I (t) = sigmoid (9)
+ωci C (t − 1) + bi
where: ωxi , ωhi and ωci are the input data for the current
moment of EV historical charging, the previous moment
LSTM cell output and the previous moment cell output con-
nection weights to the input gate, respectively, and bi is the
bias of the input gate.
The values of the forgotten gates of the LSTM network:
ωxf x (t) + ωhf h (t − 1)
 
F (t) = sigmoid (10)
+ωfi C (t − 1) + bf
where: ωxf , ωhf and ωcf are the EV historical charging input
data at that moment, the previous moment LSTM cell output
and the previous moment cell output connection weights to
the forgetting gate, respectively; bf is the bias of the forgetting
gate.
FIGURE 2. Flow chart of EV charging safety warning model. In this way, the current LSTM memory cell state value:
C (t) = F (t) ⊗ C (t − 1) + I (t) ⊗ C̃ (t) (11)
IV. EV CHARGING DATA PREDICTION MODEL where ⊗ denotes the residence product operation.
The third step of the charging safety warning model con- And the value of the output gate of the LSTM network:
structed in Section III, accurate prediction of the complete
ωxo x (t) + ωho h (t − 1)
 
change pattern of EV charging data is crucial. It is apparent O (t) = sigmoid (12)
that EV charging data series represent standard time series +ωco C (t − 1) + bo
data, and predicting the variable constitutes a regression where: ωxo , ωho and ωco are the connection weights of the
problem for time series analysis. LSTM algorithm model current moment’s input, the previous moment’s LSTM cell

VOLUME 11, 2023 55085

 X. Diao et al.: Research on EV Charging Safety Warning Based on A-LSTM Algorithm

output and the previous moment’s cell output to the output the battery pack of the EV. Each individual cell has a rated
gate, respectively, and bo is the bias of the output gate. voltage of 3.7V and a rated capacity of 2.2Ah. Therefore,
It is feasible to determine the output of the LSTM memory the rated voltage and capacity of the EV battery pack are
cell at a given time t by combining equations (8) to (13). 328.5V and 69Ah, respectively. Throughout the entire charg-
ing process, the EV adopts a typical three-stage charging
h (t) = O (t − 1) ⊗ tanh (C (t − 1)) (13) method, with slow charging before the battery pack’s SOC
In summary, the working process of LSTM can be simplified reaches 80% using constant current charging. The charging
as follows: given the input value x(t) at the current time step, termination voltage of the vehicle’s battery system is 377V,
the information is filtered through candidate memory cells, and the discharge termination voltage is 276V. The operating
under the control of the input gate, to update the current temperature range for charging is between 0 to 50◦ C, and
memory cell. The forgetting gate determines whether the for discharging, it is between −20 to 60◦ C. The maximum
current memory cell can access information from the previous continuous charging current allowed is 35A, and the max-
cell, and the valuable information retained by these two parts imum continuous discharge current is 103A. The charging
(i.e., the updated memory) is passed to the next LSTM. The protection voltage for each individual cell is 4.1V, and the
output gate controls whether the information in the memory discharge protection voltage is 3V.
cell is transmitted to the hidden state for use in the output The experimental EV operating data is presented in
layer, and h(t) is also connected to the next LSTM cell Table 2, which includes fundamental information on vehicle
module. The interaction and control of the three gates allows operation and data related to the vehicle battery system.
for the longer-term memory of the input information.
TABLE 2. Summary of EV Charging Data.

B. A-LSTM ALGORITHM
To enhance the precision of the LSTM model’s predictions
and mitigate its error, this study utilizes the error corre-
lation linear analysis approach and proposes the A-LSTM
(Adaptation LSTM) model. This model involves developing
a relationship equation (13) that correlates the input variables
with the historical prediction error.
eh = f (x1 , . . . , xn ) (14)
where: eh represents the LSTM historical prediction error;
f (x1 , . . . , xn ) is a primary function on the input, (x1 , . . . , xn
denotes the input).
The prediction model after error correction is shown in
equation (14).
hA-LSTM = f (x1 , . . . , xn ) + h (x1 , . . . , xn ) (15)
where: h(x1 , . . . , xn ) is the established LSTM prediction B. EV CHARGING DATA PRE-PROCESSING
model; f (x1 , . . . , xn ) is the error linear correction function, In this study, a total of 153,247 vehicle charging history data
and hA-LSTM is the final output of the A-LSTM algorithm. are collected, with a data time span of 1 year, specifically
containing the basic data of vehicle-pile communication and
V. EXPERIMENTAL VERIFICATION AND ANALYSIS the data type required for the model, which can depict the
To validate the proposed EV warning model, this paper vehicle charging scenario more completely. External factors,
selected two types of daily charging data from vehicles for such as measurement errors of vehicle sensors and trans-
charging characteristic analysis and experimental simulation mission errors of data, may lead to null values, abnormally
verification. The charging data of vehicles using fast charging large values, and repeated values in historical charging data.
and slow charging were collected separately. The selected Moreover, due to the high data collection frequency, multiple
vehicles in this study have a demonstrative effect and can charging data correspond to the same SOC value when the
serve as a reference for developing other EV charging safety vehicle SOC accuracy is kept to 1 decimal place. There-
warning models. fore, data de-aggregation is required. After screening and
eliminating the abnormal data and blank data, the remaining
A. EV CHARGING DATA valid charging data are 98581. Then, the data were normal-
The EV battery system used in this study consists of ized to meet the input requirements of the model prediction
18650 type ternary lithium-ion batteries. The EV is composed algorithm.
of 92 individual cells connected in series to form one group, After data pre-processing, extract the relevant charging
and a total of 32 groups are connected in parallel to form data, and equations (2) and (3) were used to calculate the

55086 VOLUME 11, 2023

X. Diao et al.: Research on EV Charging Safety Warning Based on A-LSTM Algorithm

correlation coefficients c between the real-time SOC of the the stable motor temperature and the battery temperature.
battery pack, the initial SOC of the battery pack charging, The correlation coefficient of each parameter was then recal-
the SOH of the battery pack, the charging current of the culated, as presented in Figure 5. Upon comparing Figure 3
battery pack, the maximum temperature of the single cell, and Figure 5, it is observed that the correlation coefficients
and the maximum voltage of the single cell. The purpose was between temperature (after the improved formulation) and
to investigate the degree of coupling between these factors other factors, except SOH, are significantly improved. This
and to determine the charging safety influence factor rwf . The is due to the narrow SOH range (94% to 96.5%) of the EV
results of these calculations are presented in Fig. 3. data obtained in this study, which is not enough to cover the
The strength of the correlation between two variables is entire SOH cycle.
indicated by the absolute value of the Pearson correlation Upon comprehensive observation of the correlation coef-
coefficient, with a value closer to 1 indicating a stronger ficients between the factors compared in Fig. 5, it is evident
correlation. Fig. 3, which displays the correlation coefficient that the correlation between single cell voltage and the other
heatmap, indicates that the real-time SOC of the battery pack factors is stronger when considered together. Therefore, sin-
is strongly correlated with both the EV charging current and gle cell voltage is selected as the warning factor.
the highest single cell voltage of the battery pack. The health
of the battery pack is strongly correlated with the initial SOC
of the battery pack charging. In contrast to the exposition of
related literature, the maximum temperature of the single cell
of the battery pack is weakly correlated with other factors.
Upon ranking, it was discovered that the maximum tempera-
ture of the single cell of the battery pack is more significantly
impacted by the ambient temperature, and that it changes
more under different ambient temperature conditions.

FIGURE 4. Charging temperature curve of vehicle motor and battery in

January and August.

FIGURE 3. Pearson coefficient between parameters.

In order to consider the influence of ambient tempera-

ture, this paper selects the vehicle temperature in winter
and summer when the ambient temperature difference is
large to study, and integrates the EV motor temperature and
battery temperature change as shown in Fig. 4. As evident
from the figure, the difference between temperature values FIGURE 5. Pearson coefficient after improved temperature expression.
in January and August is significant, regardless of whether
it is the battery or motor temperature. This is also why
temperature has a weak correlation with other factors. The C. CHARGING VOLTAGE PREDICTION
motor temperature curve demonstrates that it changes less To construct a sound algorithm model, it is crucial to deter-
and stabilizes gradually after a period of charging, while the mine the primary hyperparameters of the A-LSTM algorithm,
battery temperature changes continuously during charging. which includes the number of neural units in the input and
Hence, this paper uses the stable motor temperature as the output layers, the number of layers in the hidden layer, and
reference for the ambient temperature and reconstructs the the number of neural units in the hidden layer. Addition-
battery temperature by calculating the difference between ally, we have selected the root mean square error sRMSE

VOLUME 11, 2023 55087

 X. Diao et al.: Research on EV Charging Safety Warning Based on A-LSTM Algorithm

TABLE 3. Results of MAE with different parameters of A-LSTM.

TABLE 4. A-LSTM RMSE results with different parameters.

(Root Mean Square Error, RMSE) and mean absolute per- to include three hidden layers, with each layer containing
centage error sMAE (Mean Absolute Error, MAE) as the 100 units. The remaining hyperparameters of the A-LSTM
evaluation criteria to assess the prediction performance of the model are provided in Table 5.
A-LSTM algorithm under different parameters. The specific
formulas for these two evaluation criteria are shown in (16) TABLE 5. Hyperparameter setting of A-LSTM model.
and (17).
v
u n ,
uX
sRMSE = t (yi − fi )2
n (16)
i=1
n
1 X
sMAE = |yi − fi | (17)
n
i=1

where: fi is the predicted value of the data, yi is the original

value of the data, and i is the ordinal number of the data (i = Once the parameters of the A-LSTM model were deter-
1, 2, . . . , n). mined, the algorithm was utilized to fit the historical charging
The input for the A-LSTM algorithm consists of the initial voltages of the vehicle. A subset of the prediction results
SOC, real-time SOC, battery pack SOH, charging current, obtained from the test set are presented in Fig. 6, and Fig. 7
and temperature during EV charging, resulting in an input illustrates the absolute error between the predicted and actual
dimension of 5. The maximum voltage of single cell during values of the selected dataset.
charging is selected as the output, resulting in an output
dimension of 1.
To identify the optimal parameters, the number of implied
layers of A-LSTM is selected as 1-4 layers in turn, and the
number of neurons per layer of implied layers is 50, 100, 150,
200 to find the optimal parameters. The prediction results and
evaluation indexes are computed for each set of parameters.
The output results of the two evaluation indexes are presented
in Table 3 and Table 4.
The results presented in Tables 3 and 4 indicate that the FIGURE 6. EV charging data prediction error.
optimal evaluation metrics for the A-LSTM model were
achieved when the number of hidden layers was set to three From Figures 6 and 7, it is evident that the proposed
and each layer contained 100 hidden units. Based on this algorithm yields high accuracy in prediction the maximum
finding, the structure of the A-LSTM model was finalized voltage of the individual battery during the charging of EVs.

55088 VOLUME 11, 2023

X. Diao et al.: Research on EV Charging Safety Warning Based on A-LSTM Algorithm

determination (R2 ), as shown in equation (18). The specific

prediction results are shown in Fig. 9.
(fi − yi )2
P
i
R2 = 1 − 2 (18)
n

P 1 P
yi − n yi
i i=1
where fi is the predicted value of the data, yi is the original
value of the data, and i is the ordinal number of the data
(i = 1, 2, . . . , n).
As depicted in Figure 9, it can be observed that for the
EV data in the test set, the A-LSTM algorithm prediction
results are evenly distributed in the center of the graph, with
a maximum R2 value of 0.99781. This indicates that the
algorithm can more accurately predict voltage changes during
EV charging, providing reliable prediction data and residual
data for the subsequent charging warning threshold setting.
FIGURE 7. Comparison of prediction results.
To further compare the performance of LSTM and
A-LSTM algorithms, we calculated the prediction residuals
using the EV test set data. The distribution histogram was
This facilitates the prediction of the entire voltage change plotted as shown in Figure 10. It is observed that the resid-
process during vehicle charging. The variation in initial SOC uals of the A-LSTM algorithm are primarily distributed in
from 0 to 99.9% during charging is illustrated in Fig. 8. the range of −0.01238 to 0.006933, whereas the prediction
The results obtained from the A-LSTM algorithm developed residuals of the LSTM algorithm are distributed in the range
in this study demonstrate excellent prediction performance, of −0.01619 to 0.009573. This indicates that the A-LSTM
completely covering the state of charge for each type of charg- algorithm exhibits higher prediction accuracy than the LSTM
ing, thereby highlighting the exceptional predictive accuracy algorithm.
of the proposed prediction algorithm.

FIGURE 8. EV charging data forecast results.

To verify the prediction accuracy of the A-LSTM model

proposed in this paper, we selected BP neural network, RNN FIGURE 9. Comparison of R2 values tested by different algorithms.
neural network, LSTM neural network, and A-LSTM neural
network to fit the charging voltage of the vehicle. The BP
neural network was set up with 3 fully connected layers, D. EV CHARGING SAFETY WARNING EFFECT
while the RNN, LSTM, and A-LSTM neural networks were To establish the vehicle warning threshold, we utilized the
set up with 3 layers, and the number of hidden units in dynamic threshold model after obtaining the vehicle charging
each layer was set to 100, respectively. We compared the prediction value. Specifically, we selected a charging dataset
prediction results of these algorithms using the coefficient of and constructed the warning threshold based on the dynamic

VOLUME 11, 2023 55089

 X. Diao et al.: Research on EV Charging Safety Warning Based on A-LSTM Algorithm

As shown, the warning threshold of voltage changes over

time and gradually shrinks with an increase in SOC range.
Compared to a fixed threshold, a dynamic threshold is more
flexible in warning vehicles during charging while ensuring
safety, thereby maximizing the safety of the vehicle charging
point.

FIGURE 11. Dynamic warning threshold.

FIGURE 10. Histogram of prediction residuals for different algorithms.

threshold construction rules illustrated in Figure 11. This

construction approach divided the charging curve into four
regions based on voltage change characteristics.
In warring area I, where the SOC is between 10% and 30%,
the voltage rises rapidly while remaining at a low level.
To accommodate this situation, we set the warning threshold
to ±(1.12%∼1.16%) of the normal value. As in warring
area II (SOC 30%∼60%), the voltage gradually rises, and
its increase slows down. we set the warning threshold to
±(1.08%∼1.12%) of the normal value. Consequently, the
warning threshold is smaller than the previous area. And
in warring area III (SOC 60%∼80%), the voltage is high, FIGURE 12. Dynamic warning threshold after adjustment.
but its increase slows down even further, resulting in an
increased risk of overcharging. Here, we set the threshold to As the collected EV charging data did not include any fault
±(1.05%∼1.08%) of the normal value. Finally, in Region IV data, we simulated three types of vehicle fault state data to
(SOC 80%∼100%), the battery is near completion, and the assess the accuracy of the early warning model. To generate
voltage is about to reach its peak, making it susceptible to this data, we used the fault data setting method outlined
overcharging, we set the threshold to ±(1.03%∼1.05%) of in Table 6. And then we calculated the accuracy of the model
the normal value. If any abnormality is detected in the battery, using equation (19).
the warning threshold range will be further contracted.
Figure 12 illustrates how we adjusted the warning thresh- wa = (Ndet /Nerr ) ∗ 100 (19)
old after detecting abnormal data in warning area III. Verr ∈ (rand (−0.1 ∼ 0.1) ∗ Vnor ) (20)

55090 VOLUME 11, 2023

X. Diao et al.: Research on EV Charging Safety Warning Based on A-LSTM Algorithm

TABLE 6. Simulated vehicle failure data.

FIGURE 14. Fixed threshold warning accuracy.

For fault types 2, 3, and 4, the proposed warning model in

Verr ∈ (rand (−0.15 ∼ 0.15) ∗ Vnor ) (21) this paper can accurately indicate the charging faults of vehi-
Verr ∈ (rand (−0.25 ∼ 0.25) ∗ Vnor ) (22) cles, with an average accuracy rate of 99.0%. By comparing
Fig. 13 with Fig. 14, it can be concluded that the dynamic
where wa is the warning accuracy, Ndet is the number of fault warning thresholds proposed in this paper can improve the
data detected by the model, and Nerr is the number of fault accuracy of charging safety warnings for all four types of
data set in this paper, Verr represents the fault voltage data faults. For fault types 2 and 4, the dynamic threshold can
set, Vnor represents the normal charging data, and ‘rand()’ improve the warning accuracy rate by an average of 2.51%.
represents a function that generates a random number. Particularly for the third charging fault type, the dynamic
To validate the accuracy of the proposed charging warn- threshold can improve the average warning accuracy rate
ing model, we used the predicted data from the BP neural by 5.84% compared to using a fixed threshold. This is because
network, RNN neural network, LSTM neural network, and the warning model constructed in this paper can accurately
A-LSTM neural network (as discussed in section IV-A) to and strictly monitor and warn the subsequent charging pro-
construct fixed warning thresholds and dynamic warning cess by dynamically reducing the warning threshold after the
thresholds. We then ran the warning models in a sequence and fault is identified.
recorded the warning accuracy. The results of our analysis are Another element in evaluating the warning model is the
presented in Fig. 13 and Fig. 14. warning time. In this study, the 1SOC, as shown in equa-
tion (23), is chosen as the measure of the vehicle warning
time, where a lower 1SOC value indicates better warning
results.
1SOC = SOCerr − SOCdet (23)
where SOCerr represents the SOC at the beginning of the
fault, while SOCdet represents the SOC at which the fault was
detected. A smaller value of 1SOC indicates that the model
can detect vehicle faults in a timely manner.
Then we set the abnormal voltage value using equa-
tion (18), and identified the corresponding SOC interval in
which the voltage was located as shown in Table 5. Next,
we executed the warning model sequentially for each SOC
interval in Table 5, and the results are presented in Fig. 15.
As shown in Fig. 15, the dynamic threshold warning model
proposed in this paper can provide early warning for faults
in a short period of time. For the first type of fault, since
FIGURE 13. Dynamic threshold warning accuracy rate. the SOC is low at this time and in order to prevent false
alarms, the dynamic threshold is set relatively loose. There-
According to Fig. 13, for fault type 1, all four prediction fore, the required warning 1SOC is 3.8%. For fault type 3,
algorithms can accurately identify empty data and abnormal the SOC and battery voltage are both in a relatively high state.
data thanks to the proposed warning architecture in this paper. To ensure the safety of the vehicle to the greatest extent

VOLUME 11, 2023 55091

 X. Diao et al.: Research on EV Charging Safety Warning Based on A-LSTM Algorithm

possible, the warning threshold is set more strictly, with a resources, and facilitate the implementation of the national
required warning 1SOC of 0.8%. Compared with the fixed big data strategy.
threshold, the dynamic threshold proposed in this paper for The proposed method can effectively monitor various
charging safety warning has excellent warning timeliness. physical quantity data of EV charging, enable charging fault
warning of EV, and prevent false alarms caused by incorrect
charging data. However, this study has some limitations, par-
ticularly in terms of data acquisition of SOH and the lack of
complete life cycle EV charging data. Further research could
explore these areas in greater depth.

REFERENCES
[1] X. Feng, Y. Pan, X. He, L. Wang, and M. Ouyang, ‘‘Detecting the internal
short circuit in large-format lithium-ion battery using model-based fault-
diagnosis algorithm,’’ J. Energy Storage, vol. 18, pp. 26–39, Aug. 2018,
doi: 10.1016/j.est.2018.04.020.
[2] B. Mao, C. Zhao, H. Chen, Q. Wang, and J. Sun, ‘‘Experimental and
modeling analysis of jet flow and fire dynamics of 18650-type lithium-
ion battery,’’ Appl. Energy, vol. 281, Jan. 2021, Art. no. 116054, doi:
10.1016/j.apenergy.2020.116054.
[3] A. Barré, F. Suard, M. Gérard, M. Montaru, and D. Riu, ‘‘Statistical
analysis for understanding and predicting battery degradations in real-life
electric vehicle use,’’ J. Power Sources, vol. 245, pp. 846–856, Jan. 2014,
doi: 10.1016/j.jpowsour.2013.07.052.
FIGURE 15. Time-limit of warning for different models. [4] X. Feng, S. Zheng, D. Ren, X. He, L. Wang, H. Cui, X. Liu, C. Jin, F. Zhang,
C. Xu, H. Hsu, S. Gao, T. Chen, Y. Li, T. Wang, H. Wang, M. Li, and
M. Ouyang, ‘‘Investigating the thermal runaway mechanisms of lithium-
ion batteries based on thermal analysis database,’’ Appl. Energy, vol. 246,
pp. 53–64, Jul. 2019, doi: 10.1016/j.apenergy.2019.04.009.
VI. DISCUSSION
[5] G. Zhang, X. Wei, X. Tang, J. Zhu, S. Chen, and H. Dai, ‘‘Internal
The following conclusions can be drawn from the analysis of short circuit mechanisms, experimental approaches and detection meth-
the results over time: ods of lithium-ion batteries for electric vehicles: A review,’’ Renew.
Sustain. Energy Rev., vol. 141, May 2021, Art. no. 110790, doi:
1) The actual charging data of EVs obtained has abnormal 10.1016/j.rser.2021.110790.
samples due to sensor failures and transmission prob- [6] P. Bangalore and L. B. Tjernberg, ‘‘An artificial neural network approach
lems. Eliminating data anomalies through data prepro- for early fault detection of gearbox bearings,’’ IEEE Trans. Smart Grid,
vol. 6, no. 2, pp. 980–987, Mar. 2015, doi: 10.1109/TSG.2014.2386305.
cessing can improve the prediction accuracy of the
[7] M.-K. Tran and M. Fowler, ‘‘A review of lithium-ion battery fault diag-
algorithm; nostic algorithms: Current progress and future challenges,’’ Algorithms,
2) EV battery temperature is greatly influenced by the vol. 13, no. 3, p. 62, Mar. 2020, doi: 10.3390/a13030062.
external environment, by smoothing the motor tem- [8] Y. Zhang, T. Li, X. Yan, L. Wang, J. Zhang, X. Diao, and B. Li, ‘‘Electric
vehicle charging fault monitoring and warning method based on battery
perature and battery temperature at the same time, the model,’’ World Electr. Vehicle J., vol. 12, no. 1, p. 14, Jan. 2021, doi:
environmental influence can be significantly reduced, 10.3390/wevj12010014.
providing convenience for subsequent data prediction; [9] M.-K. Tran, M. Mathew, S. Janhunen, S. Panchal, K. Raahemifar, R. Fraser,
and M. Fowler, ‘‘A comprehensive equivalent circuit model for lithium-ion
3) A-LSTM algorithm for time series data prediction batteries, incorporating the effects of state of health, state of charge, and
problems: the accuracy and applicability of A-LSTM temperature on model parameters,’’ J. Energy Storage, vol. 43, Nov. 2021,
for EV voltage prediction problems is demonstrated by Art. no. 103252, doi: 10.1016/j.est.2021.103252.
[10] W. Song, ‘‘Research on integrated safety warning and protection system of
comparison with three other algorithms;
electric vehicle charging,’’ M.S. thesis, Nanjing Univ. Posts Telecommun.,
4) The construction of dynamic thresholds can signifi- 2019, vol. 2.
cantly improve the prediction accuracy and timeliness [11] L. Qian, M. Zhao, and W. Zhang, ‘‘A method to design the security early
of the model. Comparing dynamic thresholds with warning model of EV charging,’’ Adv. Power Syst. Hydroelectr. Eng.,
vol. 32, no. 12, pp. 114–119, 2016.
fixed thresholds, it was found that the dynamic thresh- [12] Y. Zhao, P. Liu, Z. Wang, L. Zhang, and J. Hong, ‘‘Fault and defect
olds had an average 4.52% higher accuracy in predict- diagnosis of battery for electric vehicles based on big data analy-
ing charging anomalies for different vehicles and the sis methods,’’ Appl. Energy, vol. 207, pp. 354–362, Dec. 2017, doi:
10.1016/j.apenergy.2017.05.139.
required 1SOC for warning was reduced by an average [13] L. Zhang, T. Gao, G. Cai, and K. L. Hai, ‘‘Research on electric vehicle
of 8.21%. charging safety warning model based on back propagation neural network
optimized by improved gray wolf algorithm,’’ J. Energy Storage, vol. 49,
May 2022, Art. no. 104092, doi: 10.1016/j.est.2022.104092.
VII. CONCLUSION
[14] B. Xia, Y. Shang, T. Nguyen, and C. Mi, ‘‘A correlation based fault
This paper proposes a state monitoring and fault warning detection method for short circuits in battery packs,’’ J. Power Sources,
method for electric vehicle charging processes based on vol. 337, pp. 1–10, Jan. 2017, doi: 10.1016/j.jpowsour.2016.11.007.
charging-side deep learning. The aim of the method is to [15] D. Gao, Y. Wang, X. Zheng, and Q. Yang, ‘‘A fault warning method
for electric vehicle charging process based on adaptive deep belief net-
ensure the safety of electric vehicle charging, promote the work,’’ World Electr. Vehicle J., vol. 12, no. 4, p. 265, Dec. 2021, doi:
integration of automotive safety technology with various data 10.3390/wevj12040265.

55092 VOLUME 11, 2023

X. Diao et al.: Research on EV Charging Safety Warning Based on A-LSTM Algorithm

[16] K. Zhang, Z. Yin, X. Yang, Z. Yan, and Y. Huang, ‘‘Quantitative assessment TIAN GAO is pursuing the master’s degree with
of electric safety protection for electric vehicle charging equipment,’’ in Northeast Electric Power University, Jilin, China.
Proc. Int. Conf. Circuits, Devices Syst. (ICCDS), Sep. 2017, pp. 89–94. His research interest includes electric vehicle
[17] Y. Jing, ‘‘Research on safety monitoring and fault warning method charging safety warning considering multi-factor
of electric vehicle charging based on data mining technology,’’ coupling.
M.S. thesis, North China Electr. Power Univ., vol. 3, 2022, doi:
10.27139/d.cnki.ghbdu.2021.000707.
[18] R. Yang, R. Xiong, S. Ma, and X. Lin, ‘‘Characterization of external
short circuit faults in electric vehicle Li-ion battery packs and predic-
tion using artificial neural networks,’’ Appl. Energy, vol. 260, Feb. 2020,
Art. no. 114253, doi: 10.1016/j.apenergy.2019.114253.
[19] L. Jiang, Z. Deng, X. Tang, L. Hu, X. Lin, and X. Hu, ‘‘Data-driven
fault diagnosis and thermal runaway warning for battery packs using real-
world vehicle data,’’ Energy, vol. 234, Nov. 2021, Art. no. 121266, doi:
10.1016/j.energy.2021.121266. LIANG ZHANG (Member, IEEE) received the
[20] J. Lamb, C. J. Orendorff, L. A. M. Steele, and S. W. Spangler, ‘‘Failure M.S. and Ph.D. degrees in electrical engineer-
propagation in multi-cell lithium ion batteries,’’ J. Power Sources, vol. 283, ing and automation from the Harbin Institute of
pp. 517–523, Jun. 2015, doi: 10.1016/j.jpowsour.2014.10.081. Technology, Harbin, China, in 2010 and 2015,
[21] J. Hou, X. Feng, L. Wang, X. Liu, A. Ohma, L. Lu, D. Ren, W. Huang, respectively. He is currently an Associate Profes-
Y. Li, M. Yi, Y. Wang, J. Ren, Z. Meng, Z. Chu, G.-L. Xu, K. Amine, sor with the Department of Electrical Engineering,
X. He, H. Wang, Y. Nitta, and M. Ouyang, ‘‘Unlocking the self-supported
Northeast Electric Power University, Jilin, China.
thermal runaway of high-energy lithium-ion batteries,’’ Energy Storage
His current research interests include cover
Mater., vol. 39, pp. 395–402, Aug. 2021, doi: 10.1016/j.ensm.2021.04.035.
vehicle-grid coupling, demand side response, and
[22] B. Liu, ‘‘Design of charging control flow based on electric vehicles,’’ Auto-
mobile Parts, vol. 147, no. 9, pp. 6–9, 2020, doi: 10.19466/j.cnki.1674- V2G intelligent information interaction.
1986.2020.09.002.
[23] Y. Wu and Y. Wang, ‘‘Review of internal short circuit of lithium-ion
battery,’’ Mach. Building Autom., vol. 49, no. 4, pp. 169–172, 2020.
[24] D. Ouyang, Y. He, M. Chen, J. Liu, and J. Wang, ‘‘Experimental study
on the thermal behaviors of lithium-ion batteries under discharge and JUNYU ZHANG is pursuing the master’s degree
overcharge conditions,’’ J. Thermal Anal. Calorimetry, vol. 132, no. 1, with Northeast Electric Power University, Jilin,
pp. 65–75, Apr. 2018, doi: 10.1007/s10973-017-6888-x. China. His research interest includes electric vehi-
[25] X. Cheng, R. Zhang, C. Zhao, F. Wei, J. Zhang, and Q. Zhang, ‘‘A review cle charging early warning safety.
of solid electrolyte interphases on lithium metal anode,’’ Adv. Sci., vol. 3,
no. 3, Mar. 2016, Art. no. 1500213, doi: 10.1002/advs.201500213.
[26] Y. Liu, C. Zhang, J. Jiang, W. Zhang, and L. Zhang, ‘‘Research on
capacity difference identification method of lithium-ion battery pack,’’
Proc. CSEE, vol. 41, no. 4, pp. 1422–1430, 2021, doi: 10.13334/J.0258-
8013.PCSEE.200483.

LONGFEI WANG is pursuing the master’s degree

XIAOHONG DIAO received the M.S. degree
with Northeast Electric Power University, Jilin,
from the China Electric Power Research Institute,
China. His research interest includes energy stor-
Beijing, China. Her research interests include elec-
age battery fault diagnosis.
tric vehicle charging technology and electric vehi-
cle charging safety warning.

LINRU JIANG received the M.S. degree from QIZHI WU is pursuing the master’s degree with
Nanjing Normal University, Jiangsu, China. Her Northeast Electric Power University, Jilin, China.
research interests include electric vehicle charging His research interest includes electric vehicle bat-
technology and electric vehicle charging safety tery charge and discharge warning.
warning.

VOLUME 11, 2023 55093