Introduction

The rapid development of EV makes lithium-ion battery technology a central element in driving the e-mobility revolution1. For the EV, the power battery system is the main source of power for EVs, and its charging and discharging current can be as high as hundreds of amperes, higher charging and discharging currents, will lead to a rapid rise in the internal temperature of the battery2, and if the temperature can not be effectively monitored, it will accelerate the decline in battery performance or even lead to safety accidents such as thermal runaway of the battery3. According to statistics, in the natural accidents of EV, the thermal runaway of lithium-ion battery charging accounts for one-third of the total accidents, and some of the electric vehicle lithium battery thermal runaway accidents are shown in Table 1. Excessive charging temperature may lead to serious consequences such as battery damage, performance degradation and even fire4, so real-time monitoring and prediction of the risk of thermal runaway during battery charging is crucial to ensure the safe operation of EV.

Table 1 EV lithium-ion battery charging thermal runaway accident table.
Full size table
Fig. 1
figure 1

Lithium-ion battery charging thermal runaway trigger conditions.

Full size image

At present, researchers at home and abroad have carried out a lot of exploration on the thermal runaway problem of lithium-ion battery, including experimental exploration, mechanism analysis, and safety protection5,6,7, and have preliminarily understood the triggering mechanism of thermal runaway8,9, the expansion mechanism, and the warning method of safety protection10,11, which lays the foundation for the large-scale promotion and application of EV12,13. At this stage, there are three main directions to solve the thermal runaway problem of lithium-ion battery charging in EV, namely, material improvement14,15,16, battery modeling17,18,19, and big data prediction20,21,22.

Early researchers utilized phase change materials to replace the internal cooling system of cylindrical battery hollow core rods15. The research results indicated that this system exhibited significant cooling effectiveness. However, the process of material development and replacement may present complexities. To reduce research and development costs, the establishment of lithium-ion battery models has become a primary approach for studying battery characteristics17. In practical applications, the construction of electric vehicle lithium-ion battery models primarily adopts comprehensive models, including the establishment of equivalent circuit models18, conjugate heat transfer models, and reaction kinetics models19. Through numerical simulations, multiple features are extracted to enable early warning of battery faults. However, there are delay defects in the sensors and the battery model construction is complicated, so the intelligent management of lithium-ion battery charging through big data has become the main way to ensure the safety of EV charging.

With the development of deep learning, combined with the nonlinear characteristics of the EV lithium-ion battery charging process, the deep learning network is applied in the field of EV lithium-ion battery fault early warning, for the integration of big data and new energy vehicles, provides a solid foundation. Based on previous studies utilizing characteristic signals of lithium-ion batteries for fault identification and diagnosis, existing big data-driven early warning methods integrate both external and internal parameters-such as surface temperature, terminal voltage, internal impedance, expansion force, and gas generation20,21,22. By combining these parameters, models can be trained to predict key indicators of thermal runaway, ultimately achieving reliable early warning outcomes23 employed a convolutional neural network (CNN) to classify and predict three distinct stages-safe operation, the critical pre-thermal runaway condition, and the actual thermal runaway occurrence-to prevent thermal runaway in lithium-ion batteries24 proposed a CNN and bidirectional recurrent neural network (BiGRU) based EV charging state monitoring and safety warning method, and the experimental results show that the method phase has higher prediction accuracy and prediction effect.

Although existing deep learning models can provide early warning of thermal runaway during the charging process of lithium-ion batteries in EVs, their accuracy still requires improvement, and the warning system needs further refinement. Based on the above research results, this paper proposes a thermal runaway early warning method for EV lithium-ion battery charging, based on a charging network. With battery temperature as the target observation, an LSTM-TCN temperature prediction model is constructed. A sliding window combined with residual analysis is employed to monitor the predicted and actual charging temperatures, and an early warning threshold is established to achieve effective thermal runaway prevention. The main contents of this paper are as follows:

  1. (1)

    The prediction accuracy is further improved by developing a hybrid LSTM-TCN prediction model based on the temporal characteristics of lithium-ion battery charging data.

  2. (2)

    Determine the warning threshold and enhance the reliability of early warning through residual analysis and monitoring of temperature residuals.

  3. (3)

    The generalizability of the proposed warning method is demonstrated through experimental validation using normal and abnormal charging data from two different types of EV lithium-ion batteries at varying temperatures.

The remaining chapters of this paper are organized as follows: the second part analyzes the thermal runaway of lithium-ion battery, including the reasons for the departure of the thermal runaway, the description of the stages, and the behavioral characteristics; the third part briefly describes the charging standards of lithium-ion battery for EV, preprocesses the experimental data, completes the construction of the prediction model, and carries out the design of the thermal runaway warning process. The fourth part describes and analyzes the overall experiment, first prepares the experiment, then uses different data to conduct the experiment and analyze the results. Chapter "Conclusion" summarizes the article.

Lithium-ion battery thermal runaway analysis

Lithium-ion battery charging thermal runaway trigger causes

EV lithium-ion battery is subject to internal and external factors with uncertainties during the charging process in the charging network, some of which can cause damage to the battery system, thus increasing the risk of thermal runaway charging25. Among them, internal factors are mainly battery aging, battery manufacturing defects, and EV BMS failures26,27; external factors are mainly caused by violent collisions, battery flooding, and external high and low temperatures28,29. The lithium-ion battery thermal runaway triggering conditions are categorized into three types: mechanical abuse, electrical abuse and thermal abuse30, and the specific causes of failure are shown in Fig. 1. Due to these unavoidable factors, the lithium-ion battery generates excessive heat inside the battery during the charging process, which leads to an increase in battery temperature and triggers serious safety problems such as fire and explosion31. Therefore, it is of great significance to propose a reasonable early warning method by studying the thermal runaway mechanism of EV lithium-ion battery.

Lithium-ion battery thermal runaway stage division

When thermal runaway occurs in lithium-ion battery, the most direct manifestation is a rapid increase in temperature, accompanied by extreme situations such as possible fire and explosion32. The instant thermal runaway occurs, the internal temperature of the battery increases dramatically, as shown in Fig. 2. Thermal runaway occurs in electric vehicle lithium batteries during the charging process, which is usually divided into four stages33.

Fig. 2
figure 2

Thermal runaway temperature variation.

Full size image

The first stage is the battery overcharging stage, which may occur during the battery charging process due to internal battery factors, failure of the charge control system, or improper operation, which may cause the battery to continuously receive electrical energy34. Overcharging will lead to excessive gas generation and pressure rise inside the battery, which will trigger safety hazards.

The second stage is the overheating stage, which may be triggered when the internal temperature of the battery exceeds the safety limit. Overheating will lead to changes in the internal materials of the battery, increasing the risk of thermal runaway, and may trigger a chain reaction of thermal runaway. As shown in Fig. 2, the temperature reaching point T1 in the overheating stage is the critical temperature at which thermal runaway occurs, and if we want to block thermal runaway from occurring, we have to warn and take measures before the temperature reaches T135.

The third section of the stage is the thermal runaway stage, that is, the dotted line interval shown in Fig. 2. On the basis of overheating, if the chemical reaction inside the battery is out of control, it may trigger more serious thermal runaway phenomena, including smoke, flame, and even explosion36. This stage of thermal runaway phenomenon may spread rapidly, posing a serious threat to the vehicles and the surrounding environment.

The fourth stage is the diffusion stage, once the thermal runaway occurs, it may lead to the successive loss of control of other single batteries within the battery pack, forming a chain reaction, so that the scope of the accident is expanded, causing a greater impact on safety.

These four stages constitute the key stages in which thermal runaway may occur in the charging process of EV lithium-ion battery, and these risks need to be prevented and avoided through scientific management and control measures.

Lithium-ion battery charging safety management and thermal runaway warning

EV lithium-ion battery charging network charging process reference basis

At present, the charging process of EV lithium-ion battery has a variety of reference specifications. Among them, the Specification for Battery System Safety Risk Monitoring and Failure Early Warning for Electric Vehicle Charging Process (T/CSAE 254-2022) analyzes the current situation of the EV charging process. It uses charging big data analysis to propose requirements for the charging system and safety early warning methods. This specification aims to take the charging process as a perspective to discover charging anomalies37. It also seeks to provide relevant early warnings before serious safety risks arise, allowing charging big data to play an effective role.

According to the standards stipulated in T/CSAE 254-2022, the charging process of EV needs to confirm the parameters and status of the charging equipment, including the input voltage, current and frequency, and the output voltage, current and power, to ensure the stability and safety of the charging process. At the same time, the physical connection and communication protocol of the charger and charging interface need to be tested to ensure the reliability and operability of the charging process. The charging information table of the whole EV is shown in Table 2.

Table 2 The main state information of EV lithium-ion battery before charging.
Full size table

Multi-parameter selection and pre-processing

In the EV lithium-ion battery charging network power process, the charging temperature change is affected by many factors. In order to select appropriate experimental data, we use the correlation coefficient method to calculate the correlation between each charging parameter and charging temperature, the calculation formula is as follows:

$$\begin{aligned} r=\frac{\sum {({{E}_{t}}-\bar{E})({{F}_{t}}-\bar{F})}}{\sqrt{\sum \nolimits _{t=1}^{n}{{{({{E}_{t}}-\bar{E})}^{2}}}}\sqrt{\sum \nolimits _{t=1}^{n}{{{({{F}_{t}}-\bar{F})}^{2}}}}} \end{aligned}$$
(1)

where r is the obtained correlation coefficient, with the value ranging from [− 1,1]; n is the total number of charging data; \({{E}_{t}}\) and \({{F}_{t}}\) are the real-time charging data at time \(\text {t}\) ; \(\bar{E}\) and \(\bar{F}\) are the average. The closer the absolute value of the obtained correlation coefficient r is to 1, the higher the correlation between E and F is38.

After calculation, the six parameters of charging demand voltage, demand current, lithium-ion battery cell voltage, actual charging voltage, actual charging current, and battery state of charge (SOC) are selected as model inputs. The specific correlation coefficient values are shown in Table 3. Where, RequireU and RequireI are the voltage and current levels required by the battery during the charging process, typically set based on the battery’s characteristics and charging strategy. SingleU refers to the voltage of an individual battery cell. ChargeU and ChargeI are the voltage and current supplied to the battery during the actual charging process. SOC is the state of charge of the battery. In the correlation coefficients presented in Table 3, the absolute values of the selected charging data’s correlation coefficients with charging temperature are all greater than 0.5, indicating a strong correlation with charging temperature. Furthermore, RequireU, SingleU, ChargeU, and SOC exhibit a positive correlation with temperature, while RequireI and ChargeI show a negative correlation with charging temperature.

Table 3 Table of correlation coefficients for EV charging data.
Full size table

According to the type of charging data, the resulting data set is preprocessed. Since the EV lithium-ion battery will cause the missing data obtained during the charging process due to the influence of hardware and environmental factors, the interpolation method is used to fill in the missing data39. It estimates the missing values through the function relationship between the known data, and the specific calculation formula is as follows:

$$\begin{aligned} x_a=x _{0}+\frac{{{x}_{1}}-{{x}_{0}}}{{{p}_{1}}-{{p}_{0}}}(p-{{p}_{0}}) \end{aligned}$$
(2)

where p is the missing charging data point, \({{p}_{0}}\) and \({{p}_{1}}\) are the known data points before and after p respectively, \({{x}_{0}}\) and \({{x}_{1}}\) are the charging data values corresponding to \({{p}_{0}}\) and \({{p}_{1}}\) respectively, \(x_a\) is the data value required for interpolation.

Since the charging parameters used in this paper have different scales and value ranges, the data set is normalized in order to speed up the convergence of the model and improve the stability of the model. The specific equations are as follows:

$$\begin{aligned} {{x}_{norm}}=\frac{x-{{x}_{\min }}}{{{x}_{\max }}-{{x}_{\min }}} \end{aligned}$$
(3)

where x is the original dataset, \({{x}_{\min }}\) and \({{x}_{\max }}\) are the maximum and minimum values in the data, \({{x}_{norm}}\) is the normalized data which is used as an input to the deep model network.

In this paper, charging data such as voltage, current, and temperature of charging piles, as well as on-board lithium-ion batteries, are collected through the EV charging network. The cycle-type data of charging lithium-ion batteries for two types of EVs are obtained by referencing T/CSAE 254-2022, which is the Specification for Safety Risk Monitoring and Fault Early Warning of Battery Systems in Electric Vehicle Charging Processes. This dataset is used in this experiment to validate the practicability of the proposed method.

Lithium-ion battery charging temperature prediction model construction

The LSTM-TCN model takes the temperature-related parameters such as EV lithium-ion battery charging demand voltage, demand current, and cell voltage as input data, extracts the short-time features through the LSTM network40, and then obtains the long-term time-series features through the TCN network41, and at the same time captures the long-term dependence and the local correlation in the time-series data, to better understand the long-term trend of the temperature change and the short-term fluctuation, so as to the LSTM-TCN structure is shown in Fig. 3.

LSTM network is a special type of recurrent neural network (RNN) with a special structural design that allows it to avoid the long-term dependency problem and remember information from earlier moments42. The key to the LSTM is the cell states, each of which controls the retention and transfer of information in the LSTM by means of forgetting gates, input gates, and output gates43 as shown by dashed portions of Fig. 3.

The current LSTM receives the cell state \({{C}_{t-1}}\) from the previous one moment and acts together with the current EV lithium-ion battery charging data input \({{x}_{t}}\) to generate a new cell state \({{C}_{t}}\). In this process, a forgetting gate is first used to determine the retention and forgetting of the information in \({{C}_{t-1}}\), with the following formula:

$$\begin{aligned} {{f}_{t}}=\sigma ({{W}_{f}}\cdot [{{h}_{t-1}},{{x}_{t}}]+{{b}_{f}}) \end{aligned}$$
(4)

where \({f}_{t}\) represents the output of the forget gate, with a range of [0, 1], indicating the retention ratio of each information unit within the cell state, \(\sigma\) is the Sigmoid activation function, \({W}_{f}\) represents the weight matrix of the forget gate, \({h}_{t-1}\) stands for the hidden state from the previous time step, \({x}_{t}\) is the input at the current time step, and \({b}_{f}\) is the bias term of the forget gate.

Fig. 3
figure 3

LSTM-TCN model.

Full size image

Then, an input gate determines the information to be retained in the current input \({{x}_{t}}\) with the following formula:

$$\begin{aligned} {{i}_{t}}&=\sigma ({{W}_{i}}\cdot [{{h}_{t-1}},{{x}_{t}}]+{{b}_{i}}) \end{aligned}$$
(5)
$$\begin{aligned} {{\tilde{C}}_{t}}&=\tanh ({{W}_{c}}\cdot [{{h}_{t-1}},{{x}_{t}}]+{{b}_{c}}) \end{aligned}$$
(6)

where \({i}_{t}\) is the output of the input gate, indicating the retention ratio of the current input information, \({\tilde{C}}_{t}\) is the candidate cell state, reflecting the candidate values of new information, \({W}_{i}\) and \({W}_{c}\) are the weight matrix for both the input gate and the candidate cell state, while \({b}_{c}\) is their respective bias terms.

Finally, the output gate integrates the above information to obtain the output signal \({{h}_{t}}\) with the following equation:

$$\begin{aligned} {{o}_{t}}&=\sigma ({{W}_{o}}\cdot [{{h}_{t-1}},{{x}_{t}}]+{{b}_{0}}) \end{aligned}$$
(7)
$$\begin{aligned} {{h}_{t}}&=\tanh ({{C}_{t}})*{{o}_{t}} \end{aligned}$$
(8)

where \(\sigma\) represents the sigmod neural network layer, and \(W_o\) and \(b_o\) represent the weight matrix and deviation vector. Through their action, the LSTM network is able to realize the weighting of input data and the updating of the memory state to better capture the dependencies in the sequence data44.

Despite the improvement of LSTM, the forgetting mechanism exhibits exponential decay of information limiting its ability to capture long time scale information. In contrast, the TCN specific dilation causal convolution and residual connection can adjust the sensory field size by the number of layers, dilation factor, and filter size, which enables the model to obtain stable results even when dealing with long-time input data45such as EV charging.

Therefore, the LSTM network output is used as an input to the TCN network. In this case, the predicted output temperature \({{y}_{\text {t}}}\) of the TCN network has to be determined by the inputs prior to the moment t. To achieve this goal, the initial layer of the TCN uses a one-dimensional dilated causal convolution (shown in Fig. 3), which takes the output of the LSTM network layer as input. In each intermediate layer, the size of the input layer is kept constant and zero padding is used to obtain the same size of the subsequent layer as the previous one46. In each layer, a convolution operation is performed to extract features between the charging data and the charging temperature. The computational formula is given below:

$$\begin{aligned} f({{h}_{t}})=\sum _{k=1}^{L}w_k\cdot h_{t-d\cdot k} \end{aligned}$$
(9)

where \({{h}_{t}}\) is the input sequence and L is the kernel size, \(w_k\) is the kernel weights, and d is the dilation rate.

However, the traditional convolution operation is a purely linear operation involving only addition and multiplication of matrices. Considering the complexity and nonlinear characteristics of EV charging data, this model introduces nonlinear units (ReLU) for nonlinear adjustment47.

Finally, in order to endow the network with the capability of cross-layer information propagation and ensure consistency between input and output, the causal convolutional component is connected to the one-dimensional convolutional network through residual connections, as indicated by the plus sign in the Fig. 3, resulting in the temperature prediction output. The output equation can be expressed as follows:

$$\begin{aligned} {{y}_{t}}=R[f({{h}_{t}})+Conv1D({{h}_{t}})] \end{aligned}$$
(10)

where \(f({{h}_{t}})\) is dilated convolution operation, R represents the ReLU activation function, and \(Conv1D({{h}_{t}})\) is the one-dimensional convolution operation.

To summarize, this paper adopts the LSTM network to extract sequence features and feeds these features into the TCN network to establish a predictive model for the temperature of EV lithium-ion battery during charging. By doing so, we obtain highly efficient and accurate predictions.

Model evaluation and validation standards

The evaluation metrics selected for the neural networks include three commonly used indicators: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R Squared, R2)48,49. The specific formulas are as follows:

$$\begin{aligned} MAE= & \frac{1}{n}\sum \limits _{t=1}^{n}{\left| {{y}_{t}}-{{{\hat{y}}}_{t}} \right| } \end{aligned}$$
(11)
$$\begin{aligned} RMSE= & \sqrt{\frac{1}{n}\sum \nolimits _{t=1}^{n}{{{({{y}_{t}}-{{{\hat{y}}}_{t}})}^{2}}}} \end{aligned}$$
(12)
$$\begin{aligned} R2= & 1-\frac{{{D}_{res}}}{{{D}_{tot}}}=1-\frac{\sum \nolimits _{t=1}^{n}{{{({{{\hat{y}}}_{t}}-{{y}_{t}})}^{2}}}}{\sum \nolimits _{t=1}^{n}{{{({{{\bar{y}}}_{t}}-{{y}_{t}})}^{2}}}} \end{aligned}$$
(13)

where \({{y}_{t}}\) represents the true values of EV charging temperature, \({{\hat{y}}_{t}}\) represents the predicted values of EV charging temperature, \({{D}_{res}}\) represents the sum of squared residuals, \({{D}_{tot}}\) represents the sum of squared deviations, and \({{\bar{y}}_{i}}\) represents the average output charging temperature. Smaller values of MAE and RMSE indicate higher accuracy of the model, while R2 ranges from 0 to 1, with values closer to 1 indicating a higher degree of model fit. These definitions are essential for evaluating the performance of the proposed model in predicting EV charging temperature.

Design of thermal runaway warning discrimination

The sliding window technique is primarily applied in the processing and analysis of time series data. By moving a fixed-sized window across the entire dataset, it enables segmenting the lithium-ion battery charging time series data. This allows for capturing key features such as trends, periodicity, and anomalies in the time series data, while minimizing the impact of erroneous data50. The process of the sliding window technique is illustrated in Fig. 4.

Fig. 4
figure 4

Sliding window process.

Full size image

The residual analysis technique is primarily employed to identify system anomalies or potential faults by monitoring the differences between actual observed values and model predicted values51. As temperature values do not exhibit clear trends over short periods of time, analyzing charging temperature using residual calculations can more effectively capture abnormal changes.

Take the sliding window with length N, calculate the mean residual value and the standard deviation of the residual between the current temperature and the temperature in the sliding window, and the formula is as follows:

$$\begin{aligned} \left\{ \begin{aligned}&\bar{X}=\frac{1}{N}\sum \nolimits _{t=1}^{n}{{{e}_{t}}} \\&S=\sqrt{\frac{1}{N-1}\sum \nolimits _{t=1}^{n}{\left( {{e}_{t}}-\bar{X} \right) }} \\ \end{aligned} \right. \end{aligned}$$
(14)

where \(\bar{X}\) is the mean residual value, \({{e}_{t}}\) is the residual value of the sliding window at time t, and S is the standard deviation of the residuals.

Fig. 5
figure 5

Early warning process of thermal runaway of lithium battery in electric vehicles.

Full size image

After calculating the residuals of the normal charging data of the lithium battery of electric vehicles, the mean value of the maximum residual and the standard deviation of the residuals are obtained, and the pre-alarm threshold is calculated. The formula is as follows:

$$\begin{aligned} \left\{ \begin{aligned}&\pm {{X}_{\text {E}}}=\pm {{\mu }_{1}}{{{\bar{X}}}_{\max }} \\&{{S}_{\text {E}}}={{\mu }_{2}}{{S}_{\max }} \\ \end{aligned} \right. \end{aligned}$$
(15)

where \({{\mu }_{1}}\) and \({{\mu }_{2}}\) are the early warning coefficients, which are usually integers obtained based on experience, \(\pm {{X}_{\text {E}}}\) is the upper and lower limits of the mean threshold of the pre-alarm residuals, and \({{S}_{\text {E}}}\) is the upper limit of the standard deviation of the pre-alarm residuals.

Design of thermal runaway warning discrimination

In this paper, the thermal runaway prediction and safety warning methods of lithium battery charging temperature for electric vehicles are combined, as shown in Fig. 5. Firstly, the charging parameters of lithium batteries are obtained by the charging network BMS system52. The charging pile and the electric vehicles use the international standard CAN communication protocol for data transmission, and the collected real-time data is preprocessed such as outlier elimination, missing value completion and normalization to obtain the time series data required for feature extraction.

Then, the LSTM-TCN temperature prediction model is designed, the time series data was input into the model training, and the structure and parameters of the model are adjusted by analyzing the prediction results, and finally the temperature prediction model with the highest accuracy is obtained. The algorithm structure is shown in Algorithm 1.

Algorithm 1
figure a

Integration of LSTM-TCN

Full size image

Then, the sliding window method is used to analyze the residuals of the normal charging data of different types of EV lithium-ion batteries, determine the temperature warning threshold, and carry out dynamic early warning. Figure 6 shows the early warning process. After processing the predicted temperature with a fixed-length sliding window, the standard deviation of the residuals and the mean of the residuals are calculated. According to Eq. (15), the early warning threshold is established. The residual standard deviation between the real-time predicted value and the actual value of the EV lithium-ion battery charging data is then compared with the mean residual value and the threshold. When the mean residual or residual standard deviation of the real-time charging data exceeds the limit, the charging device issues a thermal runaway warning. It also takes protective measures to ensure the safety of the vehicle and personnel.

Finally, the thermal runaway and normal data of different types of lithium battery charging for electric vehicles are used for experimental verification to further determine the accuracy and feasibility of the method.

Fig. 6
figure 6

Sliding window and residual analysis process.

Full size image

Experimental verification and analysis

Experiment preparation and subject selection

In this experiment, 12th Gen Intel(R) Core(TM) i5-12500H @3.10 GHz was used as the processor, Python 3.8 and TensorFlow 2.4.1 were used for programming, and the Keras framework was used to build a deep learning model.

The EV charging network is an EV charging facility management system based on the Internet and Internet of Things technology, which provides convenient charging services for EV by connecting various charging piles, charging stations, EVs and related cloud platforms. In this experiment, the relevant periodic data were collected through the EV charging network as the dataset of this experiment.

In this experiment, to achieve the most accurate predictive results, we selected MAE and RMSE for Nickel manganese cobalt battery (NMC) and Lithium Iron phosphate battery (LFP) as evaluation metrics. The model parameters were adjusted by comparing the errors of LSTM and TCN under different parameter settings. The specific accuracy metrics are presented in Tables 4 and 5.

The final model configuration selected includes an LSTM with 3 hidden layers and 90 neurons. For the TCN, the maximum dilation factor is set to 4, with 20 filters.

Taking two types of EV lithium-ion batteries, namely, NMC and LFP, as the experimental objects. The power battery capacity of the NMC is 91.5Ah, while the power battery capacity of the LFP is 207Ah. To demonstrate the universality of this method, normal charging data from the two types of electric vehicles at a normal room temperature of \(25^{\circ }\)C were used for model training. Then, abnormal charging data from the two types of electric vehicles under low temperature (\(- 15^\circ\)C) and high temperature (\(40^{\circ }\)C) conditions were utilized to validate the method.

Model comparison and analysis

In the process of thermal runaway of EV lithium-ion battery, the abnormal increase in battery temperature is one of the main characteristics. Therefore, accurate prediction of charging temperature can detect abnormal temperature increases in advance, which can help to take timely measures to avoid potential thermal runaway. At the same time, using multiple parameters as inputs can make the deep learning Xi model map to more complex features and relationships, and improve the sensitivity of the model to potential risks.

Table 4 Results of MAE and RMSE with different parameters of LSTM.
Full size table
Table 5 Results of MAE and RMSE with different parameters of TCN.
Full size table
Fig. 7
figure 7

Prediction plots of normal charging temperature for both models.

Full size image
Fig. 8
figure 8

Prediction plots of normal charging temperature for multiple models.

Full size image
Table 6 Multi-model prediction accuracy.
Full size table

Both the LSTM network and the TCN network exhibit strong processing capabilities for time series data. Therefore, the normal charging data of two different types of electric vehicle lithium batteries at \(25^{\circ }\)C are used as input for the LSTM-TCN model and the TCN-LSTM model. The prediction results show that the evaluation indexes of the LSTM-TCN model for the two types of electric vehicle lithium batteries are higher than those of the TCN-LSTM model. The specific values are presented in the model evaluation data in Table 6. At the same time, it is shown that for the charging data of EV lithium-ion battery, the short-term time series features are extracted first, and then the long-term features are summarized, and more accurate prediction results can be obtained, as shown in Fig. 7.

To verify the wide accuracy of LSTM-TCN, we compared it with other models such as CNN-LSTM, CNN-TCN, LSTM, TCN, and CNN using the same charging dataset. As shown in the multiple-model temperature prediction results comparison chart in Fig. 8, during the three charging periods, the LSTM-TCN model had a good fitting effect, while the other models were more likely to produce prediction deviations due to insufficient data points in the later stage. Meanwhile, for the evaluation indicators of the LSTM-TCN model, MAE was 0.009, RMSE was 0.012, and R2 was 0.99 for NMC battery, and MAE was 0.026, RMSE was 0.036, and R2 was 0.99 for LFP battery, all of which were superior to the other models. Specific evaluation indicator data are shown in Table 6.

In summary, we selected the LSTM-TCN model as the temperature prediction model in this paper, providing a data foundation for alarm threshold setting and method practicality verification.

Early warning analysis and verification

We selected a sliding window length of N=100 and determined the warning threshold by training and residual analysis of normal charging data at \(25^{\circ }\)C. Abnormal charging data from two types of electric vehicle lithium batteries were selected for pre-alarm experiments to determine the practicality and accuracy of the thermal runaway warning model.

According to Eq. (14), the residual standard deviation and mean of normal charging states for the two types of lithium batteries were calculated, as shown in Figs. 9 and 10. For the NMC battery, the maximum value of residual standard deviation 0.026, and the maximum absolute value of residual mean is 0.063. For the LFP battery, the maximum value of residual standard deviation is 0.206, and the maximum absolute value of residual mean is 0.011. Using Eq. (15), the thermal runaway pre-alarm thresholds are calculated, with \({{\mu }_{1}}\) selected as 2 and \({{\mu }_{2}}\) as 4.5. The pre-alarm threshold for NMC residual standard deviation is calculated as 0.052, and the pre-alarm threshold for residual mean is 0.284. The pre-alarm threshold for LFP residual standard deviation is 0.041, and the pre-alarm threshold for residual mean is 0.481.

Fig. 9
figure 9

NMC normal charge residual analysis.

Full size image
Fig. 10
figure 10

LFP normal charge residual analysis.

Full size image

Due to the significant impact of extreme environmental temperatures on thermal runaway during charging of EV lithium-ion batteries, the charging data under low temperature and high temperature conditions for NMC and LFP batteries are selected as input for the trained predictive model. The predicted results are then analyzed through residual analysis by comparing them with the actual values. Based on the analysis results, the pre-alarm moments are determined and compared with the actual thermal runaway moments.

For the selected abnormal sample data of the NMC battery at − 15 °C, the actual thermal runaway moment is observed at the 58183th sampling point. As shown in Fig. 11, the pre-alarm moment for residual standard deviation of the warning model is at the 57624th sampling point, which is 559 sampling points or 27.95 s ahead of the actual event. On the other hand, the pre-alarm moment for the residual mean is at the 57638th sampling point, 555 sampling points or 27.75 s ahead of the actual event.

For the selected abnormal sample data of the LFP battery at − 15 °C, the actual thermal runaway moment is observed at the 58246th sampling point. As shown in Fig. 12, the pre-alarm moment for residual standard deviation of the warning model is at the 57713th sampling point, which is 533 sampling points or 26.65 s ahead of the actual event. On the other hand, the pre-alarm moment for the residual mean is at the 57696th sampling point, 550 sampling points or 27.50 s ahead of the actual event.

Fig. 11
figure 11

Thermal runaway data early warning analysis of NMC charging at \(\hbox {-}15^{\circ }\)C.

Full size image
Fig. 12
figure 12

Thermal runaway data early warning analysis of FPL charging at \(\hbox {-}15^{\circ }\)C.

Full size image
Fig. 13
figure 13

Thermal runaway data early warning analysis of NMC charging at \(35^{\circ }\)C.

Full size image
Fig. 14
figure 14

Thermal runaway data early warning analysis of FPL charging at \(35^{\circ }\)C.

Full size image

For the selected abnormal sample data of the NMC battery at 35 °C, the actual thermal runaway moment is observed at the 58096th sampling point. As shown in Fig. 13, the pre-alarm moment for residual standard deviation of the warning model is at the 57656th sampling point, which is 440 sampling points or 22.00 s ahead of the actual event. On the other hand, the pre-alarm moment for the residual mean is at the 57680th sampling point, 416 sampling points or 20.80 s ahead of the actual event.

For the selected abnormal sample data of the LFP battery at 35 °C, the actual thermal runaway moment is observed at the 50884th sampling point. As shown in Fig. 14, the pre-alarm moment for residual standard deviation of the warning model is at the 50600th sampling point, which is 284 sampling points or 14.20 s ahead of the actual event. On the other hand, the pre-alarm moment for the residual mean is at the 50685th sampling point, 199 sampling points or 9.95 s ahead of the actual event.

Based on the experimental results, it is known that it is more difficult to issue warnings for lithium-ion batteries in electric vehicles when the ambient temperature is high. The shortest warning time is 9.95 s, which is sufficient to adjust the charging method to prevent thermal runaway events from occurring.

In conclusion, under conditions of excessively high or low ambient temperatures, utilizing residual analysis to issue warnings based on the predicted charging temperature values of lithium-ion batteries allows sufficient time for adjusting the charging method of electric vehicles, thereby ensuring the safety of both the individuals and the vehicles.

Conclusion

In order to address the safety concerns posed by thermal runaway events in EV lithium-ion battery during the charging process, this paper proposed a charging thermal runaway warning method based on the charging network for EV lithium-ion battery. Using the LSTM-TCN model for charging temperature prediction, experimental results demonstrate that this model reduces MAE values by up to 9.4% and RMSE by up to 10.2% compared to other models, with R2 consistently reaching 99.9%. The residual analysis-based warning system showed effective early warning times ranging from a minimum of 9.95 s to a maximum of 22.00 s across thermal runaway tests conducted on NMC and LFP Lithium-ion batteries in EVs under different temperatures. This approach enables proactive warning before thermal runaway initiation, enhancing safety for both vehicle occupants and the vehicle itself.

Although our method has made contributions in the field of charging safety for EV lithium-ion battery, there are certain limitations that should be acknowledged:

  1. (1)

    Insufficient research on charging methods for EV lithium-ion battery may result in suboptimal model training. Future studies will explore different charging configurations to enhance the effectiveness of the approach.

  2. (2)

    The availability of high-quality historical data is limited, which restricts the wide-ranging investigation of various vehicle models. Collecting more diverse charging data from different types of EV lithium-ion batteries will improve the generalizability of this method in future work.