Lithium battery pack health state estimation method based on data driving and equivalent circuit model fusion

Abstract

The lithium battery pack health state estimation method based on data driving and equivalent circuit model fusion comprises the following steps: acquiring actual operation data of a lithium battery pack; judging positive and negative current identification charging sections, and screening effective charging sections by using SOC change and duration; extracting health factors such as capacity change rate, voltage change rate, cumulative mileage, SOC statistical characteristics and temperature statistical characteristics, and screening data driving health factors with strong correlation by adopting a Pearson's correlation coefficient; a second-order RC equivalent circuit model is constructed, model parameters are dynamically identified through FFRLS, and equivalent circuit health factors are obtained; fusing the data-driven health factors and the equivalent circuit health factors into a comprehensive feature set; and constructing a GRU-Attention model, capturing a long-term and short-term dependency relationship of data by adopting the GRU, enhancing feature expression through Attention, and outputting the predicted health state of the lithium battery pack. The method achieves the high-precision estimation of the health state of the lithium battery pack, and is suitable for the intelligent health prediction of a new energy automobile battery management system and energy storage equipment.

Description

Technical Field

[0001] This invention relates to the field of lithium battery pack health management technology, specifically to a method for estimating the health status of lithium battery packs based on the fusion of data-driven and equivalent circuit models. Background Technology

[0002] Lithium-ion batteries are widely used in new energy vehicles and energy storage systems due to their high energy density, long lifespan, and excellent charge and discharge performance. However, with the increase of usage time, the capacity of lithium batteries irreversibly degrades. If the state of health (SOH) of lithium batteries cannot be accurately assessed, it will seriously affect the vehicle's range, safety, and energy efficiency.

[0003] Existing State of Health (SOH) estimation methods are mainly divided into three categories: empirical feature-based methods, model-based methods, and data-driven intelligent algorithms. Empirical feature-based methods typically rely on stringent operating conditions such as constant current charge-discharge and static testing to extract features like capacity, incremental capacity, or impedance; however, these methods are difficult to implement in real-world vehicle operating environments. Model-based methods estimate the state of health by establishing equivalent circuits or electrochemical models; however, model parameters vary significantly with state of charge, temperature, and aging degree, making online identification difficult. Data-driven intelligent algorithms can automatically learn battery degradation patterns from operational data, but they are highly dependent on large-scale, multi-condition labeled data, and the interpretability of the models is insufficient.

[0004] Among these methods, the Equivalent Circuit Model (ECM) is widely used for SOH estimation due to its clear structure and strong physical interpretation. Deep learning methods, on the other hand, have made some progress in battery health prediction in recent years due to their advantages in feature extraction and temporal modeling. However, their parameters vary significantly with SOC, temperature, and aging processes, making online identification difficult. The models are prone to mismatch under dynamic operating conditions, and high-order models have high computational costs, hindering real-time applications in vehicles. Deep learning methods can automatically extract features and model temporal relationships, demonstrating good accuracy in battery SOH prediction. However, they are highly dependent on large-scale labeled data, and there are domain differences between laboratory and real-vehicle operating conditions. Their generalization and transfer capabilities are insufficient, and they lack interpretability and uncertainty quantification, increasing the difficulty of engineering applications.

[0005] However, single methods have limitations: traditional equivalent circuit model (ECM) methods are difficult to model complex aging behaviors, whose parameters change dynamically with SOC, temperature and SOH, making online identification difficult, and they are also difficult to accurately characterize rate effects, hysteresis and nonlinear features.

[0006] While data-driven methods offer high accuracy, they suffer from weak physical interpretability and poor model generalization. These models heavily rely on large-scale, multi-condition labeled data, and the difference between laboratory data and actual vehicle conditions limits their cross-scenario generalization ability. Furthermore, deep learning models, being black-box structures, lack physical interpretability and uncertainty quantification, hindering safety assessment and online calibration.

[0007] Therefore, there is an urgent need for a SOH estimation method that can integrate the advantages of physical modeling and data-driven approaches, while also possessing both prediction accuracy and interpretability. Summary of the Invention

[0008] This invention provides a method for estimating the state of health (SOH) of lithium-ion battery packs based on a fusion of data-driven approaches and equivalent circuit models, aiming to improve the accuracy, stability, and physical interpretability of SOH estimation. The method first collects multi-dimensional battery data generated during the operation of new energy vehicles, determines the charging stage by current direction, and calculates the battery reference capacity using the ampere-hour inverse integral method. Subsequently, data-driven health factors, including capacity change rate, voltage change rate, SOC consistency, and temperature characteristics, are extracted. Simultaneously, a second-order RC equivalent circuit model is established, and the model parameters are identified using recursive least squares (FFRLS) with a forgetting factor, extracting physical health factors such as R0, R1, R2, C1, and C2 that reflect the battery aging mechanism. Finally, the multi-source fused features are input into a GRU-Attention neural network model for training and prediction, utilizing gated recurrent units (GRUs) to model temporal characteristics and a self-attention mechanism to enhance key feature responses. Experimental results show that this method significantly outperforms traditional methods in evaluation metrics such as RMSE, MAE, and MAPE. It achieves high-precision estimation of the health status of lithium battery packs and is applicable to intelligent health prediction of battery management systems and energy storage devices for new energy vehicles.

[0009] The technical solution adopted in this invention is as follows:

[0010] A method for estimating the state of health of lithium battery packs based on the fusion of data-driven approaches and equivalent circuit models includes the following steps:

[0011] Step 1. Data Acquisition:

[0012] Obtain actual operating data of lithium battery packs;

[0013] Step 2. Data Preprocessing:

[0014] Determine the positive or negative current to identify the charging segment, and use SOC changes and duration to filter effective charging segments;

[0015] Step 3. Data-driven extraction of health factors:

[0016] Based on electrical performance, temperature characteristics, battery consistency, and operating condition characteristics, health factors such as capacity change rate, voltage change rate, cumulative mileage, SOC statistical characteristics, and temperature statistical characteristics are extracted. Pearson correlation coefficient is used to screen data with strong correlation to drive health factors.

[0017] Step 4. Equivalent circuit model construction and model parameter acquisition:

[0018] A second-order RC equivalent circuit model is constructed, and the model parameters are dynamically identified using the recursive least squares method with forgetting factor (FFRLS) to obtain the health factor of the equivalent circuit.

[0019] Step 5. Data-driven fusion with equivalent circuit model parameters:

[0020] The data-driven health factors in step 3 and the equivalent circuit health factors in step 4 are fused into a comprehensive feature set.

[0021] Step 6. Construct a GRU-Attention model, taking the comprehensive feature set as input, using a gated recurrent unit (GRU) to capture the long-term and short-term dependencies of the data, and enhancing the feature expression through a self-attention mechanism (Attention) to output the predicted state of health (SOH) of the lithium battery pack.

[0022] Step 7. Model Training and Prediction:

[0023] The GRU-Attention model is trained using training data and then used to make predictions on test data. The model's performance is evaluated on metrics such as mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE).

[0024] In step 1, actual operating data of the lithium battery pack is obtained, including total voltage, total current, SOC, maximum single-cell voltage, minimum single-cell voltage, temperature, and cumulative mileage. These data are recorded by the battery management system according to timestamps to form an operating data table. The data is collected at 20-second sampling intervals.

[0025] Among them, the total voltage and total current are used to determine the charging and discharging state and the rate of capacity change;

[0026] SOC is used for segmented charge / discharge identification and capacity estimation;

[0027] The highest and lowest values ​​of a single cell voltage reflect the consistency and balance of the battery.

[0028] Temperature affects the rate of electrochemical reactions and the aging process;

[0029] Cumulative mileage or number of cycles is used to assess usage intensity and lifespan consumption.

[0030] In step 2, the battery reference capacity is calculated using the ampere-hour inverse integral method, and effective charging segments are screened by current direction. Data points with SOC change greater than 70% and time length greater than 50 are set as effective charging segments. The IQR method is used to process outliers and remove outliers.

[0031] The inverse integral method of ampere-hours is as follows:

[0032]

[0033] Among them: Q actual I(t) represents the current actual capacity of the battery (unit: Ah); I(t) represents the charging current (unit: A); ΔSOC represents the change in SOC during the charging process; and t and t' represent the start and end times of the charging process, respectively.

[0034] The IQR method handles abnormal data as follows:

[0035] Lower Bound = Q1 - 1.5 × IQR;

[0036] Upper Bound = Q3 + 1.5 × IQR;

[0037] Wherein, LowerBound represents low outliers and UpperBound represents high outliers; Q1 represents the data value located in the lower 25% position after all data are sorted by size, indicating the lower quartile of the data distribution; Q3 represents the data value located in the upper 25% position, indicating the upper quartile of the data distribution; IQR represents the middle 50% range of the data distribution, which is the difference between the third quartile and the first quartile.

[0038] In step 3, multiple operating conditions and electrical characteristics are extracted, including capacity change rate (dq / dt), voltage change rate (dv / dt), charging rate response (dv / di), temperature increment rate (dT / dSOC), SOC consistency, single-cell voltage difference, cumulative mileage, and charging / discharging temperature rise. Pearson coefficient is used for feature screening.

[0039] In step 3, the data-driven health factors include capacity change rate (dq / dt), voltage change rate (dv / dt), cumulative mileage, highest temperature statistical characteristics, lowest temperature statistical characteristics, and SOC statistical characteristics. Specifically, Pearson correlation coefficient analysis is used to assess the importance of these features, and features with a correlation greater than 0.5 with the battery health status are selected.

[0040] The Pearson correlation coefficient is used to measure the degree of linear correlation between two variables, and its formula is:

[0041]

[0042] Where: r 3,5 This represents the Pearson correlation coefficient between variables X and Y; X i Y represents the feature value of the i-th sample; i This represents the health status of the i-th sample; and r are the means of X and Y, respectively; 3,5 The value ranges from -1 to 1. The closer it is to ±1, the stronger the linear correlation, and the closer it is to 0, the weaker the correlation.

[0043] In step 4, the mathematical expression of the second-order RC equivalent circuit model is:

[0044]

[0045] In the above formula, U represents the battery terminal voltage, which is the actual output voltage of the battery pack and is the output variable predicted by the model; Uocv represents the battery open-circuit voltage, which represents the battery voltage under no load or steady state and reflects the chemical state of the battery; U1 represents the voltage of the first-order RC branch; U2 represents the voltage of the second-order RC branch; I represents the battery current; and R represents the battery internal resistance. This represents the time derivative of the voltage of the first-order RC branch. R1 and R2 represent the time derivative of the second-order RC branch voltage; C1 and C2 represent the polarization internal resistance; C1 and C2 represent the polarization capacitance.

[0046] The specific formula for parameter identification in the recursive least squares (FFRLS) method with forgetting factor is as follows:

[0047]

[0048] In the formula: This represents the parameter estimation vector at time k; y(k) represents the parameter estimation vector at time k-1, i.e., the parameters that have been updated at the previous time; K(k) represents the gain vector at time k; y(k) represents the observed output value at time k. This represents the transpose of the regression vector at time k;

[0049]

[0050] In the formula: P(k-1) represents the covariance matrix at time k-1; Represents the regression vector; μ represents the forgetting factor;

[0051]

[0052] In the formula: P(k) represents the covariance matrix at time k;

[0053] The battery internal resistance R0, polarization internal resistance R1, R2, and polarization capacitances C1 and C2 are obtained as the health factors of the equivalent circuit.

[0054] In step 6, the GRU-Attention model includes a GRU layer and an Attention layer, wherein the calculation formulas for the reset gate, update gate, and hidden state of the GRU layer are as follows:

[0055] r=σ(W r [h 81 ,x]+b r );

[0056] In the formula: r represents the reset gate vector; W r It is the weight matrix of the reset gate; h 81 The hidden state represents the previous time step; x represents the current input vector; b represents the hidden state at the previous time step. r The term σ represents the bias term; σ() represents the sigmoid activation function.

[0057] z=σ(W z [h 81 ,x]+b z );

[0058] In the formula: z represents the update gate vector; W z This is the weight matrix of the updated gate; b z Indicates the bias term;

[0059]

[0060] In the formula: Indicates the candidate hidden state; W[r⊙h 81 [x] represents the concatenation vector of the weight matrix applied to the previous hidden state and the current input after the reset; r represents the reset gate vector; b represents the bias vector;

[0061] ⊙ represents element-wise multiplication, and tanh represents the tanh activation function;

[0062]

[0063] In the formula: h represents the final hidden state of GRU at time t;

[0064] The formula for calculating self-attention is:

[0065]

[0066] In the formula: Attention(Q,K,V) represents the output vector matrix, which generates enhanced feature representations by weighting sequence features; Q represents the query vector; K represents the key vector; V represents the value vector; d kThe dimension of the key vector is represented by ; softmax() represents the normalization function.

[0067] This invention provides a method for estimating the state of health of lithium battery packs based on the fusion of data-driven approaches and equivalent circuit models. The technical advantages are as follows:

[0068] 1) High accuracy: By integrating physical model features, it effectively compensates for the instability of pure data-driven model predictions, and the RMSE of SOH estimation can be reduced to below 1.94.

[0069] 2) High interpretability: Circuit parameters such as R0 and C2 have clear physical meanings and can reflect the aging mechanism of the battery.

[0070] 3) Strong generalization: It has good adaptability to new energy vehicle data under complex actual working conditions and strong noise resistance.

[0071] 4) Wide adaptability: Applicable to various lithium battery types and structures, especially modular battery pack scenarios.

[0072] 5) The method proposed in this invention combines the advantages of data-driven and physical modeling, which significantly improves the accuracy of predicting the health status of lithium battery packs and has good prospects for practical application. Attached Figure Description

[0073] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0074] Figure 1 This is a flowchart of a lithium battery lifetime prediction method based on GRU-Attention according to the present invention.

[0075] Figure 2 The diagram shows the structure of the second-order RC equivalent circuit model.

[0076] Figure 3 Here is a diagram of the GRU-Attention model structure;

[0077] Figure 4(a) shows the LightGBM fitting curve;

[0078] Figure 4(b) shows the XGBoost fitting curve;

[0079] Figure 4(c) shows the LSTM-Attention fitting curve;

[0080] Figure 4(d) shows the GRU-Attention fitting curve. Detailed Implementation

[0081] The specific embodiments of the present invention will now be described with reference to the accompanying drawings, so that those skilled in the art can better understand the present invention.

[0082] Figure 1 This is a flowchart of a lithium battery lifetime prediction method based on GRU-Attention according to the present invention.

[0083] S1: Data Acquisition and Reference Capacity Calculation:

[0084] Operating data from eight new energy vehicles was collected at 20-second intervals. The data included fields such as total voltage, total current, state of charge (SOC), temperature, cell voltage difference, and cumulative mileage. Forward current was used to filter charging segments; intervals with SOC changes exceeding 70% and containing more than 50 sampling points were selected as valid segments. The inverse ampere-hour integral method was used to estimate the charging capacity of each segment, yielding the current reference battery capacity.

[0085] S2: Battery health factor extraction and feature fusion:

[0086] Eighteen data-driven features, such as dq / dt, dv / dt, Tchg_Cap, mean SOC, and unit consistency, are extracted from the effective data segment. The correlation between each feature and SOH is calculated, and key features with a correlation greater than 0.5 are retained. A second-order RC model is established and identified using FFRLS to obtain the model parameter sequence R0, R1, C1, R2, and C2. After normalization, these parameters are used as the circuit health factor.

[0087] Figure 2 This is a structural diagram of the second-order RC equivalent circuit model.

[0088] S3: Deep Model Training and Prediction:

[0089] S 3.1: Model Construction and Hyperparameter Settings:

[0090] In this step, the parameters of the model's input, hidden, and output layers are first determined based on the dimensions of the data after preprocessing. The specific settings are as follows:

[0091] Input Layer: The dimension (input_size) of the input layer is determined by the number of health factors input. Each health factor represents an input feature; therefore, the size of the input layer is equal to the number of extracted health factors. Based on the features extracted in step 2, the dimension of the input layer is 18.

[0092] Hidden Layer: The size of the hidden layer is determined using a grid search technique. To balance model complexity and computational resource consumption, the dimension (hidden_size) of the hidden layer is set to 128. This size choice ensures that the model can capture sufficient non-linear information when processing complex features, while avoiding overfitting.

[0093] Output Layer: The output layer has a dimension (output_size) of 1 and represents the predicted State of Health (SOH) value. This output layer is calculated using a linear function to ensure that the model output is a continuous value representing the battery's health state.

[0094] Regarding the selection of hyperparameters, other parameters such as learning rate and batch size were determined to be optimal values ​​suitable for the task after multiple experiments and optimizations.

[0095] S 3.2: GRU Temporal Feature Extraction:

[0096] In deep learning models, GRU (Gated Recurrent Unit) layers are used to extract temporal features from input data, especially long-term and short-term dependencies in battery health (SOH) prediction tasks. Compared to traditional RNNs (Recurrent Neural Networks), GRUs have better gradient propagation capabilities, and are particularly stable when processing long-term temporal data.

[0097] 3.2.1: GRU Network Structure:

[0098] Reset gate: The reset gate controls the correlation between the current state and the previous state, determining the degree of influence of the current input information on the model. Its calculation formula is:

[0099] r=σ(W r [h 81 ,x]+b r );

[0100] Among them, W r It is the weight matrix of the reset gate, b r σ is the bias term, and σ is the sigmoid activation function.

[0101] Update gate: The update gate determines how much state information from the previous time step should be retained in the current hidden state. Its calculation formula is:

[0102] z=σ(W i [h 81 ,x]+b i );

[0103] Among them, W i It is the weight matrix of the updated gate, b i It is a bias term.

[0104] Candidate hidden state: The candidate hidden state is the latent memory of the current time step. It is obtained by weighting the hidden state of the previous time step and the input of the current time step. The calculation formula is:

[0105]

[0106] Here, ⊙ represents element-wise multiplication, and tanh is the tanh activation function.

[0107] Final hidden state: The final hidden state is a weighted average of the previous time step state and the candidate hidden states, controlling the update of memory. Its calculation formula is:

[0108]

[0109] Through the above process, the GRU network can effectively capture the temporal information of battery state of health (SOH) changes and model the long-term and short-term dependencies in the input data.

[0110] S 3.3: Self-attention mechanism (Attention) feature enhancement:

[0111] After the GRU layer processes the data, a self-attention mechanism is applied to further enhance the features of its output hidden state sequence. The self-attention mechanism can help the model highlight the important relationships between features in the sequence, especially in long-term data, and can enhance the model's ability to perceive key features.

[0112] Attention weight calculation:

[0113] First, an attention weight matrix is ​​calculated to assign different weights to each element in the input sequence, highlighting the moments most critical for SOH prediction. The calculation formula is as follows:

[0114] Attention = sfotmax(HH) V );

[0115] Where H is the hidden state sequence output by the GRU layer, H V It is its transpose. The softmax function is used to transform the dot product result into a probability distribution and generate the attention weight matrix.

[0116] Feature weighting and aggregation:

[0117] The obtained attention weight matrix is ​​used to weight the hidden state sequence H output by the GRU layer, resulting in the weighted feature matrix Context. This weighting process further strengthens features closely related to changes in SOH, thereby improving the model's prediction accuracy. The calculation formula is as follows:

[0118] Context = Attention × H;

[0119] S3.4: Feature Mapping and Output Prediction:

[0120] The feature matrix Context, processed by the self-attention mechanism, is fed into a linear fully connected layer to obtain the predicted value of the battery health state (SOH).

[0121] Linear mapping:

[0122] The feature matrix Context is fed into a linear layer for final output prediction, calculated as follows:

[0123] y pre =W·Context+b;

[0124] Where W and b are the weights and biases of the linear layer, and the final output y pre This is a predicted value for the battery's state of health (SOH).

[0125] S 3.5: Definition and Calculation of Loss Function:

[0126] To evaluate the model's predictive performance, the mean squared error (MSE) is used as the loss function. The formula for calculating the loss function is:

[0127]

[0128] in, Let be the model's predicted value for the i-th sample. This is the actual value.

[0129] S 3.6: Model Parameter Optimization

[0130] The Adam optimizer is used to optimize the model's parameters. The Adam optimizer adaptively adjusts the learning rate for each parameter, helping to accelerate model convergence. The hyperparameter settings during training are as follows:

[0131] Learning rate (lr): 0.05;

[0132] ∈: 0.0001;

[0133] Training epochs: 500;

[0134] Every 10 training rounds, the current loss value is recorded and output to ensure that the model gradually converges during training.

[0135] S 3.7: Saving Optimal Model Parameters:

[0136] After completing model training and validating its performance, the optimal model parameters are saved. These parameters can be used for subsequent real-time predictions or further analysis to ensure the accuracy and stability of battery state of health (SOH) predictions.

[0137] The above sub-steps complete the entire process of deep model training and prediction. The implementation of each step effectively improves the model's performance in the battery state of health (SOH) prediction task and provides a reliable prediction tool for subsequent work.

[0138] Figure 3 This is a diagram of the GRU-Attention model structure.

[0139] S4: Model Prediction and Evaluation

[0140] Model prediction: Input the standardized test set data into the trained GRU-Attention model to obtain the remaining lifetime prediction value for each test sample.

[0141] Performance evaluation:

[0142] The mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and mean square error (MSE) are used to evaluate the predictive performance of a model.

[0143] The formula is as follows:

[0144]

[0145] Among them, y i For the true value, For predicted values, is the mean of the true values, and n is the number of samples.

[0146] To illustrate the technical effectiveness of this invention, datasets #2-#8 were used for training, and dataset #1 was used for testing. LightGBM, XGBoost, and LSTM-Attention were employed using the same training strategy to estimate the battery state of equilibrium (SOH). Table 1 lists the SOH estimation results for dataset #1.

[0147] Table 1#1 shows the SOH estimation results for the dataset.

[0148]

[0149] Figures 4(a) to 4(d) Fit curves for each prediction model.

[0150] This invention proposes a lithium battery state of health (SOH) estimation method based on a GRU-Attention deep learning model. By collecting and preprocessing actual new energy vehicle operating data, an effective data-driven feature and physical model feature fusion mechanism is constructed, significantly improving the accuracy and stability of lithium battery life prediction.

[0151] In the specific implementation process, the current reference capacity of the battery is first calculated using the ampere-hour inverse integral method. Eighteen data-driven health factors are extracted, including voltage change rate (dv / dt), capacity change rate (dq / dt), temperature change rate (Tchg / Cap), and SOC statistical characteristics. Physically meaningful model parameters (R0, R1, R2, C1, C2) are obtained through a second-order RC equivalent circuit model. Subsequently, features with significant correlation to SOH are screened through correlation analysis, and these data-driven features and physical model features are combined as model inputs.

[0152] Next, a GRU network combined with a self-attention mechanism is used to deeply explore the long-term and short-term dependencies in the time-series data and highlight key features to improve the model's prediction accuracy. During model training, mean squared error (MSE) is used as the loss function, and the Adam optimization algorithm is combined to optimize the model parameters, ultimately obtaining a set of high-performance and stable model parameters.

[0153] Experimental results show that the proposed GRU-Attention model exhibits significant advantages compared to traditional LightGBM, XGBoost, and LSTM-Attention models. Taking the actual new energy vehicle dataset #1 as an example, the proposed method achieves a significantly lower RMSE (1.947), MAE (1.544), MAPE (0.0119), and MSE (3.7914) compared to the comparative models. This demonstrates that the proposed method can more accurately predict battery health and is more suitable for complex operating conditions, providing a reliable and efficient technical means for battery management systems.

Claims

1. A lithium battery pack state of health estimation method based on data-driven and equivalent circuit model fusion, characterized in that The method comprises the following steps: Step 1: Obtain actual operation data of the lithium battery pack; Step 2: Judge the charging section according to the positive and negative identification of the current, and use the SOC change and the duration to screen the effective charging section; Step 3: Based on the electrical performance, temperature characteristics, battery consistency and operation condition characteristics, extract the capacity change rate, voltage change rate, cumulative mileage, SOC statistical characteristics and temperature statistical characteristics and other health factors, and use the Pearson correlation coefficient to screen the data-driven health factors with strong correlation; Step 4: Construct a second-order RC equivalent circuit model, and use the recursive least squares method with a forgetting factor FFRLS to dynamically identify the model parameters to obtain the equivalent circuit health factors; Step 5: Fuse the data-driven health factors in step 3 and the equivalent circuit health factors in step 4 into a comprehensive feature set; Step 6. Construct a GRU-Attention model, take the comprehensive feature set as the input, use the gated recurrent unit GRU to capture the long-term and short-term dependence relationship of the data, and enhance the feature expression through the self-attention mechanism Attention, and output the predicted lithium battery pack state of health SOH.

2. The lithium battery pack state of health estimation method based on data-driven and equivalent circuit model fusion of claim 1, wherein: It also includes step 7: model training and prediction: Use the training data to train the GRU-Attention model, and predict the test data to evaluate the performance of the model in the mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) and other indicators.

3. The lithium battery pack state of health estimation method based on data-driven and equivalent circuit model fusion of claim 1, wherein: In step 1, the actual operation data of the lithium battery pack includes total voltage, total current, SOC, single cell voltage maximum value, single cell voltage minimum value, temperature, cumulative mileage; these data are recorded by the battery management system according to the time stamp to form an operation data table; and are collected at a sampling interval of 20s; Among them, the total voltage and the total current are used to judge the charging and discharging state and the capacity change rate; The SOC is used for segmented charging and discharging identification and capacity estimation; The single cell voltage maximum value and the single cell voltage minimum value reflect the battery consistency and the balance state; Temperature affects the electrochemical reaction rate and the aging process; The cumulative mileage or the cycle number is used to evaluate the use intensity and the life consumption.

4. The lithium battery pack state of health estimation method based on data-driven and equivalent circuit model fusion of claim 3, wherein: In step 2, the ampere-hour inverse integration method is used to calculate the battery reference capacity, the effective charging section is screened through the current direction, the SOC change greater than 70% and the time length greater than 50 data points are set as the effective charging section; and the IQR method is used for abnormal value processing, and the outliers are removed.

5. The lithium battery pack state of health estimation method based on data-driven and equivalent circuit model fusion of claim 4, wherein: The ampere-hour inverse integration method is as follows: where: Q actual is the current actual capacity of the battery; I(t) is the charging current, ΔSOC is the change in SOC during the charging process, and t and t0are the start and end times of the charging process, respectively.

6. The lithium battery pack state of health estimation method based on data-driven and equivalent circuit model fusion of claim 5, wherein: The IQR method for abnormal data processing is as follows: Lower Bound=Q1-1.5×IQR; Upper Bound=Q3+1.5×IQR; Wherein, LowerBound is the low abnormal value point, UpperBound is the high abnormal value point; Q1 represents the data value located at the lower 25% position after all capacity data are sorted by size, indicating the lower quartile of data distribution; Q3 is located at the upper 25% position of the data value, indicating the upper quartile of the data distribution; IQR represents the range of the middle 50% of the data distribution, which is the difference between the third quartile and the first quartile.

7. The lithium battery pack state of health estimation method based on data-driven and equivalent circuit model fusion of claim 6, wherein: In step 3, the data-driven health factors include capacity change rate, voltage change rate, cumulative mileage, highest temperature statistical characteristics, lowest temperature statistical characteristics, and SOC statistical characteristics. Specifically, Pearson correlation coefficient analysis is used to evaluate the importance of the features, and features with a correlation greater than 0.5 with the battery health status are selected. The Pearson correlation coefficient is used to measure the degree of linear correlation between two variables, and its formula is: wherein: r X,Y represents the Pearson correlation coefficient value between the two variables X and Y; X i represents the feature value of the i-th sample; Y i represents the health state corresponding to the i-th sample; and are the mean values of X and Y, respectively; r X,Y The value of r ranges from -1 to 1, and the closer to ±1 indicates a stronger linear correlation, and the closer to 0 indicates a weaker correlation.

8. The lithium battery pack state of health estimation method based on data-driven and equivalent circuit model fusion of claim 7, wherein: In step 4, the mathematical expression of the second-order RC equivalent circuit model is: In the above formula, U represents the battery terminal voltage, i.e. the actual output voltage of the battery pack, which is an output variable predicted by the model; Uocv represents the open-circuit voltage of the battery, which represents the battery voltage under no load or steady state, reflecting the chemical state of the battery; U : represents the voltage of the first-order RC branch; U ； represents the voltage of the second-order RC branch; I represents the battery current; R0 represents the internal resistance of the battery; represents the time derivative of the voltage of the first-order RC branch; represents the time derivative of the voltage of the second-order RC branch; R : , R ； represents the polarization resistance; C : , C ； represents the polarization capacitance.

9. The lithium battery pack state of health estimation method based on data-driven and equivalent circuit model fusion of claim 8, wherein: The specific formula for identifying FFRLS parameters using recursive least squares with a forgetting factor is as follows: wherein: represents the parameter estimation vector at time k; represents the parameter estimation vector at time k-1, i.e. the updated parameter at the previous time; K(k) represents the gain vector at time k; y(k) represents the observed output value at time k; represents the transpose of the regression vector at time k; P(k - 1) represents the covariance matrix at time k - 1; represents the regression vector; v represents the forgetting factor; In the formula: P(k) represents the covariance matrix at time k; The battery internal resistance R0, polarization internal resistance R1, R2, and polarization capacitances C1 and C2 are obtained as the health factors of the equivalent circuit.

10. The lithium battery pack state of health estimation method based on data-driven and equivalent circuit model fusion of claim 9, wherein: In step 6, the GRU-Attention model includes a GRU layer and an Attention layer, wherein the calculation formulas for the reset gate, update gate, and hidden state of the GRU layer are as follows: r t = σ(W r [ h t-1 , x t ] + b r ); where: r t denotes the reset gate vector; r is the weight matrix of the reset gate; h t-1 denotes the hidden state at the previous time step; x t denotes the current input vector; b r denotes the bias term; σ() denotes the sigmoid activation function; z t = σ(W z [ h t-1 , x t ] + b z ); where: z t represents the update gate vector; W z is the weight matrix of the update gate; b z represents the bias term; In the formula: represents a candidate hidden state; W[r t-1 ,x t ] represents a weight matrix acting on the reset previous hidden state and the current input splicing vector; r represents a reset gate vector; and b represents a bias vector. ⊙ represents element-wise multiplication, and tanh represents the tanh activation function; In the formula: h represents the final hidden state of GRU at time t; The formula for calculating the self-attention mechanism (Attention) is: In the formula: Attention(Q, K, V) represents an output vector matrix, which generates an enhanced feature representation by weighting the sequence features; Q represents a query vector; K represents a key vector; V represents a value vector; d k represents the dimension of the key vector; and softmax() represents a normalization function.

Images