Method for estimating state of charge of series battery pack

Abstract

The invention discloses a method for estimating the state of charge of a series battery pack, which comprises the following specific steps of: firstly, constructing a battery pack equivalent circuit model based on online acquired battery parameters and series circuit characteristics, carrying out standardization processing on the terminal voltage Uti and the internal resistance R0i of each battery monomer, and carrying out Euclidean distance analysis to obtain the state of charge of the series battery pack; identifying the battery with the maximum difference as a representative battery monomer, then constructing a spatial state equation, and performing SOC estimation by adopting a self-adaptive extended Kalman filter with residual drive, thereby obtaining a battery pack state-of-charge base value SOCb; meanwhile, after abnormality judgment and effective data screening are carried out on battery monomer data collected offline, the voltage difference feature delta Uti and an estimated historical residual sequence eh are fused to form joint input, a CNN-GRU-SE hybrid neural network optimized by a KOA algorithm is used for estimating a battery pack charge state compensation value delta SOCb, and after processing is carried out through a moving average filter, the battery pack charge state compensation value delta SOCb is calculated. And superposing with the SOCb to obtain an estimated value SOCr of the state of charge of the series battery pack.

Description

Technical Field

[0001] The present invention belongs to the technical field of design and control of large-capacity battery energy storage systems in smart grids, and relates to a method for estimating the state of charge of a series-connected battery pack. Background Art

[0002] With the rapid development of battery technology, series-connected battery packs are widely used in electric vehicles and energy storage due to their ability to effectively improve system voltage and energy capacity. However, the problem of inter-cell inconsistency caused by multiple cells in series is becoming increasingly prominent. This is mainly manifested in differences in manufacturing, aging, and usage environment. These inconsistencies lead to different performance of each cell during the charge and discharge process, directly affecting the overall state of charge (SOC) estimation of the battery pack. Accurately estimating the SOC of series-connected battery packs is crucial for achieving efficient energy management, improving safety, and extending service life. Traditional SOC estimation methods are mostly based on cell characteristics, which are difficult to effectively address inter-cell inconsistencies and have limited estimation accuracy. Therefore, it is of great significance to develop new SOC estimation methods for series-connected battery packs that combine cell status monitoring with advanced algorithms to achieve accurate prediction of battery pack and cell SOC. This not only helps optimize the battery pack's operating strategy and maintenance management, but also improves the performance and reliability of the entire system.

[0003] The patent (CN116165546A) compensates for the inconsistency of battery cells by introducing a virtual measurement noise model, uses variational Bayesian inference to estimate noise statistics, and combines unscented Kalman filtering to perform SOC recursion, thereby improving estimation accuracy and real-time performance while maintaining low computational complexity. However, this method mainly relies on the virtual noise construction model for indirect compensation, and does not directly quantify and identify the performance differences between battery cells, resulting in insufficient physical interpretability. To address this shortcoming, the present invention standardizes the terminal voltage and internal resistance of the battery cells, constructs a similarity index, directly calculates the performance differences between cells, and clearly identifies the representative cells with the most significant inconsistencies, thus having stronger interpretability and practical adaptability. On this basis, an adaptive extended Kalman filter with residual drive is used to estimate SOC, and a neural network is introduced to dynamically compensate for the estimation error, thereby achieving a more targeted and physically based estimation mechanism for the inconsistency problem, further improving the accuracy and robustness of SOC estimation. Summary of the Invention

[0004] The problem solved by the present invention is to provide a method for estimating the state of charge (SOC) of a series battery pack. First, to address the performance inconsistency of individual cells within the battery pack, the battery terminal voltage and internal resistance are standardized to construct a multidimensional similarity index, accurately quantify individual differences, and effectively identify inconsistency characteristics, solving the problem of difficulty in evaluating performance differences within the battery pack. Second, the traditional AEKF algorithm is improved by introducing a residual drive mechanism to achieve dynamic adjustment and error correction of SOC estimation, thereby improving the stability of the filtering algorithm and the estimation accuracy. Third, an SOC compensator based on a bidirectional feedback gating mechanism is designed, which uses a deep neural network to compensate for the estimation residual, enhancing the system's adaptability to inconsistencies and dynamic changes, and significantly improving the accuracy and robustness of the battery pack SOC estimation.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for estimating the state of charge of a series battery pack, comprising the following specific steps:

[0006] S1. Measure and obtain battery pack and battery cell data (1), including battery pack current I and battery cell open circuit voltage U 0i , terminal voltage U ti And internal resistance R 0i , where i represents the i-th battery cell;

[0007] S2, based on the online collected battery parameters and series circuit characteristics (2) to build a battery pack equivalent model, the measured terminal voltage U ti 、Internal resistance R 0i Perform parameter standardization (4), calculate the Euclidean distance D (5) and convert it into a similarity value S, and select the battery cell with the lowest terminal voltage similarity coefficient and resistance similarity coefficient, that is, the battery cell with the greatest difference, as the representative battery cell (6);

[0008] S3, based on the battery pack equivalent model and the representative battery cell, a spatial state equation is constructed, and the state of charge is estimated by using an adaptive extended Kalman filter (8) with residual drive. The AEKF algorithm module (7) with residual drive first obtains the estimated residual information through the residual drive forgetting regulator, and dynamically adjusts the forgetting factor α of the process noise and observation noise based on the residual change trend. Q With α R , in order to achieve real-time adaptive update of the noise covariance matrix Q and R, and then complete the state initialization, state prediction and measurement prediction, Kalman gain calculation and other steps in sequence according to the extended Kalman EKF filter algorithm process. In the state update stage, the estimated residual between the current observation value and the predicted observation value is used, combined with the adaptively adjusted filter gain, to update the system state variable, and finally output the battery pack state of charge base value SOC b ;

[0009] S4: Input the battery offline data into the bidirectional gated SOC compensator module (9), and pass through the abnormality discriminator (10) and the valid data filter (11) in sequence to fuse the voltage difference characteristics ΔU of each battery cell. ti and the estimated historical residual series e h As input, the CNN-GRU-SE hybrid neural network (12) optimized by the Kepler algorithm KOA is used to output the battery pack state of charge compensation value ΔSOC b , and smoothed by a sliding average filter (13);

[0010] S5, the battery pack state of charge base value SOC b and compensation value ΔSOC b Superposition to obtain the estimated state of charge SOC of the series battery pack r .

[0011] The process of selecting a representative battery in step S2 is as follows:

[0012] (1) For the terminal voltage U of all battery cells ti and internal resistance R 0i , first standardize the parameters (4), and the expression is: in, and are the standardized terminal voltage and internal resistance of the ith battery cell, μ U and μ R are the average terminal voltage and internal resistance of all battery cells, σ U and σ R are the standard deviations of the terminal voltage and internal resistance of all battery cells, x i Refers to the original value of a battery terminal voltage and internal resistance;

[0013] (2) For each battery cell i, calculate the Euclidean distance (5), which is expressed as: Where j represents the jth battery cell, D Ui and D Ri They are respectively expressed as the Euclidean distance between the terminal voltage dimension and the internal resistance dimension between the i-th battery and the rest of the batteries;

[0014] (3) Calculate the similarity coefficient, the calculation formula is: The battery cell with the smallest similarity coefficient is selected as the representative battery (6), where S Ui and S Ri It is expressed as the terminal voltage similarity coefficient and internal resistance similarity coefficient between the i-th battery cell and the representative battery.

[0015] The residual-driven AEKF algorithm (7) in step S3 is designed as follows:

[0016]

[0017] Where Q k is the process noise covariance, α Q , α R is the forgetting factor, K k is the Kalman gain, R k is the measurement noise covariance, is the predicted state, y k is the measured voltage, e k The voltage residual is introduced, and a residual-driven double forgetting factor mechanism is introduced to achieve dynamic adaptive adjustment of the covariance between the filtering process and the observation noise, thereby improving the SOC estimation accuracy.

[0018] The bidirectional feedback gated SOC compensator (9) described in step S4 is designed as follows:

[0019] The battery offline data is sequentially input into the abnormality discriminator (10) and the effective data filter (11), abnormal samples are eliminated and normal data is extracted, and the voltage difference characteristics ΔU of each battery cell are integrated. ti and the estimated historical residual series e h As the input variable of the CNN-GRU-SE hybrid neural network (12), the output is the battery pack state of charge compensation value ΔSOC b The neural network training and optimization process is as follows: ① The collected battery cell voltage difference characteristics ΔU ti and the residual sequence e h ② Normalize the input sample to form a normalized one; ② Construct a deep neural network model composed of a convolutional neural network (CNN), a gated recurrent unit (GRU), and an attention guidance module (SE) to improve the local temporal feature extraction capability and the weighted discrimination capability of key dynamic features; ③ Use the KOA optimization algorithm to globally optimize the hyperparameters, including the number of CNN convolution kernels, the number of GRU layer nodes, and the learning rate, to improve the model training efficiency and prediction accuracy; ④ The normalized input sample X in Input network, complete feature extraction, time series modeling and attention weighting in sequence, and output is the battery state of charge compensation value estimation And the actual compensation value ΔSOC b For the supervision signal, calculate the loss function: ⑤ Based on the principle of error minimization, the back propagation algorithm is used to update the network parameters until the model training converges; ⑥ Repeat steps ④ and ⑤, continuously iteratively train the model until the loss function converges to the preset accuracy threshold, and finally output the trained neural network model and its optimal parameter set for the subsequent real-time prediction of the state of charge compensation value; ⑦ The network output ΔSOC b After being processed by a sliding average filter (13), the fluctuation noise is suppressed and the estimation stability is improved.

[0020] Compared with the disclosed patent (CN116165546A), the beneficial effects of the present invention are as follows: First, by standardizing the terminal voltage and internal resistance of the battery cell, calculating the Euclidean distance and converting it into a similarity value, and setting an inconsistency threshold to quantify the performance differences of the cells within the battery pack, the problem of insufficient identification of battery pack cell inconsistency is effectively solved; Second, based on the similarity analysis results, the representative battery cells with the most significant differences are determined, the spatial state equation is constructed, and the adaptive extended Kalman filter algorithm with residual drive is used to estimate the SOC base value, thereby enhancing the stability and responsiveness of the estimation; Third, a bidirectional gated neural network compensator that integrates voltage difference characteristics and residual history is introduced to dynamically correct the SOC estimation deviation, further improving the SOC estimation accuracy and system robustness under cell inconsistency and complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A structural diagram of a method for estimating the state of charge of a series battery pack;

[0022] Figure 2 It is the equivalent circuit model diagram of the battery cell;

[0023] Figure 3 This is the equivalent circuit model diagram of a series battery pack;

[0024] Figure 4 This is a structural diagram of a series lithium-ion battery pack. DETAILED DESCRIPTION

[0025] The present invention will be further described in detail below with reference to specific examples, which are intended to explain the present invention rather than to limit it.

[0026] Example: Figure 1 As shown, a method for estimating the state of charge of a series battery pack is provided. The series battery pack is formed by connecting a plurality of battery cells in series. The system is formed by connecting n battery cells in series, where n is a natural number greater than or equal to 1. Figure 4 The specific steps are as follows:

[0027] S1. Measure and obtain battery pack and battery cell data (1), including battery pack current I and battery cell open circuit voltage U 0i , terminal voltage Uti And internal resistance R 0i , where i represents the i-th battery cell;

[0028] The measured battery pack current I=1A and battery cell open circuit voltage U 0i =[3.65,3.62,3.59,3.61]V, each battery cell terminal voltage U ti =[3.6,3.61,3.7,3.62]V, battery cell internal resistance R 0i =[4.0,4.1,4.8,4.2]Ω.

[0029] S2, based on the online collected battery parameters and series circuit characteristics (2) to build a battery pack equivalent model, the measured terminal voltage U ti 、Internal resistance R 0i Perform parameter standardization (4), calculate the Euclidean distance D (5) and convert it into a similarity value S, and select the battery cell with the lowest terminal voltage similarity coefficient and resistance similarity coefficient, that is, the battery cell with the greatest difference, as the representative battery cell (6); the specific implementation steps are as follows:

[0030] (1) According to the formula for calculating mean and standard deviation, we can get:

[0031]

[0032] (2) According to the parameter standardization calculation formula, we can get:

[0033]

[0034] (3) Calculating the Euclidean distance yields:

[0035]

[0036] (4) According to the similarity coefficient calculation formula, we can get:

[0037]

[0038] Based on the above calculations, it can be seen that the similarity coefficient between the terminal voltage and internal resistance of Cell 3 is the lowest, so Cell 3 is selected as the representative battery cell.

[0039] S3, based on the battery pack equivalent model and the representative battery cell, a spatial state equation is constructed, and the state of charge is estimated by using an adaptive extended Kalman filter (8) with residual drive. The AEKF algorithm module (7) with residual drive first obtains the estimated residual information through the residual drive forgetting regulator, and dynamically adjusts the forgetting factor α of the process noise and observation noise based on the residual change trend. Q With α R, in order to achieve real-time adaptive update of the noise covariance matrix Q and R, and then complete the state initialization, state prediction and measurement prediction, Kalman gain calculation and other steps in sequence according to the extended Kalman EKF filter algorithm process. In the state update stage, the estimated residual between the current observation value and the predicted observation value is used, combined with the adaptively adjusted filter gain, to update the system state variable, and finally output the battery pack state of charge base value SOC b The specific steps for establishing the battery cell equivalent circuit model and its spatial state equation are as follows:

[0040] The series battery pack is composed of four battery cells connected in series. Figure 3 As shown in the figure, the rated voltage of the battery cell is 3.7V, the discharge cut-off voltage is 3V, and the rated capacity is 860mAh. t The functional relationship between it and the current I is:

[0041] U t (t)=U0[SOC(t)]-I(t)[R0(t)+R1(t) / R1(t)jωC1(t)+R2(t) / R2(t)jωC2(t)]

[0042] The relationship between each parameter and SOC is:

[0043]

[0044] Wherein, the model coefficients a0~a5 obtained by fitting the battery experimental data are -0.915, -40.867, 3.632, 0.537, -0.499, and 0.522, respectively; the values ​​of b0~b5 are -0.1463, -30.27, 0.1037, 0.0584, 0.1747, and 0.1288, respectively; the values ​​of c0~c2 are 0.1063, -62.49, and 0.0437, respectively; the values ​​of d0~d2 are -200, -138, and 300, respectively; the values ​​of e0~e2 are 0.0712, -61.4, and 0.0288, respectively; and the values ​​of f0~f2 are -3083, -180, and 5088, respectively.

[0045] According to the above mathematical model, the equivalent circuit model of the battery cell is established with SOC c 、U s and U l is the state vector, and the spatial state equation is obtained as:

[0046]

[0047] [U t,k ]=U o,k -R 0,k I k -Us,k -U l,k +ν k

[0048] Where, ω k is the process noise, ν k To measure noise.

[0049] The specific steps of the residual-driven AEKF algorithm (7) are as follows:

[0050] (1) State initialization:

[0051] Initialize state estimation State covariance P0, process noise covariance Q0, measurement noise covariance R0.

[0052] (2) Status prediction:

[0053]

[0054] Where: is the predicted state, F k is the Jacobian matrix of the state transfer function, P k|k-1 is the prediction covariance, Q k is the process noise covariance.

[0055] (3) Measurement prediction and residual calculation:

[0056]

[0057] Where: is the predicted state, y k is the measured voltage, e k is the voltage residual.

[0058] (4) Kalman gain calculation:

[0059]

[0060] Where: H k is the Jacobian matrix of the observation function, R k is the measurement noise covariance, S k is the residual covariance, K k is the Kalman gain.

[0061] (5) Status update:

[0062]

[0063] (6) Covariance update:

[0064] P k|k =(IK k Hk )P k|k-1

[0065] (7) Adaptive noise covariance update:

[0066]

[0067] Where: α Q , α R is the forgetting factor, which is a fixed value in the initial stage (k<50): α Q =α R =0.8; in the stable stage (k≥50), it gradually approaches: α Q =α R ≈0.95~0.99.

[0068] S4: Input the battery offline data into the bidirectional gated SOC compensator module (9), and pass through the abnormality discriminator (10) and the valid data filter (11) in sequence to fuse the voltage difference characteristics ΔU of each battery cell. ti and the estimated historical residual series e h As input, the CNN-GRU-SE hybrid neural network (12) optimized by the Kepler algorithm KOA is used to output the battery pack state of charge compensation value ΔSOC b , and smoothed by a sliding average filter (13); the specific implementation steps are as follows:

[0069] ① Collected battery cell voltage difference characteristics ΔU ti and the residual sequence e h ② Normalize the input sample to form a normalized one; ② Construct a deep neural network model composed of a convolutional neural network (CNN), a gated recurrent unit (GRU), and an attention guidance module (SE) to improve the local temporal feature extraction capability and the weighted discrimination capability of key dynamic features; ③ Use the KOA optimization algorithm to globally optimize the hyperparameters, including the number of CNN convolution kernels, the number of GRU layer nodes, and the learning rate, to improve the model training efficiency and prediction accuracy; ④ The normalized input sample X in Input network, complete feature extraction, time series modeling and attention weighting in sequence, and output is the battery state of charge compensation value estimation And the actual compensation value ΔSOC b For the supervision signal, calculate the loss function: ⑤ Based on the principle of error minimization, the back propagation algorithm is used to update the network parameters until the model training converges; ⑥ Repeat steps ④ and ⑤, continuously iteratively train the model until the loss function converges to the preset accuracy threshold, and finally output the trained neural network model and its optimal parameter set for the subsequent real-time prediction of the state of charge compensation value; ⑦ The network output ΔSOC bAfter being processed by a sliding average filter (13), the fluctuation noise is suppressed and the estimation stability is improved.

[0070] S5, the battery pack state of charge base value SOC b and compensation value ΔSOC b Superposition to obtain the estimated state of charge SOC of the series battery pack r .

[0071] Finally, it should be noted that the above description is merely an example of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for estimating the state of charge of a series battery pack, wherein the series battery pack is composed of n battery cells connected in series, where n is a natural number greater than 1, the method comprising the following steps: S1. Measure and obtain battery pack and battery cell data, including battery pack current I and battery cell open circuit voltage U 0i , terminal voltage U ti And internal resistance R 0i , where i represents the i-th battery cell; S2, based on the online collected battery parameters and series circuit characteristics, build a battery pack equivalent model, and measure the terminal voltage U ti 、Internal resistance R 0i Perform parameter standardization. By calculating the Euclidean distance D and converting it into a similarity value S, the battery cell with the lowest terminal voltage similarity coefficient and resistance similarity coefficient, that is, the battery cell with the greatest difference, is selected as the representative battery cell. S3. Based on the battery pack equivalent model and the representative battery cell, a spatial state equation is constructed, and the state of charge is estimated using an adaptive extended Kalman filter with residual drive. The AEKF algorithm module with residual drive first obtains the estimated residual information through the residual drive forgetting regulator, and dynamically adjusts the forgetting factor α of the process noise and observation noise based on the residual change trend. Q With α R , in order to achieve real-time adaptive update of the noise covariance matrix Q and R, and then complete the state initialization, state prediction and measurement prediction, Kalman gain calculation and other steps in sequence according to the extended Kalman EKF filter algorithm process. In the state update stage, the estimated residual between the current observation value and the predicted observation value is used, combined with the adaptively adjusted filter gain, to update the system state variable, and finally output the battery pack state of charge base value SOC b ; S4: Input the battery offline data into the bidirectional gated SOC compensator module, pass through the abnormal discriminator and the effective data filter in sequence, and integrate the voltage difference characteristics ΔU of each battery cell. ti and the estimated historical residual series e h As input, the CNN-GRU-SE hybrid neural network optimized by the Kepler algorithm KOA is used to output the battery pack state of charge compensation value ΔSOC b , and smoothed by a sliding average filter; S5, the battery pack state of charge base value SOC b and compensation value ΔSOC b Superposition to obtain the estimated state of charge SOC of the series battery pack r .

2. The method for estimating the state of charge of a series battery pack according to claim 1, wherein: The method for selecting representative batteries is as follows: (1) For the terminal voltage U of all battery cells ti and internal resistance R 0i , first standardize the parameters, and the expression is: in, and are the standardized terminal voltage and internal resistance of the ith battery cell, μ U and μ R are the average terminal voltage and internal resistance of all battery cells, σ U and σ R are the standard deviations of the terminal voltage and internal resistance of all battery cells, x i Refers to the original value of a battery terminal voltage and internal resistance; (2) For each battery cell i, calculate the Euclidean distance, which is expressed as: Where j represents the jth battery cell, D Ui and D Ri They are respectively expressed as the Euclidean distance between the terminal voltage dimension and the internal resistance dimension between the i-th battery and the rest of the batteries; (3) Calculate the similarity coefficient, the calculation formula is: The battery cell with the smallest similarity coefficient is selected as the representative battery, where S Ui and S Ri It is expressed as the terminal voltage similarity coefficient and internal resistance similarity coefficient between the i-th battery cell and the representative battery.

3. The method for estimating the state of charge of a series battery pack according to claim 1, wherein: The design of the AEKF algorithm with residual drive is as follows: Where Q k is the process noise covariance, α Q , α R is the forgetting factor, K k is the Kalman gain, R k is the measurement noise covariance, is the predicted state, y k is the measured voltage, e k The voltage residual is introduced, and a residual-driven double forgetting factor mechanism is introduced to achieve dynamic adaptive adjustment of the covariance between the filtering process and the observation noise, thereby improving the SOC estimation accuracy.

4. The method for estimating the state of charge of a series battery pack according to claim 1, wherein: The bidirectional feedback gated SOC compensator is designed as follows: The battery offline data is sequentially input into the abnormal discriminator and the effective data filter, abnormal samples are eliminated and normal data is extracted, and the voltage difference feature ΔU of each battery cell is integrated. ti and the estimated historical residual series e h As the input variable of the CNN-GRU-SE hybrid neural network, the output is the battery pack state of charge compensation value ΔSOC b The neural network training and optimization process is as follows: ① The collected battery cell voltage difference characteristics ΔU ti and the residual sequence e h ② Normalize the input sample to form a normalized one; ② Construct a deep neural network model composed of a convolutional neural network (CNN), a gated recurrent unit (GRU), and an attention guidance module (SE) to improve the local temporal feature extraction capability and the weighted discrimination capability of key dynamic features; ③ Use the KOA optimization algorithm to globally optimize the hyperparameters, including the number of CNN convolution kernels, the number of GRU layer nodes, and the learning rate, to improve the model training efficiency and prediction accuracy; ④ The normalized input sample X in Input network, complete feature extraction, time series modeling and attention weighting in sequence, and output is the battery state of charge compensation value estimation And the actual compensation value ΔSOC b For the supervision signal, calculate the loss function: ⑤ Based on the principle of error minimization, the back propagation algorithm is used to update the network parameters until the model training converges; ⑥ Repeat steps ④ and ⑤, continuously iteratively train the model until the loss function converges to the preset accuracy threshold, and finally output the trained neural network model and its optimal parameter set for the subsequent real-time prediction of the state of charge compensation value; ⑦ The network output ΔSOC b After being processed by the sliding average filter, the fluctuation noise is suppressed and the estimation stability is improved.

Images