Lithium battery SoH estimation method based on random constant current charging segment

Abstract

The invention relates to the field of SoH state estimation of a lithium ion battery of a battery energy storage system, in particular to a lithium battery SoH estimation method based on a random constant-current charging fragment, which comprises the following steps of: providing SoH characteristic parameters based on equal charging voltage difference time intervals by combining a constant-current charging curve so as to simulate the random constant-current charging fragment; in combination with adaptive transfer learning and a GRU neural network, a GRU network structure based on improved field adaptive transfer learning is proposed, so that negative transfer is avoided and network training load is reduced; the SoH estimation method based on the random constant-current charging segment is realized by combining SoH characteristic parameters and a GRU network structure based on improved domain adaptive transfer learning. The method provided by the invention has the advantages of high precision and fast speed, and is suitable for SoH estimation of the lithium battery of the battery energy storage system.

Description

Technical Field

[0001] The present invention belongs to the field of SoH state estimation of lithium-ion batteries in battery energy storage systems, and relates to a lithium battery SoH estimation method based on random constant current charging segments, which is suitable for SoH estimation of lithium batteries in battery energy storage systems. Background Art

[0002] As a flexible regulation resource for power systems, battery energy storage systems play an important role in new power systems that include a high proportion of renewable energy. Lithium batteries have been widely used in battery energy storage systems due to their high energy density, low self-discharge rate and long service life. SoH is an important indicator for characterizing battery aging. Accurate SoH is of great significance for the optimal performance of lithium batteries and system safety.

[0003] The SoH of lithium batteries is usually defined as the ratio of the current maximum available capacity to the rated capacity. The attenuation of the maximum available capacity is the result of the coupling of multiple electrochemical aging mechanisms and cannot be directly measured. Therefore, a large number of SoH estimation methods have been proposed. At present, the SoH estimation methods of lithium batteries mainly include direct measurement methods, model-based methods and data-driven methods. The direct measurement method obtains the SoH estimation value by offline measuring the aging characteristic indicators such as the capacity or internal resistance of the lithium battery. The model-based method uses prior knowledge to model the battery and can be further divided into degradation model-based methods and state observer methods. The degradation model-based methods include electrochemical models and semi-empirical models. The electrochemical model can provide a certain degree of theoretical explanation for the aging mechanism of the battery, but it has a large amount of calculation and does not directly solve the SoH estimation problem. The relatively simple semi-empirical models do not conduct in-depth research on the electrochemical processes that cause battery degradation, but they reveal the influence of stress factors on the battery SoH. The performance of the semi-empirical model depends largely on the quality and quantity of the test data. The state observer method is usually based on an equivalent circuit model and uses Kalman filtering or particle filtering to convert the SoH estimation problem into a parameter estimation problem. The accuracy of the filtering algorithm is very sensitive to the accuracy of the battery model and is easily affected by the complexity of the system and the uncertainty of the actual application environment.

[0004] Therefore, a lithium battery SoH estimation method based on random constant current charging segments is urgently needed to solve the above technical problems. Summary of the invention

[0005] The purpose of the present invention is to improve the accuracy and operation speed of SoH estimation based on data drive, and discloses a lithium battery SoH estimation method based on random constant current charging segments.

[0006] Data-driven methods can describe complex battery degradation processes without the need for in-depth mechanism research. This type of method first establishes a nonlinear mapping relationship between measurable electrical quantities and battery SoH through a neural network in the offline stage, and then inputs the newly observed electrical quantities into the trained neural network model in the online stage to obtain the SoH estimate. This type of method avoids the influence of model uncertainty and process noise on the estimation accuracy in model-based methods, and has a wide range of applications. The present invention is directed to a data-driven SoH estimation method.

[0007] First, the SoH characteristic parameters based on the equal charging voltage difference time interval are designed. Secondly, an improved domain adaptive transfer learning algorithm is established based on the traditional domain adaptive transfer learning algorithm, and a GRU network based on improved domain adaptive transfer learning is designed to reduce the neural network training load and achieve high-precision and low-computation estimation of SoH. Thirdly, a neural network data set is established based on the constant current charging data under various SoH conditions and different charging rates. The voltage segment and current rate of the battery are used as input, and the relative time corresponding to each voltage value under each SoH condition is used as output. Combined with the dichotomy method, the design of the lithium battery SoH estimation method based on random constant current charging segments is completed. Finally, the SoH estimation model is evaluated using indicators such as mean absolute error, maximum error, root mean square error, and error standard deviation.

[0008] A lithium battery SoH estimation method based on random constant current charging segments comprises the following steps:

[0009] S1. Combined with the constant current charging curve, the SoH characteristic parameters based on the equal charging voltage difference time interval are proposed to simulate the random constant current charging segment;

[0010] S2. Combining adaptive transfer learning with GRU (Gated Recurrent Unit) neural network, a GRU network structure based on Improved Domain Adaptive Transfer Learning (IDATL) is proposed to avoid negative transfer and reduce the network training load.

[0011] S3. Combining SoH feature parameters and the GRU network structure based on improved domain adaptive transfer learning, a SoH estimation method based on random constant current charging segments is implemented.

[0012] Preferably, in step S1, the process of obtaining the SoH characteristic parameters based on the equal charging voltage difference time interval is: by performing a cycle aging experiment on the lithium battery to obtain a complete constant current charging curve of the lithium battery under different SoH conditions, extracting the time interval corresponding to the equal charging voltage difference as the SoH characteristic parameter, so as to map the SoH.

[0013] Preferably, in step S2, an improved domain adaptive transfer learning algorithm is established based on the traditional domain adaptive transfer learning algorithm, and a GRU network based on the improved domain adaptive transfer learning is designed. In order to overcome the negative transfer problem existing in the later stage of traditional domain adaptive transfer learning network training, the present invention proposes an improved domain adaptive transfer learning algorithm, which introduces a hyperbolic tangent function to adaptively reduce the difference between the target domain and the source domain, thereby avoiding negative transfer and reducing the network training load.

[0014] Preferably, the step S3 includes:

[0015] Based on the SoH characteristic parameters obtained in step S1, a data set of corresponding relationships between voltage values ​​and relative time values ​​corresponding to each voltage value under five different SoH conditions and different charging rate conditions is established; a SoH estimation model is established based on a GRU network based on improved domain adaptive transfer learning; the current rate, starting voltage (Started Voltage, SV) and ending voltage (End Voltage, EV) of any constant current charging segment are input into the model, and the time values ​​predicted by the model under different SoH conditions are output; the time differences output by the five calibration models are solved, the five time differences are compared with the measured time differences, and the SoH corresponding to the measured time differences is estimated by dichotomy.

[0016] Preferably, S4, combining the test data of the autonomous experimental platform to evaluate the proposed estimation method.

[0017] Preferably, in step S4, the estimation method of step S3 is verified, and the prediction performance of the SoH estimation model is evaluated using four indicators: mean absolute error, maximum error, root mean square error and error standard deviation.

[0018] Preferably, in step S4, the autonomous experimental platform test data is obtained in the following manner:

[0019] Based on four rates of 0.2C, 0.5C, 0.75C and 1C, the SoH estimation model is trained using data from three of the rates, and the remaining rate is used for testing;

[0020] Firstly, data with charging rates of 0.75C and 0.5C were selected as the model test sets, and 20 aging curves were selected from the charging curves of the two rates for testing. In order to make the verification results more reliable, the SoH corresponding to the curves from each rate approximately covered the healthy cycle of this aging test. Secondly, three voltage segments were selected for each of the three calibrated voltage interval lengths to simulate arbitrary voltage segments, and the proposed SoH estimation method was verified.

[0021] Specifically, a lithium battery SoH (battery state of charge, SoH) estimation method based on random constant current charging segments includes the following steps:

[0022] S1. Combining the study of constant current charging curve, a SoH characteristic parameter based on equal charging voltage difference time interval is proposed;

[0023] The charge and discharge curve of the battery contains a wealth of features related to battery aging. Different features have different degrees of correlation with SoH. The higher the feature correlation, the higher the accuracy of SoH estimation. Since the randomness of the discharge process is significantly stronger than that of the charging process, the parameters are not easy to obtain. The present invention designs SoH feature parameters based on the charging curve.

[0024] In order to obtain the constant current charging curve under different health conditions, it is necessary to conduct a cycle aging experiment on the lithium battery. Set four constant current charging rates of 0.2C, 0.5C, 0.75C and 1C, and conduct battery aging charge and discharge experiments at 25°C. The entire experimental process is divided into two parts: constant current charging data acquisition and accelerated battery aging. After the battery undergoes a complete charge and discharge cycle at each of the four constant current charging rates, four aging cycle experiments are conducted. The two parts of the experiment are carried out alternately. The experimental process is described as follows.

[0025] (1) Constant current charging data acquisition

[0026] ⅰ. The battery was charged at a constant current (3.2A, i.e. 1C) and a constant voltage (4.2V) and discharged at a constant current (3.2A, i.e. 1C) and a constant voltage (2.5V), and the charge and discharge cut-off current was set to 0.032A (0.01C); during the charge and discharge conversion process, the battery was left to stand for 1 hour to eliminate the polarization effect.

[0027] ⅱ. Change the charging rate to 0.75C, 0.5C and 0.2C in turn to perform constant current and constant voltage charging, and repeat step ⅰ.

[0028] (2) Accelerated aging

[0029] ⅰ. The battery was charged at 5A with constant current and constant voltage (4.2V) and discharged at 2.5V with constant current and constant voltage. The charge and discharge cut-off current was set to 0.032A (0.01C). During the charge and discharge conversion process, the battery was left to stand for 1 hour to eliminate the polarization effect.

[0030] ⅱ. Repeat step ⅰ three times.

[0031] The experiment lasted for 3 months, and the maximum available capacity of the battery gradually decreased with the increase in the number of cycles. Since the mapping relationship between SoH and SoH characteristic parameters is clearer when SoH is higher or lower, in order to demonstrate the advancement of this method, the aging data of the battery SoH degradation process from 0.96 to 0.88 was studied, including parameters such as current, voltage, capacity and temperature. Theoretically, if the new estimation method has ideal estimation accuracy in the range of SoH [0.88, 0.96], the new estimation method has higher accuracy in the entire battery aging process.

[0032] The present invention defines SoH as the ratio of the average value of the charging capacity and the discharging capacity in one charge and discharge cycle to the maximum available capacity of the battery when it leaves the factory. The maximum available capacity of the battery gradually decreases with the increase of the number of cycles.

[0033] The constant current charging data when SoH is 0.96, 0.94, 0.92, 0.90 and 0.88 are taken for model calibration. The constant current charging data includes the relative charging time corresponding to each voltage value under each constant current charging rate condition. The SoH characteristic parameters are established with the constant current charging data of different rates under each SoH condition, and the moment when the terminal voltage of the battery charging process is 2.9V is defined as the zero time. The design idea is: divide the lithium battery voltage from 2.9V to 4.2V in units of 10mV, with a total of 131 voltage labels. Assuming that the battery voltage is 2.9V, the corresponding moment is the initial moment (i.e., zero moment). When the lithium battery is charged to 2.91V, the corresponding moments under different rates are t1_1, t2_1... respectively. Finally, a corresponding relationship data set between the voltage values ​​and the relative time values ​​corresponding to each voltage value under five different SoH conditions and different charging rate conditions is established. SoH is mapped using the equal charging voltage difference time interval as the SoH characteristic parameter.

[0034] S2. Combining adaptive transfer learning with GRU network characteristics, a GRU network structure based on improved domain adaptive transfer learning is proposed;

[0035] Feature parameters are the basis of neural network state estimation, and a stable and efficient neural network structure is the key to data-driven SoH estimation. In the process of transfer learning, the maximum mean discrepancy (MMD) is introduced to address the migration obstacles caused by the feature distribution of the source domain and the target domain. This non-parametric distance metric can measure the data distribution difference between the source domain and the target domain. By adjusting the MMD, the target domain can have a feature distribution similar to that of the source domain.

[0036] Assume that there is a source domain D S and the target domain D T , as shown in formula (1).

[0037]

[0038] If the source domain input is X S and the target domain input X t If it obeys different distributions, the formula definition of MMD is as shown in formula (2) and (3).

[0039]

[0040] In the above formula, m and n are the sequence lengths of the source domain and the target domain respectively.

[0041] represents the bi-norm operation in the Reproducing Kernel Hilbert Space (RKHS), is a mapping function in RKHS. The smaller the MMD, the stronger the feature correlation between the two data distributions in the transfer learning network. The loss function formula of the transfer learning network is:

[0042] Loss = MSE(h t )+λ·MMD(h s ,h t ) (4)

[0043] In formula (4), Loss and MSE (h t ) represent the total loss of the transfer learning network and the mean square error loss of the target model respectively; h s and h t are the feature information of the source domain and the target domain respectively; λ·MMD(h s ,h t ) represents the adaptive loss of the transfer learning network; λ is a non-negative adaptive regularization parameter that weighs the MMD and adaptive loss. When the number of iterations of the transfer learning network increases, the difference between the source domain and the target domain will gradually decrease. At this time, a fixed λ will keep the model at a high transfer depth, but it will also increase the possibility of negative transfer.

[0044] This invention proposes an improved domain adaptive transfer learning algorithm, which can overcome the negative transfer problem existing in the later stage of traditional transfer learning network training, and improve the MMD distance between the source domain and the target domain, so that the transfer depth can be gradually reduced as the MMD decreases. The improved MMD and its corresponding model loss function are shown in equations (5) and (6):

[0045]

[0046] Loss total =MSE(h t )+λ·IMMD(h s ,h t) (6)

[0047] Based on the maximum mean difference, a hyperbolic tangent term that also contains the MMD term is designed as the denominator. In formula (5), the denominator is kept in the interval [0,1] by the hyperbolic tangent function tanh. The denominator term gradually increases as the MMD decreases, so that the migration depth gradually decreases with the increase in the number of iterations. m is a non-negative threshold parameter not less than MMD, which controls the speed of migration depth reduction. In formula (6), Loss total is the model loss; λ and the denominator tanh[m-MMD(h s ,h t )] controls the migration depth at each iteration. The more knowledge is migrated from the source domain to the target domain, the greater the migration depth.

[0048] As a variant of recurrent neural network, GRU can effectively overcome the gradient problem in back propagation by continuously updating the information stored internally. Based on the GRU network, a neural network based on improved domain adaptive transfer learning is constructed. On the one hand, the migration idea based on fine-tuning strategy is adopted in the GRU layer to migrate the trained parameters of the source domain model to the target domain model to be trained, so that the GRU layer of the target domain model can continue to train parameters to quickly reach convergence; on the other hand, improved domain adaptive transfer learning is applied in the fully connected layer. Through the parameter sharing of the GRU layer and the domain adaptation of the fully connected layer, the features of the source domain and the target domain are moved from the low-dimensional space to the high-dimensional space, thereby minimizing the difference in the distribution of the two domains.

[0049] S3. Combining SoH feature parameters and the GRU network structure based on improved domain adaptive transfer learning, a SoH estimation method based on random constant current charging segments is designed;

[0050] An estimation model is established based on the GRU network based on improved domain adaptive transfer learning. The model input is the voltage segment and current rate of the battery, and the model output is the relative time corresponding to each voltage value under each SoH condition. Input the current rate, starting voltage (Started Voltage, SV) and ending voltage (End Voltage, EV) of any constant current charging segment into the model, and output the time value predicted by the model under different SoH conditions. Solve the time difference output by the five calibration models, compare the five time differences with the measured time difference, and estimate the SoH corresponding to the measured time difference by dichotomy. The specific design process is described as follows.

[0051] 1) Determine the time difference among the five time differences, the time difference that is the least greater than the actual time difference is recorded as T R , the time difference that is less than the actual time difference is recorded as T F , the measured time difference is T A ; T R , TF The SoH corresponding to the two time differences is calculated, and the measured time difference T A and T F Do poorly;

[0052] 2) Calculate the SoH corresponding to the measured time difference by the dichotomy method.

[0053]

[0054] In summary, a lithium battery SoH estimation method based on random constant current charging segments was designed using a GRU network based on improved domain adaptive transfer learning.

[0055] S4. Evaluate the proposed estimation method based on the test data of the independent experimental platform.

[0056] The prediction performance of the SoH estimation model was evaluated using four evaluation indicators: Mean Absolute Error (MAE), Maximum Error (MAXE), Root Mean Square Error (RMSE) and Standard Deviation of Error (SDE). In order to explore the impact of voltage segment length on the prediction performance of the model, three interval lengths of 0.2V (short charging time), 0.5V (medium charging time) and 0.8V (long charging time) were selected to test the model. The three voltage intervals were 3.50V-3.70V, 3.40V-3.90V and 3.30V-4.10V respectively.

[0057] In order to verify the estimation effect of the estimation method of the present invention within the SoH range of the entire data set, based on four rates of 0.2C, 0.5C, 0.75C and 1C, three of the rate data are used to train the model, and the remaining rate is used for testing. First, data with charging rates of 0.75C and 0.5C are selected as the model test set, and 20 aging curves are selected from the charging curves from the two rates for testing. In order to make the verification result more reliable, the SoH corresponding to the charging curves from each rate approximately covers the healthy cycle of this aging test; secondly, three voltage segments are selected for each of the three calibrated voltage interval lengths to simulate any voltage segment, and the proposed SoH estimation method is verified.

[0058] Beneficial effects:

[0059] Based on the current status of lithium battery SoH estimation research, the present invention proposes a lithium battery SoH estimation method based on random constant current charging segments, which has the advantages of high accuracy and high speed and is suitable for lithium battery SoH estimation in battery energy storage systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the principles of the present invention and the technical solutions in its implementation, the technical solutions involved in the present invention will be further introduced using the accompanying drawings. The following drawings are only some implementation examples of the present invention. For those skilled in the art, other technical solutions can be obtained based on the following drawings without paying any creative work.

[0061] Figure 1 It is a design flow chart of the SoH estimation method based on random constant current charging segments of the present invention;

[0062] Figure 2 The constant current charging curves at 1C rate and 15 different SoH conditions;

[0063] Figure 3 This is a diagram showing the change process of the maximum average difference item when training different models in Example 1 of the present invention;

[0064] Figure 4 The GRU network structure based on improved domain adaptive transfer learning in embodiment 1 of the present invention;

[0065] Figure 5 It is a technical roadmap of the SoH estimation method based on random constant current charging segments in Example 1 of the present invention;

[0066] Figure 6 It is the estimated result of Example 1 of the present invention under the condition of 0.75C rate;

[0067] Figure 7 This is the estimated result of Example 1 of the present invention under 0.5C rate condition. DETAILED DESCRIPTION

[0068] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0069] Example 1

[0070] This example uses the 18650 ternary lithium battery produced by Panasonic Corporation of Japan as an example. The lithium-ion battery experimental test platform consists of a host computer with control software installed, a middle computer, a Neware programmable electronic load, a battery, and a high and low temperature test chamber. The battery test equipment model is CT-4008T-5V6A-S1, the test accuracy is less than 0.05%, and the maximum current rise time is 1ms. The programmable high and low temperature test chamber model is BPH-060A, and the temperature fluctuation range is ±0.5℃, which is used to provide a stable operating temperature for the battery. The lithium battery parameters are shown in Table 1.

[0071] Table 1 Lithium battery parameters

[0072]

[0073] The charge and discharge curve of the battery contains many features related to battery aging. Different features have different correlations with SoH. The higher the feature correlation, the higher the accuracy of SoH estimation. Since the randomness of the discharge process is significantly stronger than that of the charging process, the parameters are not easy to obtain. In this embodiment, the SoH feature parameters are designed based on the charging curve.

[0074] In order to obtain constant current charging curves under different health conditions, this embodiment conducts a cycle aging experiment on the lithium battery based on the lithium-ion battery test platform. Four constant current charging rates of 0.2C, 0.5C, 0.75C and 1C are set, and the battery aging charge and discharge experiment is carried out at 25°C. The whole experimental process is divided into two parts: constant current charging data acquisition and accelerated battery aging. After the battery has undergone a complete charge and discharge cycle at four constant current charging rates, four aging cycle experiments are carried out. The two parts of the experiment are carried out alternately. The experimental process is described as follows.

[0075] (1) Constant current charging data acquisition

[0076] ⅰ. The battery was charged at a constant current (3.2A, i.e. 1C) and a constant voltage (4.2V) and discharged at a constant current (3.2A, i.e. 1C) and a constant voltage (2.5V), and the charge and discharge cut-off current was set to 0.032A (0.01C); during the charge and discharge conversion process, the battery was left to stand for 1 hour to eliminate the polarization effect.

[0077] ⅱ. Change the charging rate to 0.75C, 0.5C and 0.2C in turn to perform constant current and constant voltage charging, and repeat step ⅰ.

[0078] (2) Accelerated aging

[0079] ⅰ. The battery was charged at 5A with constant current and constant voltage (4.2V) and discharged at 2.5V with constant current and constant voltage. The charge and discharge cut-off current was set to 0.032A (0.01C). During the charge and discharge conversion process, the battery was left to stand for 1 hour to eliminate the polarization effect.

[0080] ⅱ. Repeat step ⅰ three times.

[0081] The experiment lasted for 3 months, and the maximum available capacity of the battery gradually decreased with the increase in the number of cycles. Since the mapping relationship between SoH and SoH characteristic parameters is clearer when SoH is higher or lower, in order to demonstrate the advancement of this method, the aging data of the battery SoH degrading from 0.96 to 0.88 was studied, including parameters such as current, voltage, capacity and temperature. Theoretically, if the new estimation method has ideal estimation accuracy in the range of SoH [0.88, 0.96], the new estimation method has higher accuracy for the entire battery aging process. In the range of SoH [0.88, 0.96], with a basic SoH difference of 0.01 as the interval, 15 constant current charging curves were selected to obtain the SoH characteristic parameters under different SoH conditions, such as Figure 2 As shown, the constant current charging curves under 15 different SoH conditions at 1C rate.

[0082] In this embodiment, SoH is defined as the ratio of the average value of the charging capacity and the discharging capacity in one charge and discharge cycle to the maximum available capacity of the battery when it leaves the factory. The calculation formula is:

[0083]

[0084] Among them, C charge and C discharge are the charge capacity and discharge capacity, C nominal This is the maximum available capacity of the battery when it leaves the factory.

[0085] The constant current charging data when SoH is 0.96, 0.94, 0.92, 0.90 and 0.88 are taken for model calibration. The constant current charging data includes the relative charging time corresponding to each voltage value under each rate constant current charging condition. The SoH characteristic parameters are established based on the constant current charging data of different rates under each SoH condition. The SoH characteristic parameter data set design corresponding to each SoH is shown in Table 2 below. The moment when the terminal voltage of the battery charging process is 2.9V is defined as zero time. The design idea is: divide the lithium battery voltage from 2.9V to 4.2V in units of 10mV, with a total of 131 voltage labels. Assuming that the battery voltage is 2.9V, the corresponding moment is the initial moment (ie, zero moment). When the lithium battery is charged to 2.91V, the corresponding moments under different rate conditions are t1_1, t2_1...

[0086] Table 2 Dataset design under different SoH conditions

[0087]

[0088] Finally, a data set of corresponding relationships between voltage values ​​and relative time values ​​corresponding to each voltage value under five different SoH conditions and different charging rate conditions was established.

[0089] In order to explore the correlation between the new aging characteristics and the currently common aging characteristics and SoH, a complete charging curve is obtained based on 0.2C and 0.5C rates. The reason for choosing the smaller 0.2C and 0.5C rates is that the smaller the current, the smaller the difference in the charging curves corresponding to different SoHs. If the SoH characteristic parameters have an ideal correlation with SoH at 0.2C or 0.5C rates, the charging curves will be more different at a larger rate. In theory, the correlation between SoH characteristic parameters and SoH is higher. Five common features, including constant current charging time (CCCT), constant current discharge time (CCDT), average voltage drop for unit discharge time (AVD-UDT), internal resistance (IR) and time interval for equal charging voltage difference (TI-ECVD), are extracted to study their correlation with SoH.

[0090] In order to determine the correlation between each selected SoH characteristic parameter and SoH, the Pearson coefficient is used as the characteristic correlation evaluation standard to extract SoH characteristic parameters from 30 aging curves of 0.2C and 0.5C constant current charging rates. The SoH of the selected aging curve covers the healthy cycle of this aging test. Among them, AVD-UDT is the ratio of the total voltage drop to the total time during discharge; TI-ECVD takes the time interval corresponding to the charging voltage in the range of 3.50V-3.70V as an example. The reason for selecting the 3.50V-3.70V voltage segment is that 3.50V-3.70V is the battery charging and discharging platform period, and the voltage changes relatively slowly within this range. If this interval can better characterize SoH, other voltage intervals can theoretically better characterize SoH; 3.50V-3.70V is a relatively small voltage interval. If this interval can better characterize SoH, other larger voltage intervals contain more curve features and better characterize SoH. For each characteristic parameter X, given the characteristic sequence X={X 1 ,X 2 ,X 3 ,…,X n} and health state sequence SoH = {SoH 1 ,SoH 2 ,SoH 3 ,…,SoH n}, the correlation calculation formula is:

[0091]

[0092] In formula (2), X represents the characteristic value of the battery. n is the length of the characteristic sequence and health status sequence; and are the average values ​​of the feature sequence and health state sequence respectively; cov(X,SoH) represents the covariance between the feature and the health state, σ X and σ Y They represent the standard deviation of the feature sequence and health status sequence respectively. The Pearson coefficient of each feature and SoH is calculated as shown in Table 3. The closer the Pearson coefficient is to 1, the higher the correlation between the feature and SoH.

[0093] Table 3 Pearson correlation coefficient between different SoH characteristic parameters and SoH

[0094]

[0095] As can be seen from Table 3, TI-ECVD, CCCT and IR have the highest correlation with SoH. However, in the actual charging process of the battery energy storage system, since there is residual power inside the battery at the beginning of charging, and the battery may end charging when the voltage has not reached the charging cut-off voltage, it is difficult to obtain accurate constant current charging time under most working conditions. In addition, the internal resistance has high requirements on the accuracy of the battery measurement equipment and is easily affected by temperature, so it is difficult to obtain accurate internal resistance. The equal charging voltage difference time interval can be flexibly obtained through different local voltage intervals, which is suitable for complex and changeable environments. The calculation results in Table 3 show that the new equal charging voltage difference time interval proposed in this embodiment has a higher correlation with SoH than other characteristic parameters.

[0096] Feature parameters are the basis of neural network state estimation, and a stable and efficient neural network structure is the key to data-driven SoH estimation. In the process of transfer learning, the maximum mean discrepancy (MMD) is introduced to address the migration obstacles caused by the feature distribution of the source domain and the target domain. This non-parametric distance metric can measure the data distribution difference between the source domain and the target domain. By adjusting the MMD, the target domain can have a feature distribution similar to that of the source domain.

[0097] Assume that there is a source domain D S and the target domain D T , as shown in formula (3).

[0098]

[0099] If the source domain input is X S and the target domain input X tIf it obeys different distributions, the formula definition of MMD is as shown in the following equations (4) and (5).

[0100]

[0101] In the above formula, m and n are the sequence lengths of the source domain and the target domain respectively. represents the bi-norm operation in the Reproducing Kernel Hilbert Space (RKHS), is a mapping function in RKHS. The smaller the MMD, the stronger the feature correlation between the two data distributions in the transfer learning network. The loss function formula of the transfer learning network is:

[0102] Loss = MSE(h t )+λ·MMD(h s ,h t ) (6)

[0103] In formula (6), Loss and MSE (h t ) represent the total loss of the transfer learning network and the mean square error loss of the target model respectively; h s and h t are the feature information of the source domain and the target domain respectively; λ·MMD(h s ,h t ) represents the adaptive loss of the transfer learning network; λ is a non-negative adaptive regularization parameter that weighs the MMD and adaptive loss. When the number of iterations of the transfer learning network increases, the difference between the source domain and the target domain will gradually decrease. At this time, a fixed λ will keep the model at a high transfer depth, but it will also increase the possibility of negative transfer.

[0104] This embodiment proposes an improved domain adaptive transfer learning algorithm, which can overcome the negative transfer problem existing in the later stage of traditional transfer learning network training, and improve the MMD distance between the source domain and the target domain, so that the transfer depth can be gradually reduced as the MMD decreases. The improved MMD and its corresponding model loss function are shown in equations (7) and (8):

[0105]

[0106] Loss total =MSE(h t )+λ·IMMD(h s ,h t ) (8)

[0107] Based on the maximum mean difference, a hyperbolic tangent term that also contains the MMD term is designed as the denominator. In formula (7), the denominator is kept in the interval [0,1] by the hyperbolic tangent function tanh. The denominator term gradually increases as the MMD decreases, so that the migration depth gradually decreases as the number of iterations increases. m is a non-negative threshold parameter not less than MMD, which controls the speed of migration depth reduction. In formula (8), Loss total is the model loss; λ and the denominator tanh[m-MMD(h s ,h t )] controls the migration depth at each iteration. The more knowledge is migrated from the source domain to the target domain, the greater the migration depth.

[0108] like Figure 3 As shown in FIG. 1 , it is a diagram showing the change process of the maximum average difference term when training different models in Example 1 of the present invention. This diagram is a comparison of the adaptive transfer learning effects based on traditional MMD and improved MMD. Figure 3 It can be seen that the improved adaptive transfer learning can reduce the difference between the source domain and the target domain to a greater extent within the same iteration cycle compared to the traditional transfer learning. This is because the improved domain adaptation method proposed in the present invention can spontaneously reduce the transfer depth according to the model training state, thereby reducing the number of negative transfers and shortening the convergence time of the overall transfer learning.

[0109] As a variant of recurrent neural network, GRU can effectively overcome the gradient problem in back propagation by continuously updating the information stored internally. Based on the GRU network, a neural network based on improved domain adaptive transfer learning is constructed. On the one hand, the migration idea based on fine-tuning strategy is adopted in the GRU layer to migrate the trained parameters of the source domain model to the target domain model to be trained, so that the GRU layer of the target domain model can continue to train parameters to quickly reach convergence; on the other hand, improved domain adaptive transfer learning is applied in the fully connected layer. Figure 4 As shown, this is the GRU network structure based on improved domain adaptive transfer learning of Example 1 of the present invention. Through parameter sharing of the GRU layer and domain adaptation of the fully connected layer, the features of the source domain and the target domain are moved from the low-dimensional space to the high-dimensional space, thereby minimizing the difference in distribution of the two domains.

[0110] An estimation model is established based on the GRU network based on improved domain adaptive transfer learning. The model input is the voltage segment and current rate of the battery, and the model output is the relative time corresponding to each voltage value under each SoH condition. The current rate, starting voltage and ending voltage of any constant current charging segment are input into the model, and the time value predicted by the model under different SoH conditions is output. The time difference output by the five calibration models is solved, and the five time differences are compared with the measured time difference. The SoH corresponding to the measured time difference is estimated by dichotomy. The specific design process is described as follows.

[0111] 1) Determine the time difference among the five time differences, the time difference that is the least greater than the actual time difference is recorded as T R , the time difference that is less than the actual time difference is recorded as T F , the measured time difference is T A ; T R , T F The SoH corresponding to the two time differences is calculated, and the measured time difference T A and T F Do poorly;

[0112] 2) Calculate the SoH corresponding to the measured time difference by the dichotomy method.

[0113]

[0114] Using the GRU neural network based on improved domain adaptive transfer learning, a lithium battery SoH estimation method based on random constant current charging segments is designed, such as Figure 5 As shown, it is a technical roadmap of the SoH estimation method based on random constant current charging segments according to Example 1 of the present invention.

[0115] In order to explore the influence of voltage segment length on the prediction performance of the model, three interval lengths of 0.2V (short charging time), 0.5V (medium charging time) and 0.8V (long charging time) were used to test the model. Based on four rates of 0.2C, 0.5C, 0.75C and 1C, three of the rate data were used to train the model, and the remaining rate was used for testing. Data with charging rates of 0.75C and 0.5C were selected as the model test set, and 20 curves were selected from the charging curves at these two charging rates for testing. In order to make the verification results more reliable, the SoH corresponding to the curves from each rate approximately covered the healthy cycle of this aging experiment; then three voltage segments were selected for each of the three calibrated voltage interval lengths to simulate arbitrary voltage segments to verify the SoH estimation method. The selected voltage segments and verification results are shown in Table 4, Figure 6 and Figure 7 As shown. Figure 6 As shown, it is the estimated result of Example 1 of the present invention under the condition of 0.75C rate; Figure 7 As shown, it is the estimation result of Example 1 of the present invention under the condition of 0.5C rate. Figure 6 and Figure 7 Each predicted value is the average of the predicted values ​​corresponding to the nine voltage intervals, and the estimated error is shown in Table 5.

[0116] Table 4 Voltage segments with different interval lengths

[0117]

[0118] Table 5 Estimation error of different magnifications

[0119]

[0120] Depend on Figure 6 , Figure 7 As shown in Table 5, the SoH estimation method has relatively ideal estimation accuracy in voltage segments of different lengths, indicating that the lithium battery SoH estimation method based on random constant current charging segments proposed in the present invention can achieve good performance in both model versatility and estimation accuracy.

[0121] To further verify the superiority of the estimation method of this embodiment in terms of computing load, this embodiment uses three computers with different configurations to simulate the above model. The core configurations of the three computers are:

[0122] C1: Memory 16G, CPU: AMD-Ryzen7-5800, GPU: 3050Ti;

[0123] C2: Memory 32G, CPU: Intel7-12700, no GPU;

[0124] C3: Memory 32G, CPU: Intel7-12700, GPU: 3070Ti.

[0125] The model running time is shown in Table 6. The simulation time in the table is the average simulation time of three computers under 0.75C and 0.5C constant current charging data.

[0126] Table 6 Model simulation time

[0127]

[0128] It can be seen from Table 6 that the average simulation time of the GRU algorithm model based on improved domain adaptive transfer learning is 120.74 seconds, indicating that the SoH estimation method proposed in the present invention has a good computing speed.

[0129] The technical solution of the present invention is not limited to the above-mentioned specific embodiments, and all technical variations made according to the technical solution of the present invention fall within the protection scope of the present invention.

Claims

1. A lithium battery SoH estimation method based on random constant current charging segments, characterized in that: The following steps are involved: S1. Combined with the constant current charging curve, the SoH characteristic parameters based on the equal charging voltage difference time interval are proposed to simulate the random constant current charging segment; S2. Combining adaptive transfer learning with GRU neural network, a GRU network structure based on improved domain adaptive transfer learning is proposed to avoid negative transfer and reduce network training load; S3. Combining SoH feature parameters and the GRU network structure based on improved domain adaptive transfer learning, a SoH estimation method based on random constant current charging segments is implemented.

2. The method for estimating SoH of a lithium battery based on random constant current charging segments according to claim 1, characterized in that: In step S1, the process of obtaining the SoH characteristic parameters based on the equal charging voltage difference time interval is as follows: a complete constant current charging curve of the lithium battery under different SoH conditions is obtained by performing a cycle aging experiment on the lithium battery, and the time interval corresponding to the equal charging voltage difference is extracted as the SoH characteristic parameter, so as to map the SoH.

3. The method for estimating SoH of a lithium battery based on random constant current charging segments according to claim 1, characterized in that: In step S2, an improved domain adaptive transfer learning algorithm is established based on the traditional domain adaptive transfer learning algorithm, and a GRU network based on improved domain adaptive transfer learning is designed. By introducing a hyperbolic tangent function, the difference between the target domain and the source domain is adaptively reduced, thereby avoiding negative transfer and reducing the network training load.

4. The method for estimating SoH of a lithium battery based on random constant current charging segments according to claim 1, characterized in that: The step S3 comprises: Based on the SoH characteristic parameters obtained in step S1, a corresponding relationship data set between the voltage values ​​under five different SoH conditions and different charging rate conditions and the relative time values ​​corresponding to each voltage value is established; a SoH estimation model is established based on the GRU network based on improved domain adaptive transfer learning; the current rate, starting voltage and ending voltage of any constant current charging segment are input into the model, and the time values ​​predicted by the model under different SoH conditions are output; the time differences output by the five calibration models are solved, the five time differences are compared with the measured time differences, and the SoH corresponding to the measured time differences is estimated by dichotomy.

5. The method for estimating SoH of a lithium battery based on random constant current charging segments according to claim 4, characterized in that: The following steps are also included: S4. Evaluate the proposed estimation method based on the test data of the independent experimental platform.

6. The method for estimating SoH of a lithium battery based on random constant current charging segments according to claim 5, characterized in that: In step S4, the estimation method of step S3 is verified, and the prediction performance of the SoH estimation model is evaluated using four indicators: mean absolute error, maximum error, root mean square error and error standard deviation.

7. The method for estimating SoH of a lithium battery based on random constant current charging segments according to claim 4, characterized in that: In step S4, the autonomous experimental platform test data is obtained in the following manner: Based on four rates of 0.2C, 0.5C, 0.75C and 1C, the SoH estimation model is trained using data from three of the rates, and the remaining rate is used for testing; Firstly, data with charging rates of 0.75C and 0.5C were selected as the model test sets, and 20 aging curves were selected from the charging curves of the two rates for testing. In order to make the verification results more reliable, the SoH corresponding to the curves from each rate approximately covered the healthy cycle of this aging test. Secondly, three voltage segments were selected for each of the three calibrated voltage interval lengths to simulate arbitrary voltage segments, and the proposed SoH estimation method was verified.

Images