Battery health estimation model construction method based on LSSVR and partial charge / discharge data and application
Through a battery health estimation model based on LSSVR and partial charging/discharge data, combined with the cuckoo search algorithm for global optimization, the problem of difficulty in estimating the battery health status in the existing technology under characteristic engineering is solved, and efficient SOH estimation is achieved.
Patent Information
- Application Number
- CN202411850634.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-16
- Publication Date
- 2025-03-28
AI Technical Summary
The prior art is difficult to perform battery health status (SOH) estimation without feature engineering, especially when facing partial charge and discharge data, predictive performance is limited.
Using a battery health estimation model based on LSSVR and partial charging/discharge data, the LSSVR is searched globally through the cuckoo search algorithm, a CS-LSSVM model is constructed, and the SOH estimation is directly used using partial charging/discharge data.
It realizes accurate estimation of the battery health status without relying on complex inputs and feature extraction, and improves the predictive performance under partial charge and discharge data conditions.
Smart Images
Figure CN119716548A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of health detection of new energy storage batteries, and specifically to a method and application of building a battery health estimation model based on LSSVR (least square support vector regression) and partial charging / discharging data. Background Art
[0002] The transformation of the global energy structure and the rapid development of energy storage technology have led to sodium-ion batteries (SIBs) becoming an increasingly important alternative to lithium-ion batteries. This is due to the advantages they offer in terms of resource abundance, low cost, and environmental friendliness. However, the performance of sodium-ion batteries will inevitably decline over time. Therefore, accurately and effectively estimating the state of health (SOH) of the battery has become a key issue in the design of battery management systems (BMS). As a performance indicator that reflects the current battery capacity compared to the initial capacity, SOH is essential for extending battery life, ensuring system safety, and improving energy efficiency.
[0003] Existing SOH estimation methods usually rely on the complete charge / discharge curve of the battery or complex electrochemical models. Xiang et al. [1] constructed three SOH estimation methods for SIBs, namely, multivariate linear regression, ridge regression, and support vector regression, using SOH-related features extracted from a third-order equivalent circuit model constructed from electrochemical impedance spectroscopy. Wei et al. [2] implemented a Gaussian process regression method for SOH evaluation of sodium-ion batteries by extracting six features that are highly correlated with SOH from incremental capacity analysis.
[0004] From the above literature research, although effective battery capacity degradation prediction has been achieved through these different methods, there are still limitations, mainly manifested in: In practical applications, due to the limitation of charging time or equipment operating conditions, only partial charge and discharge data can usually be obtained, which brings challenges to traditional SOH estimation methods. In addition, the accuracy and generalization ability of these methods depend to a certain extent on the optimization of model parameters, and their prediction performance is often limited when facing partial charge and discharge data. Therefore, a reliable method for estimating the health state of sodium-ion batteries that does not rely on complex input and feature extraction is imminent. Summary of the invention
[0005] The technical problem to be solved by the present invention is how to estimate the battery SOH without any feature engineering on the charge and discharge data.
[0006] The present invention solves the above technical problems through the following technical means:
[0007] The method for constructing a battery health estimation model based on LSSVR and partial charge / discharge data includes the following steps:
[0008] The complete charge and discharge data of the battery under different health conditions are collected, and the charge and discharge data under each health state are arbitrarily divided into multiple segments. Each segment of data is labeled with the same label to construct a training set; the labeled data is sent to the CS-LSSVM model for training, and the SOH is output to finally obtain the battery health estimation model.
[0009] Furthermore, the charge and discharge data is voltage or current data.
[0010] Furthermore, the construction method of the CS-LSSVM model is to use a cuckoo search algorithm to perform global optimization on the LSSVR.
[0011] Furthermore, the LSSVR is defined as follows:
[0012]
[0013] Where x is the input vector of the model and y is the corresponding output. It is a nonlinear mapping function used to map the input vector to a higher dimensional space. w is the weight vector of the hyperplane and b is the intercept. For a given data set D = {x i ,y i} n (x i ∈R d×1 ,y i ∈R,i=1,2,K,n), the optimization problem of LSSVR is:
[0014]
[0015] where e i is the error variable and C is the regularization parameter, which is used to balance model complexity and error. The objective function and constraints are combined through the Lagrangian function:
[0016]
[0017] where α i is the Lagrange multiplier x corresponding to i Deriving the Lagrangian function and eliminating variables, we obtain the following linear equations:
[0018]
[0019] Where K(x i ,x j ) is the kernel function. The above equation can be abbreviated to matrix form:
[0020]
[0021] 1 of them n =[1,1,K,1] n×1 α=[α1,α2,K,α n ], I n , is the n×n unit matrix. After inputting the training data, the intercept b and the Lagrange multiplier α are obtained by solving this linear equation system. The model training is completed. For the new sample x', the regression prediction value is:
[0022]
[0023] Kernel function is an important term in SVR, and it has different forms, such as polynomial, linear, and Gaussian kernel functions. In this patent, Gaussian kernel is selected as the kernel function, and its formula is as follows:
[0024]
[0025] Where σ is a parameter of the kernel function, called the bandwidth parameter (scale parameter), which controls the action range and distribution shape of the kernel function.
[0026] The present invention also provides a system for constructing a battery health estimation model based on LSSVR and partial charge / discharge data, comprising:
[0027] Training set construction module: collect complete charge and discharge data of the battery under different health conditions, divide the charge and discharge data under each health state into multiple segments at will, and label each segment with the same label;
[0028] Training module: The labeled data is sent to the CS-LSSVM model for training, SOH is output, and finally the target model is obtained.
[0029] Furthermore, the charge and discharge data is voltage or current data.
[0030] Furthermore, the construction method of the CS-LSSVM model is to use a cuckoo search algorithm to perform global optimization on the LSSVR.
[0031] Furthermore, the LSSVR is defined as follows:
[0032]
[0033] Where x is the input vector of the model and y is the corresponding output. It is a nonlinear mapping function used to map the input vector to a higher dimensional space. w is the weight vector of the hyperplane and b is the intercept. For a given data set D = {x i ,y i} n (x i ∈R d×1 ,y i ∈R,i=1,2,K,n), the optimization problem of LSSVR is:
[0034]
[0035] where e i is the error variable and C is the regularization parameter, which is used to balance model complexity and error. The objective function and constraints are combined through the Lagrangian function:
[0036]
[0037] where α i is the Lagrange multiplier x corresponding to i Deriving the Lagrangian function and eliminating variables, we obtain the following linear equations:
[0038]
[0039] Where K(x i ,x j ) is the kernel function. The above equation can be abbreviated to matrix form:
[0040]
[0041] 1 of them n =[1,1,K,1] n×1 α=[α1,α2,K,α n ], I n , is the n×n unit matrix. After inputting the training data, the intercept b and the Lagrange multiplier α are obtained by solving this linear equation system. The model training is completed. For the new sample x', the regression prediction value is:
[0042]
[0043] Kernel function is an important term in SVR, and it has different forms, such as polynomial, linear, and Gaussian kernel functions. In this patent, Gaussian kernel is selected as the kernel function, and its formula is as follows:
[0044]
[0045] Where σ is a parameter of the kernel function, called the bandwidth parameter (scale parameter), which controls the action range and distribution shape of the kernel function.
[0046] The present invention also provides a method for estimating the SOH of new energy batteries using the battery health estimation model based on LSSVR and partial charge / discharge data, which is characterized by obtaining battery charge and discharge data, directly inputting it into the health estimation model, and outputting the SOH result.
[0047] The present invention also provides a new energy battery SOH estimation system, including the above-mentioned battery health estimation model based on LSSVR and partial charge / discharge data.
[0048] The advantages of the present invention are:
[0049] 1. The partial charge and discharge data used as input features are easy to obtain and do not require feature engineering
[0050] The input of the method framework proposed in this patent can be any continuous partial data obtained during the charging and discharging process, such as voltage, etc. These data are easy to obtain, do not require feature engineering, and do not rely on data integrity. Only partial continuous data is required.
[0051] 2. The SOH assessment does not depend on the type of data condition
[0052] The method proposed in this patent does not depend on the type of battery data operating conditions. It can be a battery training data set formed by multiple operating conditions. It only needs to input the partial charging and discharging conditions of the battery under these different operating conditions to complete the SOH evaluation task. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 A schematic diagram of a flow chart of a model building method / evaluation method in an embodiment of the present invention;
[0054] Figure 2 The capacity aging path diagram of all SIBs given for a specific case in an embodiment of the present invention;
[0055] Figure 3 The charging / discharging data of different cycles given for a specific case in an embodiment of the present invention;
[0056] Figure 4 The SOH estimation result using LOOCV is given for a specific case in an embodiment of the present invention;
[0057] Figure 5 Box plots of SOH prediction errors for different input lengths for a specific case in an embodiment of the present invention. DETAILED DESCRIPTION
[0058] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0059] This embodiment proposes a method for constructing a battery health estimation model based on LSSVR and partial charge / discharge data. The technical idea is:
[0060] The complete charge and discharge data of the battery under different health conditions are collected, and the charge and discharge data under each health state are arbitrarily divided into multiple segments, and each segment of data is labeled with the same label to construct a training set; the labeled data is sent to the CS-LSSVM model for training, and the SOH is output to finally obtain the target model.
[0061] Take charging voltage data as an example. For example, the complete voltage data is as follows: 2.51, 2.52, 2.53, 2.54, 2.55, 2.56. There are six points in total, and the corresponding SOH is 90%. If N=3, the training set is {2.51, 2.52, 2.53} and 90%, {2.52, 2.53, 2.54} and 90%, {2.53, 2.54, 2.55} and 90%, {2.54, 2.55, 2.56} and 90%. The input is different and partially overlapped fragmented charging and discharging data, but the corresponding SOH is a value, that is, for a complete charging and discharging curve, if there are 100 points, assuming N=3, theoretically 100-3+1 inputs can be generated, but their outputs are all the same SOH. With multiple charging and discharging data corresponding to SOH (charging and discharging data under different health states of the battery), more training data can be constructed to meet the model training requirements.
[0062] Specifically, this embodiment is based on the least squares support vector regression (LSSVR) method of cuckoo search (CS) for directly using partial charge and discharge curves to estimate SOH. As an efficient nonlinear regression model, LSSVR reduces the computational complexity of the model by introducing the least squares error criterion, and can effectively capture the aging characteristics of SOH by combining the radial basis function (RBF). On the other hand, CS, as a global optimization algorithm, can effectively avoid local optimal traps during parameter optimization, further improving the modeling ability and prediction accuracy of LSSVR.
[0063] (1) Framework of least squares support vector regression
[0064] LSSVR is a regression method based on the extension of support vector machine (SVM). The key is to introduce a square loss function instead of the ε-insensitive loss function in the classic SVR, which transforms the optimization problem from quadratic programming to solving a linear equation system and simplifies the calculation process. The definition of LSSVR is as follows:
[0065]
[0066] Where x is the input vector of the model and y is the corresponding output. It is a nonlinear mapping function used to map the input vector to a higher dimensional space. w is the weight vector of the hyperplane and b is the intercept. For a given data set D = {x i ,y i} n (x i ∈R d×1 ,y i ∈R,i=1,2,K,n), the optimization problem of LSSVR is:
[0067]
[0068] where e i is the error variable and C is the regularization parameter, which is used to balance model complexity and error. The objective function and constraints are combined through the Lagrangian function:
[0069]
[0070] where α i is the Lagrange multiplier x corresponding to i Deriving the Lagrangian function and eliminating variables, we obtain the following linear equations:
[0071]
[0072] Where K(x i ,x j ) is the kernel function. The above equation can be abbreviated to matrix form:
[0073]
[0074] 1 of them n =[1,1,K,1] n×1 α=[α1,α2,K,α n ], I n , is the n×n unit matrix. After inputting the training data, the intercept b and the Lagrange multiplier α are obtained by solving this linear equation system. The model training is completed. For the new sample x', the regression prediction value is:
[0075]
[0076] Kernel function is an important term in SVR, and it has different forms, such as polynomial, linear, and Gaussian kernel functions. In this patent, Gaussian kernel is selected as the kernel function, and its formula is as follows:
[0077]
[0078] Where σ is a parameter of the kernel function, called the bandwidth parameter (scale parameter), which controls the action range and distribution shape of the kernel function.
[0079] (2) Parameter optimization based on Cuckoo Search
[0080] The performance of the LSSVR algorithm depends on the choice of the penalty factor C and the kernel function band parameter σ. The penalty factor is used to adjust the model's balance between the confidence coefficient and the empirical risk, while the kernel function width parameter affects the model's nonlinear mapping ability. These two hyperparameters not only determine the model's ability to fit known samples, but also significantly affect its ability to generalize to unknown samples. However, for different problem objects, these parameters usually need to be flexibly adjusted. However, in practice, the parameters are usually set empirically, which makes it challenging to guarantee the best model performance.
[0081] In order to further improve the modeling accuracy and generalization ability of LSSVR, an effective parameter optimization strategy must be adopted to promote the model to achieve the best balance between the fitting proficiency of known samples and the prediction ability of unknown samples. This patent uses an emerging heuristic optimization algorithm CS, which can automatically search and optimize the hyperparameters of LSSVR. Due to its powerful global search ability, the CS algorithm can effectively avoid the local optimal problem that may occur in traditional methods. This enables the CS algorithm to obtain excellent hyperparameter combinations, thereby improving the overall performance of LSSVR. The basic steps of CS can be summarized as shown in Table I.
[0082] Table I Cuckoo search algorithm
[0083]
[0084]
[0085] (3) SOH estimation based on LSSVR and partial charge / discharge data
[0086] The purpose of this embodiment is to propose a general, lightweight and practical method to estimate the SOH of SIBs. In this patent, partial charge / discharge data (a kind of easily accessible data) will be directly used as the input of LSSVR without any feature engineering. The proposed framework for estimating battery SOH uses partial charge data to determine the battery SOH. Specifically, for a given SIB dataset {(x1,y1),(x2,y2), K ,(x N ,y N )}, where x i =[x1,x2, K ,x k ]A piece of charge / discharge data, such as voltage. i is the corresponding SOH, and k is the segment partial charge and discharge dimension. Secondly, a training set is formed according to the proposed design ideas, and the kernel parameters are optimized using the training set, and then the CS-LSSVM model is built. Finally, partial charge / discharge data of the target test SIB is collected, and the trained model is used to predict the SOH of the battery.
[0087] The target model constructed in this embodiment has the following advantages for estimating the health status of new energy storage batteries:
[0088] 1. The partial charge and discharge data used as input features are easy to obtain and do not require feature engineering
[0089] The input of the method framework proposed in this embodiment can be any continuous partial data obtained during the charging and discharging process, such as voltage, etc. These data are easy to obtain, do not require feature engineering, and do not rely on data integrity. Only partial continuous data is required.
[0090] 2. The SOH assessment does not depend on the type of data condition
[0091] The method proposed in this embodiment does not depend on the type of battery data operating conditions. It can be a battery training data set formed by multiple operating conditions. It only needs to input partial charging and discharging conditions of the battery under these different operating conditions to complete the SOH evaluation task.
[0092] Example of implementation method
[0093] (1) Experimental data
[0094] Four commercially available sodium-ion soft-pack battery cells (82.0×61.0×7.0 mm) were used in this experiment. Each cell has a rated capacity of 3.27 Ah and has a NaNi 1 / 3 Fe 1 / 3 Mn 1 / 3O2 (NFM) cathode and hard carbon (HC) anode. The terminal discharge and charge voltages are 1.5V and 3.9V, respectively. In order to be closer to practical applications, the aging experiment uses a constant power (CP) protocol in addition to the common constant current (CC) protocol. There are four charge and discharge cycle test protocols: 0.5C / 0.5C, 1C / 1C, 0.5P / 0.5P and 1P / 1P. Among them, 0.5C / 0.5C and 1C / 1C represent CC charge and discharge modes, and 0.5P / 0.5P and 1P / 1P represent constant power CP charge and discharge modes. The maximum number of cycles is 700, and every 100 cycles is a reference test cycle. The detailed aging test steps are shown in Table II.
[0095] Table II Soft-package SIB aging performance test
[0096]
[0097] The results of the aging test are as follows Figure 2 The corresponding charging and discharging curves are shown in Figure 3 As shown. Figure 2 It can be found that under different conditions, the SOH of the NFM system is around 90% after 500 cycles, but it drops significantly in 600-700 cycles. Increasing the charge and discharge rate and increasing the charge and discharge power will have a greater impact on the cycle, resulting in faster capacity degradation. Figure 3 It can be found that the charge and discharge curves of SIBs with different SOHs show different trends, which can be used for modeling the SOH method.
[0098] Results and Discussion
[0099] In this section, we will introduce the SOH estimation results using partial charge and discharge curve data and CS-LSSVR. The results will be fully demonstrated using Leave-Out Cross-Validation (LOOCV). In order to evaluate the accuracy of the estimation, we choose the mean absolute error (MAE) and root mean square error (RMSE) to quantify the estimation error, calculated as
[0100]
[0101] where y i is the observed value, is the predicted value and n is the number of samples. Figure 4 The SOH prediction results of the test SIB using the other three batteries as training sets are shown respectively, where the length of the partial charge data used is N = 300. The efficacy of this method on SOH is shown in Figure 4 shown.
[0102] Obviously, the length of the input data will affect the results. Figure 5Box plots illustrating the prediction errors of SOH for different lengths. Figure 5 It shows that N = 300 exhibits the best results and the lowest error. This result shows that the assumption that length is proportional to quality is not necessarily valid. Although a longer length may contain more valid information, it also contains a greater proportion of negative information and reduces the efficiency of algorithm processing.
[0103] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for constructing a battery health estimation model based on LSSVR and partial charge / discharge data, characterized in that: The following steps are involved: The complete charge and discharge data of the battery under different health conditions are collected, and the charge and discharge data under each health state are arbitrarily divided into multiple segments. Each segment of data is labeled with the same label to construct a training set; the labeled data is sent to the CS-LSSVM model for training, and the SOH is output to finally obtain the battery health estimation model.
2. The method for constructing a battery health estimation model based on LSSVR and partial charge / discharge data according to claim 1, characterized in that: The charge and discharge data are voltage or current data.
3. The method for constructing a battery health estimation model based on LSSVR and partial charge / discharge data according to claim 1 or 2, characterized in that: The method for constructing the CS-LSSVM model is to use the cuckoo search algorithm to perform global optimization on the LSSVR.
4. The method for constructing a battery health estimation model based on LSSVR and partial charge / discharge data according to claim 3, characterized in that: The LSSVR is defined as follows: Where x is the input vector of the model and y is the corresponding output. It is a nonlinear mapping function used to map the input vector to a higher dimensional space. w is the weight vector of the hyperplane and b is the intercept. For a given data set D = {x i ,y i } n (x i ∈R d×1 ,y i ∈R,i=1,2,K,n), the optimization problem of LSSVR is: where e i is the error variable and C is the regularization parameter, which is used to balance model complexity and error. The objective function and constraints are combined through the Lagrangian function: where α i is the Lagrange multiplier x corresponding to i Deriving the Lagrangian function and eliminating variables, we obtain the following linear equations: Where K(x i ,x j ) is the kernel function. The above equation can be abbreviated to matrix form: 1 of them n =[1,1,K,1] n×1 i,j∈1,2,K,n,,α=[α1,α2,K,α n ], I n , is the n×n unit matrix. After inputting the training data, the intercept b and the Lagrange multiplier α are obtained by solving this linear equation system. The model training is completed. For the new sample x', the regression prediction value is: Kernel function is an important term in SVR, and it has different forms, such as polynomial, linear, and Gaussian kernel functions. In this patent, Gaussian kernel is selected as the kernel function, and its formula is as follows: Where σ is a parameter of the kernel function, called the bandwidth parameter (scale parameter), which controls the action range and distribution shape of the kernel function.
5. A system for constructing a battery health estimation model based on LSSVR and partial charge / discharge data, characterized in that: include: Training set construction module: collect complete charge and discharge data of the battery under different health conditions, divide the charge and discharge data under each health state into multiple segments at will, and label each segment with the same label; Training module: The labeled data is sent to the CS-LSSVM model for training, SOH is output, and finally the target model is obtained.
6. The system for constructing a battery health estimation model based on LSSVR and partial charge / discharge data according to claim 5, characterized in that: The charge and discharge data are voltage or current data.
7. The system for constructing a battery health estimation model based on LSSVR and partial charge / discharge data according to claim 5 or 6, characterized in that: The method for constructing the CS-LSSVM model is to use the cuckoo search algorithm to perform global optimization on the LSSVR.
8. The system for constructing a battery health estimation model based on LSSVR and partial charge / discharge data according to claim 7, characterized in that: The LSSVR is defined as follows: Where x is the input vector of the model and y is the corresponding output. It is a nonlinear mapping function used to map the input vector to a higher dimensional space. w is the weight vector of the hyperplane and b is the intercept. For a given data set D = {x i ,y i } n (x i ∈R d×1 ,y i ∈R,i=1,2,K,n), the optimization problem of LSSVR is: where e i is the error variable and C is the regularization parameter, which is used to balance model complexity and error. The objective function and constraints are combined through the Lagrangian function: where α i is the Lagrange multiplier x corresponding to i Deriving the Lagrangian function and eliminating variables, we obtain the following linear equations: Where K(x i ,x j ) is the kernel function. The above equation can be abbreviated to matrix form: 1 of them n =[1,1,K,1] n×1 i,j∈1,2,K,n,,α=[α1,α2,K,α n ], I n , is the n×n unit matrix. After inputting the training data, the intercept b and the Lagrange multiplier α are obtained by solving this linear equation system. The model training is completed. For the new sample x', the regression prediction value is: Kernel function is an important term in SVR, and it has different forms, such as polynomial, linear, and Gaussian kernel functions. In this patent, Gaussian kernel is selected as the kernel function, and its formula is as follows: Where σ is a parameter of the kernel function, called the bandwidth parameter (scale parameter), which controls the action range and distribution shape of the kernel function.
9. Use the battery health estimation model based on LSSVR and partial charge / discharge data as described in any one of claims 1 to 4 to estimate the SOH of the new energy battery, characterized in that: Obtain battery charge and discharge data, directly input it into the health estimation model, and output the SOH result.
10. A new energy battery SOH estimation system, characterized in that: The battery health estimation model comprising the LSSVR and partial charge / discharge data as described in any one of claims 1 to 4.