Lithium battery SOH estimation method based on EIS ensemble learning algorithm

Abstract

The invention relates to the technical field of lithium battery SOH estimation methods, in particular to a lithium battery SOH estimation method based on an EIS ensemble learning algorithm. Comprising the following steps: S1, collecting battery performance data; s2, collecting corresponding electrochemical impedance spectroscopy data through an EIS method; s3, obtaining feature data through ICA, DVA and DTV methods, performing normalization processing on the feature data and the data obtained in the S2, and merging the data into a feature vector; s4, calculating a capacity fading rate CAR; according to the capacity fading rate, allocating to different algorithms to carry out SOH estimation, and when the capacity fading rate is less than or equal to 10%, selecting an ELM algorithm to calculate an SOH estimation value; when the capacity fading rate is greater than 10% and less than or equal to 30%, selecting a CNN architecture to calculate an SOH estimated value; when the capacity fading rate is greater than 30%, selecting an SVM algorithm to calculate an SOH estimated value; and S5, displaying and storing the predicted SOC result. Compared with the prior art, the optimal estimation algorithm is dynamically switched based on the capacity fading rate, the full life cycle estimation precision is improved, and the calculation complexity is remarkably reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium battery state of health estimation method, in particular to a lithium battery SOH estimation method based on EIS ensemble learning algorithm. BACKGROUND

[0002] With the wide application of lithium batteries in the field of home energy storage, accurately estimating the state of health (SOH) of lithium batteries is of great significance for prolonging battery life, reducing maintenance costs and ensuring safe operation of the system. Existing SOH estimation methods mainly include model-based methods and data-driven methods. Although the model-based method can better explain the electrochemical process inside the battery, it is complex to model and has high computational cost. Although the data-driven method has high portability and robustness, it has the following problems:

[0003] Limited estimation accuracy of a single algorithm: a single machine learning algorithm often has difficulty achieving high-precision estimation when dealing with complex battery aging processes.

[0004] Large data requirement: traditional data-driven methods require a large amount of data to train the model, which is difficult to meet in practical applications, especially in the case of few battery cycles.

[0005] Unreasonable selection of frequency points: when using EIS combined with machine learning, existing research usually only selects 1 or 2 low-frequency frequency points, which cannot fully reflect the aging state of the battery. SUMMARY

[0006] To solve the problems raised in the background art, the present application provides a lithium battery SOH estimation method based on EIS ensemble learning algorithm, which includes the following steps: S1, collecting battery performance data, including current battery available capacity and nominal capacity; S2, collecting corresponding electrochemical impedance spectrum data from the battery performance data by EIS method; S3, obtaining the incremental capacity peak position by ICA method, obtaining the entropy value by DVA method, and obtaining the temperature slope by DTV method, and normalizing and merging the above feature data and the data obtained in S2 into a feature vector; S4, calculating the capacity attenuation rate CAR, whose formula is as follows: According to the capacity attenuation rate, the SOH estimation value is allocated to different algorithms, when the capacity attenuation rate is less than or equal to 10%, the ELM algorithm is selected to calculate the SOH estimation value; when the capacity attenuation rate is greater than 10% and less than or equal to 30%, the CNN architecture is selected to calculate the SOH estimation value; when the capacity attenuation rate is greater than 30%, the SVM algorithm is selected to calculate the SOH estimation value; S5, displaying and saving the predicted SOC result.

[0007] The EIS method in step S2 comprises the following steps: S21, setting specific frequency points in the electrochemical impedance spectrum data; S22, using the EIS device to measure the battery at the selected frequency points, collecting the electrochemical impedance spectrum data of the battery at different aging stages, obtaining the impedance and phase angle data of the battery, calculating the ohmic internal resistance of the battery by the direct current method, and calculating the polarization internal resistance of the battery by the difference method.

[0008] Step S21 comprises the following steps: S201, determining a plurality of frequency points in the frequency range at different aging stages according to the electrochemical principle of the battery in the frequency range of 0.01 Hz-8000 Hz of the electrochemical impedance spectrum, to form a frequency point candidate set; S202, combining the frequency points in the frequency point candidate set to randomly form 100 different frequency point combinations; S203, using a genetic algorithm to optimize the frequency point combinations and select the individual code with the highest fitness: first, set the individual code to a binary vector [b1, b2,..., bi,..., bN] of each individual, bi=1 indicates that the i-th frequency point is selected, and the fitness of the individual is calculated, and the calculation formula is as follows: m ],b i =1 represents that the i-th frequency point is selected, the fitness of the individual is calculated, and the calculation formula is as follows: Wherein SOHtrue is the actual battery health state value, and SOH pred is the set battery health state value; the selected method is used to select the individual for breeding, and the formula of the selection probability is as follows: N is the population size; finally, single-point crossover and bit flip mutation operations are adopted to perform multiple iterations to obtain the individual code with the highest fitness and the corresponding frequency point combination.

[0009] The ELM algorithm for calculating the SOH estimation value in step S4 comprises the following steps: S401, input layer to hidden layer mapping, the input layer comprises 3 nodes, and the input vector is x=[x1, x2, x3] T , wherein x1, x2, and x3 represent the impedance real part at 0.1 Hz, the impedance imaginary part, and the temperature respectively, T represents the transpose of a matrix, the hidden layer comprises 8 nodes, and the output of the i-th hidden layer node is: Wherein, f(·) is an activation function, wi is an input layer to hidden layer weight vector, and bi is a bias; S402, weight β=[β1, β2,..., β8] T of the hidden layer to the output layer is calculated, and the calculation formula is as follows: β=(H T H) -1 H T y, wherein H is a hidden layer output matrix, y is an actual SOH label, and the calculation formula of the SOH estimation value is as follows: where hi is the output of the i-th hidden layer node, and βi is the corresponding weight.

[0010] The step S4 of calculating the SOH estimation value by the CNN architecture comprises the following steps: S411, setting three 1D-CNN layers and two MaxPooling layers; S412, inputting a sequence x∈R L , a convolution kernel k∈R 5 , a width of 5, a step S=2, and an i-th convolution output being: where b is a bias; S413, performing a MaxPooling operation, for an input sequence with a length of M, a pooling window size of 2, and a step of 2, the output is: y j =max{x 2j , x 2j+1}, where j represents a position index, the feature dimension is reduced by 2 layers of MaxPooling, and key mutation information is retained; S414, mapping to obtain an SOH value, and a calculation formula thereof is as follows: SOH=f(W z +b), wherein W is a feature vector, z and b are weights and biases, and f(·) is a ReLU activation function.

[0011] The step S4 of calculating the SOH estimation value by the SVM algorithm comprises the following steps: S511, obtaining a scalar value of a sample by an RBF kernel function, and a calculation formula thereof is as follows: K(x i , x j )=exp(-γ‖x i -x j ‖ 2 ), wherein γ is a kernel parameter and is set to be 0.05; S512, optimizing an objective by a weighted loss function, introducing an abnormal value weight, and suppressing the influence of voltage jitter noise, and a formula thereof is as follows: where w is a weight vector, b is a bias term, ξ is a relaxation variable vector, C is a penalty parameter and is set to be 2.5, δi is an indication variable, is set to be 0.7 if the sample is an abnormal value, is set to be if the sample is not an abnormal value, and ξi is a relaxation variable of the i-th sample; S513, calculating an SOH estimation, and a formula thereof is as follows: where α i is a Lagrange multiplier, and x is an input feature.

[0012] Compared with the prior art, the present application integrates the advantages of multiple algorithms through an ensemble learning framework, introduces EIS collected data features, dynamically switches an optimal estimation algorithm based on a capacity attenuation rate, improves estimation accuracy in a whole life cycle, and significantly reduces computational complexity. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A method flowchart of the present application. DETAILED DESCRIPTION

[0014] The application will be further described in the following with reference to the accompanying drawings.

[0015] As Figure 1 A lithium battery SOH estimation method based on an EIS ensemble learning algorithm, the lithium battery SOH estimation method comprising the following steps: S1, collecting battery performance data, including current battery available capacity and nominal capacity;

[0016] S2, collecting corresponding electrochemical impedance spectrum data from the battery performance data by an EIS method; S3, obtaining an incremental capacity peak position by an ICA method, an entropy value by a DVA method, and a temperature slope by a DTV method, and performing normalization processing on the above feature data and the data obtained in S2 to combine the data into a feature vector; S4, calculating a capacity attenuation rate CAR, and the calculation formula is as follows: According to the capacity attenuation rate, SOH estimation values are distributed to different algorithms, when the capacity attenuation rate is less than or equal to 10%, an ELM algorithm is selected to calculate the SOH estimation value, the single prediction time consumption is short, and the efficiency is significantly improved; when the capacity attenuation rate is greater than 10% and less than or equal to 30%, a CNN architecture is selected to calculate the SOH estimation value, which can capture the local mutation characteristics of the charge-discharge curve, and the fitting degree of the nonlinear relationship of the medium-term data is higher; when the capacity attenuation rate is greater than 30%, an SVM algorithm is selected to calculate the SOH estimation value, which can suppress the influence of voltage jitter noise, and the late prediction is more stable; S5, displaying and saving the predicted SOC result.

[0017] The EIS method in step S2 comprises the following steps: S21, setting specific frequency points in the electrochemical impedance spectrum data; S22, using an EIS device to measure the battery at the selected frequency points, collecting the electrochemical impedance spectrum data of the battery at different aging stages, obtaining the impedance and phase angle data of the battery, and calculating the ohmic internal resistance of the battery by the direct current method, and the polarization internal resistance of the battery by the difference method, as the reaction battery internal state characteristics, reducing the dependence of machine learning on a large amount of charge-discharge cycle data. According to the results, the parameters of the algorithm are adjusted, the model performance is optimized, and the accuracy of SOH estimation is improved.

[0018] Step S21 includes the following steps: S201, determining a plurality of frequency points in different aging stage frequency ranges according to the electrochemical principle of the battery in a frequency range of 0.01 Hz-8000 Hz of the electrochemical impedance spectrum, to form a frequency point candidate set; S202, combining the frequency points in the frequency point candidate set to randomly form 100 different frequency point combinations; S203, using a genetic algorithm to optimize the frequency point combinations and select the individual code with the highest fitness: first, set the individual code of each individual to a binary vector [b1, b2,..., bN], where b i =1 indicates that the i th frequency point is selected, and the fitness of the individual is calculated, and the calculation formula is as follows: m ], b i i = 1 indicates that the i th frequency point is selected, and the fitness of the individual is calculated, and the calculation formula is as follows: wherein SOH true is the actual battery state of health value, and SOH pred is the set battery state of health value; the selected method is used to select the individual for breeding, and the formula of the selection probability is as follows: N is the population size, and the value range is 20-50; finally, the single-point crossover and bit-flip mutation operations are adopted, the crossover and mutation crossover probability is selected between 0.8-0.9, the gene exchange frequency is controlled, the high probability is avoided to damage the high-quality individual or the low impact is avoided to affect the search efficiency, the mutation probability is selected between 0.01-0.05, the new gene generation probability is controlled, the algorithm is avoided to be premature, and the optimal frequency point combination is obtained after multiple iterations within 50-100 iterations.

[0019] The ELM algorithm calculates the SOH estimation value in step S4, including the following steps: S401, input layer to hidden layer mapping, the input layer contains 3 nodes, and the input vector is x=[x1, x2, x3] T , wherein x1, x2, and x3 represent the 0.1 Hz impedance real part, impedance imaginary part, and temperature respectively, T represents the transpose of the matrix, the hidden layer contains 8 nodes, the model fitting capability is determined, the calculation efficiency and accuracy are balanced, and the output of the i th hidden layer node is: wherein f(·) is an activation function, the activation function is selected as a ReLU function or a sigmoid function, the hidden layer output nonlinear mapping is controlled, wi is the weight vector from the input layer to the hidden layer, and bi is the bias; S402, weight β=[β1, β2,..., β8] T from the hidden layer to the output layer is calculated, and the calculation formula is as follows: β=(H T H) -1 H T y, wherein H is the hidden layer output matrix, y is the actual SOH label, and the SOH estimation value is calculated, and the calculation formula is as follows: wherein hi is the output of the i th hidden layer node, and βi is the corresponding weight.

[0020] The CNN architecture in step S4 calculates the SOH estimation value, including the following steps: S411, three 1D-CNN layers are set to extract multi-scale features and cover nonlinear changes in different time windows, and two MaxPooling layers are set to compress the feature dimension each time to enhance the robustness of the features; S412, the input sequence is x∈R L , the convolution kernel is k∈R 5 , the width is 5, the step S is 2, 5 consecutive data points are processed each time, local mutations are captured, the amount of calculation is reduced, and interval information between features is retained, and the i-th convolution output is: where b is a bias; S413, a MaxPooling operation is performed, for an input sequence with a length of M, the pooling window size is 2, the step is 2, and the output is: y j =max{x 2j , x 2j+1}, where j represents a position index, the feature dimension is reduced by 2 layers of MaxPooling, and key mutation information is retained; S414, the SOH value is mapped, and the calculation formula is as follows: SOH=f(wz+b), where W is a feature vector, z and b are weights and biases, f(·) is a ReLU activation function, the gradient vanishing problem is alleviated, and the fitting ability of nonlinear data is improved.

[0021] The SVM algorithm in step S4 calculates the SOH estimation value, including the following steps: S511, the scalar value of the sample is obtained by using the RBF kernel function, the nonlinear relationship such as the sudden increase of high-frequency impedance in the later period is processed, and the model generalization ability is improved, and the calculation formula is as follows: K(x i , x j )=exp(-γ||x i -x j || 2 ) where γ is a kernel parameter and is set to 0.05 to avoid overfitting; S512, a weighted loss function is used to optimize the objective, an abnormal value weight is introduced, and the influence of voltage jitter noise is suppressed, and the formula is as follows: where w is a weight vector, b is a bias term, ξ is a relaxation variable vector, C is a penalty parameter and is set to 2.5, δi is an indicator variable, and is set to 0.7 if the sample is an abnormal value, the contribution of the voltage jitter abnormal value to the loss function is reduced, the MAE is stabilized at 1.2%, and is set to if the sample is not an abnormal value, ξi is the relaxation variable of the i-th sample; S513, the SOH estimation is calculated, and the formula is as follows: where α i is a Lagrange multiplier, and x is an input feature.

[0022] Implementation case: As for calculating the current SOH of lithium iron phosphate battery, S1, the available capacity of the battery is collected as 18 Ah, and the nominal capacity is 20 Ah, S2, the corresponding electrochemical impedance spectrum data is collected from the battery performance data by EIS method, specifically, in the frequency range of 0.01 Hz-8000 Hz, through correlation analysis, 2 points with correlation coefficient >0.85 related to capacity attenuation are selected, and according to electrochemical principle, 2 points in low frequency region (0.01-1 Hz) are selected to reflect diffusion process, 2 points in medium frequency region (1-100 Hz) are selected to reflect charge transfer process, and 3 points in high frequency region (100-8000 Hz) are selected to reflect ohmic impedance, to obtain the frequency point candidate set [0.01, 0.05, 0.1, 5, 50, 100, 1000, 3000, 5000], and 100 groups of 5-point combinations are randomly generated according to the candidate set, such as Comb1=[0.05, 0.1, 5, 100, 1000, 5000], Comb2=[0.01, 50, 100, 3000, 5000], Comb3=[0.01, 0.1, 5, 100, 5000]…Comb100=[0.1, 5, 50, 100, 3000, 5000]. Set the population size as 50, the iteration number as 100, the crossover probability as 0.8, and the mutation probability as 0.05, set the coding of each individual, such as the binary vector V=[1, 0, 1, 1, 0, 1, 0, 0, 1] corresponding to Comb3, calculate the fitness of each individual, such as 0.988 for Comb100, calculate the selection probability of each individual, such as 0.021 for Comb100, randomly generate 50 individuals, through single-point crossover and bit flip mutation, after 100 iterations, the individual with the highest fitness is obtained, the coding is [0, 1, 1, 1, 0, 1, 0, 0, 1], and the corresponding frequency points are [0.05 Hz, 0.1 Hz, 5 Hz, 100 Hz, 5000 Hz]. Measure and calculate the internal resistance, the frequency tends to infinity, such as at 5000 Hz, the impedance real part is 0.025 Ω, the ohmic internal resistance is 0.025 Ω, such as at 0.05 Hz, the impedance real part is 0.15 Ω, the total internal resistance is 0.15 Ω, and the polarization internal resistance is 0.125 Ω. S3, input features, EIS features include ohmic internal resistance 0.025 Ω, polarization internal resistance 0.125 Ω, and impedance real part at 0.5 Hz 0.085 Ω. Through ICA method, the incremental capacity peak position is 0.12 V, through DVA method, the entropy value is 1.85, through DTV method, the temperature slope is 0.05 ℃ / min, combined with ohmic internal resistance, polarization internal resistance, and impedance real part at 0.5 Hz, normalization processing is carried out, and combined into feature vector: [0.025, 0.125, 0.085, 0.12, 1.85, 0.05]. S4, the capacity attenuation rate CAR is calculated as 10%.According to the capacity attenuation rate assigned to different algorithms for SOH estimation value, the capacity attenuation rate is 10%, the ELM algorithm is selected to calculate the SOH estimation value. Set the input vector: x = [0.085, -0.035, 25], the weight vector is w1 = [0.1, -0.2, 0.3], the bias b1 is 0.5, the activation function is the sigmoid function, the hidden layer node output is 0.87, the output layer weight is 0.95, and the SOH estimation value is 90.5%. S5, the current SOH estimation value is 90.5% and is saved.

Claims

1. A lithium battery SOH estimation method based on EIS ensemble learning algorithm, characterized in that: The lithium battery SOH estimation method includes the following steps: S1, collecting battery performance data, including the current usable capacity and nominal capacity; S2, collecting corresponding electrochemical impedance spectroscopy data from the battery performance data using the EIS method; S3, obtaining the incremental capacity peak position using the ICA method, the entropy value using the DVA method, and the temperature slope using the DTV method. These feature data and the data obtained in S2 are then normalized and merged into a feature vector; S4, calculating the capacity decay rate (CAR), the formula of which is as follows: Based on the capacity decay rate, different algorithms are assigned to calculate the SOH estimate. When the capacity decay rate is less than or equal to 10%, the ELM algorithm is selected to calculate the SOH estimate; when the capacity decay rate is greater than 10% and less than or equal to 30%, the CNN architecture is selected to calculate the SOH estimate; when the capacity decay rate is greater than 30%, the SVM algorithm is selected to calculate the SOH estimate; S5, display and save the predicted SOC results.

2. The lithium battery SOH estimation method based on the EIS ensemble learning algorithm according to claim 1, characterized in that: The EIS method in step S2 includes the following steps: S21, setting a specific frequency point in the electrochemical impedance spectroscopy data; S22, using an EIS device to measure the battery at the selected frequency point, collecting electrochemical impedance spectroscopy data of the battery at different aging stages, obtaining the battery's impedance and phase angle data, and using the DC method, when the frequency approaches infinity, using the real part of the impedance as equivalent to the ohmic internal resistance to calculate the battery's ohmic internal resistance, and using the difference method, subtracting the ohmic internal resistance from the total internal resistance to calculate the battery's polarization internal resistance.

3. The lithium battery SOH estimation method based on the EIS ensemble learning algorithm according to claim 2, characterized in that: Step S21 includes the following steps: S201, within the frequency range of 0.01Hz-8000Hz in the electrochemical impedance spectroscopy, based on the electrochemical principle of the battery, determine multiple frequency points within different aging stages to form a candidate set of frequency points; S202, combine the frequency points in the candidate set to randomly form 100 different frequency point combinations; S203, use a genetic algorithm to optimize the frequency point combinations and select the individual code with the highest fitness: first, set the individual code so that each individual is a binary vector [b1, b2, ..., b...]. m ], b i =1 indicates that the i-th frequency point is selected, and the fitness of the individual is calculated using the following formula: in SOH true represents the actual battery health status value, and SOH pred represents the set battery health status value. Individuals are selected for breeding using a selection method, and the selection probability formula is as follows: N is the population size; finally, single-point crossover and bit-flip mutation operations are used to perform multiple iterations to obtain the individual code with the highest fitness and the corresponding frequency point combination.

4. The lithium battery SOH estimation method based on the EIS ensemble learning algorithm according to claim 1, characterized in that: The SOH estimate calculated by the ELM algorithm in step S4 includes the following steps: S401. Input layer to hidden layer mapping: The input layer contains 3 nodes. Let the input vector be x = [x1, x2, x3]. T Where x1, x2, and x3 represent the real part of the 0.1Hz impedance, the imaginary part of the impedance, and the temperature, respectively; T represents the transpose of the matrix; the hidden layer contains 8 nodes; and the output of the i-th hidden layer node is: Where f(·) is the activation function, wi is the weight vector from the input layer to the hidden layer, and bi is the bias; S402, the weights from the hidden layer to the output layer β=[β1,β2,...,β8] T The output layer weights are calculated using the following formula: β=(H T H) -1 H T y, where H is the hidden layer output matrix and y is the actual SOH label. The SOH estimate is calculated using the following formula: Where hi is the output of the i-th hidden layer node, and βi is the corresponding weight.

5. The lithium battery SOH estimation method based on the EIS ensemble learning algorithm according to claim 1, characterized in that: The step S4 of calculating the SOH estimate using the CNN architecture includes the following steps: S411, setting up three 1D-CNN layers and two MaxPooling layers; S412, the input sequence is x∈R L The convolution kernel is k∈R 5 With a width of 5 and a stride of S = 2, the output of the i-th convolution is: Where b is the bias; S413, perform MaxPooling operation, for an input sequence of length M, the pooling window size is 2, the stride is 2, and the output is: y j =max{x 2j x 2j+1 }, where j represents the position index, the feature dimension is reduced by two layers of MaxPooling, and key mutation information is preserved; S414, the SOH value is obtained by mapping, and its calculation formula is as follows: SOH=f(Wz+b), where W is the feature vector, z and b are the weights and biases, and f(·) is the ReLU activation function.

6. The lithium battery SOH estimation method based on the EIS ensemble learning algorithm according to claim 1, characterized in that: The SVM algorithm in step S4 calculates the SOH estimate using the following steps: S511, obtaining the scalar value of the sample through the RBF kernel function, the calculation formula of which is as follows: K(x i x j )=exp(-γ||x i -x j || 2 Where γ is the kernel parameter and is set to 0.05; S512, the weighted loss function optimization objective, introduces outlier weights to suppress the influence of voltage jitter noise, and its formula is as follows: Where w is the weight vector, b is the bias term, ξ is the slack variable vector, C is the penalty parameter set to 2.5, δi is the indicator variable, set to 0.7 if the sample is an outlier, and set to 0.7 if the sample is not an outlier, and ξi is the slack variable of the i-th sample; S513, calculate the SOH estimate, the formula is as follows: Where α i Let x be a Lagrange multiplier and x be the input feature.

Images