Lithium-ion battery soh estimation method based on life prognosis and transfer learning

Abstract

The application discloses a lithium ion battery SOH estimation method based on life prognosis and migration learning, extracts battery life characteristic values, establishes a battery SOH gated recurrent unit (GRU) prediction model by using the historical cycle data of two clustering center batteries, the input is the 1C rate constant-current charging time and 1C rate constant-current charging average voltage of the battery, and the output is the SOH of the battery. For the battery to be estimated, the life characteristics are extracted from the early cycle data, the life category is judged, and the corresponding SOH prediction model is matched; then, the early cycle data is used for migration learning of the prediction model, and then, on the basis of the model after migration learning, the 1C rate constant-current charging time and 1C rate constant-current charging average voltage of the battery are taken as the input, and SOH estimation is carried out. The estimation method is established on the basis of an existing model, does not need to model the data of each battery, saves the time cost under the premise of guaranteeing the estimation precision.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power battery management, and is used for estimating the state of health of lithium ion batteries in charge and discharge cycles. BACKGROUND

[0002] During use, the performance of a lithium ion battery gradually deteriorates with charge and discharge cycles, mainly manifested as attenuation of available capacity and decline of power performance. Therefore, in order to ensure that the battery can meet the power demand of the power consumption device, it is necessary to estimate the state of health (SOH) represented by the available capacity. The estimation of the state of health is used to ensure the safety and reliability of the battery use, and is an important function of the battery management system.

[0003] The existing data-driven SOH estimation method establishes a data model according to the past charge and discharge cycle data of the battery, and estimates the SOH of the target battery. This method has good prediction effect for the SOH estimation of the same batch of batteries and the same charge and discharge cycle mode of the batteries, but when the target battery is different from the data model battery and the battery is in different charge and discharge cycle modes, the estimation accuracy of this method will be greatly affected. Therefore, the data-driven estimation method has the problems of poor generalization performance and adaptability. SUMMARY

[0004] To solve the problems of low SOH estimation accuracy, poor generalization performance and adaptability of the data-driven method for different batches and different charge and discharge cycle mode batteries, the application proposes a battery SOH estimation method based on the combination of life prognosis and transfer learning.

[0005] The specific application content is as follows: first, the early charge and discharge cycle data (such as the first 100 charge and discharge cycle data, and the battery can have different charge and discharge cycle modes) of a batch of batteries are needed, and two characteristic values of battery life diagnosis are extracted; then the characteristic values are taken as inputs, and a clustering algorithm is used to classify the batteries, for example, into two categories, long-life and short-life batteries; the clustering algorithm can also output two cluster center batteries, and based on the historical charge and discharge data of the cluster center battery, a health factor is extracted, and a gated recurrent unit (GRU) is used to establish two data models, a long-life battery data model and a short-life battery data model, for SOH estimation of the battery.

[0006] For the target battery to be estimated, its early cycle data, such as the first 100 charge and discharge cycle data, are needed, and the life diagnosis characteristic values are extracted, and the expected life of the battery is determined according to the characteristic values. According to the length of the expected life, the corresponding data model is selected, and the early cycle data is used for transfer learning of the model to estimate the SOH of the target battery.

[0007] In order to make the technical solutions of the present application clearer, the present application will be further described in detail below in combination with the drawings and specific implementation steps.

[0008] Step 1, life diagnosis characteristic value extraction. In the data of the first 100 cycles of a selected batch of batteries, two life diagnosis characteristic values of each battery are extracted, which are LPF_1 and LPF_2 respectively. The calculation of LPF_1 is shown in formula (1) as follows:

[0009]

[0010] In the formula, V and Q are both coordinate points on the Q(V) curve, and the Q(V) curve is a curve with the cumulative discharge capacity (Ah) as the vertical coordinate and the corresponding discharge voltage (V) as the horizontal coordinate. Formula (1) requires the Q(V) curve of the 10th and 100th cumulative discharge capacity of each battery as a function of discharge voltage. In the formula, N is the number of discrete points on the Q(V) curve, and N is determined according to the actual data collection points, and the number of collection points of the two curves is required to be the same.

[0011] The life diagnosis characteristic value LPF_2 is calculated as shown in formula (2) as follows:

[0012]

[0013] In the formula, is ΔQ 10-100 the i-th value of the sequence, i∈1,2,…,N, is ΔQ 10-100 the average value of the sequence.

[0014] In the formula, ΔQ 10-100 is calculated as shown in formula (3) as follows,

[0015] ΔQ 10-100 =Q 10 -Q 100 (3)

[0016] In the formula, Q 10 is the capacity sequence of the 10th cycle Q(V) curve of the battery, Q 100 is the capacity sequence of the 100th cycle Q(V) curve of the battery, Q 10 and Q 100 have the same sequence length N.

[0017] Step 2, apply the k-means clustering algorithm, input the two life diagnosis characteristic values of each battery extracted in step 1, set the number of clusters to 2, divide the selected batteries into two categories of long-life batteries and short-life batteries, and output two cluster center batteries.

[0018] Step 3, based on the historical data of 2 cluster center batteries, the 1C rate constant current charging time and the 1C rate constant current charging average voltage are input into the GRU neural network as health factors, and the battery SOH is output, and the data training is carried out based on the GRU neural network, and the SOH estimation model for long and short life battery prediction is obtained respectively, and the SOH estimation model library is formed.

[0019] Step 4, the first 100 charge and discharge cycle data of the battery to be estimated are obtained, two life characteristic identification values are extracted according to formula (1) and formula (2), and are input into the life identification clustering algorithm of step 2, and the distance between the battery to be estimated and the cluster center battery is determined, wherein L1 is the distance between the battery to be estimated and the long life battery cluster center battery, and L2 is the distance between the battery to be estimated and the short life battery cluster center battery. Determine the category of the target battery and match the appropriate SOH estimation model.

[0020] Step 5, the 1C rate constant current charging time, 1C rate constant current charging average voltage and SOH sequence are extracted from the data collected in the early life of the battery to be estimated to form retraining data, and the migration learning is carried out on the matched SOH data model, only the weights and biases of the fully connected layer in the GRU neural network are adjusted in the training process, and the parameters of other layers are kept unchanged, the migration model is obtained, and the adaptability training of the model is realized.

[0021] Step 6, the health factors of 1C rate constant current charging time and 1C rate constant current charging average voltage are extracted online from the current cycle of the battery to be estimated and input into the migration model trained in step 5, and the SOH estimation value can be output.

[0022] Compared with the prior art, the method first diagnoses the life of the battery, so as to apply the data model meeting the life expectation, solves the generalization or the need for re-modeling of the model; secondly, when the data model is applied, the early data is used for migration learning of the model, on the basis of spending very small time cost, the prediction accuracy of the model is improved, and the applicability of the model is solved. In specific practice, good prediction effect is obtained in the prediction of different batches of batteries and different charge and discharge cycle mode batteries. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 It is a battery life classification diagram based on K-Means algorithm.

[0024] Figure 2 It is a diagram for establishing a battery SOH prediction model library.

[0025] Figure 3A diagram for battery life estimation, diagnosis and model matching. In the diagram, L1 is the distance between the battery to be estimated and the cluster center battery of long-life batteries. L2 is the distance between the battery to be estimated and the cluster center battery of short-life batteries

[0026] Figure 4 A diagram for SOH prediction of the battery to be estimated.

[0027] Figure 5 A flowchart of the overall process of the method of the present application. DETAILED DESCRIPTION

[0028] The following detailed implementation method includes the following steps:

[0029] In order to more clearly illustrate the technical solutions of the present application, the present application will be further described in detail in combination with the drawings and specific implementation steps.

[0030] Step 1, life diagnosis characteristic value extraction. In the data of the first 100 cycles of a selected batch of batteries, two life diagnosis characteristic values of each battery are extracted, which are LPF_1 and LPF_2 respectively. The calculation of LPF_1 is shown in formula (1):

[0031]

[0032] In the formula, V and Q are coordinate points on the Q(V) curve, and the Q(V) curve is a curve with cumulative discharge capacity (Ah) as the vertical coordinate and corresponding discharge voltage (V) as the horizontal coordinate. Formula (1) requires the Q(V) curve of the 10th and 100th cumulative discharge capacity of each battery as a function of discharge voltage. In the formula, N is the number of discrete points on the Q(V) curve, and N is determined according to the actual data collection points, and the number of collection points of the two curves is required to be the same.

[0033] The life diagnosis characteristic value LPF_2 is calculated as shown in formula (2):

[0034]

[0035] In the formula, is ΔQ 10-100 is the ith value of the sequence, i∈1,2,…,N, is ΔQ 10-100 is the average value of the sequence.

[0036] In the formula, ΔQ 10-100 is calculated as shown in formula (3),

[0037] ΔQ 10-100 = Q 10 - Q 100 (3)

[0038] In the formula, Q10 Q is the capacity sequence of the Q(V) curve of the battery at the 10th cycle, 100 Q is the capacity sequence of the Q(V) curve of the battery at the 100th cycle, 10 and Q is the capacity sequence of the Q(V) curve of the battery at the 1000th cycle. 100 The sequence lengths of Q, Q and Q are unified as N.

[0039] Step 2, apply the k-means clustering algorithm, input each battery two life identification feature values extracted in step 1, set the cluster number to 2, divide the selected batteries into long-life batteries and short-life batteries, and output two cluster center batteries. The specific operation is shown in Figure 1 .

[0040] Step 3, taking the historical data of the two cluster center batteries as the basis data, taking the 1C rate constant current charging time and the 1C rate constant current charging average voltage as the health factor input into the GRU neural network, taking the battery SOH as the output, based on the GRU neural network, data training is carried out, and the SOH estimation model for long and short life battery prediction is obtained respectively, forming the SOH estimation model library. The specific operation is shown in Figure 2 .

[0041] Step 4, obtain the previous 100 charge and discharge cycle data of the battery to be estimated, extract two life feature identification values according to formula (1) and formula (2), and input them into the life identification clustering algorithm of step 2, according to the distance between the battery to be estimated and the cluster center battery, wherein L1 is the distance between the battery to be estimated and the cluster center battery of the long-life battery, and L2 is the distance between the battery to be estimated and the cluster center battery of the short-life battery. Determine the category of the target battery, and match the appropriate SOH estimation model. The operation is shown in Figure 3 .

[0042] Step 5, extract the 1C rate constant current charging time, 1C rate constant current charging average voltage and SOH sequence from the data collected in the early life of the battery to be estimated to form retraining data, and perform transfer learning on the matched SOH data model. In the training process, only the weights and biases of the fully connected layer in the GRU neural network are adjusted, and the parameters of other layers are kept unchanged, to obtain a transfer model and realize adaptive training of the model.

[0043] Step 6, extract the health factors, i.e. 1C rate constant current charging time and 1C rate constant current charging average voltage from the current cycle of the battery to be estimated, and input them into the transfer model trained in step 5, to output the SOH estimation value. The operation is shown in Figure 4 . In order to illustrate the implementation effect of the method of the present application, the SOH of a batch of batteries will be estimated, and the estimation accuracy of the method will be verified by root mean square error (RMSE).

[0044] The 124 batteries are collectively charged using one-step or two-step fast charging strategies, and a total of 71 charging modes are included. First, 94 of the 124 batteries are cycled data as a training data set, and the remaining 30 battery data are used as a test data set.

[0045] Using the steps in the present application, the training data set is classified using the k-means clustering algorithm, and the batteries are divided into long-life batteries and short-life batteries. The cluster centers of each type of battery are output, and a life expectancy recognition model is established. The historical data of the cluster centers of different categories are used to train the GRU model, and long-life battery prediction models and short-life battery prediction models are obtained, forming a prediction model library.

[0046] The battery life recognition features extracted from the battery to be estimated are input into the life expectancy recognition model, and the battery life recognition result of the battery to be tested is output. According to the recognition result, a data model matching the life is selected from the prediction model library. That is, if the battery to be tested is identified as a short-life battery, the short-life battery prediction model is automatically matched, and vice versa, the long-life battery model is automatically matched.

[0047] Then, the early cycle data of the battery to be estimated is extracted as transfer data, and the matched prediction model is trained to obtain a transfer prediction model.

[0048] Finally, the 1C rate constant current charging time and the 1C rate constant current charging average voltage are extracted from the current test data of the battery to be estimated and input into the transfer model, and the SOH prediction result is output.

[0049] The SOH of 30 batteries is predicted and verified by discharge capacity. The root mean square error (RMSE) of 40% of the batteries is less than 0.01, and the root mean square error (RMSE) of 60% of the batteries is between 0.01 and 0.02. The variance of the RMSE of the model for predicting 30 batteries is 1.91x10 -5 , with less fluctuation. It shows that this method can be used for different charging modes with good generalization and adaptability.

Claims

1. A lithium-ion battery SOH estimation method based on life prognosis and transfer learning, characterized in that: Comprising the following steps: Step 1, life diagnosis characteristic value extraction; select a batch of battery 100 cycle data, extract two life diagnosis characteristic values of each battery, respectively LPF_1 and LPF_2; wherein the calculation of LPF_1 is shown as formula (1): In the formula, V and Q are both coordinate points on the Q(V) curve, and the Q(V) curve is a curve with cumulative discharge capacity as the vertical coordinate and the corresponding discharge voltage as the horizontal coordinate; formula (1) requires Q(V) curve of the 10th and 100th cumulative discharge capacity of each battery with respect to discharge voltage; in the formula, N is the number of discrete points on the Q(V) curve, N is determined according to the actual data collection points, and the number of collection points of the two curves is required to be the same; The calculation of life diagnosis characteristic value LPF_2 is shown as formula (2): wherein is ΔQ 10-100 the i-th value of the sequence, i ∈ 1, 2,..., N, is ΔQ 10-100 the average value of the sequence; where ΔQ 10-100 The calculation of the sequence is shown in equation (3), ΔQ 10-100 = Q 10 - Q 100 (3) wherein Q 10 is the sequence of capacities of the Q(V) curve of the 10th cycle of the battery, Q 100 is the sequence of capacities of the Q(V) curve of the 100th cycle of the battery, Q 10 and the sequence length of Q 100 is unified to N; Step 2, applying k-means clustering algorithm, taking the two life diagnosis characteristic values of each battery extracted in step 1 as input, setting the number of clusters to 2, and dividing the selected batteries into long life battery and short life battery two categories, and outputting two cluster center batteries; Step 3, taking the historical data of the two cluster center batteries as the basis data, taking the 1C rate constant current charging time and 1C rate constant current charging average voltage as the health factor input into the GRU neural network, and taking the battery SOH as the output, based on the GRU neural network for data training, respectively obtaining the SOH estimation model for long and short life battery prediction, forming the SOH estimation model library; Step 4, obtaining the previous 100 charge and discharge cycle data of the battery to be estimated, extracting two life characteristic recognition values according to formula (1) and formula (2), and inputting them into the life recognition clustering algorithm of step 2, determining the category of the target battery according to the algorithm output, and matching the appropriate SOH estimation model; Step 5, extracting 1C rate constant current charging time, 1C rate constant current charging average voltage and SOH sequence from the data collected in the early life of the battery to be estimated to form retraining data, and performing transfer learning on the matched SOH data model, only adjusting the weights and biases of the fully connected layer in the GRU neural network during the training process, keeping the parameters of other layers unchanged, obtaining the transfer model, and realizing the adaptive training of the model; Step 6, extracting the health factors, i.e. 1C rate constant current charging time and 1C rate constant current charging average voltage from the current cycle of the battery to be estimated online, and inputting them into the transfer model trained in step 5, so as to output the SOH estimation value. 2.The lithium-ion battery SOH estimation method based on life prognosis and transfer learning according to claim 1, wherein, The calculation of life characteristic value includes the calculation of LPF_1 and LPF_2.

Images