Health factor selection method for battery health state estimation

Abstract

The invention discloses a health factor selection method for battery health state estimation, which is characterized in that health factors of a battery are screened by applying a random forest algorithm, and the number of decision trees is determined to be 100. And extracting in-bag data from the health factor data set by using a sampling method with putting back to construct and train a decision tree. Out-of-bag data is applied to carry out original data error and rearrangement error calculation of the decision tree model on the health factors, importance scoring is carried out on the health factors according to the difference value of the two errors, and therefore five health factors most relevant to the health state can be selected for the battery. The algorithm not only can reveal the nonlinear relationship between the battery health state and the health factor, takes the selected health factor as input, and adopts the long short-term memory network model to predict the health state value of each battery, but also can obtain a more accurate estimation result. The method is of great significance for predicting the development of the battery health state.

Description

Technical Field

[0001] The present invention belongs to the field of power battery management, and specifically relates to selecting health factors using a random forest algorithm. The selected health factors can improve the accuracy of battery health state estimation. Background Art

[0002] Lithium-ion batteries have become the mainstream energy storage device, including electric vehicles, due to their high energy density, excellent cycle stability and low self-discharge rate. During the charge and discharge cycle of lithium-ion batteries, the battery will gradually age, which is usually reflected in the reduction of battery capacity and the increase of internal resistance. When charging and discharging at high currents, it may cause safety problems such as temperature rise and even fire and explosion. Therefore, accurate assessment of the battery's state of health (SOH) is crucial to ensure battery safety and reliability. It is a core function of the battery management system (BMS).

[0003] At present, the data-driven method for predicting battery SOH has the advantage of not being limited to the specific type of battery and has good applicability. The neural network model is widely used in the data-driven method, and the battery SOH is estimated by establishing a nonlinear relationship between the battery health factor and the battery SOH. The health factor is a parameter or state variable that describes the internal characteristics of the battery, including directly measurable battery internal resistance, temperature or battery terminal voltage, and physical quantities calculated through direct measurement. Using the health factor as input data, the battery SOH can be estimated through a neural network.

[0004] Due to the important role of health factors, the current extraction of health factors not only includes the battery voltage and temperature at different stages that can be directly measured, but also can obtain the battery's charging and discharging energy, information entropy and other parameters at certain stages through analysis and calculation. Therefore, there are dozens of health factors, which reflect the health status of the battery from different aspects.

[0005] The input of up to dozens of health factors may lead to the complexity of the neural network model and reduce the training efficiency of the model. In addition, due to the complex information relationship reflected by the health factors, it affects the accuracy of the network model estimation. To solve this problem, some researchers use the Pearson correlation coefficient to screen health factors and reduce network input. However, due to the complexity and inconsistency of the battery, this method will encounter the problem of being unable to screen health factors with high correlation coefficients. In addition, since the association between SOH and health factors may contain nonlinear characteristics, and the Pearson correlation coefficient can only evaluate linear relationships, this limits its ability to select appropriate health factors.

[0006] To overcome these limitations, this study proposed an innovative method to extract 34 health factors based on the data of lithium-ion battery charging and discharging process, and select 5 health factors from them using Random Forest (RF) algorithm. These health factors were then input into the neural network to predict the battery's SOH.

[0007] Practice has proved that this method of selecting health factors can eliminate the interference and redundant information of health factors, thereby achieving more accurate SOH prediction than the Pearson correlation coefficient method. Summary of the invention

[0008] The purpose of the present invention is to provide a method for selecting health factors, which are used as inputs of the neural network for estimating SOH. The method screens the health factors of the battery by applying the RF algorithm, wherein the number of decision trees is determined to be 100. The in-bag data is extracted from the health factor data set using a sampling method with replacement to construct and train the decision tree. The original data error and permutation error of the decision tree model are calculated for the health factors using out-of-bag data, and the importance of the health factors is scored according to the difference between the two errors, so that the five health factors most relevant to the SOH can be selected for the battery. The algorithm can not only reveal the nonlinear relationship between the battery SOH and the health factors, but also use the selected health factors as inputs and use the Long Short-Term Memory network (LSTM) model to predict the SOH value of each battery, and more accurate estimation results can be obtained.

[0009] The specific implementation method includes the following steps:

[0010] In order to more clearly illustrate the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation steps.

[0011] Step 1: Based on the battery charging and discharging data, 34 health factors are extracted. Each health factor is a vector containing a time series, as shown in Table 1.

[0012] Step 2: Normalize the 34 health factor sequences separately and map the values ​​to the [0,1] interval.

[0013]

[0014] Among them, x i is the value of the ith health factor, min(x) is the minimum value in the health factor vector, and max(x) is the maximum value in the health factor vector. i The data value after normalization.

[0015] Step 3: Set the number of decision trees to 100 and use the sampling method with replacement to build and train the decision tree regression algorithm. For each decision tree, the sampled data is called the in-bag data. Since it is a sampling method with replacement, there are samples that are not extracted, and these samples are called out-of-bag data.

[0016] Step 4: Perform out-of-bag error analysis on all health factor data. Input the out-of-bag data containing the health factor sequence corresponding to the decision tree in step 3 into the decision tree for SOH prediction, and calculate the error between the predicted value and the out-of-bag sample SOH test value, which is given by formula (2).

[0017]

[0018] Among them, OOB Error t k represents the out-of-bag error of the health factor relative to the t-th tree. t is the number of health factor samples in the out-of-bag data of the t-th tree, y i is the SOH test value corresponding to the i-th sample of the health factor, is the i-th predicted value of the health factor by the t-th tree.

[0019] Since the same health factor may be out-of-bag data for n decision trees, it is necessary to apply formula (2) to calculate the out-of-bag error for n trees. By averaging these errors, the average out-of-bag error of the health factor can be obtained. The calculation is given by formula (3)

[0020]

[0021] Where n is the number of decision trees, OOB Error t is the health factor relative to the out-of-bag error of the t-th tree.

[0022] Step 5: Randomly rearrange the values ​​of the health factor vector to disrupt the order of the values ​​in the vector. Then repeat step 4 to obtain a new out-of-bag error.

[0023]

[0024] Step 6: For each health factor, calculate the difference IMP of the out-of-bag errors of the two arrangements. The calculation is given by formula (5):

[0025] IMP=E OOB,permuted -E OOB,origin (5)

[0026] Among them, E OOB,origin is the average out-of-bag error of n trees on the original dataset, E j,OOB,permuted is the average out-of-bag error after the health factor is permuted.

[0027] Step 7: The IMP of each health factor is the importance score of the health factor. The higher the score, the more relevant it is to SOH. Based on the score, select the five health factors with the highest IMP.

[0028] Step 8: SOH prediction output. The selected health factor is used as the model input to the LSTM neural network, and the battery SOH is used as the output to train the model. Figure 1 The LSTM model shown is used to predict the battery SOH. The hyperparameters of the LSTM neural network model are 512 neurons, 64 neurons in the fully connected layer, 200 iterations, 16 batch sizes, and a learning rate of 0.001. For new test data, input it into the trained model to obtain the estimated value of the battery SOH.

[0029] In order to evaluate the prediction performance of the model, the mean squared percentage error (MSPE), root mean square error (RMSE) and root mean squared percentage error (RMPSE) are used as indicators, and the error is given by the formula.

[0030]

[0031] Among them, Y i is the battery SOH measurement value, is the predicted value given by the model. N is the number of samples for estimation.

[0032] The whole process of this method can be used Figure 2 express. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The structure of the neural network (LSTM) used for SOH estimation

[0034] Figure 2 Flowchart of the method

[0035] Figure 3 Battery SOH estimation results. Specific embodiments

[0036] The present invention is further described in detail below with reference to the accompanying drawings and examples, and the SOH of a lithium-ion battery undergoing a charge and discharge cycle is estimated.

[0037] The test battery material is LiFePO4 / graphite, with a nominal capacity of 1.1Ah. The battery is charged with a two-stage high current, with a maximum charging current of 6C, followed by a constant current charge at a rate of 1C, and the battery is fully charged with constant voltage charge. After charging, the battery is left to stand, and then discharged at a rate of 4C. The ratio of the discharged capacity to the nominal capacity is used as the battery SOH measurement value. The above charge and discharge mode is repeated until the discharge capacity of the battery reaches 80% of the nominal capacity. According to the experimental results, the life of the battery is recorded as the number of cycles. A batch of test data of charge and discharge cycle batteries were randomly selected, and their lifespans ranged from 788 to 1227 times.

[0038] In the first step, all health factors of the battery are extracted as shown in Table 1, and the SOH value of each battery per cycle is calculated based on the discharge capacity of the battery.

[0039] In the second step, the 34 health factors of each battery are normalized using formula (1).

[0040] The third step is to determine the number of decision trees to be 100. Use the sampling method with replacement to build a decision tree, and extract 34 times. The extracted health factors are the in-bag data, and the unextracted ones are called out-of-bag data. The in-bag data are used for decision tree splitting and regression training. The decision tree splitting follows the minimum variance method. The out-of-bag data are used to calculate the error of the health factor.

[0041] Step 4: Calculate the health importance score for each health factor using formula (2) to formula (5). Select the top 5 health factors according to the importance score.

[0042] In the fifth step, the above steps are repeated for the selected batteries, and 5 health factors are selected for each battery. The health factors selected for the test batteries are shown in Table 2.

[0043] In the sixth step, the five health factors of each battery are input into the LSTM model to predict the battery SOH, and its prediction performance is evaluated by applying formulas (7)-(9). Figure 3 The training data and prediction results of two of the batteries are shown. This shows that the method can estimate the SOH of the battery using the current cycle data. The horizontal axis of the figure is the number of cycles, and the vertical axis is the SOH value. The black line in the figure is the measured value of the battery SOH, the green line is the training data SOH value, and the red line is the predicted SOH value.

[0044] In order to show the advancedness of this method, five health factors were selected by using the Pearson correlation coefficient method for comparison with the health factor selection method. The health factors selected for each battery are shown in Table 3. The health factors were input into the LSTM model with the same hyperparameter configuration, and the estimated SOH was compared for accuracy. The comparison results are shown in Table 4.

[0045] Table 1 Battery health factors

[0046]

[0047]

[0048] Table 2 Health factors selected by RF algorithm

[0049]

[0050] Table 3 Pearson correlation coefficients for selected health factors

[0051]

[0052] Table 4 Error comparison between the two methods

[0053]

[0054] The present invention uses the Random Forest (RF) algorithm to screen the health factors of electric vehicle batteries, which can more accurately estimate the SOH of the battery. The present invention can effectively identify key factors closely related to the battery state of health (SOH) from a large number of health factors. Compared with the traditional Pearson correlation coefficient method, it can significantly improve the accuracy of health factor selection and the accuracy of the prediction model.

Claims

1. A health factor selection method for lithium-ion battery health status estimation, characterized in that: This method is based on 34 health factors extracted during the battery charging and discharging process. Based on the long short-term memory neural network, it is proposed to use the random forest algorithm to select health factors, and input the selected health factors into the LSTM neural network as a model to achieve high-precision prediction of the battery health status. The ratio of the available discharge capacity to the nominal capacity is the battery health status. The specific implementation methods are as follows: Step 1: Based on the battery charging and discharging data, 34 health factors are extracted, each of which is a vector containing a time series; Step 2: Normalize the 34 health factor sequences respectively and map the values ​​to the interval [0, 1]; Among them, x i is the value of the ith health factor, min(x) is the minimum value in the health factor vector, and max(x) is the maximum value in the health factor vector; D i Take the value of the normalized data; Step 3: Set the number of decision trees to 100, and use the sampling method with replacement to build and train the decision tree regression algorithm. For each decision tree, the sample extracted is called the in-bag data. Since it is a sampling method with replacement, there are samples that are not extracted, which are called out-of-bag data. Step 4: Perform out-of-bag error analysis on all health factor data; input the out-of-bag data containing the health factor sequence corresponding to the decision tree in step 3 into the decision tree for SOH prediction, and calculate the error between the predicted value and the out-of-bag sample SOH test value, which is given by formula (3); Among them, OOB Error t represents the out-of-bag error of the health factor relative to the t-th tree; k t is the number of health factor samples in the out-of-bag data of the t-th tree, yi is the SOH test value corresponding to the ith sample of the health factor, is the i-th predicted value of the health factor by the t-th tree; Since the same health factor may be out-of-bag data for n decision trees, formula (2) is used to calculate the out-of-bag error of n trees. These out-of-bag errors are averaged to obtain the average out-of-bag error of the health factor. The calculation is given by formula (3): Where n is the number of decision trees, OOB Error t is the out-of-bag error of the health factor relative to the t-th tree; Step 5: Randomly rearrange the values ​​of the health factor vector to disrupt the order of the values ​​in the vector; then repeat step 4 to obtain a new out-of-bag error. Step 6: For each health factor, calculate the difference IMP of the out-of-bag errors of the two arrangements. The calculation is given by formula (5): IMP=E OOB,permuted -E OOB,origin (5) Among them, E OOB,origin is the average out-of-bag error of n trees on the original dataset, E j，OOB,permuted is the average out-of-bag error after the health factor is rearranged; Step 7: The IMP of each health factor is the importance score of the health factor. The higher the score, the more relevant it is to SOH. Select the five health factors with the highest IMP based on the score. Step 8, SOH prediction output; input the selected health factor as the model into the LSTM neural network, and use the battery SOH as the output to train the model; for new test data, input it into the trained model to obtain the estimated value of the battery SOH.

2. The health factor selection method for lithium-ion battery health status estimation according to claim 1, characterized in that: In step eight, the hyperparameters of the LSTM neural network model used to predict battery SOH are 512 neurons, 64 neurons in the fully connected layer, 200 iterations, 16 batch sizes, and 0.001 learning rates.

Images