Battery model construction method and battery deterioration prediction device

Abstract

To solve the above problem, the battery model construction method of the present application includes: a step (ST2) of constructing a battery model using learning data; a step (ST3-ST4) of evaluating a prediction error between a measured value of SOH and a predicted value based on the battery model for each sample battery, and determining the presence or absence of a deviation in the prediction error inherent in each sample battery; a step (ST6-ST7) of constructing a first error prediction model that associates an explanatory variable defined based on a use history parameter with a target variable corresponding to the predicted value of the average of the prediction error, and determining the presence or absence of a first correlation between the measured value of the average of the prediction error obtained in the step (ST3-ST4) and a predicted value based on the first error prediction model; and a step (ST7) of reconstructing the battery model when the presence of a deviation is determined in the step (ST3-ST4) and the presence of the first correlation is determined in the step (ST6-ST7).

Description

Technical Field

[0001] This invention relates to a battery model construction method and a battery degradation prediction device. More specifically, it relates to a battery model construction method for constructing a battery model to calculate predicted values ​​of battery degradation indicators, and a battery degradation prediction device for calculating predicted values ​​of battery degradation indicators using the battery model constructed by the battery model construction method. Background Technology

[0002] Secondary batteries used in electric or hybrid vehicles are characterized by deterioration with use. Once deteriorated, secondary batteries cannot perform at their full potential; therefore, appropriate measures are needed to slow down the progression of deterioration. Furthermore, to implement such measures, the degree of deterioration needs to be accurately estimated. For example, Patent Document 1 discloses a technique for estimating the degree of deterioration of a secondary battery based on usage history data (e.g., evolution of current, voltage, and temperature, or the elapsed lifespan of the secondary battery). Additionally, the deterioration characteristics of a secondary battery are also related to its manufacturing data; therefore, the degree of deterioration is sometimes estimated based on both usage history data and manufacturing data, as described above.

[0003] In addition, among battery models that estimate the degree of degradation of secondary batteries based on usage history data and manufacturing data, besides linear regression models, a regression model using neural networks and gradient boosting trees (hereinafter referred to as "GBDT") has also been proposed.

[0004] [Previous Technical Documents]

[0005] (Patent Documents)

[0006] Patent Document 1: International Publication No. 2020 / 044713 Summary of the Invention

[0007] [The problem the invention aims to solve]

[0008] However, linear regression models often struggle to achieve higher prediction accuracy than regression models based on neural networks or Gaussian DBT. In contrast, regression models based on neural networks or Gaussian DBT have more complex structures than linear regression models, making it difficult to determine the impact of factors such as historical or manufacturing-time data on the predicted values. Therefore, in regression models based on neural networks or Gaussian DBT, it is challenging to tune the existing model to improve prediction accuracy.

[0009] The purpose of this invention is to provide a method for constructing a battery model that can predict battery degradation with high accuracy, and a battery degradation prediction device.

[0010] [Technical means to solve the problem]

[0011] (1) The battery model construction method of the present invention constructs a battery model, wherein the battery model associates the values ​​of a plurality of first input parameters and a plurality of second input parameters related to the degradation index of the battery with the predicted value of the aforementioned degradation index. The battery model construction method is characterized by comprising: a data acquisition step, which acquires the aforementioned first and second input parameters and data related to the measured value of the aforementioned degradation index using a plurality of sample batteries; a construction step, which uses at least a portion of the data acquired in the aforementioned data acquisition step as learning data to construct the aforementioned battery model; and an error tendency determination step, which evaluates the measured value of the aforementioned degradation index and the value of the aforementioned degradation index based on the aforementioned battery model for each of the aforementioned sample batteries. The prediction error between the predicted values ​​is determined, and the presence or absence of the bias inherent in the prediction error in each of the aforementioned sample batteries is determined; a first correlation determination step is to construct a first error prediction model by using the aforementioned learning data, which associates the explanatory variable defined based on the aforementioned first input parameter with the target variable corresponding to the predicted value of the aforementioned prediction error, and to determine whether there is a first correlation between the measured value of the aforementioned prediction error obtained in the aforementioned error tendency determination step and the predicted value of the aforementioned prediction error based on the aforementioned first error prediction model; and a first reconstruction step is to reconstruct the aforementioned battery model when it is determined in the aforementioned error tendency determination step that the aforementioned bias exists and in the aforementioned first correlation determination step that the aforementioned first correlation exists.

[0012] (2) Preferably, the aforementioned battery model construction method further includes: a second correlation determination step, which uses the aforementioned learning data to construct a second error prediction model that associates the explanatory variable defined based on the aforementioned second input parameter with the target variable that corresponds to the predicted value of the aforementioned prediction error, and determines whether there is a second correlation between the measured value of the aforementioned prediction error obtained in the aforementioned error tendency determination step and the predicted value of the aforementioned prediction error based on the aforementioned second error prediction model; and a second reconstruction step, which reconstructs the aforementioned battery model when it is determined in the aforementioned error tendency determination step that there is the aforementioned deviation and in the aforementioned second correlation determination step that there is the aforementioned second correlation.

[0013] (3) Preferably, the first input parameter is a usage history parameter defined based on the time-series data of the current, voltage and temperature of the battery, and the second input parameter is a manufacturing history parameter defined based on the manufacturing data of the battery.

[0014] (4) Preferably, in the first reconstruction step, the battery model is reconstructed by using the first error prediction model.

[0015] (5) Preferably, in the second reconstruction step, the battery model is reconstructed by using the second error prediction model.

[0016] (6) Preferably, the first and second input parameters and the time-series data related to the measured values ​​of the aforementioned degradation index obtained in the aforementioned data acquisition step are divided into the aforementioned learning data within the specified learning period and the verification data within the verification period after the aforementioned learning period. In the aforementioned error tendency determination step, the presence or absence of the aforementioned deviation is determined based on the evaluation results of each of the aforementioned sample batteries based on the distribution of the aforementioned prediction error within the aforementioned learning period and the evaluation results related to the aforementioned prediction error within the aforementioned learning period and the aforementioned prediction error within the aforementioned verification period.

[0017] (7) Preferably, the aforementioned battery model construction method further includes a third reconstruction step, wherein when the aforementioned error tendency determination step determines that the aforementioned deviation exists and the aforementioned first and second correlation determination steps determine that neither of the aforementioned first and second correlations exists, the aforementioned battery model is reconstructed by adding an offset term of a constant value corresponding to the cumulative usage time of the aforementioned battery.

[0018] (8) The battery degradation prediction device of the present invention is characterized in that it includes: an input parameter value acquisition unit, which acquires the values ​​of a plurality of first input parameters and a plurality of second input parameters that are related to the degradation index of the battery; and a model prediction unit, which inputs the values ​​of the first and second input parameters acquired by the aforementioned input parameter value acquisition unit into a battery model constructed using the battery model construction method described in any one of (1) to (7), thereby calculating the predicted value of the aforementioned degradation index.

[0019] (The effect of the invention)

[0020] (1) The battery model construction method of the present invention includes: a construction step, constructing a battery model using learning data obtained by using multiple sample batteries; an error tendency determination step, evaluating the prediction error between the measured value of a degradation index and the predicted value of a degradation index based on the battery model for each sample battery, and determining whether there is a bias in the prediction error inherent in each sample battery; a first correlation determination step, constructing a first error prediction model that associates an explanatory variable defined based on a first input parameter with a target variable corresponding to the predicted value of the prediction error, and determining whether there is a first correlation between the measured value of the prediction error and the predicted value of the prediction error based on the first error prediction model; and a first reconstruction step, reconstructing the battery model when it is determined in the error tendency determination step that there is a bias in the prediction error inherent in each sample battery and in the first correlation determination step that there is a first correlation. Here, when it is determined that there is a bias in the prediction error inherent in each sample battery and that there is a first correlation, it can be said that the main reason for the bias in the prediction error of each sample battery is the first input parameter. In this invention, by reconstructing the battery model through such an error tendency determination step and a first correlation determination step, the main causes of the prediction error can be determined while the battery model is reconstructed to improve the prediction accuracy.

[0021] (2) The battery model construction method of the present invention further includes: a second correlation determination step, which constructs a second error prediction model that associates an explanatory variable defined based on a second input parameter with a target variable corresponding to a prediction error, and determines whether there is a second correlation between the measured value of the prediction error obtained in the error tendency determination step and the predicted value of the prediction error based on the second error prediction model; and a second reconstruction step, which reconstructs the battery model when it is determined in the error tendency determination step that there is an inherent bias in the prediction error of each sample battery and in the second correlation determination step that there is a second correlation. Here, when it is determined that there is an inherent bias in the prediction error of each sample battery and that there is a second correlation, it can be said that the main reason for the deviation in the prediction error of each sample battery is the second input parameter. In the present invention, by reconstructing the battery model through such an error tendency determination step and a second correlation determination step, the main reason for the deviation in the prediction error can be determined while the battery model is reconstructed to improve the prediction accuracy.

[0022] (3) In this invention, the first input parameter is defined as a usage history parameter based on time-series data of the battery's current, voltage, and temperature, and the second input parameter is defined as a manufacturing history parameter based on data from the battery's manufacturing process. Therefore, according to this invention, the main causes of prediction error deviation for each sample battery can be determined using the usage history parameter and the manufacturing history parameter. Thus, the battery model can be reconstructed to improve prediction accuracy while appropriately determining the main causes of prediction error deviation.

[0023] (4) In this invention, in the first reconstruction step, if it is determined that there is an inherent bias in each sample cell in the prediction error and a first correlation exists, the battery model is reconstructed by using a first error prediction model. Thus, the battery model can be reconstructed in a way that reduces the bias of the prediction error for each sample cell, i.e., in a way that improves prediction accuracy.

[0024] (5) In this invention, in the second reconstruction step, if it is determined that there is an inherent bias in each sample cell in the prediction error and a second correlation exists, the battery model is reconstructed by using a second error prediction model. Thus, the battery model can be reconstructed in a way that reduces the bias of the prediction error for each sample cell, i.e., in a way that improves prediction accuracy.

[0025] (6) In this invention, in the error tendency determination step, the evaluation results of each sample battery based on the distribution of prediction errors during the learning period and the evaluation results of the correlation between the prediction errors during the learning period and the prediction errors during the verification period are used to determine whether the prediction errors have inherent biases in each sample battery. Therefore, it is possible to determine with good accuracy whether the prediction errors have inherent biases in each sample battery, and thus, a battery model with high prediction accuracy can be constructed.

[0026] (7) The battery model construction method of the present invention further includes a third reconstruction step. This third reconstruction step involves reconstructing the battery model by adding an offset term corresponding to a constant value of the battery's cumulative usage time when the prediction error is determined to have a bias in the error tendency determination step and both the first and second correlations are determined to be absent in the first and second correlation determination steps. Here, when the prediction error is determined to have a bias and both the first and second correlations are determined to be absent, it can be said that the main reason for the prediction error bias is that the battery model cannot pick up some input parameter other than the first and second input parameters. In the present invention, in this case, the battery model is reconstructed by adding an offset term, thereby improving the prediction accuracy of the battery model without adding new input parameters.

[0027] (8) The battery degradation prediction device of the present invention includes: an input parameter value acquisition unit that acquires the values ​​of a plurality of first and second input parameters; and a model prediction unit that inputs the values ​​of the first and second input parameters acquired by the input parameter value acquisition unit into a battery model constructed using the battery model construction method described above, thereby calculating a predicted value of the battery degradation index. According to the present invention, the degradation evolution of a battery during use can be predicted with high accuracy. Attached Figure Description

[0028] Figure 1 This is a diagram illustrating the structure of a battery degradation prediction device according to an embodiment of the present invention.

[0029] Figure 2 It is a schematic diagram illustrating multiple constructions using historical parameters.

[0030] Figure 3 It is a flowchart illustrating the specific process of constructing the battery model.

[0031] Figure 4 This is a flowchart illustrating the specific process of error tendency determination and processing.

[0032] Figure 5 It is a schematic diagram illustrating the prediction error of the State of Health (SOH).

[0033] Figure 6 This is an example of an evaluation result showing the distribution of prediction errors for SOH during the learning period.

[0034] Figure 7 The graph (R) is obtained by plotting the average prediction error for each sample battery on a plane with the average prediction error during the learning period as the vertical axis and the average prediction error during the validation period as the horizontal axis. 2 (≒0.64).

[0035] Figure 8 It is a flowchart illustrating the specific process of the first relevant determination and processing.

[0036] Figure 9 It is a schematic diagram illustrating the construction of the first error prediction model and the first error learning data.

[0037] Figure 10 This is a flowchart illustrating the specific process of the second related determination and processing.

[0038] Figure 11 It is a schematic diagram illustrating the construction of the second error prediction model and the second error learning data. Detailed Implementation

[0039] Hereinafter, a battery degradation prediction apparatus according to an embodiment of the present invention and a method for constructing a battery model used in the battery degradation prediction apparatus will be described with reference to the drawings.

[0040] Figure 1 This is a diagram illustrating the structure of the battery degradation prediction device 1 according to this embodiment.

[0041] The battery degradation prediction device 1 predicts the degree of degradation of the battery 2 based on time-series data of the current, voltage, and temperature of the battery 2, as well as manufacturing data of the battery 2. The following description pertains to the case where the battery degradation prediction device 1 is installed in an electric vehicle (not shown) that uses the power of the battery 2 for operation, and predicts the degree of degradation of the battery 2 in that electric vehicle; however, the present invention is not limited thereto. All or part of the components of the battery degradation prediction device 1 may also be constituted by a server communicatively connected to the electric vehicle.

[0042] Battery 2 is a secondary battery capable of both discharging (converting chemical energy into electrical energy) and charging (converting electrical energy into chemical energy). Hereinafter, a lithium-ion battery, which uses the movement of lithium ions between electrodes for charging and discharging, will be described as battery 2; however, the invention is not limited to this. Battery 2 is connected to an electrical load (not shown), such as an inverter or a drive motor, and charging and discharging occur between the battery and the electrical load.

[0043] The battery degradation prediction device 1 is a computer composed of hardware such as a central processing unit (CPU), auxiliary storage devices such as hard disk drives (HDDs) or solid-state drives (SSDs) that store various programs, and main storage devices such as random access memory (RAM) that store data temporarily needed by the processing unit when executing programs. In the battery degradation prediction device 1, this hardware structure enables the various functions of the data acquisition unit 11, the input parameter value acquisition unit 13, and the model prediction unit 14.

[0044] The data acquisition unit 11 acquires timing data of the current, voltage, and temperature of the battery 2 based on the output from the battery sensor (not shown) installed on the battery 2.

[0045] The input parameter value acquisition unit 13 acquires the values ​​of multiple input parameters that are related to the degradation index of the battery 2 based on the time-series data of current, voltage and temperature acquired by the data acquisition unit 11 and the manufacturing data of the battery 2 stored in the storage medium (not shown), and inputs these input parameter values ​​to the model prediction unit 14.

[0046] like Figure 1 As shown, the multiple input parameters defined in the input parameter value acquisition unit 13 include: multiple usage history parameters defined based on time-series data of current, voltage, and temperature acquired by the data acquisition unit 11; and multiple manufacturing history parameters defined based on manufacturing data of the battery 2. The input parameter value acquisition unit 13 calculates the values ​​of the aforementioned usage history parameters at a predetermined period based on the time-series data of current, voltage, and temperature acquired by the data acquisition unit 11, and inputs these usage history parameters and manufacturing history parameters to the model prediction unit 14. Hereinafter, the case where the calculation period for the usage history parameters in the input parameter value acquisition unit 13 is set to two weeks will be described, but the present invention is not limited thereto.

[0047] Here, the manufacturing history parameters include the material property values ​​of the various materials used in the manufacture of battery 2, the manufacturing process data values ​​that characterize the manufacturing process of battery 2, and the inspection values ​​for various checks on battery 2. That is, the manufacturing history parameters are different from the usage history parameters; they are essentially fixed values ​​that do not change over time.

[0048] Figure 2 This is a schematic diagram illustrating the construction of multiple historical parameters defined in the input parameter value acquisition unit 13. For example... Figure 2 As shown, the multiple usage history parameters include multiple voltage factor parameters, multiple temperature factor parameters, and multiple current factor parameters.

[0049] The voltage factor parameter is a parameter that takes the voltage of battery 2 as a factor. In other words, the voltage factor parameter is the parameter that has the highest correlation with the voltage among the current, voltage, and temperature of battery 2. In this embodiment, the voltage factor parameter is defined as the cumulative value of the dwell time within a specified range of the state of charge (SOC) that is approximately proportional to the open voltage of battery 2. More specifically, the first SOC cumulative time is the cumulative value of the dwell time within a specified range of battery 2 when the SOC is in the range of 0 to 10%, the second SOC cumulative time is the cumulative value of the dwell time within a specified range of battery 2 when the SOC is in the range of 10 to 20%, the third SOC cumulative time is the cumulative value of the dwell time within a specified range of battery 2 when the SOC is in the range of 20 to 30%, the fourth SOC cumulative time is the cumulative value of the dwell time within a specified range of battery 2 when the SOC is in the range of 30 to 40%, and the fifth SOC cumulative time is the cumulative value of the dwell time within a specified range of battery 2 when the SOC is in the range of 40 to 50%. The sixth SOC cumulative time is the cumulative value of the specified interval of time that the SOC of battery 2 remains in the range of 50-60%, the seventh SOC cumulative time is the cumulative value of the specified interval of time that the SOC of battery 2 remains in the range of 60-70%, the eighth SOC cumulative time is the cumulative value of the specified interval of time that the SOC of battery 2 remains in the range of 70-80%, the ninth SOC cumulative time is the cumulative value of the specified interval of time that the SOC of battery 2 remains in the range of 80-90%, and the tenth SOC cumulative time is the cumulative value of the specified interval of time that the SOC of battery 2 remains in the range of 90-100%.

[0050] The temperature factor parameter is a parameter with the temperature of battery 2 as a factor. In other words, the temperature factor parameter is the parameter most correlated with the temperature among the current, voltage, and temperature of battery 2. In this embodiment, the temperature factor parameter is defined as the cumulative value of the residence time within a specified interval of the temperature range of battery 2 when the operating temperature range of battery 2 is divided into ten equal parts. More specifically, the first temperature cumulative time is the cumulative value of the specified interval of the residence time of the temperature of battery 2 within the first temperature range, the second temperature cumulative time is the cumulative value of the specified interval of the residence time of the temperature of battery 2 within the second temperature range which is higher than the first temperature range, the third temperature cumulative time is the cumulative value of the specified interval of the residence time of the temperature of battery 2 within the third temperature range which is higher than the second temperature range, the fourth temperature cumulative time is the cumulative value of the specified interval of the residence time of the temperature of battery 2 within the fourth temperature range which is higher than the third temperature range, the fifth temperature cumulative time is the cumulative value of the specified interval of the residence time of the temperature of battery 2 within the fifth temperature range which is higher than the fourth temperature range, and so on. The sixth temperature cumulative time is the cumulative value of the specified interval of time that the temperature of battery 2 remains in the sixth temperature range, which is higher than the fifth temperature range. The seventh temperature cumulative time is the cumulative value of the specified interval of time that the temperature of battery 2 remains in the seventh temperature range, which is higher than the sixth temperature range. The eighth temperature cumulative time is the cumulative value of the specified interval of time that the temperature of battery 2 remains in the eighth temperature range, which is higher than the seventh temperature range. The ninth temperature cumulative time is the cumulative value of the specified interval of time that the temperature of battery 2 remains in the ninth temperature range, which is higher than the eighth temperature range. The tenth temperature cumulative time is the cumulative value of the specified interval of time that the temperature of battery 2 remains in the tenth temperature range, which is higher than the ninth temperature range.

[0051] The current factor parameter is a parameter with the current of battery 2 as a factor. In other words, the current factor parameter is the parameter that has the highest correlation with the current among the current, voltage, and temperature of battery 2. In this embodiment, the cumulative value of the product of the charging current and time of battery 2 over a specified interval, i.e., the cumulative charging current capacity, and the cumulative value of the product of the discharging current and time of battery 2 over a specified interval, i.e., the cumulative discharging current capacity, are defined as the current factor parameter. As described above, this embodiment focuses on the case where the cumulative value of the product of current and time is defined as the current factor parameter, but the present invention is not limited to this. For example, the number of charge-discharge cycles in battery 2, which corresponds to the number of times charging and discharging are switched, can also be defined as the current factor parameter.

[0052] return Figure 1 The input parameter value acquisition unit 13 calculates the values ​​of multiple usage history parameters, which are constructed as described above based on the time-series data sent from the data acquisition unit 11, at a predetermined period, and inputs these usage history parameters and manufacturing history parameters to the model prediction unit 14.

[0053] The model prediction unit 14 has a battery model that uses multiple input parameters, consisting of the aforementioned multiple usage history parameters and manufacturing history parameters, as explanatory variables, and associates these explanatory variables with the predicted value of the degradation index of battery 2, i.e., the target variable. The predicted value of the degradation index of battery 2 is calculated by inputting the values ​​of the multiple input parameters obtained from the input parameter value acquisition unit 13 into the battery model. In this embodiment, the degradation index of battery 2 is described using the state of health (SOH), which represents the percentage of the full charge capacity at degradation when the initial full charge capacity [Ah] of battery 2 is set to 100%, but the present invention is not limited to this. Here, the battery model is used by referring to... Figure 3 The model is constructed using the battery model construction method described above.

[0054] Next, we will explain the method of using computers to build battery models.

[0055] Figure 3 This is a flowchart illustrating the specific process of the battery model construction method in this embodiment.

[0056] First, in step ST1, the battery model designer uses multiple (e.g., several hundred) sample batteries, each assigned an identification (ID), throughout the entire specified testing period (e.g., 66 weeks) to acquire time-series data of multiple usage history parameters and State of Health (SOH) measurements for each sample battery, as well as measurement values ​​of multiple manufacturing history parameters for each sample battery. This time-series data of usage history parameters, manufacturing history parameters, and SOH throughout the entire testing period (hereinafter, collectively referred to as "test data") is stored in a storage medium in a state associated with its respective ID. Furthermore, the test data acquired in step ST1 throughout the entire testing period is divided into learning data belonging to the specified learning period (e.g., week 2 to week 60) and verification data belonging to the verification period following the learning period (e.g., week 62 to week 66).

[0057] Next, in step ST2, the designer uses the learning data obtained in step ST1 to construct a battery model based on known model building methods. This model uses multiple historical usage parameters and multiple historical manufacturing parameters as explanatory variables, and the predicted SOH value as the target variable. Furthermore, this battery model can be constructed based on a linear regression model, a neural network, or GBDT, etc.

[0058] Next, in step ST3, the designer performs error bias determination processing and proceeds to step ST4. The error bias determination processing evaluates the prediction error between the measured SOH value and the predicted SOH value based on the battery model for each sample battery, and determines whether there is a bias of the inherent prediction error in each sample battery.

[0059] Figure 4 This is a flowchart illustrating the specific process of error tendency determination and processing.

[0060] First, in step ST21, the designer calculates the prediction error of SOH during the learning period for each sample battery using learning data and a battery model. Here, the prediction error of SOH during the learning period is calculated by subtracting the predicted value of SOH from the measured values ​​of SOH contained in the learning data. The predicted value of SOH is obtained by inputting the values ​​of multiple usage history parameters and multiple manufacturing history parameters contained in the learning data into the battery model (see [reference]). Figure 5 ).

[0061] Next, in step ST22, the designer evaluates the distribution of the prediction error for SOH during the learning period for each sample cell using the prediction error calculation results from step ST21. More specifically, the designer uses the calculation results from step ST21, such as... Figure 6 As shown, for each sample cell, the sample mean, partial variance, standard error, and 95% confidence interval of the population mean prediction error for SOH are calculated. For each sample cell, the presence of significant bias in the SOH prediction error during the learning period is evaluated. Here, the presence of significant bias in the SOH prediction error can be determined by evaluating the position of the upper and lower limits of the calculated 95% confidence interval of the population mean relative to 0. More specifically, if both the upper and lower limits of the 95% confidence interval of the population mean are below or above 0, it can be determined that there is a significant bias in the SOH prediction error with a 95% confidence level.

[0062] Next, in step ST23, the designer refers to the evaluation results in step ST22 to determine whether there is a significant deviation in the prediction error of SOH during the learning period for a specified proportion (e.g., 90%) of all sample batteries.

[0063] If the decision result in step ST23 is negative (NO), the designer determines that there is no inherent bias in the prediction error of SOH based on the battery model for each sample battery (step ST24), and then proceeds to... Figure 3Step ST4. Alternatively, if the result of step ST23 is yes (YES), the designer proceeds to step ST25.

[0064] In step ST25, the designer uses the calculation results from step ST21 to calculate the average prediction error of SOH during the learning period for each sample battery.

[0065] In step ST26, the designer calculates the average prediction error of SOH during the verification period for each sample battery using validation data and a battery model. Here, the prediction error of SOH during the verification period is calculated by subtracting the predicted SOH value from the measured SOH value contained in the validation data. The predicted SOH value is obtained by inputting the values ​​of multiple usage history parameters and multiple manufacturing history parameters contained in the validation data into the battery model. Furthermore, in step ST26, the average prediction error of SOH during the verification period is calculated for each sample battery using the calculation results of these prediction errors during the verification period.

[0066] In step ST27, the designer uses the calculation results from steps ST25 and ST26 to calculate a parameter representing the strength of the correlation between the average prediction error of SOH during the learning period and the average prediction error of SOH during the validation period thereafter. In this embodiment, a coefficient of determination (hereinafter referred to as "R") is used as the parameter representing the strength of the correlation between the two parameters. 2 The invention is described using the abbreviation "(this abbreviation), but is not limited thereto."

[0067] In step ST28, the designer, referring to the evaluation results in step ST27, determines whether the correlation strength between the average prediction error during the learning period and the average prediction error during the validation period is above a specified threshold. Using R as described above... 2 When the relevant strength is evaluated, this judgment value is set to, for example, 0.49.

[0068] If the determination result in step ST28 is negative, the designer determines that there is no inherent bias in the prediction error of SOH based on the battery model for each sample battery (refer to step ST24), and then proceeds to... Figure 3 Step ST4. Additionally, if the determination result in step ST28 is yes, the designer determines that there is an inherent bias in the SOH prediction error based on the battery model for each sample battery (refer to step ST29), and then proceeds to... Figure 3 Step ST4.

[0069] Figure 7This is a graph obtained by plotting the average prediction error for each sample battery on a plane with the average prediction error during the learning period as the vertical axis and the average prediction error during the validation period as the horizontal axis. Furthermore, in Figure 7 The figure in the middle shows when a strong correlation (R) is confirmed in the average of these prediction errors. 2 Examples of ≒0.64). For example... Figure 7 As shown, the stronger the correlation, the closer the distribution of the average prediction error is to the straight line with the specified slope shown by the dashed line.

[0070] As described above, in the error tendency determination process, the evaluation results of each sample cell based on the distribution of prediction error during the learning period (refer to step ST21) and the evaluation results of the correlation between the average prediction error during the learning period and the average prediction error during the verification period (refer to step ST27) are used to determine whether the prediction error of SOH has an inherent bias for each sample cell.

[0071] return Figure 3 In step ST4, the designer determines whether the result of the error tendency assessment process indicates an inherent bias in each sample cell within the SOH prediction error. If the result of step ST4 is negative, the designer determines that the fabricated cell model requires no improvement and terminates the process. Figure 3 The process is shown below. Additionally, if the determination result in step ST4 is yes, the designer proceeds to step ST5 to determine the main cause of the prediction error deviation.

[0072] Next, in step ST5, the designer performs a first correlation determination process to determine whether the deviation of the above prediction error is caused by the use of historical parameters, and then proceeds to step ST6.

[0073] Figure 8 It is a flowchart illustrating the specific process of the first relevant determination and processing.

[0074] First, in step ST31, the designer prepares the first error learning data for constructing the first error prediction model described later.

[0075] Figure 9 This is a schematic diagram illustrating the construction of the first error prediction model and the first error learning data. For example... Figure 9As shown, the first error prediction model is a regression model that uses multiple compressed usage history parameters defined based on multiple usage history parameters as explanatory variables, the predicted value of the average prediction error of SOH as the target variable, and correlates these explanatory variables with the target variable. Therefore, in step ST31, first error learning data is generated by combining input data and output data. The input data consists of the values ​​of multiple compressed usage history parameters defined based on time-series data of multiple usage history parameters from the learning data obtained in step ST1 throughout the entire learning period. The output data consists of the measured value of the average prediction error of SOH during the learning period, calculated for each sample battery in step ST25. Here, each compressed usage history parameter is defined, for example, by the average or variance of each usage history parameter throughout the entire learning period, as defined above as a cumulative value over a specified interval.

[0076] return Figure 8 In step ST32, the designer uses the first error learning data prepared in step ST31 to construct a first error prediction model based on known model building methods. This model associates multiple compressed historical parameters (i.e., explanatory variables) with the predicted values ​​(i.e., the target variable) of the average prediction error. Furthermore, this first error prediction model can be constructed based on a linear regression model, a neural network, or GBDT, etc.

[0077] Next, in step ST33, the designer calculates a parameter (as described above, the coefficient of determination R in this embodiment) representing the strength of the first correlation between the measured value of the average prediction error of SOH contained in the first error learning data prepared in step ST31 and the predicted value of the average prediction error of SOH obtained by inputting the first error learning data into the first error prediction model constructed in step ST32. 2 ), and transferred to Figure 3 Step ST6.

[0078] return Figure 3 In step ST6, the designer refers to the evaluation results of the first correlation in step ST33 to determine whether the strength of the first correlation is above the specified first judgment value. This is achieved using the determination coefficient R as described above. 2When evaluating the correlation strength, the first judgment value is, for example, set to 0.49. Here, a first correlation strength greater than or equal to the first judgment value means that a first error prediction model, using multiple historical parameters as input, can predict the prediction error contained in the battery model, where these historical parameters are partly the same as the input to the battery model. Therefore, a first correlation strength greater than or equal to the first judgment value means that the main cause of the prediction error deviation is the use of historical parameters, thus implying that there is room for reconstructing the battery model by re-examining the input system using historical parameters within the input system for the battery model, thereby eliminating the deviation in the prediction error of SOH. Conversely, a first correlation strength less than the first judgment value means that the main cause of the prediction error deviation is not the use of historical parameters, thus implying that even re-examining the input system using historical parameters within the input system for the battery model cannot eliminate the prediction error deviation.

[0079] Therefore, if the result of the determination in step ST6 is negative, the designer proceeds to step ST8; if the result of the determination in step ST6 is positive, the designer proceeds to step ST7.

[0080] In step ST7, the designer, based on the determination that there is a deviation in the prediction error and that the first correlation is above the first determination value, reconstructs the battery model by re-examining the input system using historical parameters in the battery model's input system. Then, the process returns to step ST3. More specifically, a new battery model is reconstructed by combining the battery model before reconstruction with the first error prediction model constructed in the first correlation determination process of step ST5. That is, the battery model is reconstructed by adding the output of the battery model before reconstruction to the output of the first error prediction model.

[0081] Next, in step ST8, the designer performs a second correlation determination process to determine whether the deviation of the above prediction error is caused by manufacturing history parameters, and then proceeds to step ST9.

[0082] Figure 10 This is a flowchart illustrating the specific process of the second related determination and processing.

[0083] First, in step ST41, the designer prepares second error learning data for constructing the second error prediction model described later.

[0084] Figure 11 This is a schematic diagram illustrating the construction of the second error prediction model and the second error learning data. For example... Figure 11As shown, the first error prediction model is a regression model that uses multiple manufacturing history parameters as explanatory variables, the predicted value of the average prediction error of SOH as the target variable, and correlates these explanatory variables with the target variable. This differs from the input of the first error prediction model described above. Therefore, in step ST41, second error learning data is generated by combining the input data and output data. The input data consists of the values ​​of multiple manufacturing history parameters from the learning data obtained in step ST1, and the output data consists of the measured value of the average prediction error of SOH during the learning period, calculated for each sample battery in step ST25.

[0085] return Figure 10 In step ST42, the designer uses the second error learning data prepared in step ST41 to construct a second error prediction model that associates multiple manufacturing history parameters (explanatory variables) with the predicted value of the average prediction error (target variable) based on known model building methods. Furthermore, this second error prediction model can be constructed based on a linear regression model, a neural network, or GBDT, etc.

[0086] Next, in step ST43, the designer calculates a parameter (as described above, the coefficient of determination R in this embodiment) representing the strength of a second correlation between the measured value of the average prediction error of SOH contained in the second error learning data prepared in step ST41 and the predicted value of the average prediction error of SOH obtained by inputting the second error learning data into the second error prediction model constructed in step ST42. 2 ), and transferred to Figure 3 Step ST9.

[0087] return Figure 3 In step ST9, the designer refers to the evaluation results of the second correlation in step ST43 to determine whether the strength of the second correlation is above the specified second judgment value. This is achieved using the determination coefficient R as described above. 2When evaluating the strength of the second correlation, the second judgment value is, for example, set to 0.49. Here, a second correlation strength greater than or equal to the second judgment value means that a second error prediction model, using multiple historical usage parameters as input, can predict the prediction error contained in the battery model, where these parameters are identical to a portion of the input to the battery model. Therefore, a second correlation strength greater than or equal to the second judgment value means that the main cause of the prediction error deviation is the manufacturing historical parameters, thus implying that there is room for reconstructing the battery model by re-examining the input system for the manufacturing historical parameters in the input system for the battery model, thereby eliminating the deviation in the prediction error of SOH. Conversely, a second correlation strength less than the second judgment value means that the main cause of the prediction error deviation is not the manufacturing historical parameters, thus implying that even re-examining the input system for the manufacturing historical parameters in the input system for the battery model cannot eliminate the prediction error deviation.

[0088] Therefore, if the designer's determination result in step ST9 is yes, the process proceeds to step ST10. In step ST10, the designer, based on the determination that there is a deviation in the prediction error and that the second correlation is above the second determination value, reconstructs the battery model by re-examining the input system of the manufacturing history parameters in the battery model's input system, and then returns to step ST3. More specifically, a new battery model is reconstructed by combining the battery model before reconstruction with the second error prediction model constructed in the second correlation determination process in step ST8. That is, the battery model is reconstructed by adding the output of the battery model before reconstruction with the output of the second error prediction model.

[0089] Furthermore, if the determination result in step ST9 is negative, i.e., if it is determined that there is a deviation in the prediction error and the strengths of the first and second correlations are both less than the first and second determination values, the designer proceeds to step ST11. At this point, it can be said that the main reason for the deviation in the prediction error is some input parameter that cannot be picked up in the battery model, other than the usage history parameters and manufacturing history parameters. Therefore, in step ST11, the designer reconstructs the battery model by adding an offset term to the battery model before reconstruction, the offset of which corresponds to a constant value corresponding to the cumulative usage time of the battery. Afterwards, the process ends. Figure 3 The processing is shown below. Here, the constant value corresponding to the cumulative usage time can be, for example, the average prediction error of SOH calculated for each sample cell in steps ST25 and ST26 relative to the average of all sample cells.

[0090] The battery model construction method and battery degradation prediction device 1 according to this embodiment achieve the following effects.

[0091] (1) The battery model construction method of this embodiment includes: a construction step ST1, which constructs a battery model using learning data obtained by using multiple sample batteries; error tendency determination steps ST3 to ST4, which evaluate the prediction error between the measured value of SOH and the predicted value of SOH based on the battery model for each sample battery, and determine whether there is a deviation of the inherent prediction error in each sample battery; first correlation determination steps ST5 to ST6, which constructs a first error prediction model that uses compressed historical parameters generated by compressing historical parameters over time as explanatory variables and the predicted value of the average prediction error of SOH as the target variable, and determines whether the strength of the first correlation between the measured value of the average prediction error and the predicted value of the average prediction error based on the first error prediction model is above a first determination value; and a first reconstruction step ST7, which reconstructs the battery model when it is determined in the error tendency determination steps ST3 to ST4 that there is a deviation of the inherent prediction error in each sample battery and in the first correlation determination steps ST5 to ST6 that the strength of the first correlation is above the first determination value. Here, if it is determined that there is an inherent bias in each sample cell within the prediction error and the strength of the first correlation is above the first determination value, it can be said that the main reason for the bias in the prediction error of each sample cell is the use of historical parameters. In this embodiment, by reconstructing the battery model through these error tendency determination steps ST3 to ST4 and the first correlation determination steps ST5 to ST6, the main reason for the bias in the prediction error can be determined while reconstructing the battery model to improve the prediction accuracy.

[0092] (2) The battery model construction method of this embodiment further includes: a second correlation determination step ST8-ST9, which constructs a second error prediction model that uses manufacturing history parameters as explanatory variables and the predicted value of the average prediction error as the target variable, and determines whether the strength of the second correlation between the measured value of the average prediction error obtained in the error tendency determination step ST3-ST4 and the predicted value of the average prediction error based on the second error prediction model is greater than or equal to a second determination value; and a second reconstruction step ST10, which reconstructs the battery model when it is determined in the error tendency determination step ST3-ST4 that there is an inherent deviation in the prediction error of each sample battery and in the second correlation determination step ST8-ST9 that the strength of the second correlation is greater than or equal to the second determination value. Here, when it is determined that there is an inherent deviation in the prediction error of each sample battery and the strength of the second correlation is greater than or equal to the second determination value, it can be said that the main reason for the deviation in the prediction error of each sample battery is the manufacturing history parameters. In this embodiment, by reconstructing the battery model through such error tendency determination steps ST3-ST4 and second correlation determination steps ST8-ST9, the main reason for the deviation in the prediction error can be determined while the battery model is reconstructed to improve the prediction accuracy.

[0093] (3) In this embodiment, in the first correlation determination steps ST5 to ST6, it is determined whether the deviation of the prediction error is caused by the use of historical parameters, and in the second correlation determination steps ST8 to ST9, it is determined whether the deviation of the prediction error is caused by manufacturing historical parameters. Thus, according to this embodiment, the main causes of the deviation of the prediction error for each sample battery can be determined by using historical parameters and manufacturing historical parameters. Therefore, the battery model can be reconstructed to improve the prediction accuracy while appropriately determining the main causes of the deviation of the prediction error.

[0094] (4) In the first reconstruction step ST7, if it is determined that there is an inherent bias in each sample cell in the prediction error and the strength of the first correlation is above the first determination value, the battery model is reconstructed by using the first error prediction model. Thus, the battery model can be reconstructed in a way that reduces the bias of the prediction error of each sample cell, that is, in a way that improves the prediction accuracy.

[0095] (5) In the second reconstruction steps ST8 to ST9, if it is determined that there is an inherent bias in each sample cell in the prediction error and the strength of the second correlation is determined to be above the second determination value, the battery model is reconstructed by using the second error prediction model. Thus, the battery model can be reconstructed in a way that reduces the bias of the prediction error of each sample cell, that is, in a way that improves the prediction accuracy.

[0096] (6) In the error tendency determination steps ST3 to ST4, based on the evaluation results of each sample battery of the prediction error distribution during the learning period, and the evaluation results of the correlation between the average prediction error during the learning period and the average prediction error during the validation period, it is determined whether there is an inherent bias in the prediction error of each sample battery. Thus, it is possible to determine with good accuracy whether there is an inherent bias in the prediction error of each sample battery, so the result can construct a battery model with high prediction accuracy.

[0097] (7) The battery model construction method further includes a third reconstruction step ST11. ST11 is performed when, in error tendency determination steps ST3-ST4, a deviation is determined in the prediction error, and in the first and second correlation determination steps ST5-ST6 and ST8-ST9, both the first and second correlations are determined to be less than the first and second determination values. In this case, the battery model is reconstructed by adding an offset term corresponding to the cumulative usage time of the battery to the battery model before reconstruction. Here, when a deviation is determined in the prediction error and the strength of both the first and second correlations is determined to be less than the first and second determination values, it can be said that the main reason for the deviation in the prediction error is that the battery model cannot pick up certain input parameters other than usage history parameters and manufacturing history parameters. In this embodiment, in this situation, the battery model is reconstructed by adding an offset term to the battery model before reconstruction. Therefore, the prediction accuracy of the battery model can be improved without adding new input parameters.

[0098] (8) The battery degradation prediction device 1 of this embodiment includes: an input parameter value acquisition unit 13, which acquires the values ​​of multiple usage history parameters and multiple manufacturing history parameters; and a model prediction unit 14, which inputs the values ​​of the usage history parameters and manufacturing history parameters acquired by the input parameter value acquisition unit 13 into a battery model constructed using the battery model construction method described above, thereby calculating the predicted value of the SOH of the battery 2. According to this embodiment, the degradation evolution of the battery 2 during use can be predicted with high accuracy.

[0099] The above description pertains to one embodiment of the present invention, but the invention is not limited thereto. Appropriate modifications to the details are possible within the scope of the present invention.

[0100] Figure Labels

[0101] 1: Battery degradation prediction device

[0102] 11: Data Acquisition Department

[0103] 13: Input Parameter Value Acquisition Section

[0104] 14: Model Prediction Department

[0105] 2: Battery

Claims

1. A battery model construction method, wherein the method constructs a battery model, the battery model associating the values ​​of a plurality of first input parameters and a plurality of second input parameters that are related to battery degradation indicators with the predicted values ​​of the aforementioned degradation indicators, the battery model construction method being characterized by comprising: The data acquisition step involves using multiple sample batteries to acquire the aforementioned first and second input parameters and data related to the measured values ​​of the aforementioned degradation indicators; The construction step uses at least a portion of the data acquired in the aforementioned data acquisition step as learning data to construct the aforementioned battery model; The error tendency determination step evaluates the prediction error between the measured value of the aforementioned degradation index and the predicted value of the aforementioned degradation index based on the aforementioned battery model for each of the aforementioned sample batteries, and determines whether there is a deviation of the aforementioned prediction error inherent in each of the aforementioned sample batteries. The first correlation determination step involves constructing a first error prediction model using the aforementioned learning data, which associates the explanatory variable defined based on the aforementioned first input parameters with a target variable corresponding to the predicted value of the aforementioned prediction error; and determining whether there is a first correlation between the measured value of the aforementioned prediction error obtained in the aforementioned error tendency determination step and the predicted value of the aforementioned prediction error based on the aforementioned first error prediction model; and, In the first reconstruction step, when it is determined in the aforementioned error tendency determination step that the aforementioned deviation exists and in the aforementioned first correlation determination step that the aforementioned first correlation exists, the aforementioned battery model is reconstructed.

2. The battery model construction method according to claim 1, wherein, Also includes: The second correlation determination step involves constructing a second error prediction model using the aforementioned learning data, which associates the explanatory variable defined based on the aforementioned second input parameters with the target variable corresponding to the predicted value of the aforementioned prediction error. Furthermore, it determines whether there is a second correlation between the measured value of the aforementioned prediction error obtained in the aforementioned error tendency determination step and the predicted value of the aforementioned prediction error based on the aforementioned second error prediction model; and... The second reconstruction step involves reconstructing the battery model when the aforementioned deviation is determined to exist in the aforementioned error tendency determination step and the aforementioned second correlation is determined to exist in the aforementioned second correlation determination step.

3. The battery model construction method according to claim 2, wherein, The aforementioned first input parameter is a usage history parameter defined based on the time-series data of the battery's current, voltage, and temperature. The aforementioned second input parameter is a manufacturing history parameter defined based on the manufacturing data of the aforementioned battery.

4. The battery model construction method according to claim 3, wherein, In the aforementioned first reconstruction step, the aforementioned battery model is reconstructed by using the aforementioned first error prediction model.

5. The battery model construction method according to claim 3, wherein, In the aforementioned second reconstruction step, the aforementioned battery model is reconstructed by using the aforementioned second error prediction model.

6. The battery model construction method according to claim 4, wherein, In the aforementioned second reconstruction step, the aforementioned battery model is reconstructed by using the aforementioned second error prediction model.

7. The battery model construction method according to any one of claims 3 to 6, wherein, The first and second input parameters and the time-series data related to the measured values ​​of the aforementioned degradation index obtained in the aforementioned data acquisition steps are divided into the aforementioned learning data within the specified learning period and the verification data within the verification period after the aforementioned learning period. In the aforementioned error tendency determination step, the presence or absence of the aforementioned deviation is determined based on the evaluation results of each of the aforementioned sample batteries based on the distribution of the aforementioned prediction error during the aforementioned learning period, and the evaluation results of the correlation between the aforementioned prediction error during the aforementioned learning period and the aforementioned prediction error during the aforementioned verification period.

8. The battery model construction method according to any one of claims 3 to 6, wherein, It also includes a third reconstruction step, which is to reconstruct the battery model by adding an offset term of a constant value corresponding to the cumulative usage time of the battery when the aforementioned error tendency determination step determines that the aforementioned deviation exists and the aforementioned first and second correlation determination steps determine that neither of the aforementioned first nor second correlation exists.

9. The battery model construction method according to claim 7, wherein, It also includes a third reconstruction step, which is to reconstruct the battery model by adding an offset term of a constant value corresponding to the cumulative usage time of the battery when the aforementioned error tendency determination step determines that the aforementioned deviation exists and the aforementioned first and second correlation determination steps determine that neither of the aforementioned first nor second correlation exists.

10. A battery degradation prediction device, characterized in that, include: The input parameter value acquisition unit acquires the values ​​of multiple first input parameters and multiple second input parameters that are related to the battery degradation index; and, The model prediction unit inputs the values ​​of the first and second input parameters obtained by the aforementioned input parameter value acquisition unit into the battery model constructed using the battery model construction method according to any one of claims 1 to 6, thereby calculating the predicted value of the aforementioned degradation index.

Images