Method and device for predicting health degree of battery and computer equipment

Abstract

The invention discloses a battery health degree prediction method and device and computer equipment, relates to the technical field of batteries, and solves the problem that the detection reliability of the battery health degree is relatively low in a battery health degree offline prediction task at present. The method comprises the steps of performing abnormal point detection based on a battery training set, determining position information of each abnormal point, then updating each target parameter in an obtained additive model by using a gradient descent method based on the position information of each abnormal point and the battery training set, and obtaining a target additive model for predicting the health degree, and then determining a plurality of abnormal discharge cycles and a plurality of normal discharge cycles based on the battery test set, predicting a first health degree corresponding to each abnormal discharge cycle and a second health degree corresponding to each normal discharge cycle by using the target additive model, and finally determining a prediction curve of the health degrees changing along with the discharge cycles.

Description

Technical Field

[0001] The present application relates to the field of battery technology, and in particular to a method, device and computer equipment for predicting battery health. Background Art

[0002] With my country's increasing emphasis on clean energy and the increasing demand for energy storage, the battery industry has developed rapidly in China. Modeling and predicting the state of health (SOH) of single cells is crucial to estimating battery performance and is also one of the core algorithm supports for battery management systems.

[0003] At present, data-driven models are mainly used for modeling and prediction. Data-driven models can complete model training and prediction based on discharge data, and have a wider range of applications. At the same time, data-driven models can also model changes in environmental factors such as temperature during the discharge process, and have good robustness. However, the above data-driven model mainly uses the sliding window method to predict the health of the battery online, that is, the value of the point is determined based on the data of several cycles before the prediction point, and for the offline prediction of battery health, that is, based on the known battery health of several discharge cycles, the battery health in the future for a long period of time is predicted. Since the data-driven model does not consider the local attenuation trend caused by abnormal points, the reliability of the battery health prediction result is low. Summary of the invention

[0004] The embodiments of the present application provide a method, apparatus and computer device for predicting battery health, the main purpose of which is to solve the problem of low reliability of battery health prediction results in the current offline battery health prediction task, because the data-driven model does not take into account the local attenuation trend caused by abnormal points.

[0005] According to the first aspect of the present application, a method for predicting battery health is provided, comprising:

[0006] Acquire a battery training set, perform outlier detection based on the battery training set, and determine location information of each outlier;

[0007] Determine an additive model, and based on the location information of each abnormal point and the battery training set, use a gradient descent method to update each target parameter in the additive model to obtain a target additive model for predicting health;

[0008] Acquire a battery test set, and determine a plurality of abnormal discharge cycles and a plurality of normal discharge cycles based on the battery test set;

[0009] Predicting a first health degree corresponding to each abnormal discharge cycle in the plurality of abnormal discharge cycles and a second health degree corresponding to each normal discharge cycle in the plurality of normal discharge cycles by using the target additive model;

[0010] A prediction curve of the change of health degree with discharge cycle is determined based on the first health degree corresponding to each abnormal discharge cycle and the second health degree corresponding to each normal discharge cycle.

[0011] Optionally, before obtaining the battery training set, the method further includes:

[0012] Acquire a battery data set, wherein the battery data set includes a plurality of discharge cycles and a cumulative discharge amount and a health level corresponding to each discharge cycle;

[0013] The battery data set is divided according to a preset ratio to determine the battery training set and the battery test set.

[0014] Optionally, performing outlier detection based on the battery training set to determine location information of each outlier includes:

[0015] Querying a plurality of discharge cycles and the accumulated discharge amount corresponding to each discharge cycle from the battery training set;

[0016] Calculating the difference between the cumulative discharge amount corresponding to each of the discharge cycles and the cumulative discharge amount corresponding to the last adjacent discharge cycle;

[0017] When it is detected that the difference exceeds a difference threshold, determining that the point corresponding to the discharge cycle is an abnormal point and the discharge cycle is an abnormal discharge cycle;

[0018] The abnormal health degree corresponding to the abnormal point is obtained from the battery training set, and the abnormal discharge cycle and the abnormal health degree corresponding to the abnormal point are used as the position information of the abnormal point.

[0019] Optionally, the determining the additive model comprises:

[0020] Obtaining a first corresponding relationship between the gain of the abnormal point to the main trend and a plurality of first target parameters and an abnormal discharge cycle, and using the first corresponding relationship as an exponential decay function corresponding to the abnormal point;

[0021] Acquire a second corresponding relationship between the health degree and a plurality of second target parameters and the discharge cycle, and use the second corresponding relationship as a main trend fitting function;

[0022] The additive model is generated based on the exponential decay function corresponding to the abnormal point and the main trend fitting function.

[0023] Optionally, based on the location information of each abnormal point and the battery training set, each target parameter in the additive model is updated by using a gradient descent method to obtain a target additive model for predicting health, including:

[0024] Calculate a mean square error based on the exponential decay function and the main trend fitting function to obtain a loss function, wherein the loss function is related to the multiple first target parameters and the multiple second target parameters;

[0025] Using the battery training set to train the multiple second target parameters in the loss function to obtain a trained intermediate loss function;

[0026] The plurality of first target parameters in the trained intermediate loss function are trained using the position information of the abnormal points to obtain a trained target loss function.

[0027] Taking partial derivatives of each target parameter in the trained target loss function to obtain a gradient function corresponding to each target parameter, and determining a parameter value of each target parameter by using the gradient descent method based on the gradient function corresponding to each target parameter, wherein the target parameter is a first target parameter or a second target parameter;

[0028] The corresponding target parameters in the additive model are updated based on the parameter values ​​of the target parameters to obtain the target additive model for predicting health status.

[0029] Optionally, the target parameter corresponding to the additive model is updated based on the parameter value of each target parameter to obtain the target additive model for predicting health, including:

[0030] Obtaining a correspondence between a parameter value of each of the first target parameters and a recovery value, and updating the corresponding first target parameter in the exponential decay function based on each of the correspondences to obtain a target exponential decay function;

[0031] Using the parameter value of each second target parameter to update the corresponding second target parameter in the main trend fitting function, to obtain a target main trend fitting function;

[0032] The target additive model for predicting health is generated based on the target exponential decay function and the target main trend fitting function.

[0033] Optionally, the using the target additive model to predict a first health degree corresponding to each abnormal discharge cycle in the multiple abnormal discharge cycles and a second health degree corresponding to each normal discharge cycle in the multiple normal discharge cycles includes:

[0034] Based on each of the abnormal discharge cycles and the corresponding preset recovery value, the prediction gain corresponding to each of the abnormal discharge cycles is predicted using the target exponential decay function in the target additive model, and based on each of the abnormal discharge cycles, the health degree corresponding to each of the abnormal discharge cycles is determined using the target main trend fitting function in the target additive model, and the sum of the prediction gain and the health degree corresponding to each of the abnormal discharge cycles is calculated to obtain a first health degree corresponding to each of the abnormal discharge cycles;

[0035] Based on each of the normal discharge cycles, a second health degree corresponding to each of the normal discharge cycles is determined using a target main trend fitting function in the target additive model.

[0036] Optionally, after determining the prediction curve of the health degree changing with the discharge cycle based on the first health degree corresponding to each abnormal discharge cycle and the second health degree corresponding to each normal discharge cycle, the method further includes:

[0037] Extracting the actual health corresponding to each discharge cycle from the battery test set, and extracting the predicted health corresponding to each discharge cycle from the prediction curve of the health varying with the discharge cycle;

[0038] The mean square error corresponding to each discharge cycle is determined based on the actual health and predicted health corresponding to each discharge cycle. When the mean square error corresponding to a discharge cycle exceeds the target threshold, it is determined that the prediction result of the target additive model is unreliable. When the mean square errors corresponding to each discharge cycle do not exceed the target threshold, it is determined that the prediction result of the target additive model is reliable.

[0039] According to a second aspect of the present application, a battery health prediction device is provided, comprising:

[0040] A first acquisition module is used to acquire a battery training set, perform abnormal point detection based on the battery training set, and determine the location information of each abnormal point;

[0041] An updating module, used for determining an additive model, and updating each target parameter in the additive model by using a gradient descent method based on the location information of each abnormal point and the battery training set, so as to obtain a target additive model for predicting health;

[0042] A second acquisition module, configured to acquire a battery test set, and determine a plurality of abnormal discharge cycles and a plurality of normal discharge cycles based on the battery test set;

[0043] A prediction module, configured to predict, by using the target additive model, a first health degree corresponding to each of the plurality of abnormal discharge cycles and a second health degree corresponding to each of the plurality of normal discharge cycles;

[0044] A determination module is used to determine a prediction curve of the health degree changing with the discharge cycle based on the first health degree corresponding to each abnormal discharge cycle and the second health degree corresponding to each normal discharge cycle.

[0045] According to a third aspect of the present application, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and wherein the processor implements the steps of any one of the methods described in the first aspect when executing the computer program.

[0046] By means of the above technical scheme, the present application provides a method, device and computer equipment for predicting battery health. The present application can determine the discharge cycle in which the abnormal SOH rebound occurs and the abnormal health corresponding to the discharge cycle by performing abnormal point detection on the battery data in the battery training set, that is, the location information of each abnormal point can be determined, and then the additive model is determined so that each abnormal point can correspond to an attenuation characteristic curve and a set of parameters. Subsequently, the target parameters in the additive model are updated according to the location information of each abnormal point and the battery training set, that is, the target parameters in the additive model are solved to complete the fitting of the battery SOH curve. This battery SOH fitting curve is the specific form of the target additive model. The target additive model can be used to predict the battery health of each discharge cycle in the future, and then the battery health of each discharge cycle can be used to generate a prediction curve of the change of health with the discharge cycle; through the above process, the local attenuation characteristics brought by the abnormal points are taken into account in the target additive model used to predict the health. In the subsequent offline prediction task, that is, the battery test set is predicted using the target additive model, and a battery health prediction result with higher reliability will be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1A A flowchart of a method for predicting battery health provided by an embodiment of the present application is shown;

[0048] Figure 1B A schematic diagram showing an abnormal point detection result of a battery health prediction method provided in an embodiment of the present application;

[0049] Figure 2A A flowchart of another method for predicting battery health provided in an embodiment of the present application is shown;

[0050] Figure 2B A schematic diagram showing a fitting result and a prediction result in another battery health prediction method provided in an embodiment of the present application is shown;

[0051] Figure 2C A scatter plot of each first target parameter and a recovery value in another battery health prediction method provided in an embodiment of the present application is shown;

[0052] Figure 3A A schematic diagram of the structure of a battery health prediction device provided in an embodiment of the present application is shown;

[0053] Figure 3B A schematic diagram of the structure of a battery health prediction device provided in an embodiment of the present application is shown;

[0054] Figure 3C A schematic diagram of the structure of a battery health prediction device provided in an embodiment of the present application is shown;

[0055] Figure 4 A schematic diagram of the device structure of a computer device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0056] Various aspects and features of the present application are described herein with reference to the accompanying drawings.

[0057] It should be understood that various modifications may be made to the embodiments of the present application. Therefore, the above description should not be considered as limiting, but only as an example of an embodiment. Other modifications within the scope and spirit of the present application will occur to those skilled in the art.

[0058] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0059] These and other characteristics of the present application will become apparent from the following description of a preferred form of embodiment given as a non-limiting example with reference to the accompanying drawings.

[0060] It should also be understood that although the present application has been described with reference to some specific examples, those skilled in the art will be able to readily implement many other equivalent forms of the present application.

[0061] The above and other aspects, features and advantages of the present application will become more apparent in view of the following detailed description when taken in conjunction with the accompanying drawings.

[0062] Specific embodiments of the present application are described hereinafter with reference to the accompanying drawings; however, it should be understood that the embodiments applied for are merely examples of the present application, which may be implemented in a variety of ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that obscure the present application. Therefore, the specific structural and functional details applied for herein are not intended to be limiting, but merely serve as a basis and representative basis for the claims to teach those skilled in the art to use the present application in a variety of ways with substantially any suitable detailed structure.

[0063] This specification may use the phrases "in one embodiment," "in another embodiment," "in yet another embodiment," or "in other embodiments," all of which may refer to one or more of the same or different embodiments according to the present application.

[0064] The present application embodiment provides a method for predicting battery health. Figure 1A As shown, including:

[0065] 101. Obtain a battery training set, perform outlier detection based on the battery training set, and determine location information of each outlier.

[0066] In the embodiment of the present application, the execution subject is a detection system that can execute the battery health prediction method. The detection system needs to first obtain a battery training set, which includes multiple discharge cycles and the cumulative discharge amount and health corresponding to each discharge cycle. After that, the detection system performs abnormal point detection based on the battery training set to determine the location information of each abnormal point.

[0067] It should be noted that anomalies can be understood as data points that are significantly different from other points in the data and exhibit outlier characteristics. If these anomalies are deleted directly, it is likely to cause deviations between the sample distribution and the overall distribution, thus affecting the results of the data analysis. The correct approach to dealing with anomalies in the data is to establish an anomaly detection mechanism, analyze the causes of the anomalies, and then select a processing method: if it is a manual recording error, change the data to the correct value; if the error is caused by a disturbance of some factor, consider incorporating this disturbance into the model and calculating the direction and magnitude of the disturbance. For example, for the 18650 cobalt oxide lithium ion battery dataset of the NASA PCoE laboratory, Figure 1B As shown in the figure, the SOH change curve of the single battery does not conform to the accelerated decay characteristics under ideal conditions, but rises abnormally in certain discharge cycles. This is because during the experiment, the battery charging and discharging are not completely continuous, and there are multiple charging operations. After a short rebound, the impact of the rebound will gradually disappear in the next few discharge cycles, and the decay rate of the curve will gradually return to normal levels. Therefore, the data points where the battery SOH increases abnormally are identified as abnormal points, and these abnormal points can be mathematically modeled later.

[0068] 102. Determine the additive model, based on the location information of each abnormal point and the battery training set, use the gradient descent method to update each target parameter in the additive model, and obtain the target additive model for predicting health.

[0069] In the embodiment of the present application, Figure 1BIt can be seen that each rebound of the battery SOH will have an impact on several subsequent discharge processes. In the short term, the battery SOH will gradually return to the main trend within a few cycles after the abnormal point occurs. The increase of each abnormal point on the future battery SOH can be roughly regarded as an exponential decay process; in the long term, although the battery SOH will gradually return to the main trend with the passage of time, each abnormal point will cause a slight rebound in the main trend, and the more obvious the mutation of the battery SOH, the more obvious the rebound of the main trend in the subsequent discharge process. Therefore, it is necessary to consider the long-term impact of abnormal points. Figure 1B The dotted line in the figure is the subsequent SOH estimation without abnormality. By comparing the difference between the solid line and the dotted line, it can be seen that a higher mutation value corresponds to faster local attenuation and greater long-term gain. Based on this, a parameter model can be established to characterize the long-term impact of abnormal points. This parameter model is an additive model, which can correspond abnormal points to an attenuation characteristic curve and a set of parameters. After determining the additive model, the target parameters in the additive model can be updated using the gradient descent method based on the location information of each abnormal point and the battery training set to obtain a target additive model for predicting health.

[0070] 103. Obtain a battery test set, and determine a plurality of abnormal discharge cycles and a plurality of normal discharge cycles based on the battery test set.

[0071] In an embodiment of the present application, after determining the target additive model for predicting health, a battery test set can be obtained, which includes multiple discharge cycles and the cumulative discharge amount corresponding to each discharge cycle. Through the battery test set, multiple abnormal discharge cycles and multiple normal discharge cycles can be determined. The abnormal discharge cycle can be understood as the difference between the corresponding cumulative discharge amount and the cumulative discharge amount corresponding to the previous adjacent discharge cycle exceeds a fixed value, while the normal discharge cycle can be understood as the difference between the corresponding cumulative discharge amount and the cumulative discharge amount corresponding to the previous adjacent discharge cycle does not exceed a fixed value. It should be noted that the battery test set also includes the actual health corresponding to each discharge cycle, but the actual health is not used in the prediction stage, and is only used in the subsequent verification of the reliability of the prediction results of the target additive model.

[0072] 104. Predict, by using a target additive model, a first health degree corresponding to each abnormal discharge cycle in a plurality of abnormal discharge cycles and a second health degree corresponding to each normal discharge cycle in a plurality of normal discharge cycles.

[0073] In the embodiment of the present application, after determining a plurality of abnormal discharge cycles and a plurality of normal discharge cycles, the target additive model can be used to predict the first health degree corresponding to each of the plurality of abnormal discharge cycles and the second health degree corresponding to each of the plurality of normal discharge cycles. It should be noted that the first and second above have no physical meaning and are only used to distinguish the health degree.

[0074] 105. Determine a prediction curve of health varying with discharge cycle based on the first health corresponding to each abnormal discharge cycle and the second health corresponding to each normal discharge cycle.

[0075] In an embodiment of the present application, a prediction curve of the change of health with discharge cycle can be generated based on the first health corresponding to each abnormal discharge cycle and the second health corresponding to each normal discharge cycle, so that the health corresponding to each discharge cycle in the future period of time can be predicted.

[0076] The method provided in the embodiment of the present application can determine the discharge cycle where the SOH abnormal rebound occurs and the abnormal health corresponding to the discharge cycle by detecting abnormal points on the battery data in the battery training set, that is, it can determine the position information of each abnormal point, and then determine the additive model so that each abnormal point can correspond to an attenuation characteristic curve and a set of parameters, and then update the target parameters in the additive model according to the position information of each abnormal point and the battery training set, that is, solve each target parameter in the additive model to complete the fitting of the battery SOH curve. This battery SOH fitting curve is the specific form of the target additive model. The target additive model can be used to predict the battery health of each discharge cycle in the future, and then the battery health of each discharge cycle can be used to generate a prediction curve of the health changing with the discharge cycle; through the above process, the local attenuation characteristics brought by the abnormal points are taken into account in the target additive model used to predict the health. In the subsequent offline prediction task, that is, the battery test set is predicted using the target additive model, and a battery health prediction result with higher reliability will be obtained.

[0077] Further, as a refinement and extension of the specific implementation of the above embodiment, in order to fully illustrate the specific implementation of this embodiment, the embodiment of the present application provides another method for predicting battery health, such as Figure 2A As shown, including:

[0078] 201. Obtain a battery data set, and divide the battery data set.

[0079] In an embodiment of the present application, the battery data set can select a lithium cobalt oxide battery data set, and the data set is divided into a battery training set and a battery test set in a ratio of 7:3. The first 70% of the data is used to train the fitting function of the main trend, and the last 30% of the data is used to test the accuracy of the model or the reliability of the prediction results. Since the model proposed in this application has the function of correcting the main trend according to the location information of the abnormal point, the occurrence time (discharge cycle) and the size of the recovery value of the abnormal point are known during the test (but the exact health is not known). Through battery testing, not only the predicted health of each discharge cycle in the future can be predicted, but also the reliability of the prediction results of the target additive model can be determined in combination with the actual health included in the battery test set.

[0080] 202. Perform outlier detection based on the battery training set to determine the location information of each outlier.

[0081] In the embodiment of the present application, multiple discharge cycles and the cumulative discharge amount corresponding to each discharge cycle can be first queried from the battery training set, and then the difference between the cumulative discharge amount corresponding to each discharge cycle and the cumulative discharge amount corresponding to the previous adjacent discharge cycle is calculated. Then, when it is detected that the difference exceeds the difference threshold, the point corresponding to the discharge cycle is determined to be an abnormal point and the discharge cycle is an abnormal discharge cycle. Finally, the abnormal health corresponding to the abnormal point is obtained from the battery training set, and the abnormal discharge cycle and abnormal health corresponding to the abnormal point are used as the location information of the abnormal point. In other words, when the cumulative discharge amount of a battery in a certain discharge cycle is higher than the cumulative discharge amount of the battery in the previous discharge cycle by a fixed value, the previous discharge cycle is detected as an abnormal point.

[0082] For example, a battery data set of three batteries, namely, batteries No. 5, No. 6, and No. 7, can be obtained for abnormal point detection. Since the SOH decay speed and recovery amplitude of these three batteries are different, a threshold is set for each battery separately in order to accurately identify all abnormal points. For batteries with a larger SOH recovery amplitude, a larger threshold is set; and for batteries with a smaller SOH recovery amplitude, a smaller threshold is set. In other words, the difference threshold can be flexibly set according to the characteristics of each battery.

[0083] 203. Determine the additive model.

[0084] In the embodiment of the present application, first, a first correspondence between the gain of the abnormal point to the main trend and multiple first target parameters and the abnormal discharge period is obtained, and the first correspondence is used as an exponential decay function corresponding to the abnormal point, wherein the specific form of the exponential decay function is:

[0085] y=ce -rx +d,

[0086] Among them, x is the discharge cycle corresponding to the abnormal point, y is the gain of the abnormal point to the main trend at x after the abnormal point occurs; c, r, d are collectively referred to as multiple first target parameters, and their values ​​are related to the recovery value of the battery SOH at the abnormal point. Note that this formula is applicable to all discharge cycles after the current discharge cycle.

[0087] Then, a second corresponding relationship between the health degree and a plurality of second target parameters and the discharge cycle is obtained, and the second corresponding relationship is used as the main trend fitting function, wherein the specific form of the main trend fitting function is:

[0088] y'=a1x' 3 +a2x' 2 +a3x'+b,

[0089] Among them, y' is the health corresponding to a discharge cycle, x' is the discharge cycle, a1, a2, a3, b are collectively referred to as multiple second target parameters. It should be noted that for the main trend, the parameter models currently used to predict battery SOH include polynomial functions, exponential functions, double exponential functions, etc. The embodiment of the present application uses a cubic polynomial function as the main trend fitting function.

[0090] Finally, an additive model is generated based on the exponential decay function corresponding to the outliers and the main trend fitting function.

[0091] 204. Based on the location information of each abnormal point and the battery training set, each target parameter in the additive model is updated using the gradient descent method to obtain a target additive model for predicting health.

[0092] In an embodiment of the present application, first, a mean square error is calculated based on an exponential decay function and a main trend fitting function to obtain a loss function, and the loss function is related to a plurality of first target parameters and a plurality of second target parameters.

[0093] It should be noted that the loss function is updated using the gradient descent method to obtain the parameter value. For the simplicity of the model form, we believe that the short-term impact of the abnormal point on the battery SOH (expressed by the exponential decay function) only works before the next abnormal point occurs, while the long-term impact of the abnormal point on the battery SOH (expressed by the constant d) works for all discharge cycles after the abnormal point occurs. The mean square error (MSE) is used as the loss function, and the specific form of the loss function is:

[0094]

[0095] Among them, m represents the total number of discharge cycles, x represents the time when the jth abnormal point occurs, i ,y iThey represent the i-th discharge cycle and the corresponding battery SOH respectively. The entire loss function is divided into three lines and can be seen as three parts. The first part corresponds to the cycle 1 to t1-1, during which there is only the main trend; the second part corresponds to the cycle t j to j+1 -1, j ranges from 1 to n-1, during this period, there is a superposition of the main trend, the short-term impact of the j-th abnormal point, and the long-term impact of the previous j abnormal points; the third part corresponds to the period t n From 0 to m, the components of this period are the same as those of the second part, but since the next abnormal point has not yet appeared, the loss corresponding to this part of the period needs to be expressed separately.

[0096] Subsequently, the battery training set is used to train multiple second target parameters in the loss function to obtain a trained intermediate loss function, and the position information of each abnormal point is used to train multiple first target parameters in the trained intermediate loss function to obtain a trained target loss function.

[0097] Specifically, due to the existence of the polynomial form, when the number of cycles is large, the parameters of the higher-order terms are prone to gradient explosion problems, resulting in the inability to effectively update other parameters a1 and a2. Therefore, the pre-training method can be used to first train the parameters a1, a2, a3 and b of the main trend, and then set their learning rates to 0, and the other parameters c j , r j , d j The pre-training process is completed using the classic Matlab polyfit curve fitting tool.

[0098] It should be noted that the learning rates of the three parameters c, r, and d are set differently for an internal reason. Due to the existence of the exponential function, the parameters c and r are given higher learning rates to prevent the gradient from disappearing; the learning rate of d is relatively small because d is a global parameter and the update of parameter d should be relatively cautious.

[0099] Next, the partial derivative of each target parameter in the trained target loss function is calculated to obtain the gradient function corresponding to each target parameter, and based on the gradient function corresponding to each target parameter, the parameter value of each target parameter is determined using the gradient descent method, and the target parameter is the first target parameter or the second target parameter.

[0100] Specifically, by taking the partial derivative of each target parameter in the loss function, the corresponding gradient form can be obtained, and the gradient descent method can be used to update the parameters to continuously reduce the value of the loss function. The parameter value of each target parameter can be determined through the above process. For example, there are 118 discharge cycles in the battery training set, that is, 1 to 118. Among them, through the outlier detection, it can be seen that there are multiple abnormal discharge cycles, which are also the locations of the outliers. Through the above process, the mutation of the outlier point and the subsequent attenuation characteristics can be accurately detected. The specific fitting curve of the target additive model for the battery training set is as follows: Figure 2B As shown in the figure, there are two curves, curve 1 and curve 2. Curve 1 is the decay of the real health, and curve 2 is the total fitting result (with correction added). See the fitting result part before the vertical dotted line. It can be seen from curve 2 that the target additive model has some defects in fitting the part before the first abnormal point and the part near the fourth abnormal point. This is caused by the fact that the form of the main trend fitting function is too simple and cannot fit the local characteristics of the curve well. It is worth emphasizing that although the overly simple main trend function causes problems in local fitting, it is not advisable to select too complex function forms in the fitting stage. This is because more complex functions have stronger fitting ability and may have replaced the above-mentioned exponential decay model to fit the abnormal points first. It is meaningless to establish an additive model on this basis, because the main trend already contains local characteristics, and it is naturally impossible to predict the local decay characteristics of the battery SOH by modeling the information of the abnormal points. Here, a simple cubic function is selected as the main trend fitting function, which can give full play to the role of the exponential decay model and obtain satisfactory results in the prediction stage.

[0101] Finally, the corresponding target parameters in the additive model are updated based on the parameter values ​​of each target parameter to obtain a target additive model for predicting health.

[0102] Specifically, to obtain the target additive model, it is necessary to first obtain the corresponding relationship between the parameter value of each first target parameter and the recovery value, and update the corresponding first target parameter in the exponential decay function based on each corresponding relationship to obtain the target exponential decay function.

[0103] In the embodiment of the present application, the fitting result of the battery SOH curve is obtained by gradient descent, and the c, r, d parameter values ​​corresponding to each abnormal point are also obtained. Next, the regression method is adopted to analyze the statistical relationship between the recovery value and the parameters c, r, d, so as to estimate the specific parameter values ​​of c, r, d by the size of the preset recovery value. It should be noted that the recovery value corresponding to each discharge cycle is known, so that each abnormal point corresponds to a set of c, r, d parameter values ​​and recovery values. Based on this, a scatter plot of the recovery value and the fitted parameter value can be generated, such as Figure 2CAs shown in the scatter plot, it can be seen that there is a relatively obvious positive / negative correlation between the SOH recovery value and the values ​​of each parameter. If the points with higher recovery values ​​are removed, the related features will no longer be so obvious. In fact, points with higher recovery values ​​are crucial to determining the regression equation and can be regarded as strong influencing points. Therefore, in the fitting process, attention should be paid to the fitting effect of the curve near the abnormal points with high recovery values, and no large errors should occur, otherwise it will have an adverse effect on the subsequent SOH prediction. After removing the abnormal points, a simple linear regression is performed on the recovery value and each parameter based on the positive / negative correlation characteristics reflected in the image. The results are shown in the figure. Figure 2C As shown (the regression line is a straight line). After obtaining the regression equation (that is, the corresponding relationship), the values ​​of the parameters c, r, and d can be predicted based on the preset recovery values, thereby determining the attenuation curve characteristics after the abnormal point occurs.

[0104] Afterwards, the corresponding second target parameters in the main trend fitting function are updated using the parameter values ​​of the second target parameters to obtain the target main trend fitting function.

[0105] It should be noted that, in the above curve fitting stage, for example, a1 is 12, a2 ​​is 20, a3 is 50, and b is 15, then in the subsequent prediction stage, a1 is still 12, a2 ​​is still 20, a3 is still 50, and b is still 15. And c, r, d will be determined according to the preset recovery value corresponding to each abnormal discharge cycle.

[0106] Finally, a target additive model for predicting health is generated based on the target exponential decay function and the target main trend fitting function.

[0107] 205. Determine a plurality of abnormal discharge cycles and a plurality of normal discharge cycles based on the battery test set.

[0108] 206. Predict, using a target additive model, a first health degree corresponding to each abnormal discharge cycle in a plurality of abnormal discharge cycles and a second health degree corresponding to each normal discharge cycle in a plurality of normal discharge cycles.

[0109] In an embodiment of the present application, after determining the target additive model, the abnormal discharge cycle can be determined through the battery test set, and each abnormal discharge cycle corresponds to a preset recovery value, which is known. Then, based on each abnormal discharge cycle and the corresponding preset recovery value, the target exponential decay function in the target additive model is used to predict the prediction gain corresponding to each abnormal discharge cycle, and based on each abnormal discharge cycle, the target main trend fitting function in the target additive model is used to determine the health corresponding to each abnormal discharge cycle, and the sum of the prediction gain and the health corresponding to each abnormal discharge cycle is calculated to obtain the first health corresponding to each abnormal discharge cycle; then, based on each normal discharge cycle, the second health corresponding to each normal discharge cycle is determined using the target main trend fitting function in the target additive model.

[0110] 207. Determine a prediction curve of health varying with discharge cycle based on the first health corresponding to each abnormal discharge cycle and the second health corresponding to each normal discharge cycle.

[0111] Specifically, based on the first health degree corresponding to each abnormal discharge cycle and the second health degree corresponding to each normal discharge cycle, a prediction curve of the health degree changing with the discharge cycle can be determined. The second target parameter value in the main trend fitting function in the above-mentioned target additive model is unchanged in the prediction stage. In fact, the linear extrapolation method is used here, that is, the starting point of the extrapolation is taken as the passing point, and the numerical derivative of the point is taken as the slope to construct a linear function for extrapolation. The reason for doing this is that directly extrapolating the cubic function may predict an upward trend that is completely inconsistent with common sense, causing unnecessary errors. In addition, due to the gain effect of abnormal points on the main trend, directly fitting the original curve to the main trend will cause the future SOH prediction to be too optimistic and the predicted value to be too high. To this end, it is necessary to first make empirical corrections to the training data and then train the main trend.

[0112] For example, taking the No. 5 battery in the above example as an example, see Figure 2B , the part before the vertical dotted line is the fitting curve obtained using the battery training set, and the part after the vertical dotted line is the prediction curve obtained using the battery test set, that is, the discharge cycle 1-118 is used to train the parameters of the main trend simulation, and the discharge cycle 119-168 is used to test the prediction effect of the model. The dividing line is represented by a dotted line. Figure 2B It can be seen that in the prediction stage, the prediction of the attenuation trend after the abnormal point is based on the corresponding relationship between each first target parameter and the recovery value, and in the fitting stage, the parameter value of each first parameter is obtained using the gradient descent method, which can effectively avoid the overfitting problem. In this way, the final prediction result of the battery health is more in line with common sense and there is no abnormal deviation.

[0113] It should be further explained that the method proposed in the embodiment of the present application is equivalent to considering the impact of abnormal points on battery SOH on the basis of the traditional method.

[0114] 208. Conduct reliability test on the health prediction results of the target additive model.

[0115] In an optional implementation of the embodiments of the present application, the actual health corresponding to each discharge cycle can be extracted from the battery test set, and the predicted health corresponding to each discharge cycle can be extracted from the prediction curve of the change of health with the discharge cycle; then, the mean square error corresponding to each discharge cycle is determined based on the actual health and predicted health corresponding to each discharge cycle. When the mean square error corresponding to a discharge cycle exceeds the target threshold, it is determined that the prediction result of the target additive model is unreliable. When the mean square errors corresponding to each discharge cycle do not exceed the target threshold, it is determined that the prediction result of the target additive model is reliable.

[0116] In another optional implementation of the embodiment of the present application, for each discharge cycle, the absolute value of the difference between the actual health and the predicted health corresponding to the discharge cycle can also be calculated, and the absolute value of the difference is compared with the target difference. If the absolute value of the difference exceeds the target difference, it is determined that the prediction result of the target additive model is unreliable. If the absolute value of the difference does not exceed the target difference, it is determined that the prediction result of the target additive model is reliable.

[0117] In another optional implementation of the embodiment of the present application, the mean absolute error corresponding to each discharge cycle can also be determined based on the actual health and predicted health corresponding to each discharge cycle. When the mean absolute error corresponding to a discharge cycle exceeds the target threshold, it is determined that the prediction result of the target additive model is unreliable. When the mean absolute errors corresponding to each discharge cycle do not exceed the target threshold, it is determined that the prediction result of the target additive model is reliable.

[0118] The method provided in the embodiment of the present application can determine the discharge cycle where the SOH abnormal rebound occurs and the abnormal health corresponding to the discharge cycle by detecting abnormal points on the battery data in the battery training set, that is, it can determine the position information of each abnormal point, and then determine the additive model so that each abnormal point can correspond to an attenuation characteristic curve and a set of parameters, and then update the target parameters in the additive model according to the position information of each abnormal point and the battery training set, that is, solve each target parameter in the additive model to complete the fitting of the battery SOH curve. This battery SOH fitting curve is the specific form of the target additive model. The target additive model can be used to predict the battery health of each discharge cycle in the future, and then the battery health of each discharge cycle can be used to generate a prediction curve of the health changing with the discharge cycle; through the above process, the local attenuation characteristics brought by the abnormal points are taken into account in the target additive model used to predict the health. In the subsequent offline prediction task, that is, the battery test set is predicted using the target additive model, and a battery health prediction result with higher reliability will be obtained.

[0119] Further, as a specific implementation of the method described in FIG. 1, as shown in FIG. Figure 3A As shown, an embodiment of the present invention provides a battery health prediction device, including: a first acquisition module 301, an update module 302, a second acquisition module 303, a prediction module 304 and a determination module 305.

[0120] The first acquisition module 301 is used to acquire a battery training set, perform abnormal point detection based on the battery training set, and determine the location information of each abnormal point;

[0121] The updating module 302 is used to determine the additive model, based on the location information of each abnormal point and the battery training set, using the gradient descent method to update each target parameter in the additive model to obtain a target additive model for predicting health;

[0122] The second acquisition module 303 is used to acquire a battery test set, and determine a plurality of abnormal discharge cycles and a plurality of normal discharge cycles based on the battery test set;

[0123] The prediction module 304 is used to predict the first health degree corresponding to each abnormal discharge cycle in the multiple abnormal discharge cycles and the second health degree corresponding to each normal discharge cycle in the multiple normal discharge cycles by using the target additive model;

[0124] The determination module 305 is used to determine a prediction curve of the change of health degree with the discharge cycle based on the first health degree corresponding to each abnormal discharge cycle and the second health degree corresponding to each normal discharge cycle.

[0125] In specific application scenarios, such as Figure 3B As shown, the device also includes: a division module 306.

[0126] The division module 306 is also used to obtain a battery data set, which includes multiple discharge cycles and the cumulative discharge capacity and health status corresponding to each discharge cycle; divide the battery data set according to a preset ratio to determine the battery training set and the battery test set.

[0127] In a specific application scenario, the first acquisition module 301 is also used to query multiple discharge cycles and the cumulative discharge amount corresponding to each discharge cycle from the battery training set; calculate the difference between the cumulative discharge amount corresponding to each discharge cycle and the cumulative discharge amount corresponding to the previous adjacent discharge cycle; when it is detected that the difference exceeds the difference threshold, determine that the point corresponding to the discharge cycle is an abnormal point and the discharge cycle is an abnormal discharge cycle; obtain the abnormal health corresponding to the abnormal point from the battery training set, and use the abnormal discharge cycle and the abnormal health corresponding to the abnormal point as the location information of the abnormal point.

[0128] In a specific application scenario, the update module 302 is also used to obtain a first correspondence between the gain of the abnormal point to the main trend and multiple first target parameters and abnormal discharge cycles, and use the first correspondence as the exponential decay function corresponding to the abnormal point; obtain a second correspondence between the health degree and multiple second target parameters and the discharge cycle, and use the second correspondence as the main trend fitting function; generate the additive model based on the exponential decay function corresponding to the abnormal point and the main trend fitting function.

[0129] In a specific application scenario, the update module 302 is also used to calculate the mean square error based on the exponential decay function and the main trend fitting function to obtain a loss function, wherein the loss function is related to the multiple first target parameters and the multiple second target parameters; the multiple second target parameters in the loss function are trained using the battery training set to obtain a trained intermediate loss function; the multiple first target parameters in the trained intermediate loss function are trained using the location information of each abnormal point to obtain a trained target loss function. Partial derivatives are taken for each target parameter in the trained target loss function to obtain a gradient function corresponding to each target parameter, and based on the gradient function corresponding to each target parameter, the parameter value of each target parameter is determined using the gradient descent method, wherein the target parameter is the first target parameter or the second target parameter; the corresponding target parameter in the additive model is updated based on the parameter value of each target parameter to obtain the target additive model for predicting health.

[0130] In a specific application scenario, the update module 302 is also used to obtain the correspondence between the parameter value and the recovery value of each of the first target parameters, and update the corresponding first target parameters in the exponential decay function based on each of the corresponding relationships to obtain a target exponential decay function; use the parameter value of each of the second target parameters to update the corresponding second target parameters in the main trend fitting function to obtain a target main trend fitting function; and generate the target additive model for predicting health based on the target exponential decay function and the target main trend fitting function.

[0131] In a specific application scenario, the prediction module 304 is also used to predict the prediction gain corresponding to each abnormal discharge cycle based on each abnormal discharge cycle and the corresponding preset recovery value, using the target exponential decay function in the target additive model, and based on each abnormal discharge cycle, using the target main trend fitting function in the target additive model to determine the health corresponding to each abnormal discharge cycle, calculate the sum of the prediction gain and the health corresponding to each abnormal discharge cycle, and obtain the first health corresponding to each abnormal discharge cycle; based on each normal discharge cycle, using the target main trend fitting function in the target additive model to determine the second health corresponding to each normal discharge cycle.

[0132] In specific application scenarios, such as Figure 3C As shown, the device further includes: a testing module 307.

[0133] The inspection module 307 is used to extract the actual health corresponding to each discharge cycle from the battery test set, and to extract the predicted health corresponding to each discharge cycle from the prediction curve of the change of the health with the discharge cycle; based on the actual health and predicted health corresponding to each discharge cycle, the mean square error corresponding to each discharge cycle is determined; when the mean square error corresponding to a discharge cycle exceeds the target threshold, it is determined that the prediction result of the target additive model is unreliable; when the mean square errors corresponding to each discharge cycle do not exceed the target threshold, it is determined that the prediction result of the target additive model is reliable.

[0134] The device provided in the embodiment of the present application can determine the discharge cycle where the SOH abnormal rebound occurs and the abnormal health corresponding to the discharge cycle by detecting abnormal points on the battery data in the battery training set, that is, it can determine the position information of each abnormal point, and then determine the additive model so that each abnormal point can correspond to an attenuation characteristic curve and a set of parameters, and then update the target parameters in the additive model according to the position information of each abnormal point and the battery training set, that is, solve each target parameter in the additive model to complete the fitting of the battery SOH curve. This battery SOH fitting curve is the specific form of the target additive model. The target additive model can be used to predict the battery health of each discharge cycle in the future, and then the battery health of each discharge cycle can be used to generate a prediction curve of the health changing with the discharge cycle; through the above process, the local attenuation characteristics brought by the abnormal points are taken into account in the target additive model used to predict the health. In the subsequent offline prediction task, that is, the battery test set is predicted using the target additive model, and a battery health prediction result with higher reliability will be obtained.

[0135] It should be noted that for other corresponding descriptions of the functional units involved in the battery health prediction device provided in the embodiment of the present invention, reference can be made to Figure 1A and FIG. 3A to FIG. 3C The corresponding description in will not be repeated here.

[0136] In an exemplary embodiment, see Figure 4 , a computer device is also provided, the computer device includes a bus, a processor, a memory and a communication interface, and may also include an input and output interface and a display device, wherein each functional unit can communicate with each other through the bus. The memory stores a computer program, and the processor is used to execute the program stored in the memory and execute the battery health prediction method in the above embodiment.

[0137] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by hardware, or by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each implementation scenario of the present application.

[0138] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily required for implementing the present application.

[0139] Those skilled in the art will appreciate that the modules in the device in the implementation scenario can be distributed in the device in the implementation scenario according to the implementation scenario description, or can be changed accordingly and located in one or more devices different from the implementation scenario. The modules in the above implementation scenario can be combined into one module, or can be further split into multiple sub-modules.

[0140] The above application serial numbers are for description only and do not represent the advantages or disadvantages of the implementation scenarios.

[0141] The above embodiments are only exemplary embodiments of the present application and are not intended to limit the present application. The protection scope of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and protection scope of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present application.

Claims

1. A method for predicting battery health, characterized in that: include: Acquire a battery training set, perform outlier detection based on the battery training set, and determine location information of each outlier; Determine an additive model, and based on the location information of each abnormal point and the battery training set, use a gradient descent method to update each target parameter in the additive model to obtain a target additive model for predicting health; Acquire a battery test set, and determine a plurality of abnormal discharge cycles and a plurality of normal discharge cycles based on the battery test set; Predicting a first health degree corresponding to each abnormal discharge cycle in the plurality of abnormal discharge cycles and a second health degree corresponding to each normal discharge cycle in the plurality of normal discharge cycles by using the target additive model; A prediction curve of the change of health degree with discharge cycle is determined based on the first health degree corresponding to each abnormal discharge cycle and the second health degree corresponding to each normal discharge cycle.

2. The method according to claim 1, characterized in that Before obtaining the battery training set, the method further includes: Acquire a battery data set, wherein the battery data set includes a plurality of discharge cycles and a cumulative discharge amount and a health level corresponding to each discharge cycle; The battery data set is divided according to a preset ratio to determine the battery training set and the battery test set.

3. The method according to claim 1, characterized in that: The performing outlier detection based on the battery training set to determine the location information of each outlier includes: Querying a plurality of discharge cycles and the accumulated discharge amount corresponding to each discharge cycle from the battery training set; Calculating the difference between the cumulative discharge amount corresponding to each of the discharge cycles and the cumulative discharge amount corresponding to the last adjacent discharge cycle; When it is detected that the difference exceeds a difference threshold, determining that the point corresponding to the discharge cycle is an abnormal point and the discharge cycle is an abnormal discharge cycle; The abnormal health degree corresponding to the abnormal point is obtained from the battery training set, and the abnormal discharge cycle and the abnormal health degree corresponding to the abnormal point are used as the position information of the abnormal point.

4. The method according to claim 1, characterized in that The determined additive model comprises: Obtaining a first corresponding relationship between the gain of the abnormal point to the main trend and a plurality of first target parameters and an abnormal discharge cycle, and using the first corresponding relationship as an exponential decay function corresponding to the abnormal point; Acquire a second corresponding relationship between the health degree and a plurality of second target parameters and the discharge cycle, and use the second corresponding relationship as a main trend fitting function; The additive model is generated based on the exponential decay function corresponding to the abnormal point and the main trend fitting function.

5. The method according to claim 4, characterized in that Based on the location information of each abnormal point and the battery training set, each target parameter in the additive model is updated by using the gradient descent method to obtain a target additive model for predicting health, including: Calculate a mean square error based on the exponential decay function and the main trend fitting function to obtain a loss function, wherein the loss function is related to the multiple first target parameters and the multiple second target parameters; Using the battery training set to train the multiple second target parameters in the loss function to obtain a trained intermediate loss function; The plurality of first target parameters in the trained intermediate loss function are trained using the position information of the abnormal points to obtain a trained target loss function. Taking partial derivatives of each target parameter in the trained target loss function to obtain a gradient function corresponding to each target parameter, and determining a parameter value of each target parameter by using the gradient descent method based on the gradient function corresponding to each target parameter, wherein the target parameter is a first target parameter or a second target parameter; The corresponding target parameters in the additive model are updated based on the parameter values ​​of the target parameters to obtain the target additive model for predicting health status.

6. The method according to claim 5, characterized in that The updating of the corresponding target parameters in the additive model based on the parameter values ​​of each target parameter to obtain the target additive model for predicting health includes: Obtaining a correspondence between a parameter value of each of the first target parameters and a recovery value, and updating the corresponding first target parameter in the exponential decay function based on each of the correspondences to obtain a target exponential decay function; Using the parameter value of each second target parameter to update the corresponding second target parameter in the main trend fitting function, to obtain a target main trend fitting function; The target additive model for predicting health is generated based on the target exponential decay function and the target main trend fitting function.

7. The method according to claim 1, characterized in that The using the target additive model to predict a first health degree corresponding to each abnormal discharge cycle in the plurality of abnormal discharge cycles and a second health degree corresponding to each normal discharge cycle in the plurality of normal discharge cycles includes: Based on each of the abnormal discharge cycles and the corresponding preset recovery value, the prediction gain corresponding to each of the abnormal discharge cycles is predicted using the target exponential decay function in the target additive model, and based on each of the abnormal discharge cycles, the health degree corresponding to each of the abnormal discharge cycles is determined using the target main trend fitting function in the target additive model, and the sum of the prediction gain and the health degree corresponding to each of the abnormal discharge cycles is calculated to obtain a first health degree corresponding to each of the abnormal discharge cycles; Based on each of the normal discharge cycles, a second health degree corresponding to each of the normal discharge cycles is determined using a target main trend fitting function in the target additive model.

8. The method according to claim 1, characterized in that After determining the prediction curve of health degree changing with discharge cycle based on the first health degree corresponding to each abnormal discharge cycle and the second health degree corresponding to each normal discharge cycle, the method further includes: Extracting the actual health corresponding to each discharge cycle from the battery test set, and extracting the predicted health corresponding to each discharge cycle from the prediction curve of the health varying with the discharge cycle; The mean square error corresponding to each discharge cycle is determined based on the actual health and predicted health corresponding to each discharge cycle. When the mean square error corresponding to a discharge cycle exceeds the target threshold, it is determined that the prediction result of the target additive model is unreliable. When the mean square errors corresponding to each discharge cycle do not exceed the target threshold, it is determined that the prediction result of the target additive model is reliable.

9. A battery health prediction device, characterized in that: include: A first acquisition module is used to acquire a battery training set, perform abnormal point detection based on the battery training set, and determine the location information of each abnormal point; An updating module, used for determining an additive model, and updating each target parameter in the additive model by using a gradient descent method based on the location information of each abnormal point and the battery training set, so as to obtain a target additive model for predicting health; A second acquisition module, configured to acquire a battery test set, and determine a plurality of abnormal discharge cycles and a plurality of normal discharge cycles based on the battery test set; A prediction module, configured to predict, by using the target additive model, a first health degree corresponding to each of the plurality of abnormal discharge cycles and a second health degree corresponding to each of the plurality of normal discharge cycles; A determination module is used to determine a prediction curve of the health degree changing with the discharge cycle based on the first health degree corresponding to each abnormal discharge cycle and the second health degree corresponding to each normal discharge cycle.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images