Prediction Method, System, Electronic Device and Storage Medium for Battery Aging Trajectory

Abstract

The present invention discloses a method, a system, an electronic device and a storage medium for predicting the aging trajectory of a battery, relating to the technical field of machine learning. The method includes: testing the battery for a predetermined number of cycles to obtain cycle test data; predicting the battery aging end-point index based on the cycle test data to obtain an aging prediction result, where the aging prediction result includes the cycle life and the life capacity of the battery; performing error verification on the aging prediction result and selecting a feature combination according to the verification result to obtain feature combination data; using the feature combination data to predict the aging process trajectory of the battery to obtain a future aging trajectory of the battery capacity changing with the increase in the number of cycles; and being capable of improving the accuracy of the prediction results of the battery life and the future aging trajectory.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and particularly to a method, a system, an electronic device and a storage medium for predicting the aging trajectory of a battery. Background Art

[0002] With the increasingly wide application of batteries, the prediction of battery life and future aging trajectory is also of great significance for battery applications.

[0003] Currently, personnel in related fields usually predict the life and future aging trajectory of a battery by means of an electrochemical model or a data-driven model. However, there are many physical and chemical parameters in the electrochemical model, and many internal mechanism processes have not been fully explained. The prediction ability is extremely limited and the adaptability is poor. Therefore, it is more accurate to use a data-driven model for prediction.

[0004] However, the existing prediction technologies using data-driven models for battery life and future aging trajectory are not yet mature, and the accuracy of prediction results needs to be improved. Summary of the Invention

[0005] The main purpose of the present invention is to provide a method, a system, an electronic device and a storage medium for predicting the aging trajectory of a battery, aiming to improve the accuracy of prediction results of battery life and future aging trajectory.

[0006] To achieve the above object, in a first aspect of the present invention, a method for predicting the aging trajectory of a battery is provided, including: testing the battery for a predetermined number of cycles to obtain cycle test data; predicting the battery aging end-point index according to the cycle test data to obtain an aging prediction result, where the aging prediction result includes the cycle life and life capacity of the battery; verifying the error of the aging prediction result, and selecting a feature combination according to the verification result to obtain feature combination data; using the feature combination data to predict the aging process trajectory of the battery to obtain a future aging trajectory of the battery capacity changing with the increase of the number of cycles.

[0007] Further, the prediction of the battery aging end-point index includes: inputting the cycle test data into a pre-established multi-output Gaussian process model, and receiving the aging prediction result output by the multi-output Gaussian process model; the establishment method of the multi-output Gaussian process model includes: adding a dimension input value to the input end of a pre-established Gaussian model, where the dimension input value is the mean value corresponding to the modeling sample data of the Gaussian model; correspondingly designing the input end and output end of the Gaussian model according to the dimension input value to enable the Gaussian model to simultaneously predict multiple output variables, so as to obtain a multi-output Gaussian process model, where the input of the multi-output Gaussian process model is the feature of the target task and the added value, and the output is the aging prediction result corresponding to each added value.

[0008] Further, performing error verification on the aging prediction result and selecting a feature combination according to the verification result to obtain feature combination data, including:

[0009] Step 1: Search for all combinations of a single feature and record the verification error of each combination , where is the number of single features in this combination. In this step ; ; represents different combinations, ;

[0010] Step 2: Sort the verification errors in ascending order of their values, and select at most groups of the smallest error values corresponding feature combinations, denoted as , where is the combination corresponding to the smallest value, and so on, to obtain each feature combination; ;

[0011] Step 3: When , all combinations composed of features are . Then only select the combinations in that completely contain the features in for searching. The number of all search combinations is denoted as , and record the verification error corresponding to each combination ;

[0012] Step 4: Let in Step 2, and continue to execute Step 2 and Step 3 until when all combinations of features are searched;

[0013] Step 5: Sort the verification errors recorded for each combination in Steps 1 to 4 to obtain a sequence , and the corresponding combinations to obtain an arranged optimal combination set. When , it corresponds to the smallest verification error, and so on, to obtain the feature combination data.

[0014] Further, the prediction of the aging process trajectory of the battery using the feature combination data includes: using a pre-trained early aging trajectory predictor model to perform early aging trajectory prediction on the feature combination data to obtain an early prediction result; using a linear correction method to process the early prediction result to obtain a future aging trajectory in which the battery capacity changes with the increase of the number of cycles after correction to remove the cumulative error.

[0015] Further, the training method of the early aging trajectory predictor model includes:

[0016] Performing coordinate transformation on the battery capacity degradation curve, and introducing an intermediate variable polar angle When it corresponds to the starting point of the battery capacity degradation curve, and when It corresponds to the end point of the battery capacity degradation curve;

[0017] Designing and constructing a basic aging trajectory predictor model based on a long short-term memory recurrent neural network, and the model form is:

[0018]

[0019] Among them, among them Is the predicted capacity trajectory sequence of a single battery cell , which is composed of a cycle number sequence Obtained according to the polar angle And a capacity sequence Composed, and the predicted capacity trajectory sequence is the predicted output of the basic aging trajectory predictor model;

[0020] Using pre-selected source domain data to perform hyperparameter optimization on the cross-validation of the basic aging trajectory predictor model, and training based on the source domain data to obtain an early aging trajectory predictor model.

[0021] Further, the coordinate transformation of the battery capacity degradation curve includes:

[0022] Introducing an intermediate variable polar angle , and converting the battery capacity degradation curve from the "capacity-cycle number" coordinate system to the "capacity-polar angle " coordinate system and the "cycle number-polar angle " coordinate system, and sampling within the Range of the polar angle To obtain a cycle number sequence And a capacity sequence .

[0023] Further, the processing of the early prediction result using the linear correction method includes: correcting the cycle life sequence According to the cycle life output by the multi-output Gaussian process modelTo remove the cumulative error. Then, for the predicted cycle number sequence and the capacity sequence perform coordinate inverse transformation, using the cycle number sequence as the abscissa and the capacity sequence as the ordinate, and eliminate the introduced intermediate variable, the polar angle to obtain the predicted target battery capacity degradation curve trajectory.

[0024] A second aspect of the present invention provides a prediction system for battery aging trajectory, including: a cycle test module for testing the battery for a predetermined number of cycles to obtain cycle test data; an aging end early prediction module for predicting battery aging end indicators based on the cycle test data to obtain an aging prediction result, where the aging prediction result includes the cycle life and life capacity of the battery; a feature combination module for verifying the error of the aging prediction result and selecting a feature combination according to the verification result to obtain feature combination data; an aging prediction module for using the feature combination data to predict the aging process trajectory of the battery to obtain the future aging trajectory of the battery capacity changing with the increase of the cycle number.

[0025] A third aspect of the present invention provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the prediction method for battery aging trajectory according to any one of the above is implemented.

[0026] A fourth aspect of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the prediction method for battery aging trajectory according to any one of the above is implemented.

[0027] The present invention provides a prediction method, system, electronic device, and storage medium for battery aging trajectory. The beneficial effects are as follows: The present invention proposes a new prediction technology for battery aging trajectory. This technology can well predict the future trajectory of the battery aging process through the battery cycle life and life capacity, as well as a unique feature combination strategy, thereby improving the accuracy of the prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative efforts.

[0029] Figure 1Flow chart of the prediction method for the battery aging trajectory in the embodiment of the present invention;

[0030] Figure 2 Joint modeling framework diagram for the early prediction of battery cycle life and aging trajectory;

[0031] Figure 3 Schematic diagram of the feature combination optimization strategy for Bayesian search;

[0032] Figure 4 Schematic diagram of the coordinate transformation of the capacity degradation curve;

[0033] Figure 5 Flow chart of the early prediction of the battery based on the joint machine learning modeling of MOGP and LSTM-FC RNN;

[0034] Figure 6 Schematic diagram of the verification dataset summary;

[0035] Figure 7 Schematic diagram of the optimal selection result of the feature combination

[0036] Figure 8 Schematic diagram of the MAPE of the cycle life of the top 10 groups of BCs in the optimal selection result of the feature combination;

[0037] Figure 9a Schematic diagram of the MAPE of the cycle life of the top 10 groups of BCs in the optimal selection result of the feature combination;

[0038] Figure 9b Schematic diagram of the MAPE of the life capacity of the top 10 groups of BCs in the optimal selection result of the feature combination;

[0039] Figure 10 Schematic diagram showing the training and test results of the cycle life and life capacity of three representative BCs (BC1, BC5, and BC10);

[0040] Figure 11a Schematic diagram of the first prediction result of the battery capacity degradation trajectory by the joint modeling scheme in the Prim test set;

[0041] Figure 11b Schematic diagram of the second prediction result of the battery capacity degradation trajectory by the joint modeling scheme in the Prim test set;

[0042] Figure 11c Schematic diagram of the third prediction result of the battery capacity degradation trajectory by the joint modeling scheme in the Prim test set;

[0043] Figure 11d Schematic diagram of the first prediction result of the battery capacity degradation trajectory by the joint modeling scheme in the Sec test set;

[0044] Figure 11e Schematic diagram of the second prediction result of the battery capacity degradation trajectory in the Sec test set for the joint modeling scheme;

[0045] Figure 11f Schematic diagram of the third prediction result of the battery capacity degradation trajectory in the Sec test set for the joint modeling scheme;

[0046] Figure 11g Schematic diagram of the first prediction result of the battery capacity degradation trajectory in the Add test set for the joint modeling scheme;

[0047] Figure 11h Schematic diagram of the second prediction result of the battery capacity degradation trajectory in the Add test set for the joint modeling scheme;

[0048] Figure 11i Schematic diagram of the third prediction result of the battery capacity degradation trajectory in the Add test set for the joint modeling scheme;

[0049] Figure 12 Schematic diagram of the average error of the cycle number corresponding to the predicted capacity degradation curve;

[0050] Figure 13 Frame diagram of the prediction system for the battery aging trajectory in the embodiment of the present invention;

[0051] Figure 14 Schematic block diagram of the structure of the electronic device in the embodiment of the present invention. Detailed implementation manners

[0052] To make the objectives, features, and advantages of the present invention more obvious and understandable, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present invention.

[0053] Accurate early prediction of battery aging is helpful for new product optimization and end - application management. Based on such a background, in view of the significant differences in the aging processes of battery cells caused by conditions such as battery manufacturing tolerances, fast charging, and high - current discharging, the present invention provides a joint modeling scheme for decoupling the interaction between battery inconsistency differences and non - linear aging, which can accurately and adaptively predict the cycle life and future aging trajectories of batteries under different working conditions to solve one or more of the following problems:

[0054] (1) As the design parameters increase, battery development is time-consuming and expensive. The traditional development process of new batteries requires continuous trial and error of manual empirical knowledge and a large number of experimental accumulations. Aging tests take months to years. An early prediction model method is needed to achieve early prediction of batteries, thereby shortening the battery development cycle and saving costs.

[0055] (2) Under the existing methods, accurate prediction of the aging trajectory depends on the availability of cycle data. Early cycle data can predict the aging end-point indicators, but it is difficult to meet the requirements of rapid optimization of battery products and end-use management.

[0056] (3) There is a strong internal correlation between different aging variable indicators of the battery. If co-modeling prediction of these variable indicators can be achieved, it will provide a more accurate and efficient prediction model and prediction results.

[0057] (4) The feature combinations required by the prediction model currently rely more on manual experience and battery mechanism knowledge. For data-driven methods, features are selected because of their prediction ability rather than their physical meaning. Excellent feature combinations are crucial for the performance of the prediction model. An efficient optimal feature combination strategy needs to be provided for the selection of high-dimensional feature combinations.

[0058] (5) For the battery aging prediction model based on machine learning, due to the feature gap between the source domain and the target domain, the prediction model trained well in the source domain has difficult-to-meet requirements in terms of prediction performance when generalized to the target domain applications under multiple working conditions.

[0059] Please refer to Figure 1 for a prediction method of a battery aging trajectory provided by the present invention, including:

[0060] S101. Test the battery for a predetermined number of cycles to obtain cycle test data;

[0061] S102. Predict the battery aging end-point indicators based on the cycle test data to obtain an aging prediction result, where the aging prediction result includes the cycle life and life capacity of the battery;

[0062] S103. Verify the error of the aging prediction result and select a feature combination according to the verification result to obtain feature combination data;

[0063] S104. Use the feature combination data to predict the aging process trajectory of the battery to obtain the future aging trajectory of the battery capacity changing with the increase in the number of cycles.

[0064] A new battery aging trajectory prediction technology proposed by the present invention can well predict the aging process trajectory of the battery by means of the battery cycle life, life capacity, and a unique feature combination strategy, thereby improving the accuracy of the prediction results.

[0065] When specifically implemented, as Figure 2 shown, the solution proposed by the present invention. First, an end-to-end early aging endpoint predictor is constructed based on the multiple output Gaussian process (MOGP) method for predicting battery aging endpoint indicators. Second, based on the long short-term memory (LSTM) network and combined with the output of the early aging endpoint predictor, an aging process trajectory predictor in the "prompt learning" paradigm is designed for predicting the non-linear aging process trajectory of the battery. Then, the present invention proposes a feature combination optimization search strategy based on the Bayesian principle to provide the optimal feature combination input for the predictor model. Finally, based on the early cycle data features, the aging endpoint predictor provides a "prompt" input for the aging trajectory predictor to achieve accurate and adaptive prediction of the battery cycle life and future aging trajectory.

[0066] Specifically, in one embodiment, predicting the battery aging endpoint indicators includes: inputting the cycle test data into a pre-established multiple output Gaussian process model and receiving the aging prediction results output by the multiple output Gaussian process model.

[0067] The method for establishing the multiple output Gaussian process model includes: adding a dimension input value to the input end of the pre-established Gaussian model, where the dimension input value is the mean corresponding to the modeling sample data of the Gaussian model; correspondingly designing the input end and output end of the Gaussian model according to the dimension input value to enable the Gaussian model to simultaneously predict multiple output variables, obtaining the multiple output Gaussian process model, where the input of the multiple output Gaussian process model is the features of the target task and the added value, and the output is the aging prediction result corresponding to each added value.

[0068] Specifically, in this embodiment, the MOGP method is used to establish an early aging endpoint predictor for the inconsistency differences between battery cells. MOGP is modeled based on the classical Gaussian process (GP) method and realizes the simultaneous prediction of multiple output variables by adding a dimension information at the input end. The MOGP model established in this study is an improvement of the traditional single output GP to achieve knowledge transfer and learning across multiple output variables.

[0069] Suppose it is necessary to model with the same input features and simultaneously map to variables , then the input and output of the model are marked as ,in , is the number of data samples related to modeling. The modeling steps are as follows:

[0070] (1) By adding a dimension input at the input Realize the multi-output function of the GP model, where According to the output variable The range is determined by taking the mean value of the modeling sample data as the input value of the added dimension:

[0071]

[0072] (2) Adding dimensions The value indicates the corresponding output , and then improve the design of the model input, the input of the GP model training and output for:

[0073]

[0074] in, For input, is the output.

[0075] (3) The trained GP model inputs the characteristics of the target task and add value , that is, input ,when The predicted value of the model output corresponding to different values ​​is the corresponding , thus realizing the function of MOGP. Constructing MOGP prediction model based on traditional GP method:

[0076]

[0077] The outputs in the present invention are respectively the cycle life of the battery and life capacity ,therefore and . Using the same input features , set the additional input to ,Right now Corresponding to output cycle life , Corresponding to output life capacity .

[0078] In one embodiment, for a data-driven method, features are selected because of their predictive ability rather than their physical meaning. An excellent feature combination is crucial for the performance of a prediction model. Aiming at the deficiency of the traditional method of selecting feature combinations based on manual experience and trial-and-error, the present invention proposes an optimization strategy for feature combination based on Bayesian search. The corresponding solution is as Figure 3 shown. Suppose there are initially extracted original features, denoted as , , ……, . Specifically, the error verification is performed on the aging prediction results, and the feature combinations are selected according to the verification results to obtain the feature combination data, including:

[0079] Step 1: Search for all combinations of a single feature , record the verification error corresponding to each combination . Among them, is the number of single features in this combination. In this step, ; represents different combinations, ;

[0080] Step 2: Sort the verification errors in ascending order, and take at most groups of the smallest error values corresponding feature combinations, denoted as . Among them, is the smallest when is the value corresponding to the combination , and so on to obtain each feature combination;

[0081] Step 3: When , all combinations composed of features are . Then only select the combinations that completely contain the features in for search. The number of all search combinations is denoted as , record the verification error corresponding to each combination ;

[0082] Step 4: Let in Step 2, and continue to execute Step 2 and Step 3 until when all combinations of features are searched;

[0083] Step 5: Sort the verification errors corresponding to each combination recorded in Steps 1 to 4 to obtain a sequence , and the corresponding combinations , to obtain an arranged optimal combination set. When , it corresponds to the minimum verification error, and so on, to obtain feature combination data.

[0084] In one embodiment, using the feature combination data for predicting the aging process trajectory of a battery includes: using a pre-trained early aging trajectory predictor model to perform early aging trajectory prediction on the feature combination data to obtain an early prediction result; using a linear correction method to process the early prediction result to obtain a future aging trajectory in which the battery capacity changes with the increase in the number of cycles after correction to remove the cumulative error.

[0085] The training method of the early aging trajectory predictor model includes: performing coordinate transformation on the battery capacity degradation curve, and introducing an intermediate variable, the polar angle corresponds to the starting point of the battery capacity degradation curve when corresponds to the ending point of the battery capacity degradation curve when

[0086] Designing and constructing a basic aging trajectory predictor model based on a long short-term memory recurrent neural network. The model form is:[[]]

[0087]

[0088] wherein is the predicted capacity trajectory sequence of a single battery cell , which is composed of a cycle number sequence obtained according to the polar angle and a capacity sequence . The predicted capacity trajectory sequence is the predicted output of the basic aging trajectory predictor model.

[0089] Using pre-selected source domain data to perform hyperparameter optimization on the cross-validation of the basic aging trajectory predictor model, and training based on the source domain data to obtain an early aging trajectory predictor model.

[0090] In this embodiment, as Figure 4 shown Figure 4 is the coordinate transformation of the capacity degradation curve. Performing coordinate transformation on the battery capacity degradation curve includes: introducing an intermediate variable, the polar angle , converting the battery capacity degradation curve from the "capacity - cycle number" coordinate system to the "capacity - polar angle " coordinate system and the "cycle number - polar angle " coordinate system, and sampling within the range of the polar angle to obtain a cycle number sequence and a capacity sequence .

[0091] Among them, consists of as the "hint" information and the sampled polar angle sequence , that is .

[0092] As Figure 5 shown, for the prediction method of the battery aging trajectory provided by the present invention, the aging end early predictor and the aging trajectory early predictor model are jointly used for the early prediction of the target battery. The features extracted from the early cycle data of the target battery are input, and the aging end early predictor outputs the cycle life and the life capacity of the battery; then, the sum of the predicted cycle life and the polar angle sequence is used as the input of the aging trajectory early predictor, and finally the predicted cycle number sequence and the capacity sequence are obtained.

[0093] Finally, the linear correction method is used to process the early prediction results. The cycle number sequence is linearly corrected according to the cycle life output by the multi-output Gaussian process model to remove the cumulative error. Then, coordinate inverse transformation is performed on the predicted cycle number sequence and the capacity sequence . Using the cycle number sequence as the abscissa and the capacity sequence as the ordinate, the intermediate variable polar angle introduced is eliminated, and finally the predicted capacity degradation curve trajectory of the target battery is obtained.

[0094] In one embodiment, after the "feature combination optimization strategy based on Bayesian search", that is, after step S103, for the problem of insufficient prediction accuracy caused by the gap between the source domain and target domain features, the prediction method of the battery aging trajectory proposed by the present invention further includes a feature alignment scheme to narrow the feature gap between the target domain and the source domain.

[0095] Specifically, first, battery samples from the target domain that cycle to failure are required. The cycle life and features of a single cell are respectively represented as and . Next, select the two samples in the source domain that are closest to , and represent the corresponding cycle lives of these two samples as and (where ), and the corresponding features are represented as and Then, the monomers in the target domain are calculated by interpolation method. Feature gap to source domain :

[0096]

[0097]

[0098] Single feature The offset value Obtained by averaging:

[0099]

[0100] Finally, for any single feature in the target domain , new and better features can be generated through alignment operations :

[0101]

[0102] After processing, the features Replace the original to complete the operation.

[0103] In one embodiment, the method for predicting battery aging trajectory provided by the present invention further includes a step of correcting the preliminary prediction.

[0104] Specifically, since the main framework proposed by the present invention is composed of an aging endpoint predictor and an aging trajectory predictor connected in series, the present invention uses a linear correction method to correct the initial prediction for the error accumulation problem inherent in the model series connection. The linear correction is based on the predicted cycle life of the aging endpoint predictor. calculate.

[0105] First, the correction factor is defined as:

[0106]

[0107] in and They are the cycle life predicted by the endpoint predictor model (MOGP) and trajectory predictor model (LSTM-FC), Corresponding to A single battery.

[0108] Then, we get a calibration vector :

[0109]

[0110] Finally, for the first The cycle number sequence of the battery , the calibration sequence is as follows:

[0111]

[0112] Obtained after correction is the linear correction of the output of the aging process trajectory predictor . This step can be used after the final prediction output of the aging process trajectory predictor. Whether to use it depends on the size of the series structure error. When the series error is large, this linear correction scheme can be used to eliminate the error.

[0113] Therefore, the battery aging trajectory prediction method proposed in the present invention first proposes an early predictor model for battery aging end-point indicators based on a multi-output Gaussian process, which can achieve simultaneous accurate prediction of multiple end-point indicator variables of aging, and is more efficient and superior than separately modeling and predicting multiple indicator variables;

[0114] Secondly, based on the neural network method, an aging process trajectory predictor model in the 'prompt learning' paradigm is proposed, which is jointly modeled and predicted with the aging end-point predictor, can achieve accurate prediction of the battery aging process trajectory, and can effectively avoid the cumulative error caused by the series model by combining with the linear correction method;

[0115] Then, a feature optimization scheme based on Bayesian search is proposed. For the combination selection of high-dimensional features, it can avoid relying on artificial experience and battery mechanism knowledge, and efficiently and accurately obtain the optimal feature combination;

[0116] Finally, a feature alignment strategy is proposed, which can effectively reduce the feature gap between the source domain and the target domain and significantly improve the prediction performance of battery aging modeling.

[0117] To better prove the technical effects of the present invention, in an embodiment, the battery aging trajectory prediction method is also verified as follows:

[0118] This embodiment uses the battery dataset publicly available from the Massachusetts Institute of Technology and Stanford University to verify the beneficial effects. This dataset consists of 169 commercial lithium iron phosphate batteries produced by A123 Systems, divided into four batches, tested starting from different dates, with 40 - 45 batteries in each batch. Table 1 lists the key technical indicators of these batteries. The experiment aims to explore the battery capacity degradation caused by complex aging mechanisms and manufacturing variability under different fast - charging conditions. All batteries were cycled under 81 different fast - charging conditions (two - step fast charging for batches 1 - 3, four - step fast charging for batch 4, charging to 80% state of charge in about 10 minutes), but the discharge conditions were the same (4C to 2.0V, where 1C is 1.1A). The cycle life of the tested batteries varied from 150 to 2300, and the cycle life was defined as the number of cycles before the rated capacity (1.1Ah) decreased by 20%. Although under the same cycling conditions, due to the high current rate amplifying manufacturing defects, the batteries still showed significant degradation variability.

[0119] Table 1 Technical indicators of the experimental batteries

[0120]

[0121] In this embodiment, only the training dataset ( Figure 6 the "Train" training set in Figure 6 is used to develop and train all models. The other three test datasets ( Figure 6 the "Prim" test set, "Sec" test set, and "Add" test set in

[0122] are used to comprehensively evaluate the prediction performance of the proposed joint - modeling scheme.

[0122] Verification is as follows:

[0123] 1. Performance metrics

[0124] The metrics root mean square error (RMSE), mean absolute percentage error (MAPE), absolute percentage error (APE), and absolute error (AE) are selected to evaluate the model performance. RMSE is defined as:

[0125]

[0126] where is the observed variable, is the predicted variable and is the total number of samples.

[0127] MAPE is defined as:

[0128]

[0129] APE is defined as:

[0130]

[0131] AE is defined as:

[0132]

[0133] The definitions of all variables are as above.

[0134] 2.1. Performance of the Endpoint Prediction Model Based on Artificial Feature Combinations

[0135] To verify the performance of the early prediction of the aging endpoint of the present invention, comparisons were made with both the dataset benchmark model and the traditional SOGP method. Two SOGP models were constructed to predict cycle life and life capacity respectively, but all other optimization procedures were consistent with MOGP. The present invention maintained the same feature combination ([F1, F2, F3, F4, F7, F8]) of the "Discharge" model of this dataset. Figure 7 The training and prediction results of cycle life and life capacity are shown, and Table 2 summarizes the numerical errors.

[0136] Table 2 Numerical Results of Different Models

[0137]

[0138] In Table 2, one battery in the "Prim" test set failed quickly and did not match other observed patterns. Therefore, the "Prim" test results in parentheses ( ) correspond to excluding this battery. "+" indicates the sum and average operation.

[0139] For the "Prim" and "Sec" test sets with the same two-step fast charging protocol as the "Tran" training set in the source domain, the RMSE of the SOGP model for predicting cycle life is 134 cycles, and the MAPE is 8.01%. Under fair comparison, this result is better than the 9.4% MAPE of the "Discharge" model proposed by the dataset benchmark model and also better than the 9.1% RMSE of the "Full" model with more features. The RMSE and MAPE of the SOGP model for predicting life capacity are 146 Ah and 7.97% respectively. In addition, the MAPE of the MOGP model of the present invention for predicting cycle life and life capacity are 7.87% and 7.90% respectively, which are better than the SOGP model, and the RMSE indicators are the same, as shown in Table 2. These results prove the correctness and feasibility of the MOGP early prediction model and optimization strategy of the present invention, as well as the superior performance of the MOGP model.

[0140] However, for the "Add" test set with a four-step fast charging protocol, whether modeled by the SOGP or MOGP method, based on this artificial feature combination, relatively large prediction errors will occur. As shown in Table 2, the MAPE of both cycle life and life capacity exceeds 30%, which is unacceptable. This result illustrates the gap between the target and source domain units, mainly caused by the differences in fast charging protocols. Although the same type of battery is described in Table 1, due to the inappropriate feature combination, the problem of insufficient generalization ability of the prediction model is exposed.

[0141] 2.2. Performance of the end-point prediction model based on optimized feature combinations

[0142] After demonstrating the superior performance of the MOGP model in Section 2.1 above, Section 2.2 focuses on feature optimization to improve the modeling generalization ability. The corresponding techniques include: a feature combination optimization strategy based on Bayesian search and a feature alignment scheme.

[0143] Execute a feature combination optimization program based on Bayesian search on the MOGP model, verify the error by selecting the maximum mean absolute error (Absolute error, AE) of the cycle life, and sort the feature combinations according to AE. Figure 8 The top 10 best feature combinations (Best feature combination, BC) after optimization are shown. Figures 9(a) and 9(b) are the predicted MAPE of cycle life and life capacity corresponding to the top 10 BCs respectively. First, for the "Prim" and "Sec" test sets that are the same as the training set with a two-step fast charging protocol, the errors of 7 BCs are less than 8%, and the maximum error of the top 10 BCs does not exceed 12%. The present invention emphasizes the error metrics of these two test sets because these BCs are better than the benchmarks without relying on manual selection, and thus represent a progress in data-driven modeling.

[0144] Then, for the "Add" test set with a four-step fast charging protocol, the errors of 3 BCs are less than 15%, and the errors of the other 7 BCs are greater than 20%. The problem is that the BCs that are more suitable for the "Add" test set will lead to an increase in errors in the "Sec" test set. This indicates that it is limited to ensure that the prediction model establishes cross-domain mapping relationships only based on the combination of original individual features, or at least it is a labor-intensive exploration process. For this reason, the present invention proposes a feature alignment method to transform the original features. Taking the "Add" test set as an example, we randomly select 5 out of 45 battery cells to calculate the feature offset of the source domain. Then, we align the original individual features. This method makes the errors of the top 10 BCs of the "Add" test set lower than 15%, and the deviation of 5 BCs lower than 10%, which is a very satisfactory result.

[0145] Figure 10 The training and test results of the cycle life and life capacity of three representative BCs (BC1, BC5, and BC10) are shown. Table 3 summarizes the numerical errors. For the "Prim" and "Sec" test sets, the errors in the cycle life and life capacity of BC1 combined with five individual features can reach 6.75% and 6.67% respectively. This indicates that increasing the number of features to improve the prediction accuracy is not necessary, as fewer but better features perform better. For the "Add" test set, the errors in the cycle life and life capacity of BC10 are 10.95% and 11.70% respectively, and reach 8.38% and 7.51% after the feature alignment operation. BC5 is overall the best, ensuring that the errors in "Prim" and "Sec" are less than 7%, and the "Add" after feature alignment is less than 10%.

[0146] Table 3 Numerical error results of the end-point predictors of three representative BCs

[0147]

[0148] In Figure 10 the top 10 best feature combinations, each column is a set of feature BCs, the grid filling color indicates that the feature is selected, and the blank is the opposite. The shading of the color reflects the maximum validation error.

[0149] Figure 10 The cycle life and life capacity prediction results of three representative BCs based on the MOGP aging end-point predictor are also listed.

[0150] 2.3. Performance of the trajectory prediction model

[0151] In this section, the life predicted by the MOGP end-point predictor model based on BC5 in the previous section is selected as a hint. The prediction results under three conditions of priori (P), uncorrected (UC), and corrected (C) are used to verify the performance of the trajectory predictor of the present invention. Priori prediction means using the true life capacity. In contrast, uncorrected prediction is based on the life capacity predicted by MOGP, while corrected prediction means using the MOGP predicted cycle life to further linearly correct the preliminary prediction results. Based on the 95% confidence level of the life capacity predicted value of the MOGP model, the prediction trajectory uncertainty is obtained. In addition to the cycle life at the failure threshold, the trajectory prediction results are also evaluated according to the cycle number error corresponding to the capacity degradation interval, where the capacity interval is equally divided into 1000 parts.

[0152] According to the cycle life, after feature alignment, three representative units are selected in the "Prim", "Sec" test sets and the "Add" test set respectively. Figure 11a - Figure 11ishows their trajectory prediction results. For all test units, Figure 12 and Table 4 respectively illustrate the cycle number error distribution and numerical results corresponding to the predicted capacity trajectory. First, the prior prediction results of each unit fit well with its actual capacity curve. For all test sets, the MAPE of the cycle life is 2.13%, and the average MAPE of the cycle number corresponding to the predicted curve is 3.58%. This demonstrates the great potential of the early trajectory predictor of the present invention, which can predict the aging trajectory accurately enough once there is a hint input with small errors. Then, there are only intuitive small differences between the uncorrected and corrected prediction curves. Important information such as the capacity degradation trend, inflection point, and end of life can be correctly reflected in the prediction curve. The MAPE of the cycle life is 8.13% and 8.03% respectively, that is, only a cumulative error of 0.10% is generated. This error can be eliminated by linear correction. As shown in Table 4, for three different data sets, the capacity trajectory prediction error is always very low. Generally speaking, the average MAPE of the cycle number corresponding to the predicted curve is 6.88% and 7.09% respectively. It should be emphasized that such results were predicted based on the first 100 cycle data when the battery had not yet shown capacity degradation. The joint model of the present invention provides strong adaptability and generalization ability for the early prediction of batteries and accurately predicts their capacity aging trajectory.

[0153] Table 4 Numerical results of the error in predicting the capacity degradation trajectory using the proposed joint modeling scheme

[0154]

[0155] The prediction method of the battery aging trajectory proposed by the present invention can be used for the online monitoring and evaluation of the performance state of new energy vehicle batteries: In the model proposed by the present invention, first, the optimal models of the aging end predictor and the aging process trajectory predictor are independently developed offline based on the source domain data respectively, and then the developed independent models are jointly embedded into the on-vehicle battery management system to realize online application. With the accumulation of online measurement data, the early prediction of the cycle life and aging trajectory is realized;

[0156] It can also be used for the development, testing, and inspection of battery products: The joint scheme proposed by the present invention is applicable to the early prediction of the aging behavior characteristics of batteries under different factor conditions including temperature, charge and discharge conditions, battery materials, etc., so as to be able to analyze the influence relationship of different factors on battery aging only through a small amount of cycle data, without time-consuming and laborious test tests, saving costs for battery development.

[0157] It can also be used to adapt to the future development trends of cloud technology and the Internet of Things: The basic part of the model of the present invention can be shared and migrated. Even if the charging and discharging methods of the battery are different, the model can be migrated and adapted. Therefore, the vehicle only needs to measure the charging and discharging data during the battery usage process online and upload it to the cloud. All vehicles share the model parameters, and with a large amount of working condition data and the powerful computing and storage capabilities of cloud computing, a more accurate prediction effect can be achieved.

[0158] Therefore, the battery aging trajectory prediction method proposed by the present invention has universality and is applicable not only to the battery management system of electric vehicles, but also to other devices that need to use batteries, such as electric mining vehicles, drones, laptops, etc., and can also evaluate and predict the aging performance and health status of their batteries.

[0159] Please refer to Figure 13 , in an embodiment, the present invention further provides a battery aging trajectory prediction system, including: a cycle test module 1, an early aging end prediction module 2, a feature combination module 3, and an aging prediction module 4; the cycle test module 1 is used to test the battery for a predetermined number of cycles to obtain cycle test data; the early aging end prediction module 2 is used to predict the battery aging end index according to the cycle test data to obtain an aging prediction result, and the aging prediction result includes the cycle life and life capacity of the battery; the feature combination module 3 is used to verify the error of the aging prediction result and select a feature combination according to the verification result to obtain feature combination data; the aging prediction module 4 is used to use the feature combination data to predict the battery aging process trajectory to obtain the future aging trajectory of the battery capacity changing with the increase in the number of cycles.

[0160] In an embodiment, the early aging end prediction module 2 includes a multi-output Gaussian process model training unit and an aging result prediction unit: input the cycle test data into a pre-established multi-output Gaussian process model and receive the aging prediction result output by the multi-output Gaussian process model;

[0161] The multi-output Gaussian process model training unit includes: an input end design sub-unit and an output end design sub-unit; the input end design sub-unit is used to add a dimension input value to the input end of the pre-established Gaussian model, and the dimension input value is the mean value corresponding to the modeling sample data of the Gaussian model; the output end design sub-unit is used to correspondingly design the input end and output end of the Gaussian model according to the dimension input value to enable the Gaussian model to simultaneously predict multiple output variables to obtain a multi-output Gaussian process model, and the input of the multi-output Gaussian process model is the feature of the target task and the added value, and the output is the aging prediction result corresponding to each added value.

[0162] The feature combination module 3 includes: a first search unit, a sorting unit, a recording unit, a second search unit, and a feature combination data analogizing unit;

[0163] The first combined search unit is used to search for combinations of all individual features and record each combination corresponding verification error , where is the number of individual features of the combination. In this step ; represents different combinations, ;

[0164] The sorting unit is used to sort the values of the verification error from smallest to largest, and take at most groups of the smallest error values corresponding feature combinations, denoted as , where is the smallest when the value corresponds to the combination , and so on, to obtain each feature combination;

[0165] The recording unit is used when , all combinations composed of features are , then only select the combinations that completely contain the features in for search. The number of all search combinations is denoted as , and record each combination corresponding verification error ;

[0166] The second search unit is used to make in the sorting unit and continue to execute Step 2 and Step 3 until all combinations of features are searched;

[0167] The feature combination data analog unit is used to sort the verification errors corresponding to each combination recorded by the first search unit, the sorting unit, the recording unit, and the second search unit to obtain the sequence , and the corresponding combinations to obtain the sorted optimal combination set. When corresponds to the smallest verification error, and so on, to obtain the feature combination data.

[0168] The aging prediction module 4 includes: an early prediction unit and an aging trajectory prediction unit; the early prediction unit is used to perform early aging trajectory prediction on the feature combination data using a pre-trained early aging trajectory predictor model to obtain an early prediction result; the aging trajectory prediction unit is used to process the early prediction result using a linear correction method to obtain a future aging trajectory in which the battery capacity changes with the increase in the number of cycles after correction to remove the cumulative error.

[0169] The aging prediction module 4 further includes an early aging trajectory predictor model training unit, and the early aging trajectory predictor model training unit includes: a coordinate transformation subunit, a basic model modeling unit, and a hyperparameter optimization unit;

[0170] The coordinate transformation subunit is used to perform coordinate transformation on the battery capacity degradation curve, introducing an intermediate variable polar angle The starting point of the battery capacity degradation curve corresponding to the battery when The end point of the battery capacity degradation curve corresponding to the battery when

[0171] The basic model modeling unit is used to design and construct a basic aging trajectory predictor model based on a long short-term memory recurrent neural network, and the model form is:

[0172]

[0173] Wherein, Is the predicted capacity trajectory sequence of the battery cell , which is composed of a cycle number sequence Obtained according to the polar angle And a capacity sequence The predicted capacity trajectory sequence is the predicted output of the basic aging trajectory predictor model;

[0174] The hyperparameter optimization unit is used to perform hyperparameter optimization on the cross-validation of the basic aging trajectory predictor model using pre-selected source domain data and train based on the source domain data to obtain an early aging trajectory predictor model.

[0175] The coordinate transformation subunit is specifically used to introduce an intermediate variable polar angle , converting the battery capacity degradation curve from the "capacity - cycle number" coordinate system to the "capacity - polar angle " coordinate system and the "cycle number - polar angle " coordinate system, and sampling within the Range of the polar angle To obtain a cycle number sequence And a capacity sequence .

[0176] The aging trajectory prediction unit is specifically used to perform On the predicted cycle number sequence Perform coordinate inverse transformation with the cycle number sequence as the abscissa and the capacity sequence as the ordinate to eliminate the intermediate variable polar angle , and obtain the predicted target battery capacity degradation curve trajectory.

[0177] The embodiment of the present application also provides an electronic device. Please refer to Figure 14 . The electronic device includes: a memory 601, a processor 602, and a computer program stored on the memory 601 and executable on the processor 602. When the processor 602 executes the computer program, the prediction method of the battery aging trajectory described above is implemented.

[0178] Furthermore, the electronic device further includes: at least one input device 603 and at least one output device 604.

[0179] The above-mentioned memory 601, processor 602, input device 603, and output device 604 are connected through a bus 605.

[0180] Among them, the input device 603 can specifically be a camera, a touch panel, a physical button, or a mouse, etc. The output device 604 can specifically be a display screen.

[0181] The memory 601 can be a high-speed random access memory (RAM, Random Access Memory) or a non-volatile memory, such as a disk memory. The memory 601 is used to store a set of executable program codes, and the processor 602 is coupled to the memory 601.

[0182] Furthermore, the embodiment of the present application also provides a computer-readable storage medium. The computer-readable storage medium can be disposed in the electronic device in the above-mentioned embodiments, and the computer-readable storage medium can be the memory 601 described above. A computer program is stored on the computer-readable storage medium, and when the program is executed by the processor 602, the prediction method of the battery aging trajectory described in the foregoing embodiments is implemented.

[0183] Furthermore, the computer-readable storage medium can also be various media such as a USB flash drive, a mobile hard disk, a read-only memory 601 (ROM, Read-Only Memory), a RAM, a magnetic disk, or an optical disc that can store program codes.

[0184] In several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of modules is only a logical function division. In actual implementation, there can be other division methods. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling, direct coupling, or communication connection to each other can be through some interfaces. The indirect coupling or communication connection of devices or modules can be in electrical, mechanical, or other forms.

[0185] The modules described as separate components may or may not be physically separated. The components shown as modules may or may not be physical modules, that is, they can be located in one place or distributed to multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0186] In addition, in each embodiment of the present invention, the functional modules can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software functional modules.

[0187] If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in each embodiment of the present invention. It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily all essential to the present invention.

[0188] In the above embodiments, the descriptions of each embodiment have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0189] The above is the description of a method, system, electronic device, and storage medium for predicting the aging trajectory of a battery. For those skilled in the art, according to the idea of the embodiments of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for predicting the aging trajectory of a battery, characterized in that, Including: Testing the battery for a predetermined number of cycles to obtain cycle test data; Predicting the battery aging end-point index based on the cycle test data to obtain an aging prediction result, where the aging prediction result includes the cycle life and life capacity of the battery; Verifying the error of the aging prediction result and selecting a feature combination according to the verification result to obtain feature combination data; Using the feature combination data to predict the aging process trajectory of the battery to obtain a future aging trajectory of the battery capacity changing with the increase of the cycle number; The verifying the error of the aging prediction result and selecting a feature combination according to the verification result to obtain feature combination data includes: Step 1: Search for all combinations of individual features and record each combination along with the corresponding verification error , where is the number of individual features in the combination. In this step ; represents different combinations, ; Step 2: Sort the values of the verification error in ascending order, and take at most groups of the minimum error values and their corresponding feature combinations, denoted as , where is the minimum value and its corresponding combination , and so on, to obtain each feature combination; Step 3: When occurs, all combinations formed by features are . Among them, if completely contains features in , only the combinations that meet this condition are selected for search. The number of all search combinations is denoted as , and the verification error corresponding to each combination is recorded; Step 4: Let the in Step 2, and continue to execute Step 2 and Step 3 until is reached, and all combinations of features are searched for; Step Five: Record each combination in Steps One to Four corresponding verification error Perform sorting to obtain a sequence , as well as the corresponding combination , to obtain an arranged optimal combination set. When , it corresponds to the minimum verification error, and so on, to obtain feature combination data.

2. The method for predicting the battery aging trajectory according to claim 1, characterized in that The predicting the battery aging end-point index includes: Inputting the cycle test data into a pre-established multi-output Gaussian process model and receiving the aging prediction result output by the multi-output Gaussian process model; The method for establishing the multi-output Gaussian process model includes: Adding a dimension input value to the input end of a pre-established Gaussian model, where the dimension input value is the mean value corresponding to the modeling sample data of the Gaussian model; Correspondingly designing the input end and output end of the Gaussian model according to the dimension input value to enable the Gaussian model to simultaneously predict multiple output variables, thereby obtaining a multi-output Gaussian process model. The input of the multi-output Gaussian process model is the feature of the target task and the added value, and the output is the aging prediction result corresponding to each added value.

3. The method for predicting the battery aging trajectory according to claim 1, characterized in that The using the feature combination data to predict the aging process trajectory of the battery includes: Using a pre-trained aging trajectory early predictor model to perform early prediction of the aging trajectory on the feature combination data to obtain an early prediction result; Using a linear correction method to process the early prediction result to obtain a future aging trajectory of the battery capacity changing with the increase of the cycle number after correcting and removing the cumulative error.

4. The method for predicting the battery aging trajectory according to claim 3, characterized in that The training method of the aging trajectory early predictor model includes: Perform coordinate transformation on the battery capacity degradation curve. The introduced intermediate variable, the polar angle corresponds to the starting point of the battery capacity degradation curve when corresponds to the ending point of the battery capacity degradation curve when Designing and constructing an aging trajectory predictor basic model based on a long short-term memory recurrent neural network, and the model form is: Among them, is the predicted capacity trajectory sequence of the battery cell , which is composed of the cycle number sequence obtained according to the polar angle and the capacity sequence . The predicted capacity trajectory sequence is the predicted output of the basic model of the aging trajectory predictor; Using pre-selected source domain data to perform hyperparameter optimization on the aging trajectory predictor basic model through cross-validation and training based on the source domain data to obtain an aging trajectory early predictor model.

5. The method for predicting the battery aging trajectory according to claim 4, characterized in that The performing coordinate transformation on the battery capacity degradation curve includes: Introduce an intermediate variable, the polar angle , and convert the battery capacity degradation curve from the "capacity - cycle number" coordinate system to the "capacity - polar angle " coordinate system and the "cycle number - polar angle " coordinate system, and sample within the range of the polar angle to obtain the cycle number sequence and the capacity sequence . .

6. The method for predicting the battery aging trajectory according to claim 5, characterized in that The using a linear correction method to process the early prediction result includes: The sequence of corrected cycle numbers of the cycle life output according to the multi-output Gaussian process model to remove the cumulative error; Then, perform inverse coordinate transformation on the predicted cycle number sequence and the capacity sequence using the cycle number sequence as the abscissa and the capacity sequence as the ordinate, eliminating the intermediate variable polar angle to obtain the predicted trajectory of the target battery capacity degradation curve.

7. A prediction system for battery aging trajectory, characterized in that, Including: A cycle test module for testing the battery for a predetermined number of cycles to obtain cycle test data; An early prediction module for the end point of aging, which is used to predict the end point index of battery aging according to the cyclic test data to obtain an aging prediction result, where the aging prediction result includes the cycle life and life capacity of the battery; A feature combination module, which is used to verify the error of the aging prediction result and select a feature combination according to the verification result to obtain feature combination data; An aging prediction module, which is used to predict the aging process trajectory of the battery by using the feature combination data to obtain a future aging trajectory of the battery capacity changing with the increase of the number of cycles; The feature combination module includes: a first search unit, a sorting unit, a recording unit, a second search unit and a feature combination data analogizing unit; The first combined search unit is used to combine all individual features for searching, and record each combination corresponding verification error , where is the number of individual features of this combination. In this step ; represents different combinations, ; The sorting unit is used to sort in ascending order according to the value of the verification error and take at most groups of the smallest error values corresponding feature combinations, denoted as , where is the smallest when the value corresponds to the combination , and so on, to obtain each feature combination; The recording unit is used to when When All combinations composed of Among them If Fully included in Search for combinations of features in Record each combination The corresponding verification error ; The second search unit is used to make the in the sorting unit continue to execute Step 2 and Step 3 until all combinations of features are found; The feature combination data analog unit is used to record each combination by the first search unit, sorting unit, recording unit, and second search unit The corresponding verification error Sort them to obtain a sequence And the corresponding combinations To obtain the sorted optimal combination set. When Corresponds to The verification error is the smallest, and so on, to obtain the feature combination data.

8. An electronic device, comprising: A memory, a processor and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the method according to any one of claims 1 to 6 is implemented.

Images