Generative adversarial network model-based battery calendar aging capacity prediction method

Abstract

The invention discloses a generative adversarial network model-based battery calendar aging capacity prediction method, which belongs to the field of battery calendar aging prediction and comprises the following steps of S1, establishing a calendar aging semi-experience prediction model of a lithium ion battery; s2, establishing and training a battery calendar aging capacity prediction generative adversarial network; and S3, evaluating. According to the battery calendar aging capacity prediction method based on the generative adversarial network model, the CFGAN (capacity prediction generative adversarial network) is designed by taking electrochemical knowledge of the battery as a guidance basis, so that the knowledge and the data keep satisfactory consistency, and the theoretical strength and the prediction performance are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of battery calendar aging prediction, and in particular to a battery calendar aging capacity prediction method based on a generative adversarial network model. Background Art

[0002] Lithium-ion batteries are considered one of the most promising energy storage systems and are widely used in smart grids, electric vehicles and other related fields. However, due to the high nonlinearity of battery aging and the complexity of the actual electrochemical process, effective and accurate battery aging prediction remains a difficult task.

[0003] Lithium-ion battery aging usually includes two parts: cycle aging and calendar aging. Different modeling solutions have been studied in the prior art for battery calendar aging, which can be mainly divided into two categories: knowledge-driven models (KDMs) and data-driven models (DDMs). An effective KDM is to establish electrochemical models (EMs), which focus on the detailed description of the internal chemical mechanisms related to battery calendar aging; however, this type of KDM usually requires the processing of a large number of partial differential equations, which greatly increases the computational workload of the prediction process. Another type of KDM is the semi-empirical model (SEMs), which captures the potential mapping relationship between the dynamic process of battery calendar aging and storage conditions based on electrochemical knowledge (such as Arrhenius law). However, the open-loop nature of SEM exhibits poor generalization, so its prediction performance depends significantly on the experimental quality of the relevant calendar aging tests.

[0004] With the rapid development of machine learning and computing platforms, more and more machine learning algorithms are being applied to battery aging prediction, such as neural networks, Gaussian process regression (GPR), relevance vector machine (RVM), support vector regression (SVR) and linear quantile regression (probabilistic prediction), etc. However, existing models still face the following challenges: Challenge 1: Limited integration of knowledge and data: The first challenge of battery calendar aging prediction is that there is a non-negligible gap between the modeling paradigms of KDM and DDM. That is, KDM pays great attention to maintaining rigorous mathematical consistency with the electrochemical theory of lithium-ion battery calendar aging, but their ability to utilize large-scale measured data sets is limited, and because they rely on optimization-based parameter identification, they cannot properly handle noise or multimodality in the data set. As for DDM in the field of calendar aging prediction, prior knowledge of battery electrochemical theory is rarely considered in the modeling process. Therefore, the existing KDM and DDM are both overly biased towards domain knowledge or data statistics to a certain extent, and there is currently no way to merge KDM and DDM together.

[0005] Challenge 2: Limited approximation range: Existing multi-step probabilistic forecasting models have limited approximation capabilities, and the joint distributions they predict cannot cover all possible arbitrarily complex distributions.

[0006] Almost all existing univariate probability prediction models, including parametric models (such as autoregression, generalized autoregression conditional heteroskedasticity, deep learning-based models, etc.) and non-parametric models (such as kernel density estimation, adaptive resampling, memory-based ordered regression deep neural networks, etc.), have limited approximation capabilities and their prediction results must obey certain specific constraints. Some existing multi-step prediction models are multivariate extensions of the above univariate models (which have been shown to have limited approximation capabilities), such as vector autoregression and multivariate kernel density estimation; others are essentially combinations of several univariate models, where the joint distribution of different time steps is not fully considered.

[0007] In actual engineering, due to the nonlinear characteristics of the calendar aging process and its complex correlation with other influencing factors (such as temperature, time, SOC, etc.), there is no evidence to support that future battery capacity must conform to certain types of distribution restrictions. Summary of the invention

[0008] In order to address the gap between the KDM and DDM modeling paradigms, the present invention provides a battery calendar aging capacity prediction method based on a generative adversarial network model, and designs a knowledge + data-driven generative adversarial network model that can fully balance battery electrochemical theory and data set statistical characteristics. Since battery electrochemical knowledge provides theoretical support for calendar aging modeling, the fusion of domain knowledge and data-driven in the present invention can significantly improve the performance of calendar aging prediction; At the same time, in order to better estimate the joint distribution of the prediction target, a multivariate "complete" prediction model is required, and the distribution generated by it should be able to cover all possible distributions of arbitrary complexity. The present invention adopts a generative adversarial network to solve this challenge. Based on the ability of the generative adversarial network to learn arbitrary distributions, it can simulate arbitrarily complex multivariate distributions.

[0009] To achieve the above object, the present invention provides a battery calendar aging capacity prediction method based on a generative adversarial network model, comprising the following steps: S1. Establish a semi-empirical prediction model for calendar aging of lithium-ion batteries; S2. Establish and train a generative adversarial network for battery calendar aging capacity prediction; S3. Evaluation.

[0010] Preferably, step S1 specifically includes the following steps: S11. Set prediction tasks: (1); In the formula, is the explanatory variable; Store time for batteries; The temperature at which the battery is stored; is the stored state of charge SOC; for The capacity of the battery at all times; is the lag interval of the prediction task, for The capacity of the battery at all times; S12. Set the target variable: (2); In the formula, represents all target variables; is the maximum prediction time step, for The capacity of the battery at all times; for The capacity of the battery at all times; S13. Simultaneously establish multi-step point prediction and probability capacity prediction models: Formula (1) is abstracted into the following mapping: (3); In the formula, express In the conditions The joint distribution under is the prediction target of the calendar aging prediction model for lithium-ion batteries, based on Get the moment The marginal distribution of , which is In the conditions Under the conditional distribution, then from Infer further prediction results from S14. Construct a semi-empirical prediction model for calendar aging of lithium-ion batteries based on electrochemical principles: (4); In the formula, Indicates the nominal capacity of the battery; , represents the unknown parameters that need to be identified.

[0011] Preferably, the battery calendar aging capacity prediction generative adversarial network described in step S2 includes a generator , Discriminator and conditional ; It specifically includes the following steps: S21. Training conditioner Become a conditional information transformer to convert the explanatory variables Convert to conditional features ,Right now ; S211, Generator and the discriminator Conditioner The obtained conditional features As a condition, the training objective of the generative adversarial network for battery calendar aging capacity prediction is: (5); In the formula, Represents the optimization generator with the minimization objective ; Denotes the maximum objective optimization discriminator ; Indicates a given Lower Discriminator Output for observation probability; Indicates a given Lower Discriminator For the generator Generated fake samples Probability; Represents a false sample ;in, and Represents an observation sample and false samples The conditional distribution of S212. Solving formula (5) is equivalent to minimizing and The Jensen Shannon divergence of , formula (5) reaches the global minimum if and only if the following formula is satisfied: (6); By simultaneously inputting As additional information, For the generator , Discriminator The adjustment will and Re-record them as and , formula (5) can be rewritten as follows: (7); conditioner Will Convert to conditional feature ; If No. The components are expressed as ,and =1,2,3,4, then It is composed of: (8); (9); (10); (11); (12); In the formula, Represents a conditioner; All represent learnable parameters; S213, respectively, the conditioner , Generator and the discriminator Train, and each time the generator and the discriminator After training, the conditioner Training, Conditioner The training objectives are: (13); In the formula, Represents the optimization condition for maximizing the objective ; Indicates conditional characteristics; S214, using conditioner Execute from arrive The equivalent transformation of As a condition, it is equivalent to As a condition, we obtain: (14); (15); Then we get: (16); S22, Generator-based and conditional The prediction results are generated: Approximate sampling: Given and conditional What you get , then from the known given distribution Sampling a low-dimensional random noise , the low-dimensional random noise and Input to the generator at the same time In, get A sample of : (17); because , so Considered as At the same time, after the battery calendar aging capacity prediction generative adversarial network is trained, Considered as Sampling in ; Mean estimation based on repeated sampling: From Multiple samplings are performed and all are input into formula (17), and we get An approximate sample of ,in represents the number of samples, The empirical distribution formed is estimates; Calculation of prediction results for each time step: The point prediction and probability prediction for each prediction time step are calculated as follows; , estimated through experience The average value determines the point prediction result : (18); In the formula, express Point prediction results of ; Calculated by the following empirical algorithm The probability distribution prediction results are: Define any The empirical nominal proportions are as follows: (19); In the formula, represents the indicator function; Representation and The corresponding empirical nominal ratio; Then, build a dictionary , and get a given nominal proportion from the dictionary The quantile approximation of : (20); In the formula, Indicates nominal proportion The approximate quantile of .

[0012] Preferably, Select values ​​from {5%, 10%, 15%, …, 95%} in sequence.

[0013] Preferably, in the battery calendar aging capacity prediction generative adversarial network described in step S2, the generator and the discriminator Both use LSTM and bidirectional LSTM based on recurrent networks as the basic building layers, stack multiple BiLSTM layers in the opposite direction of LSTM, and use the hidden state of the previous BiLSTM as the input of the next BiLSTM; and use residual technology to add short-circuit links between non-adjacent neural network layers.

[0014] Preferably, a short-circuit link is added every three BiLSTM or fully connected layers.

[0015] Preferably, the evaluation in step S3 includes the following evaluation indicators: point prediction indicators and probability prediction indicators; Among them, for point prediction, the maximum absolute error and root mean square error are used as evaluation indicators, which are defined as: (twenty one); (twenty two); In the formula, and Represents the time step The maximum absolute error and root mean square error of represents the number of samples; It means taking the maximum value among all samples; Further, take and

[0016] (twenty three); In the formula, number of time steps; The probability prediction is expressed in quantile form, and the quantile prediction evaluation indicators mainly include the following two: mean proportional deviation and skill score; The average proportional deviation is the core evaluation indicator based on reliability, which is defined as follows: (twenty four); In the formula, Indicates nominal proportion and time step The average proportional deviation of The average absolute proportional deviation over all nominal proportions is taken as the average absolute proportional deviation for the prediction time step Reliability evaluation of probability prediction results: (25); In the formula, is the time step The mean absolute proportional deviation, is the total number of nominal proportions; The skill score is obtained by fusing reliability and sharpness, which is expressed as follows: (26); In the formula, Represents the time step skill score.

[0017] The present invention has the following beneficial effects: 1. The battery calendar aging capacity prediction generative adversarial network (CFGAN) is used for calendar aging point prediction and probability prediction of lithium-ion batteries. The ability of generative adversarial networks to learn arbitrarily complex distributions enables CFGAN to have the potential to generate (predict) all possible unconstrained joint distributions of arbitrary shapes; 2. CFGAN achieves a deep fusion of electrochemical knowledge and data features, takes battery calendar aging knowledge as its design basis, maintains the consistency and complementarity between knowledge and data, and its knowledge-driven + data-driven modeling paradigm significantly improves the theoretical strength and prediction performance of CFGAN; 3. CFGAN has better versatility than existing models. It can still maintain satisfactory performance under zero-sample conditions, while other existing models cannot maintain prediction accuracy, which shows that CFGAN has fully grasped the complex multimodal nature of the calendar aging prediction task.

[0018] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a framework diagram of the battery calendar aging capacity prediction method based on the generative adversarial network model of the present invention; Figure 2 It is a result diagram of the best model for predicting four points in the simulation experiment of the present invention; Figure 3 Result diagram of four optimal probability prediction models of the simulation experiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical scheme and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention are further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions.

[0021] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or inherent to these processes, methods, products or devices.

[0022] Like reference numerals and letters denote similar items in the following drawings, and thus, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0023] In the description of the present invention, it should be noted that the terms "upper", "lower", "inside", "outside", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, or are directions or positional relationships in which the product of the invention is usually placed when in use. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as a limitation on the present invention.

[0024] In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "setting", "installation" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0025] GANs, which consist of adversarially trained generators and discriminators, have recently attracted increasing attention for their powerful ability to learn arbitrarily complex data distributions, which opens up the possibility of building GAN-based prediction models whose prediction distributions can theoretically be of arbitrary shapes. This ability to learn arbitrarily complex distributions has been fully exploited in this work to overcome the limited approximation range of prediction results, where the generator of CFGAN is trained to generate the joint distribution of future capacities at different time steps, and the generated multi-dimensional distribution can theoretically cover all possible distributions.

[0026] Based on the above analysis, the present invention is designed: a battery calendar aging capacity prediction method based on a generative adversarial network model, comprising the following steps: S1. Establish a semi-empirical prediction model for calendar aging of lithium-ion batteries; Step S1 specifically includes the following steps: S11. Setting prediction tasks: (1); In the formula, is the explanatory variable; Store time for batteries; The temperature at which the battery is stored; is the stored state of charge SOC; for The capacity of the battery at all times; is the lag interval of the prediction task, for The capacity of the battery at all times; S12. Set the target variable: (2); In the formula, represents all target variables; is the maximum prediction time step, for The capacity of the battery at all times; for The capacity of the battery at all times; S13. Simultaneously establish multi-step point prediction and probability capacity prediction models: Formula (1) is abstracted into the following mapping: (3); In the formula, express In the conditions The joint distribution under is the prediction target of the calendar aging prediction model for lithium-ion batteries, based on Get the moment The marginal distribution of , which is In the conditions Under the conditional distribution, then from Infer further prediction results from S14. Construct a semi-empirical prediction model for calendar aging of lithium-ion batteries based on electrochemical principles: (4); In the formula, Indicates the nominal capacity of the battery; , represents the unknown parameters that need to be identified.

[0027] S2. Establish and train a generative adversarial network for battery calendar aging capacity prediction; like Figure 1 As shown, the battery calendar aging capacity prediction generative adversarial network described in step S2 includes a generator , Discriminator and conditional ; It specifically includes the following steps: S21. Training conditioner Become a conditional information transformer to convert the explanatory variables Convert to conditional feature ,Right now ; S211, Generator and the discriminator Conditioner The obtained conditional features As a condition, the training objective of the generative adversarial network for battery calendar aging capacity prediction is: (5); In the formula, Represents the optimization generator with the minimization objective ; Denotes the maximum objective optimization discriminator ; Indicates a given Lower Discriminator Output for observation probability; Indicates a given Lower Discriminator For the generator Generated fake samples Probability; Represents a false sample ;in, and Represents an observation sample and false samples The conditional distribution of S212. Solving formula (5) is equivalent to minimizing and The Jensen Shannon divergence of , formula (5) reaches the global minimum if and only if the following formula is satisfied: (6); By simultaneously inputting As additional information, For the generator , Discriminator The adjustment will and Re-record them as and , formula (5) can be rewritten as follows: (7); conditioner Will Convert to conditional feature , ; If No. The components are expressed as ,and =1,2,3,4, then It is composed of: (8); (9); (10); (11); (12); In the formula, Represents a conditioner; All represent learnable parameters; S213, respectively, the conditioner , Generator and the discriminator Train, and each time the generator and the discriminator After training, the conditioner Training, Conditioner The training objectives are: (13); In the formula, Represents the optimization condition for maximizing the objective ; Indicates conditional characteristics; S214, using conditioner Execute from arrive The equivalent transformation of As a condition, it is equivalent to As a condition, we obtain: (14); (15); Then we get: (16); S22, Generator-based and conditional The prediction results are generated: Approximate sampling: Given and conditional What you get , then from the known given distribution Sampling a low-dimensional random noise , the low-dimensional random noise and Input to the generator at the same time In, get A sample of : (17); because = , so Considered as At the same time, after the battery calendar aging capacity prediction generative adversarial network is trained, Considered as Sampling in ; Mean estimation based on repeated sampling: From Multiple samplings are performed and all are input into formula (17), and we get An approximate sample of ,in represents the number of samples, The empirical distribution formed is estimates; Calculation of prediction results for each time step: The point prediction and probability prediction for each prediction time step are calculated as follows; , estimated through experience The average value determines the point prediction result : (18); In the formula, express Point prediction results of ; Calculated by the following empirical algorithm The probability distribution prediction results are: Define any The empirical nominal proportions are as follows: (19); In the formula, represents the indicator function; Representation and The corresponding empirical nominal ratio; Then, build a dictionary , and get a given nominal proportion from the dictionary The quantile approximation of : (20); In the formula, Indicates nominal proportion The approximate quantile of .

[0028] In the battery calendar aging capacity prediction generative adversarial network described in step S2, the generator and the discriminator Both use LSTM and bidirectional LSTM based on recurrent networks as basic construction layers, that is, stacking multiple BiLSTM layers in the opposite direction of LSTM, and using the hidden state of the previous BiLSTM as the input of the next BiLSTM; and using residual technology to add short-circuit links between non-adjacent neural network layers. A short-circuit link is added every three BiLSTM or fully connected layers.

[0029] S3. Evaluation.

[0030] The evaluation described in step S3 includes the following evaluation indicators: point prediction indicators and probability prediction indicators; Among them, for point prediction, the maximum absolute error and root mean square error are used as evaluation indicators, which are defined as: (twenty one); (twenty two); In the formula, and They correspond to the time steps The maximum absolute error and root mean square error of represents the number of samples; It means taking the maximum value among all samples; Further, take and Average value over all steps: (twenty three); In the formula, represents the number of time steps; The probability prediction is expressed in quantile form, and the quantile prediction evaluation indicators mainly include the following two: mean proportional deviation and skill score; The average proportional deviation is the core evaluation indicator based on reliability, which is defined as follows: (twenty four); In the formula, Indicates nominal proportion and time step The average proportional deviation of The average absolute proportion deviation (APD) over all nominal proportions is taken as the average proportion deviation for the prediction time step Reliability evaluation of probability prediction results: (25); In the formula, is the time step The average absolute proportional deviation (APD) of is the total number of nominal proportions; The skill score is obtained by fusing reliability and sharpness, which is expressed as follows: (26); In the formula, Represents the time step A higher skill score means better prediction performance.

[0031] Simulation experiment Data preprocessing: The battery calendar aging capacity prediction generative adversarial network (CFGAN) was tested on real data of commercial Panasonic NCRBD batteries. The dataset contains 27 capacity aging sequences, which were tested under 9 different storage conditions (SOC and temperature), where SOC was set to {20%, 50%, 90%}, T (temperature) was set to {10°C, 25°C, 45°C}, and measured every 30 days. Each sequence contains 17 values. In each case (such as [20%, 10°C]), three capacity series are included. Before CFGAN training, the 27 capacity sequences are divided into 4 subsets, namely training set, validation set, test set 1 and test set 2; Two datasets are formulated, test set 1 and test set 2. The reason for establishing test set 2 is to test whether CFGAN can be well generalized to a condition that has never been seen before. Because a large number of samples are required to train deep networks such as CFGAN, cubic spline interpolation is used so that each sequence in the training set has a resolution of 1 minute and the length of each sequence is 691201.

[0032] Comparative test of point forecasts: Several advanced point forecasting methods are tested in the benchmark, including time series-based models and data-driven models, namely, linear vector autoregression (VAR), Gaussian process regression (GPR), multivariate deep convolutional neural network (DCNN), and multivariate deep BiLSTM (DBiLSTM) constructed by stacking multiple BiLSTM layers; among them, DBiLSTM, DCNN, and CFGAN are based on deep learning, while the others are not.

[0033] Table 1 Performance comparison results of point prediction

[0034] As shown in Table 1, CFGAN performs the best, achieving the lowest MAE and RMSE for all time steps on both test sets. It also shows that: (1) VAR and GPR both have satisfactory performance in 1-step and 2-step prediction. However, as the time step increases, their prediction performance gradually deteriorates (especially VAR). The reason is that the deep structure and high nonlinearity of DBiLSTM, DCNN and CFGAN enable them to more accurately predict the joint distribution of the target, while the relatively shallow structure of GPR and VAR makes their prediction effect poor. (2) Although DBiLSTM, DCNN and CFGAN are all models based on deep learning, the results show that DCNN performs worse than the other two, which means that the recurrent structure may be more suitable than the convolutional structure for time series tasks. (3) Among the best four models (DBiLSTM, DCNN, CFGAN, GPR), the performance of CFGAN on test set 2 is close to that on test set 1, while the performance of other models on test set 2 is significantly lower than that on test set 1, indicating that the performance of other models is limited and the zero-sample generalization ability is insufficient.

[0035] Comparative test of probabilistic forecasts: In this test, several cutting-edge parametric and non-parametric models were benchmarked. (1) Probabilistic generalization of the persistence model; (2) Normal distribution generalization of VAR; (3) Linear quantile regression (LQR); (4) Multivariate version of Deep Autoregressive Recurrent Network (DeepAR); (5) Quantile prediction model based on Extreme Learning Machine and Particle Swarm Optimization (ELM-PSO); (6) Multivariate version of Direct Quantile Regression (DQR).

[0036] Table 2 Performance comparison results of probability prediction

[0037] From Table 2, we can see that: (1) CFGAN has the most satisfactory performance in terms of reliability and sharpness, achieving the lowest and the highest ; (2) CFGAN All are below 3%, which is a fairly superior prediction performance; (3) The linear nature of VAR again limits its approximation ability, making it the second worst model on the test set (only better than Pers). The performance of DeepAR, ELM-PSO and DQR on test set 2 is significantly lower than that on test set 1, showing worse zero-shot capabilities than CFGAN; (4) ELM-PSO and DQR are both based on shallow (three-layer) extreme learning machines, and the shallow number of layers limits their approximation ability and generalizability. Therefore, their performance is not as good as CFGAN with a deep structure; (5) Similar to point prediction, the performance of CFGAN on test set 1 is very close to that on test set 2; (6) From a parametric and non-parametric perspective, non-parametric models (CFGAN, ELM-PSO, DQR) are generally better than parametric models (Pers, VAR and DeepAR), especially on test set 2.

[0038] like Figure 2 and Figure 3 As shown, (1) DeepAR tends to underestimate most quantiles. (2) ELM-PSO tends to underestimate quantiles with low nominal ratios and overestimate quantiles with high nominal ratios. (3) On the contrary, DQR tends to overestimate quantiles with low nominal ratios and underestimate quantiles with high nominal ratios. (4) CFGAN has the slightest underestimation or overestimation, and also has the slightest curve vibration.

[0039] In summary, CFGAN outperforms other existing models in both point prediction and probability prediction, which not only proves that CFGAN has satisfactory numerical prediction performance, but also shows that the deep-structured CFGAN captures the complex multimodal mechanism of the battery calendar aging process, thereby proving that the CFGAN described in the present invention is superior in prediction and promotion.

[0040] The present invention adopts the above-mentioned battery calendar aging capacity prediction method based on the generative adversarial network model, and designs CFGAN (capacity prediction generative adversarial network) based on the electrochemical knowledge of the battery, so that the knowledge and data maintain a satisfactory consistency, and significantly improve its theoretical strength and prediction performance.

[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A battery calendar aging capacity prediction method based on a generative adversarial network model, characterized in that: The following steps are involved: S1. Establish a semi-empirical prediction model for calendar aging of lithium-ion batteries; S2. Establish and train a generative adversarial network for battery calendar aging capacity prediction; S3. Evaluation.

2. The battery calendar aging capacity prediction method based on the generative adversarial network model according to claim 1 is characterized in that: Step S1 specifically includes the following steps: S11. Set prediction tasks: （1）； In the formula, is the explanatory variable; Store time for batteries; The temperature at which the battery is stored; is the stored state of charge SOC; for The capacity of the battery at all times; is the lag interval of the prediction task, for The capacity of the battery at all times; S12. Set the target variable: （2）； In the formula, represents all target variables; is the maximum prediction time step, for The capacity of the battery at all times; for The capacity of the battery at all times; S13. Simultaneously establish multi-step point prediction and probability capacity prediction models: Formula (1) is abstracted into the following mapping: （3）； In the formula, express In the conditions The joint distribution under is the prediction target of the calendar aging prediction model for lithium-ion batteries, based on Get the moment The marginal distribution of , which is In the conditions Under the conditional distribution, then from Infer further prediction results from S14. Construct a semi-empirical prediction model for calendar aging of lithium-ion batteries based on electrochemical principles: （4）； In the formula, Indicates the nominal capacity of the battery; , represents the unknown parameters that need to be identified.

3. The battery calendar aging capacity prediction method based on the generative adversarial network model according to claim 2 is characterized in that: The battery calendar aging capacity prediction generative adversarial network described in step S2 includes a generator , Discriminator and conditional ; It specifically includes the following steps: S21. Training conditioner Become a conditional information transformer to convert the explanatory variables Convert to conditional feature ,Right now ; S211, Generator and the discriminator Conditioner The obtained conditional features As a condition, the training objective of the generative adversarial network for battery calendar aging capacity prediction is: （5）； In the formula, Represents the optimization generator with the minimization objective ; Denotes the maximum objective optimization discriminator ; Indicates a given Lower Discriminator Output for observation The probability of Indicates a given Lower Discriminator For the generator Generated fake samples Probability; Represents a false sample ;in, and Represents an observation sample and false samples The conditional distribution of S212. Solving formula (5) is equivalent to minimizing and The Jensen Shannon divergence of , formula (5) reaches the global minimum if and only if the following formula is satisfied: （6）； By simultaneously inputting As additional information, For the generator , Discriminator The adjustment will and Re-record them as and , formula (5) can be rewritten as follows: (7)； conditioner Will Convert to conditional feature , ; If No. The components are expressed as ,and =1,2,3,4, then It is composed of: （8）； （9）； （10）； （11）； （12）； In the formula, Represents a conditioner; All represent learnable parameters; S213, respectively, the conditioner , Generator and the discriminator Train, and each time the generator and the discriminator After training, the conditioner Training, Conditioner The training objectives are: （13）； In the formula, Represents the optimization condition for maximizing the objective ; Indicates conditional characteristics; S214, using conditioner Execute from arrive The equivalent transformation of As a condition, it is equivalent to As a condition, we obtain: （14）； （15）； Then we get: （16）； S22, Generator-based and conditional The prediction results are generated: Approximate sampling: Given and conditional What you get , then from the known given distribution Sampling a low-dimensional random noise , the low-dimensional random noise and Input to the generator at the same time In, get A sample of : （17）； because , so Considered as At the same time, after the battery calendar aging capacity prediction generative adversarial network is trained, Considered as Sampling in ; Mean estimation based on repeated sampling: From Multiple samplings are performed and all are input into formula (17), and we get Approximate sample of ,in represents the number of samples, The empirical distribution formed is estimates; Calculation of prediction results for each time step: The point prediction and probability prediction for each prediction time step are calculated as follows; , estimated through experience The average value determines the point prediction result : (18)； In the formula, express Point prediction results of ; Calculated by the following empirical algorithm The probability distribution prediction results are: Define any The empirical nominal proportions are as follows: (19)； In the formula, represents the indicator function; Representation and The corresponding empirical nominal ratio; Then, build a dictionary , and get a given nominal proportion from the dictionary The quantile approximation of : （20）； In the formula, Indicates nominal proportion The approximate quantile of .

4. The battery calendar aging capacity prediction method based on the generative adversarial network model according to claim 3 is characterized in that: Select values ​​from {5%, 10%, 15%, …, 95%} in sequence.

5. The battery calendar aging capacity prediction method based on the generative adversarial network model according to claim 4 is characterized in that: In the battery calendar aging capacity prediction generative adversarial network described in step S2, the generator and the discriminator Both use LSTM and bidirectional LSTM based on recurrent networks as basic building layers, stack multiple BiLSTM layers in the opposite direction of LSTM, and use the hidden state of the previous BiLSTM as the input of the next BiLSTM; And through residual technology, short-circuit links are added between non-adjacent neural network layers.

6. The battery calendar aging capacity prediction method based on the generative adversarial network model according to claim 5 is characterized in that: A short-circuit link is added every three BiLSTM or fully connected layers.

7. The battery calendar aging capacity prediction method based on the generative adversarial network model according to claim 1 is characterized in that: The evaluation described in step S3 includes the following evaluation indicators: point prediction indicators and probability prediction indicators; Among them, for point prediction, the maximum absolute error and root mean square error are used as evaluation indicators, which are defined as: （21）； （22）； In the formula, and Represents the time step The maximum absolute error and root mean square error; represents the number of samples; It means taking the maximum value among all samples; Further, take and Average value over all steps: （23）； In the formula, represents the number of time steps; The probability prediction is expressed in quantile form, and the quantile prediction evaluation indicators mainly include the following two: mean proportional deviation and skill score; The average proportional deviation is the core evaluation indicator based on reliability, which is defined as follows: （24）； In the formula, Indicates nominal proportion and time step The average proportional deviation of The average absolute proportional deviation over all nominal proportions is taken as the average absolute proportional deviation for the prediction time step Reliability evaluation of probability prediction results: （25）； In the formula, is the time step The mean absolute proportional deviation, is the total number of nominal proportions; The skill score is obtained by fusing reliability and sharpness, which is expressed as follows: （26）； In the formula, Represents the time step skill score.

Images