A method and system for predicting the remaining cycle life and inflection point of a power battery

Abstract

The application discloses a kind of power battery remaining cycle life and inflection point prediction method and system, it is related to the technical field of chargeable battery health management, solve the technical problem that the remaining effective life of battery and inflection point occurrence probability prediction accuracy is lower, its technical scheme key points are to construct inflection point probability model KPP by the nonlinear trend of battery capacity recession curve, then construct KPP prediction model by inflection point probability model KPP, on the basis that KPP prediction model outputs inflection point occurrence probability, obtain the prediction of battery remaining cycle life.By end-to-end input voltage sequence data, multiple steps are simultaneously predicted in advance Health status and inflection point probability, and using bayesian uncertainty perception technology and adversarial transfer learning method Realize the method of general learning under different operating environments and battery types, provide guarantee for battery safety and effective operation, and be conducive to establishing comprehensive general battery management system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rechargeable battery health management, and particularly relates to a method and system for predicting the remaining cycle life and inflection point of a power battery. BACKGROUND

[0002] Power lithium ion batteries, as important new energy storage devices, have the advantages of long cycle service life, high first charge-discharge capacity and low environmental pollution in practical applications, and have become the first choice of green power today. Due to the rapid market expansion in the fields of energy storage and transportation, people have put forward higher requirements for the operation life and safety of lithium ion batteries. The performance degradation of lithium ion batteries limits their energy storage and power output capability, and also affects the cost, life and other performances of electric vehicles, which is a key scientific problem in the research of new energy storage batteries. With a new round of technological revolution and industrial reform, artificial intelligence technology has become a scientific technology that can truly change the production process of human society. In the field of power battery health management, compared with physical electrochemical models that involve complex parameter expressions and difficult solutions, deep learning methods can directly realize end-to-end learning from health factors to capacity state estimation.

[0003] In the existing deep learning modeling method for predicting lithium ion batteries, the prediction of the "inflection point" in the capacity degradation trend has not been paid attention to. The occurrence of the inflection point is related to the rapid capacity decline and the use of secondary batteries, and needs to be accurately predicted in advance. In addition, deep learning methods with uncertainty perception have been paid more and more attention in battery health management. Through the Bayesian method, the uncertainty of the weights in neural network training can be quantified, and finally the prediction confidence interval of the remaining cycle life is given. How to accurately predict the capacity inflection point and the confidence interval of the remaining cycle life at the same time is a problem to be solved. SUMMARY

[0004] The present application provides a method and system for predicting the remaining cycle life and inflection point of a power battery, and the technical purpose is to establish an end-to-end method for simultaneously predicting the remaining effective life and the occurrence probability of the inflection point of the battery, so as to realize a general prediction method under different operating conditions and battery types.

[0005] The above technical purpose of the present application is realized by the following technical scheme:

[0006] A method for predicting the remaining cycle life and inflection point of a power battery, comprising:

[0007] S1: obtaining a voltage curve sequence X n =[V n-l+1 ,V n-l+2 ,...,V nThe input is fed into the first feature extraction module for feature extraction, resulting in the battery health feature H. n Among them, V i The voltage curve represents the i-th charge-discharge cycle.

[0008] S2: Battery health characteristics H n The data are respectively input to the first KPP prediction module and the first RUL prediction module. The first KPP prediction module predicts the battery health characteristics H. n Output KPP prediction sequence KPP n+1～n+m KPP n+1～n+m That is, the probability of the inflection point occurring; where m represents the predicted sequence length and l represents the length of the input voltage curve sequence;

[0009] S3: Transform the KPP prediction sequence KPP n+1～n+m The input to the first RUL prediction module is the battery health feature H. n and KPP prediction sequence KPP n+1～n+m That is, [H] n KPP n+1～n+m The output of the first RUL prediction module is the RUL prediction sequence, which is the confidence interval of the remaining cycle lifetime (RUL) predicted in advance over multiple steps. n+1～n+m ;

[0010] S4: Predict the KPP sequence based on KPP n+1～n+m and RUL prediction sequence RUL n+1～n+m Predict the remaining cycle life and inflection point of the source battery; where the remaining cycle life is the remaining number of charge and discharge cycles.

[0011] The first KPP prediction module predicts inflection points using the inflection point probability model KPP. The construction process of the inflection point probability model KPP includes:

[0012] By linearly fitting the values ​​of battery health status based on capacity within the range of 100% to 80%, a fitting curve is obtained. The non-linear trend of the capacity degradation curve is then obtained by subtracting the fitting curve from the original degradation curve.

[0013] The capacity-based battery health status is represented as follows:

[0014]

[0015] n represents the current charge / discharge cycle number, Q0 and Q n These represent the initial capacity and the capacity of the nth discharge, respectively.

[0016] The nonlinear trend of the capacity decay curve is calculated using an exponential weighting method, which is expressed as:

[0017]

[0018] wherein, denotes the normalized nonlinear trend; R(n) is related to the remaining number of charge-discharge cycles to the end of life, normalized to 0-1; in order to maintain the same range as the battery health state, the calculation result of formula (2) is normalized to [0.8, 1], and the first KPP prediction module is made to take the value of 1 at the inflection point and after the inflection point, so as to obtain the final inflection point probability model KPP.

[0019] Further, the method further comprises predicting the remaining cycle life and the probability of inflection point occurrence of the target battery by a transfer learning method, comprising:

[0020] S5: copying the first feature extraction module, the first KPP prediction module and the first RUL prediction module of the source battery to the prediction model of the target battery to obtain the prediction model of the target battery, comprising a second feature extraction module, a second KPP prediction module and a second RUL prediction module;

[0021] S6: inputting the battery health features of the source battery and the battery health features of the target battery into the classifier for training to perform feature alignment;

[0022] S7: inputting the battery health features of the target battery after feature alignment into the second feature extraction module for training;

[0023] S8: repeating steps S6 to S7 until the verification error is less than a set threshold to obtain the final prediction model of the target battery, and predicting the remaining cycle life and the probability of inflection point occurrence of the target battery by the final prediction model of the target battery.

[0024] Further, in steps S6 and S7, the classifier is trained three times and the second feature extraction module is trained once at each iteration.

[0025] Further, the loss function of the classifier is an improved Wasserstein distance, denoted as:

[0026]

[0027] wherein, denotes the battery health features of the source battery; H denotes the battery health features of the target battery; denotes the mixed features, λ gp denotes the loss weight of the gradient penalty factor, and D denotes the classifier.

[0028] Further, the first feature extraction module and the second feature extraction module are both composed of a convolution layer and a Dropout layer, and the Dropout layer is used for weight distribution statistics.

[0029] Further, the method further comprises outputting a prediction interval of uncertain perception, the method calculating a prediction mean and a confidence interval by adding a Dropout layer in the feature extraction module and performing Monte Carlo sampling when testing the output.

[0030] A prediction system for the remaining cycle life and the inflection point of a power battery, the system comprising a prediction model, the prediction model comprising:

[0031] a first feature extraction module for extracting features from a voltage sequence of a source battery to obtain a battery health feature;

[0032] a first KPP prediction module for outputting a KPP prediction sequence from the battery health feature, the KPP prediction sequence being an inflection point occurrence probability, and predicting the inflection point occurrence probability of the source battery through the KPP prediction sequence;

[0033] a first RUL prediction module for outputting an RUL prediction sequence from the battery health feature and the KPP prediction sequence, the RUL prediction sequence being a confidence interval of the remaining cycle life predicted in advance by multiple steps, and predicting the remaining cycle life of the source battery through the RUL prediction sequence.

[0034] Further, the system further comprises a transfer learning model for evaluating the health state of a target battery through the trained prediction model, the transfer learning model comprising a second feature extraction module, a second KPP prediction module, a second RUL prediction module and a classifier, the second feature extraction module, the second KPP prediction module and the second RUL prediction module being obtained by copying the first feature extraction module, the first KPP prediction module and the first RUL prediction module;

[0035] the classifier being used for training the battery health feature of the source battery and the battery health feature of the target battery to perform feature alignment, inputting the battery health feature of the target battery after the feature alignment to the second feature extraction module for training; when the validation error of the classifier and the second feature module after the training is less than a set threshold, obtaining a final prediction model of the target battery, and predicting the remaining cycle life and the inflection point occurrence probability of the target battery through the final prediction model of the target battery.

[0036] The power battery remaining cycle life and inflection point prediction method and system have the following beneficial effects: the nonlinear trend of the battery capacity degradation curve is used to predict the occurrence of the inflection point in advance, and the double indicators of the battery health state and the inflection point occurrence probability are simultaneously predicted; the deep transfer learning method is used to predict the remaining cycle life and the inflection point occurrence probability of the battery under different operating environments and battery types; and the confidence interval of the battery cycle life is obtained through uncertain perception. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 A schematic diagram of battery state of health and inflection point occurrence probability;

[0038] Figure 2 A schematic diagram of a system for predicting the remaining useful life and inflection point occurrence probability of a power battery;

[0039] Figure 3 A schematic diagram of a transfer learning method. DETAILED DESCRIPTION

[0040] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings.

[0041] The power battery remaining cycle life and inflection point prediction method described in the present application comprises:

[0042] S1: input the voltage curve sequence X of the source battery n = [V n-l+1 , V n-l+2 ,..., V n ] to the first feature extraction module for feature extraction to obtain the battery health feature H n ; wherein V i represents the voltage curve of the i-th charge and discharge.

[0043] S2: input the battery health feature H n to the first KPP prediction module and the first RUL prediction module respectively, the first KPP prediction module outputs the KPP prediction sequence KPP n according to the battery health feature H n+1～n+m , KPP n+1～n+m , which is the inflection point occurrence probability; wherein m represents the sequence length of the advance prediction, and l represents the length of the input voltage curve sequence.

[0044] Specifically, the first KPP prediction module performs inflection point prediction through the inflection point probability model KPP, and the construction process of the inflection point probability model KPP comprises:

[0045] By linear fitting the values of the capacity-based state of health (SOH) within 100% to 80%, a fitting curve is obtained, and then the fitting curve is subtracted from the original degradation curve to obtain the nonlinear trend of the capacity degradation curve.

[0046] Wherein, the capacity-based state of health is represented as:

[0047]

[0048] Wherein, n represents the current charge and discharge cycle number, Q0 and Q nrespectively represent the initial capacity and the capacity of the nth discharge.

[0049] The nonlinear trend of the capacity degradation curve is calculated by exponential weighting, denoted as:

[0050]

[0051] wherein, represents the normalized nonlinear trend; R(n) is related to the remaining number of charge-discharge cycles to the end of life, normalized to 0-1; in order to maintain the same range as the battery health state, the calculation result of formula (2) is normalized to [0.8, 1], and the first KPP prediction module is set to 1 at the inflection point and after the inflection point, thereby obtaining the inflection point probability model KPP as shown in formula (3). Figure 1 The inflection point probability model KPP has the highest value of 1 at the inflection point, and the overall value range of the model is 0-1.

[0052] S3: input the KPP prediction sequence KPP n+1～n+m to the first RUL prediction module, and the input of the first RUL prediction module is the battery health feature H n and the KPP prediction sequence KPP n+1～n+m , i.e. [H n , KPP n+1～n+m ], and the output of the first RUL prediction module is the RUL prediction sequence, i.e. the remaining cycle life confidence interval RUL n+1～n+m predicted in advance by multiple steps.

[0053] S4: according to the KPP prediction sequence KPP n+1～n+m and the RUL prediction sequence RUL n+1～n+m , the remaining cycle life and the inflection point of the source battery are predicted; wherein the remaining cycle life is the remaining number of charge-discharge cycles.

[0054] In the above method, the input of the first feature extraction module is the voltage curve sequence X n = [V n-l+1 , V n-l+2 ,..., V n ], and the output H n of the first feature extraction module is the extracted feature sequence.

[0055] The input of the first KPP prediction module is H n , and the output is the inflection point occurrence probability KPP n+1～n+m ; the input of the first RUL prediction module is the feature sequence value and the inflection point occurrence probability value, i.e. [H n , KPP n+1～n+m ], and the output is the remaining cycle life confidence interval RUL n+1～n+mwherein m represents the sequence length of the early prediction, l represents the length of the input voltage curve sequence, V i represents the voltage curve of the i-th charge and discharge.

[0056] When training the prediction model of the source battery, the mean square error can be used as the loss function. The overall loss function L RUL-KPP The loss function L RUL and L KPP of the two prediction modules (the first RUL prediction module and the first KPP prediction module) are weighted to form:

[0057] L RUL-KPP = (1-λ) L RUL + λL KPP ;

[0058] wherein λ represents the weighting weight.

[0059] Correspondingly, the power battery remaining cycle life and inflection point prediction system described in the present application comprises a prediction model, which comprises a first feature extraction module, a first RUL prediction module and a first KPP prediction module, as shown in Figure 2 As a specific embodiment, since the running environment and type of the battery are different, how to use the trained prediction model (i.e. the prediction model of the source battery) to evaluate the health state of the target battery under different battery types and running conditions, so as to realize the prediction of the battery state only relying on the stored source battery model and the voltage data collected by the sensor in the case of no label value. Based on this, in order to further predict the remaining cycle life of the target battery of different running environment and battery type, the system is also provided with a transfer learning model, which evaluates the remaining cycle life confidence interval and inflection point occurrence probability of the target battery through the trained prediction model.

[0060] The transfer learning model comprises a second feature extraction module, a second KPP prediction module and a second RUL prediction module and a classifier, and the second feature extraction module, the second KPP prediction module and the second RUL prediction module are obtained by copying the first feature extraction module, the first KPP prediction module and the first RUL prediction module. That is, during initialization, the second feature extraction module, the second KPP prediction module and the second RUL prediction module are the same as the first feature extraction module, the first KPP prediction module and the first RUL prediction module.

[0061] The classifier is used to train the battery health features of the source battery and the battery health features of the target battery for feature alignment, and the battery health features of the target battery after feature alignment are input to the second feature extraction module for training; when the verification error after training of the classifier and the second feature module is less than a set threshold, the final prediction model of the target battery is obtained, and the remaining useful life confidence interval and the inflection point occurrence probability of the target battery are predicted through the final prediction model of the target battery.

[0062] In the final prediction model of the target battery, the specific parameters of the second KPP prediction module and the second RUL prediction module are set according to the parameters of the target battery, and the setting mode is similar to that of the first KPP prediction module and the first RUL prediction module of the source battery.

[0063] Correspondingly, the present application can predict the remaining cycle life and the inflection point occurrence probability of the target battery through the transfer learning method, as shown in Figure 3 , specifically comprising:

[0064] S5: copying the first feature extraction module, the first KPP prediction module and the first RUL prediction module of the source battery into the prediction model of the target battery to obtain the prediction model of the target battery including the second feature extraction module, the second KPP prediction module and the second RUL prediction module.

[0065] S6: inputting the battery health features of the source battery and the battery health features of the target battery to the classifier for training to perform feature alignment. That is, the same subspace in the battery health features of the target battery and the battery health features of the source battery is aligned, and then the classifier discriminates the battery health features of the target battery after feature alignment.

[0066] S7: inputting the battery health features of the target battery after feature alignment to the second feature extraction module for training. After the classifier discriminates the battery health features of the target battery after feature alignment, the battery health features are input to the second feature extraction module for training.

[0067] In steps S6 and S7, each time iteration, the classifier is trained three times, and then the second feature extraction module is trained once. That is, in step S6, after the classifier is trained three times, it is transferred to step S7 to train the second feature extraction module once.

[0068] S8: repeating steps S6 to S7 until the verification error is less than the set threshold to obtain the final prediction model of the target battery, and the remaining useful life confidence interval and the inflection point occurrence probability of the target battery are predicted through the final prediction model of the target battery.

[0069] As a specific embodiment, the loss function of the classifier is an improved Wasserstein distance, denoted as:

[0070]

[0071] wherein, represents the battery health feature of the source battery; H represents the battery health feature of the target battery; represents a mixed feature, λ gp represents a loss weight of a gradient penalty factor, and D represents a classifier.

[0072] As a specific embodiment, the first feature extraction module and the second feature extraction module are both composed of a convolution layer and a Dropout layer. By adding a Dropout layer to the feature extraction module and performing Monte Carlo sampling when testing output, a prediction mean and a confidence interval are calculated to give a more reliable cycle life prediction scheme.

[0073] The above is a demonstrative embodiment of the present application, and the protection scope of the present application is defined by the claims and their equivalents.

Claims

1. A method for predicting the remaining cycle life and inflection point of a power battery, characterized in that, include: S1: The voltage curve sequence X of the source battery n =[V n-l+1 V n-l+2 ,...,V n The input is fed into the first feature extraction module for feature extraction, resulting in the battery health feature H. n ; Among them, V i The voltage curve represents the i-th charge-discharge cycle. S2: Battery health characteristics H n The data are respectively input to the first KPP prediction module and the first RUL prediction module. The first KPP prediction module predicts the battery health characteristics H. n Output KPP prediction sequence KPP n+1～n+m KPP n+1～n+m That is, the probability of the inflection point occurring; where m represents the predicted sequence length and l represents the length of the input voltage curve sequence; S3: Transform the KPP prediction sequence KPP n+1～n+m The input to the first RUL prediction module is the battery health feature H. n and KPP prediction sequence KPP n+1～n+m That is, [H] n KPP n+1～n+m The output of the first RUL prediction module is the RUL prediction sequence, which is the confidence interval of the remaining cycle lifetime (RUL) predicted in advance over multiple steps. n+1～n+m ; S4: Predict the KPP sequence based on KPP n+1～n+m and RUL prediction sequence RUL n+1～n+m Predict the inflection point and remaining cycle life of the source battery; where the remaining cycle life is the remaining number of charge-discharge cycles. The first KPP prediction module predicts inflection points using the inflection point probability model KPP. The construction process of the inflection point probability model KPP includes: By linearly fitting the values ​​of battery health status based on capacity within the range of 100% to 80%, a fitting curve is obtained. The non-linear trend of the capacity degradation curve is then obtained by subtracting the fitting curve from the original degradation curve. The capacity-based battery health status is represented as follows: n represents the current charge / discharge cycle number, Q0 and Q n These represent the initial capacity and the capacity of the nth discharge, respectively. The nonlinear trend of the capacity decay curve is calculated using an exponential weighting method, which is expressed as: in, The normalized nonlinear trend is represented by R(n), which is related to the remaining number of charge and discharge cycles before the end of the lifespan and is normalized to 0 to 1. In order to maintain the same range as the battery health state, the calculation result of formula (2) is normalized to [0.8,1], and the value of the first KPP prediction module is set to 1 at the inflection point and thereafter, thus obtaining the final inflection point probability model KPP.

2. The method as described in claim 1, characterized in that, The method also includes predicting the remaining cycle life and inflection point probability of the target battery using transfer learning methods, including: S5: Copy the first feature extraction module, the first KPP prediction module and the first RUL prediction module of the source battery to the prediction model of the target battery, and obtain the prediction model of the target battery including the second feature extraction module, the second KPP prediction module and the second RUL prediction module. S6: Input the battery health features of the source battery and the target battery into the classifier for training to perform feature alignment; S7: Input the battery health features of the target battery after feature alignment into the second feature extraction module for training; S8: Repeat steps S6 to S7 until the verification error is less than the set threshold to obtain the final prediction model of the target battery. The remaining cycle life and inflection point probability of the target battery are predicted by the final prediction model of the target battery.

3. The method as described in claim 2, characterized in that, In steps S6 and S7, the classifier is trained three times in each iteration, and then the second feature extraction module is trained once.

4. The method as described in claim 2, characterized in that, The loss function of the classifier is the improved Wasserstein distance, expressed as: in, H represents the battery health characteristics of the source battery; H represents the battery health characteristics of the target battery. Representing mixed features, λ gp The loss weights represent the gradient penalty factor, and D represents the classifier.

5. The method as described in claim 2, characterized in that, Both the first feature extraction module and the second feature extraction module consist of convolutional layers and dropout layers. The dropout layer is used to perform weight distribution statistics.

6. The method as described in claim 5, characterized in that, The method also includes output uncertainty-aware prediction intervals, which are calculated by adding a Dropout layer to the feature extraction module and performing Monte Carlo sampling on the test output to calculate the prediction mean and confidence interval.

7. A prediction system for the remaining cycle life and inflection point of a power battery, the system being used in the method described in any one of claims 1-6, characterized in that, The system includes a prediction model, which includes: The first feature extraction module extracts features from the voltage sequence of the source battery to obtain battery health features; The first KPP prediction module outputs a KPP prediction sequence based on battery health characteristics. The KPP prediction sequence is the probability of the inflection point occurring. The inflection point of the source battery is predicted by the KPP prediction sequence. The first RUL prediction module outputs an RUL prediction sequence based on battery health characteristics and KPP prediction sequence. The RUL prediction sequence is the confidence interval of remaining cycle life predicted in advance in multiple steps. The remaining cycle life of the source battery is predicted by the RUL prediction sequence.

8. The system as described in claim 7, characterized in that, The system also includes a transfer learning model that evaluates the health status of the target battery using a trained prediction model. The transfer learning model includes a second feature extraction module, a second KPP prediction module, a second RUL prediction module, and a classifier. The second feature extraction module, the second KPP prediction module, and the second RUL prediction module are obtained by copying the first feature extraction module, the first KPP prediction module, and the first RUL prediction module. The classifier is used to train the battery health features of the source battery and the target battery to perform feature alignment. The battery health features of the target battery after feature alignment are input into the second feature extraction module for training. When the verification error after training the classifier and the second feature module is less than a set threshold, the final prediction model of the target battery is obtained. The remaining cycle life and the probability of the inflection point of the target battery are predicted by the final prediction model of the target battery.

Images